CN-122024761-A - GIS voiceprint-vibration collaborative diagnosis system integrating MFCC feature extraction and ConvNeXt-T network

CN122024761ACN 122024761 ACN122024761 ACN 122024761ACN-122024761-A

Abstract

The invention relates to a GIS voiceprint-vibration collaborative diagnosis system integrating MFCC feature extraction and ConvNeXt-T network, which belongs to the technical field of power equipment monitoring and comprises the steps of synchronously collecting geographic position information, voiceprint data and vibration data of GIS equipment, performing pre-emphasis, framing and windowing preprocessing on the voiceprint data to generate voiceprint data frames, adopting an SRAM (static random Access memory) integrated architecture to extract MFCC features and HHT features, carrying out fusion of the MFCC features and the HHT features according to the geographic position information to obtain a multi-parameter feature matrix containing geographic labels, carrying out end-to-end reasoning on the multi-parameter feature matrix by adopting the ConvNeXt-T network, outputting fault types and confidence, determining dynamic weights by adopting a hierarchical analysis method in combination with fault confidence, equipment operation duration, maintenance records, fault frequency, severity and spatial position information, calculating equipment health scores in real time, classifying health grades and generating maintenance suggestions. The invention obviously improves the GIS fault recognition accuracy and reduces the operation and maintenance cost of GIS large-scale application.

Inventors

ZHAO YINGYING
SI WENRONG
FU CHENZHAO
DENG XIANQIN
GAO KAI
JIANG ANFENG
HU ZHENGYONG

Assignees

国网上海市电力公司
华东电力试验研究院有限公司

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (10)

1. A GIS voiceprint-vibration collaborative diagnostic system that fuses MFCC feature extraction with ConvNeXt-T networks, comprising: The multi-parameter acquisition unit is used for synchronously acquiring geographic position information, voiceprint data and vibration data of the GIS equipment, and performing pre-emphasis, framing and windowing preprocessing on the acquired voiceprint data to generate voiceprint data frames; The integrated memory-calculation feature processing unit adopts an SRAM integrated memory-calculation architecture, receives voiceprint data frames, extracts MFCC features, receives vibration data, extracts HHT features, and performs fusion and storage of the MFCC features and the HHT features according to received geographic position information to obtain a multi-parameter feature matrix containing geographic tags; The intelligent diagnosis and health assessment unit performs end-to-end reasoning on the multi-parameter feature matrix by adopting ConvNeXt-T network, outputs fault type and confidence coefficient, and determines dynamic weight by adopting a hierarchical analysis method in combination with fault confidence coefficient, equipment operation time length, maintenance record, fault frequency, severity and space position information, calculates equipment health score in real time, classifies health grade and generates maintenance advice.
2. The MFCC feature extraction and ConvNeXt-T network-integrated GIS voiceprint-vibration co-diagnostic system of claim 1, The multi-parameter acquisition unit comprises: the voiceprint acquisition module is used for acquiring broadband voiceprint data of the GIS equipment by adopting a high-sensitivity microphone, and carrying out preliminary MFCC (frequency division multiplexing) feature extraction pretreatment on the voiceprint data to generate an initial acoustic feature vector; the vibration acquisition module is used for acquiring time-varying data of vibration acceleration of the GIS equipment by adopting a piezoelectric acceleration sensor; And the GIS positioning module is used for acquiring the geographic position information of the GIS equipment by adopting a GPS/Beidou dual-mode positioning technology.
3. The MFCC feature extraction and ConvNeXt-T network-integrated GIS voiceprint-vibration co-diagnostic system of claim 2, The voiceprint acquisition module comprises an MFCC feature extraction front unit; The MFCC feature extraction pre-unit executes pre-emphasis, framing and windowing in the MFCC algorithm, hardens the pre-emphasis, framing and windowing into micro logic of time domain running water and zero SRAM, and outputs voiceprint data frames for subsequent Mel frequency analysis.
4. The MFCC feature extraction and ConvNeXt-T network-integrated GIS voiceprint-vibration co-diagnostic system of claim 2, The integrated memory and calculation feature processing unit adopts an integrated memory and calculation SRAM architecture and comprises: The MFCC feature extraction subunit is used for extracting Mel frequency cepstrum coefficient from the voiceprint data frame output by the MFCC feature extraction front unit, and generating MFCC features through processing such as fast Fourier transform, mel filtering and cepstrum transform; The vibration time-frequency characteristic extraction subunit is used for carrying out time-frequency analysis on the time-varying data of the collected vibration acceleration of the GIS equipment by adopting Hilbert-Huang transform (HHT) to extract the HHT characteristics, wherein the HHT characteristics comprise vibration time-frequency characteristics including instantaneous frequency and amplitude envelope; The multi-feature fusion subunit is used for carrying out fusion of the MFCC features and the HHT features according to the received geographic position information, constructing a multi-parameter feature matrix in a feature splicing or feature mapping mode to obtain a multi-parameter feature matrix containing geographic labels, and realizing multi-dimensional collaborative association of voiceprint, vibration and position features.
5. The MFCC feature extraction and ConvNeXt-T network-integrated GIS voiceprint-vibration co-diagnostic system of claim 4, The memory and calculation integrated SRAM chip comprises a plurality of 8TiC calculation units and 6T-SRAM storage units; in the MFCC feature extraction subunit, an 8TiC calculation unit completes in-situ pipeline calculation on an input windowed real frame, and data does not need to be moved away from a storage area; In the vibration time-frequency characteristic extraction subunit, an 8TiC computing unit performs in-situ Hilbert-Huang transformation on the vibration acceleration data stream, the data does not need to be moved away from a storage array, EMD empirical mode decomposition, hilbert transformation and instantaneous characteristic calculation are sequentially completed, and instantaneous frequency, instantaneous amplitude envelope and Hilbert spectrum statistics are output; In the multi-feature fusion subunit, the 6T-SRAM array latches longitude, latitude and altitude of the GIS positioning module according to the corresponding sampling time stamp while receiving the acoustic vibration feature vector, and the 8TiC unit encodes the position into a grid ID through WL pulse and completes splicing with the feature vector in a charge domain to form an acoustic-vibration-bit integrated multi-parameter feature matrix.
6. The MFCC feature extraction and ConvNeXt-T network-integrated GIS voiceprint-vibration co-diagnostic system of claim 4, The intelligent diagnosis and health assessment unit comprises: The intelligent diagnosis subunit is used for carrying out three-dimensional depth separable convolution on the multi-parameter feature matrix containing the geographic tag by adopting ConvNeXt-T main network, extracting voiceprint-vibration-space coupling features and outputting fault types and confidence; The health evaluation subunit is used for fusing fault confidence, operation days, maintenance times, fault frequency and GIS position information at set time intervals by using an AHP dynamic weight model, quantitatively scoring and dividing four-level health states, and synchronously generating targeted maintenance suggestions.
7. The MFCC feature extraction and ConvNeXt-T network-integrated GIS voiceprint-vibration co-diagnostic system of claim 6, The intelligent diagnosis subunit includes: the depth feature extraction module adopts ConvNeXt-T three-layer depth separable convolution and residual blocks to extract voiceprint-vibration-space coupling features of the multi-parameter feature matrix containing the geographic tag; the global pooling and classifying module is used for obtaining 256-dimensional feature vectors through global average pooling, and outputting N-class GIS fault probabilities through a full connection layer and Softmax, wherein the GIS faults comprise complex faults such as loose contacts, metal particle discharge, component dislocation and the like; And the diagnosis register module is used for latching the identified fault type and the corresponding fault confidence level for the health assessment subunit to read in real time.
8. The MFCC feature extraction and ConvNeXt-T network-integrated GIS voiceprint-vibration co-diagnostic system of claim 6, A health assessment subunit comprising: the dynamic weight calculation module is used for fusing fault confidence, operation days, maintenance times, fault frequency and GIS position information every 100ms by using the AHP model and updating the five weights in real time; The health scoring module is used for quantifying and multiplying each index by corresponding weight according to a normalized interval [0,1], and obtaining equipment health score S, S epsilon [0,100]; The grading module is used for grading the health grade of the equipment into four grades, namely health, sub-health, fault early warning and serious fault according to the grading threshold value, and outputting corresponding maintenance suggestions.
9. The MFCC feature extraction and ConvNeXt-T network-integrated GIS voiceprint-vibration co-diagnostic system of claim 6, The intelligent diagnosis and health evaluation unit further comprises a storage unit which is used for adopting an edge-cloud hierarchical storage architecture, wherein real-time data are stored in a local storage of the integrated storage and calculation architecture, and history data are synchronized to a distributed database of the cloud to realize local quick access and cloud backup of the data.
10. The MFCC feature extraction and ConvNeXt-T network combined GIS voiceprint-vibration co-diagnostic system of any one of claims 1-9, The multi-parameter acquisition unit, the memory integrated characteristic processing unit and the intelligent diagnosis and health evaluation unit are interconnected by adopting TSN time sensitive Ethernet, the end-to-end jitter is less than or equal to 1 mu s, and the deterministic synchronous transmission of voiceprint, vibration and positioning data in a distributed link is ensured.

Description

GIS voiceprint-vibration collaborative diagnosis system integrating MFCC feature extraction and ConvNeXt-T network Technical Field The invention relates to the technical field of power equipment monitoring, in particular to a GIS voiceprint-vibration collaborative diagnosis system integrating MFCC feature extraction and ConvNeXt-T network. Background GIS (gas insulated switchgear) is a core device in a high-voltage power system, and mainly comprises the steps of integrally packaging a plurality of electric elements such as a circuit breaker, a disconnecting switch and the like, filling gas with excellent insulating property (such as sulfur hexafluoride), realizing safe control, reliable isolation and efficient transmission of high-voltage electric energy, simultaneously greatly reducing the volume of the device, and being suitable for scenes with limited space or high requirements on power supply reliability such as urban substations, hydropower stations and the like. The existing GIS monitoring technology has obvious defects that firstly sound vibration monitoring and calculation are separated, characteristic processing delay is high, energy efficiency is poor, transient faults are difficult to deal with, secondly single algorithm or single parameter diagnosis accuracy is insufficient, robustness to complex fault identification is poor, thirdly health assessment lacks dynamic weight adaptation and cannot accurately match a multi-working condition scene, and thirdly system architecture modularization is insufficient, multi-link cooperativity is poor, and full-link closed-loop management is difficult to form. Disclosure of Invention In view of the above analysis, the present invention aims to disclose a GIS voiceprint-vibration collaborative diagnosis system which fuses MFCC feature extraction and ConvNeXt-T network, so as to solve the problems set forth in the above background art. The invention discloses a GIS voiceprint-vibration collaborative diagnosis system which fuses MFCC feature extraction and ConvNeXt-T network, comprising: The multi-parameter acquisition unit is used for synchronously acquiring geographic position information, voiceprint data and vibration data of the GIS equipment, and performing pre-emphasis, framing and windowing preprocessing on the acquired voiceprint data to generate voiceprint data frames; The integrated memory-calculation feature processing unit adopts an SRAM integrated memory-calculation architecture, receives voiceprint data frames, extracts MFCC features, receives vibration data, extracts HHT features, and performs fusion and storage of the MFCC features and the HHT features according to received geographic position information to obtain a multi-parameter feature matrix containing geographic tags; The intelligent diagnosis and health assessment unit performs end-to-end reasoning on the multi-parameter feature matrix by adopting ConvNeXt-T network, outputs fault type and confidence coefficient, and determines dynamic weight by adopting a hierarchical analysis method in combination with fault confidence coefficient, equipment operation time length, maintenance record, fault frequency, severity and space position information, calculates equipment health score in real time, classifies health grade and generates maintenance advice. Further, the multi-parameter acquisition unit includes: the voiceprint acquisition module is used for acquiring broadband voiceprint data of the GIS equipment by adopting a high-sensitivity microphone, and carrying out preliminary MFCC (frequency division multiplexing) feature extraction pretreatment on the voiceprint data to generate an initial acoustic feature vector; the vibration acquisition module is used for acquiring time-varying data of vibration acceleration of the GIS equipment by adopting a piezoelectric acceleration sensor; And the GIS positioning module is used for acquiring the geographic position information of the GIS equipment by adopting a GPS/Beidou dual-mode positioning technology. Further, the voiceprint acquisition module comprises an MFCC feature extraction front unit; The MFCC feature extraction pre-unit executes pre-emphasis, framing and windowing in the MFCC algorithm, hardens the pre-emphasis, framing and windowing into micro logic of time domain running water and zero SRAM, and outputs voiceprint data frames for subsequent Mel frequency analysis. Further, the integrated memory feature processing unit adopts an integrated memory SRAM architecture, and includes: The MFCC feature extraction subunit is used for extracting Mel frequency cepstrum coefficient from the voiceprint data frame output by the MFCC feature extraction front unit, and generating MFCC features through processing such as fast Fourier transform, mel filtering and cepstrum transform; The vibration time-frequency characteristic extraction subunit is used for carrying out time-frequency analysis on the time-varying data of the collected vibration accelerati