The Scientific Basis and Mechanism of Acoustic Fingerprint Monitoring
Understanding and mastering the generation mechanism of audible sound signals from transformers is a prerequisite for all diagnosis and early warning.
These signals mainly originate from frequencies ranging from 20Hz to 20kHz in transformer noise, which are generated by various abnormalities in the windings, iron cores, and rotating cooling fan devices.
During the operation of transformers, structures such as iron cores and windings vibrate due to excitation effects such as magnetostriction, electromagnetic forces, and torque forces, generating mechanical waves. These mechanical waves are then transmitted to the outer casing of the equipment or the surrounding ambient air through solid and liquid media within the equipment. These vibrations and audible sound signals contain rich information about equipment fault states.
Current Status and Technical Challenges of Acoustic Fingerprint Monitoring
Currently, the main approach for abnormal sound detection is based on algorithms using positive sample data deviations. However, the effectiveness of further identifying fault types is limited, and there is limited exploitation of the application value of positive sample data.
During the collection of audible sound data on site, there are inevitably various types of interfering noises, such as bird calls, horns, on-load tap-changer operation sounds, corona discharge sounds, fan sounds, etc., which hinder widespread application.
A common problem in data-driven fault diagnosis within the industry is the insufficient diversity of fault sample types and the imbalance between positive and negative sample data volumes.
Currently, several monitoring methods coexist, including contact, non-contact, and ultrasonic methods, making it difficult to choose the most suitable approach for practical applications.
Technical Advantages of Penghe’s Solution
We categorize positive samples and conduct model training separately based on transformer manufacturers, equipment models, and monitoring methods (sound, vibration) to ensure that the extracted equipment health status characteristic indicators are generalizable.
We classify noise interference types, identifying both transient and steady-state interference, and apply targeted signal processing algorithms for denoising. Spectral subtraction is used to remove average interference (such as fan interference), wavelet time-frequency denoising is used to remove non-stationary transient interference, and our unique acoustic fingerprint signal-to-noise ratio metric is used to remove excessive abnormal non-stationary interference. We also leverage the advantages of both audible sound and vibration for comprehensive diagnosis.
Feature value diagnosis is the primary approach, with deep learning serving as a supplement. We select appropriate acoustic fingerprint characteristic values for judgment, extract core acoustic fingerprint features based on the distribution characteristics of existing fault samples, and use lightweight fault classification algorithms to diagnose faults such as DC bias, winding deformation, and iron core loosening.
Penghe Intelligence is one of the few service providers in the industry that simultaneously offers non-contact microphones, flush-mounted acoustic fingerprints, MEMS acoustic fingerprint arrays, and vibration sensor monitoring solutions. These solutions are suitable for high-precision acoustic fingerprint measurement and diagnosis, the collection and construction of acoustic fingerprint sample libraries, and various scenarios requiring acoustic fingerprint visual positioning.