Spectrogram Anomaly Detection Using Attentive Neural Processes

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for detecting anomalous sound in audio signals are inefficient and inaccurate, particularly for non-stationary sounds, due to the lack of frequency domain information and the need for lengthy processing of audio signals with few anomalous sound occurrences, leading to high computing resource consumption and missed detections.

Innovation Solution

The use of an attentive neural process architecture that partitions spectrograms into context and target regions, employing techniques like random RowCol selection, middle frame selection, and likelihood bootstrapping to identify anomalous sounds by comparing reconstructed and original spectrogram regions, leveraging prior knowledge and signal properties for improved detection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If the entire audio signal is processed to detect anomalous sound, then detection coverage is improved, but processing time and computing resource consumption increase significantly

Engineering Contradiction:
Improvedetection coverageVSAvoidprocessing time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The audio signal is transformed into a spectrogram and divided into multiple time-frequency regions. The system selectively processes only those regions that contain anomalous sound characteristics, rather than processing the entire audio signal. This segmentation approach maintains detection coverage while significantly reducing processing time and computational resources.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If traditional auto-encoder reconstructs the entire audio signal, then reconstruction accuracy is improved, but computing resource consumption increases

Engineering Contradiction:
Improvereconstruction accuracyVSAvoidcomputing resource consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system extracts only the relevant time-frequency regions from the spectrogram that contain anomalous sound characteristics. The auto-encoder then processes and reconstructs only these extracted regions rather than the entire audio signal. This extraction approach maintains reconstruction accuracy for anomalous regions while significantly reducing computing resource consumption.

Inventive Principle:
Principle #2Taking out (Extraction)

3Speed

If abrupt temporal changes are used for anomaly detection, then detection speed is improved, but frequency domain information is lost leading to inaccurate outcomes

Engineering Contradiction:
Improvedetection speedVSAvoiddetection accuracy
Core Design Contradiction:
SpeedVSMeasurement precision

Solution Approach 1:

The system transforms the audio signal from the time domain to the time-frequency domain using spectrogram analysis. This dimensional transformation allows the system to simultaneously detect abrupt temporal changes and analyze frequency domain characteristics. By operating in the time-frequency domain rather than purely temporal domain, the system maintains both detection speed and frequency information for accurate anomaly detection.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP4402679B1Method and system for detecting anomalous sound
Publication Date: 2026.02.18 MITSUBISHI ELECTRIC CORP
  • EP4402679B1 patent drawingFigure 1
  • EP4402679B1 patent drawingFigure 2
  • EP4402679B1 patent drawingFigure 3

AI summary

A system and method for detecting anomalous sound are disclosed. The method includes receiving a spectrogram of an audio signal with elements defined by values in a time-frequency domain of the spectrogram. Each of the values corresponds to an element of the spectrogram that is identified by a coordinate in the time-frequency domain. The time-frequency domain of the spectrogram is partitioned into a context region and a target region. The context region and the target region are processed by a neural network using an attentive neural process to recover values of the spectrogram for elements with coordinates in the target region. The recovered values of the elements of the target region are compared with values of elements of the partitioned target region. An anomaly score is determined based on the comparison. The anomaly score is used for performing a control action.