Abnormal Sound Detection via Statistical Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In environments with limited power supply, existing systems cannot transmit raw or compressed sound data due to restricted bandwidth, making it impossible to report abnormal sounds effectively from facilities to remote servers.
Innovation Solution
An abnormal sound detection system that uses a terminal to compute and transmit statistical data from sound recordings, allowing a server to learn a normal sound model and recreate artificial sounds for confirmation, even with limited bandwidth, by employing a logarithmic mel spectrogram and pseudo-spectrogram reconstruction techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If raw sound data or typical compressed sound data is transmitted, then the quality of sound transmission is improved, but the transmission traffic exceeds the limited battery-driven bandwidth
Solution Approach 1:
The patent extracts only the essential statistical features (spectral centroid, spectral rolloff, spectral flux, zero-crossing rate, and chroma features) from the complete sound signal, transmitting only these extracted parameters instead of the full sound data. This extraction approach maintains sufficient information for abnormal sound detection while dramatically reducing transmission traffic to fit within battery-driven bandwidth limits.
Solution Approach 2:
The patent creates a simplified representation (copy) of the sound data in the form of statistical feature vectors that capture the essential characteristics needed for abnormal sound detection. These feature vectors serve as a compressed copy that preserves diagnostic information while occupying minimal transmission bandwidth.
2Quantity of substance
If audio encoder is used for compression, then the sound data is compressed, but the encoding process consumes excessive power for long-term battery operation
Solution Approach 1:
The patent replaces the complex mechanical/audio encoding process (FFT, DCT, quantization) with a simpler statistical feature extraction approach. Instead of using traditional audio encoders that require significant computational resources, the system calculates basic statistical parameters directly from the sound signal, dramatically reducing power consumption while achieving effective data compression.
3Quantity of substance
If only abnormality presence/absence is reported, then the transmission traffic is reduced, but the user cannot hear and confirm the actual abnormal sound
Solution Approach 1:
The patent introduces statistical feature vectors as an intermediary representation that bridges the gap between complete sound data and simple abnormality flags. These feature vectors serve as a middle ground, providing sufficient information for both automated detection and human auditory confirmation without requiring transmission of full-resolution sound data.
Data Source
AI summary
Confirmation can be made what sound has been made under a restriction in which transmittable traffic is small. An abnormal sound detection system including an artificial sound creating function is configured, the abnormal sound detection system including a statistic calculation unit configured to calculate a statistic set expressing sizes of a direct current component, an alternating current component, and a noise component in an amplitude time series at each of frequencies of a sound inputted at a terminal, a statistic transmitting unit configured to transmit the statistic set from the terminal to a server, a statistic receiving unit configured to receive the statistic set in the server, and an artificial sound reproducing unit configured to reproduce a cyclostationary artificial sound based on the statistic set received in the server.