Machine sound anomaly detection method, system and equipment based on spectrogram and medium

By employing multi-scale spectrum analysis based on spectrograms and deep learning methods, the accuracy problem of abnormal machine sound detection has been solved, enabling efficient and intelligent diagnosis and real-time monitoring of machine sounds, applicable to multiple industrial fields.

CN121096367APending Publication Date: 2025-12-09DUKE KUNSHAN UNIVERSITY
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202511098551.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-08-06
Publication Date
2025-12-09

AI Technical Summary

Technical Problem

Existing methods for detecting abnormal machine sounds have limitations in accuracy, especially when dealing with complex operating conditions and high-dimensional sound features, making it difficult to accurately identify abnormal equipment states.

Method used

A multi-scale spectrum analysis method based on spectrograms is adopted. Spectral feature vectors are generated through Fourier transform, adaptive multi-scale scanning, lightweight residual neural network and statistical pooling layer. Clustering is performed using the k-means++ algorithm, and anomaly detection is performed by combining cosine distance for similarity calculation.

Benefits of technology

It achieves full-time and full-frequency domain feature analysis of machine sound, significantly improving the ability to capture and detect early subtle faults, supporting real-time parallel analysis, reducing unplanned downtime, and is applicable to multiple industrial fields such as power, petrochemical, and metallurgy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121096367A_ABST
    Figure CN121096367A_ABST
Patent Text Reader

Abstract

The invention discloses a spectrogram-based machine sound anomaly detection method, system and device and a medium, and the method comprises the steps: firstly collecting audio data when a machine works normally, carrying out the preprocessing of noise reduction, normalization and the like, inputting a multi-scale spectrum analysis model, and extracting a spectrum feature vector; clustering the feature vectors to obtain a clustering center vector set, and constructing a normal state feature template; to-be-detected audio data in machine operation are collected in real time, corresponding feature vectors are generated through the multi-scale spectrum analysis model, and similarity calculation is carried out on the feature vectors and the clustering center vector set; and when any similarity result exceeds a preset threshold, determining that the to-be-detected audio data is normal, and otherwise, determining that the to-be-detected audio data is abnormal. According to the method, the time-frequency features of the sound are accurately captured through multi-scale spectrum analysis, the normal state model is established in combination with clustering analysis, sound abnormity in the machine operation process can be effectively recognized, early warning of equipment faults is achieved, and the method has the advantages of being high in detection accuracy and high in anti-interference capacity.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Cited By

  • Sound anomaly detection method based on prototype network

    CN121938409A

  • Self-supervision equipment operation abnormal sound detection method

    CN122130404A