Intelligent Processing and Health Assessment Method and System for Multi-parameter Acoustic-Vibration Data in GIS

Abstract

This invention relates to a method and system for intelligent processing and health assessment of multi-parameter acoustic-vibration data in GIS, belonging to the field of power equipment monitoring technology. The method includes: collecting acoustic signature data, vibration data, and geographic location information from GIS equipment to form a raw acoustic-vibration data stream with geographic tags; writing the raw acoustic-vibration data stream into a storage-computing architecture, sequentially extracting and fusing features from the acoustic signature data and vibration data to obtain a multi-parameter feature matrix with geographic tags; using a lightweight convolutional neural network to infer and label suspected faults on the multi-parameter feature matrix; using a support vector machine (SVM) model to identify faults in the multi-parameter feature matrix with suspected fault labels, obtaining the final fault type and confidence level; and combining equipment runtime, maintenance records, fault frequency, and GIS location information to perform a health assessment of the GIS equipment and classify fault levels. This invention significantly improves the accuracy of GIS fault identification and reduces the operation and maintenance costs of large-scale GIS applications.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, and in particular to a method and system for intelligent processing and health assessment of GIS multi-parameter acoustic-vibration data. Background Technology

[0002] Gas-insulated switchgear (GIS) is a critical piece of equipment in the power grid. Its failures mainly stem from mechanical defects (such as poor contact of contacts or loose parts causing vibration) and insulation defects (such as partial discharge caused by metal particles). These two types of defects interact and exacerbate the failures, threatening the stability of the power system. Traditional monitoring relies on "vibration + sound" as the core method, including manual inspection and devices such as vibration sensors and microphones. GIS, due to its long-distance and large-scale characteristics, is suitable for distributed detection. Furthermore, there is already a research foundation for related technologies. In vibration monitoring, the differences in vibration frequencies and mechanisms between normal and fault conditions have been clearly defined, and algorithm research results have been achieved. Voiceprint recognition has expanded from personnel identification to power equipment monitoring. The mainstream approach uses MFCC to extract features and combines them with AI models and dimensionality reduction algorithms to improve the accuracy of GIS fault identification.

[0003] Existing GIS monitoring technologies have significant drawbacks. Manual inspection is difficult to identify faults and is inefficient. Traditional sensors can only collect data but cannot analyze it quickly, requiring extensive manual processing, resulting in poor timeliness and difficulty in responding to sudden faults. Although GIS is suitable for distributed detection, traditional wired sensors require complex cable layouts. Current detection is still mainly based on live detection, which is labor-intensive and time-consuming for probe installation and cable laying, limiting system efficiency. Even if multi-sensor online monitoring systems are installed in critical areas, on-site cable adjustments and manual data interpretation are still required, and daily maintenance costs are high, making economical operation difficult. At the same time, traditional monitoring systems have a "storage-computation separation" structure, with long distances between data storage and computation, resulting in low computational throughput and poor energy efficiency, making it impossible to efficiently process acoustic-vibration data in real time. Summary of the Invention

[0004] Based on the above analysis, the present invention aims to disclose a method and system for intelligent processing and health assessment of GIS multi-parameter acoustic-vibration data, in order to solve the problems mentioned in the background art.

[0005] This invention discloses a method for intelligent processing and health assessment of multi-parameter acoustic-vibration data in GIS, comprising:

[0006] S1. Use the sound sensor, vibration sensor and GIS positioning module set at the edge to collect the acoustic data, vibration data and geographical location information of the GIS device respectively, and add a unified timestamp to the three types of data to form a raw sound-vibration data stream with geographical tags;

[0007] S2. Write the original acoustic-vibration data stream into the in-memory computing architecture at the edge, and sequentially complete the feature extraction and fusion of acoustic data and vibration data within the same chip to obtain a multi-parameter feature matrix with geographic tags; and perform lightweight convolutional neural network inference on the multi-parameter feature matrix to mark suspected faults.

[0008] S3. The multi-parameter feature matrix with suspected fault markers is used to identify faults through a support vector machine (SVM) model to obtain the final fault type and confidence level.

[0009] S4. Combining equipment runtime, maintenance records, fault frequency, and GIS location information, the Analytic Hierarchy Process (AHP) is used to dynamically assign weights to conduct a health assessment of the GIS equipment and classify fault levels.

[0010] Furthermore, in S1, the sound sensor, vibration sensor, and GIS positioning module at the edge are deployed in the disconnecting switch, busbar air chamber, or grounding switch of the GIS equipment;

[0011] The sound sensor is a high-sensitivity microphone with a sampling frequency ≥44.1kHz, used to acquire broadband acoustic data of GIS equipment;

[0012] The vibration sensor is a piezoelectric accelerometer with a sampling frequency of ≥10kHz, used to acquire time-varying vibration acceleration data of GIS equipment;

[0013] The GIS positioning module adopts GPS / BeiDou dual-mode positioning technology, which provides accurate geographic location information and uses GPS / BeiDou 1PPS pulse as a reference unified timestamp for the collection of voiceprint data and vibration data. It integrates voiceprint data, vibration data and geographic location information into a raw sound-vibration data stream with geographic tags.

[0014] Further, S2 includes:

[0015] S2-1. Write the original acoustic-vibration data stream into the in-memory SRAM chip;

[0016] S2-2. Sequentially complete the extraction of voiceprint MFCC features, vibration HHT time-frequency features, and GIS location fusion within the in-memory computing SRAM chip to obtain a multi-parameter feature matrix containing geographic tags.

[0017] S2-3. Use a lightweight convolutional neural network to infer the feature matrix and output suspected fault markers.

[0018] Furthermore, the in-memory computing SRAM chip contains multiple 8TiC computing units and 6T-SRAM storage units;

[0019] In the voiceprint MFCC feature extraction in S2-2, the 8TiC computing unit performs in-situ pipelined computation on the input voiceprint data stream, and the data does not need to be moved out of the storage area; 12-20 dimensional MFCC feature vectors are extracted through pre-emphasis / framing, FFT power spectrum, Mel filtering and DCT transformation.

[0020] In the vibration HHT time-frequency feature extraction in S2-2, the 8TiC computing unit performs in-situ Hilbert-Huang transform on the vibration acceleration data stream, and sequentially completes EMD empirical mode decomposition, Hilbert transform and instantaneous feature calculation, outputting instantaneous frequency, instantaneous amplitude envelope and Hilbert spectral statistics;

[0021] In S2-2, the GIS location fusion uses a 6T-SRAM array to receive acoustic and vibration feature vectors while simultaneously latching the latitude, longitude, and altitude of the GIS positioning module according to the corresponding sampling timestamp. The 8TiC unit encodes the location as a grid ID through WL pulses and splices it with the feature vector in the charge domain to form an integrated "acoustic-vibration-position" multi-parameter feature matrix.

[0022] Furthermore, in the S2-3 lightweight convolutional neural network inference, the in-memory computing array treats the multi-parameter feature matrix as the input feature map and performs MobileNet depthwise separable convolution in situ: suspected faults are marked; when a suspected fault is identified, the fault flag position is set to "1", otherwise the fault flag position is set to "0".

[0023] Furthermore, S3, the multi-parameter feature matrix with suspected fault markers is used to identify faults through a support vector machine (SVM) model to obtain the final fault type and confidence level; including:

[0024] S3-1. Select samples marked as 1 to form a subset of suspected fault features, and standardize the subset so that the mean of each dimension is 0 and the variance is 1.

[0025] S3-2. Input the standardized feature subset into a pre-trained RBF kernel SVM classifier, which has determined the optimal penalty parameter C and kernel width γ through grid search.

[0026] S3-3. Employ a one-to-many voting strategy to output the label of the fault type with the highest probability, including three categories: loose contacts, metal particle discharge, and component misalignment.

[0027] S3-4. Based on Platt scaling, the SVM decision value is transformed into a posterior probability, which is used as the confidence level of the fault type and output along with the fault type.

[0028] Furthermore, S4 includes the following sub-steps:

[0029] S4-1. Construct a set of health assessment indicators: use failure frequency, failure severity, equipment runtime, and maintenance record missingness as assessment indicators.

[0030] S4-2. Establish the AHP judgment matrix using the 1–9 scaling method, and dynamically calculate the weight vector W of each indicator through consistency test;

[0031] S4-3. Quantize each indicator according to the normalized interval [0,1] and multiply it by the corresponding weight, and then sum the weighted values ​​to obtain the equipment health score S, S∈[0,100].

[0032] S4-4. Based on the scoring threshold, the equipment health level is divided into four levels: healthy, sub-healthy, fault warning, and serious fault; and corresponding maintenance suggestions are output.

[0033] Furthermore, the edge device supports LoRa, NB-IoT and Ethernet communication methods and has a "fault data priority transmission" mechanism: when a suspected fault is identified and the fault flag position is "1", a high-priority transmission channel is triggered to ensure that the fault data is uploaded in real time before fault identification and health assessment are performed.

[0034] Furthermore, an edge-cloud tiered storage architecture is adopted: real-time data is stored in the local storage of the edge computing architecture, while historical data is synchronized to the distributed database in the cloud, enabling fast local access and cloud backup of data.

[0035] This invention also discloses a system for implementing the intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data as described above, comprising:

[0036] The system comprises a data acquisition unit, an edge computing and storage processing unit, and an analysis, diagnosis, and health assessment unit; among which,

[0037] The data acquisition unit uses sound sensors, vibration sensors, and GIS positioning modules set at the edge to collect acoustic data, vibration data, and geographic location information of the GIS device, and adds a unified timestamp to the three types of data to form a raw acoustic-vibration data stream with geographic labels.

[0038] The edge-end in-memory computing unit writes the raw acoustic-vibration data stream into the edge-end in-memory computing architecture, and sequentially completes the feature extraction and fusion of acoustic and vibration data within the same chip to obtain a multi-parameter feature matrix with geographic tags; and performs lightweight convolutional neural network inference on the multi-parameter feature matrix to mark suspected faults.

[0039] The analysis, diagnosis, and health assessment unit uses a support vector machine (SVM) model to identify faults and obtain the final fault type and confidence level. It also uses the analytic hierarchy process (AHP) to dynamically assign weights based on equipment runtime, maintenance records, fault frequency, and GIS location information to conduct a health assessment of the GIS equipment and classify fault levels.

[0040] This invention can achieve one of the following beneficial effects:

[0041] This invention is adapted to the distributed nature of GIS, and the integrated storage and computing architecture reduces the need for cable laying. Automated analysis reduces manual workload and significantly reduces the operation and maintenance costs of large-scale GIS applications.

[0042] Among them, by simultaneously collecting multi-dimensional data of "soundprint + vibration + geographical location", it breaks through the limitations of traditional single-parameter monitoring and provides a basis for accurate fault diagnosis; it directly locates the specific equipment or component of GIS faults, solves the problem of "knowing the fault exists but not knowing the location" in traditional monitoring, and improves the efficiency of fault tracing.

[0043] It adopts an in-memory computing SRAM architecture to eliminate the latency and energy consumption of traditional in-memory computing separation, improve computing throughput and energy efficiency, and ensure the real-time performance of high-frequency data processing;

[0044] It can complete acoustic-vibration feature extraction, multi-parameter fusion and lightweight neural network preliminary identification at the edge, without relying on remote computing power, and can respond to GIS instantaneous faults in milliseconds.

[0045] A two-level algorithm architecture of "lightweight neural network for preliminary screening + parameter-optimized SVM model for accurate classification" is adopted, which significantly improves the accuracy of GIS fault identification.

[0046] Multiple indicators are quantified into health scores and levels, and maintenance suggestions are generated to promote the transformation of GIS operation and maintenance from "passive emergency repair" to "proactive prevention".

[0047] It supports multiple communication standards and has a "fault data priority transmission" mechanism, which can prioritize uploading fault data when the network is congested, thus saving time for timely maintenance.

[0048] By adopting a "edge-cloud" tiered storage approach, local storage ensures real-time data access speed while cloud backups guarantee historical data security and multi-device collaborative analysis. Attached Figure Description

[0049] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0050] Figure 1 This is a flowchart of the intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data in an embodiment of the present invention;

[0051] Figure 2 This is a typical SRAM in-memory computing architecture diagram in an embodiment of the present invention;

[0052] Figure 3 This is a comparison diagram of the von Neumann architecture and in-memory computing architecture in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram showing the composition and connection of the GIS multi-parameter acoustic-vibration data intelligent processing and health assessment system in an embodiment of the present invention. Detailed Implementation

[0054] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and, together with the embodiments of the present invention, serve to illustrate the principles of the present invention.

[0055] Example 1

[0056] One embodiment of the present invention discloses a method for intelligent processing and health assessment of GIS multi-parameter acoustic-vibration data, such as... Figure 1 As shown,

[0057] S1. Use the sound sensor, vibration sensor and GIS positioning module set at the edge to collect the acoustic data, vibration data and geographical location information of the GIS device respectively, and add a unified timestamp to the three types of data to form a raw sound-vibration data stream with geographical tags;

[0058] S2. Write the original acoustic-vibration data stream into the in-memory computing architecture at the edge, and sequentially complete the feature extraction and fusion of acoustic data and vibration data within the same chip to obtain a multi-parameter feature matrix with geographic tags; and perform lightweight convolutional neural network inference on the multi-parameter feature matrix to mark suspected faults.

[0059] S3. The multi-parameter feature matrix with suspected fault markers is used to identify faults through a support vector machine (SVM) model to obtain the final fault type and confidence level.

[0060] S4. Combining equipment runtime, maintenance records, fault frequency, and GIS location information, the Analytic Hierarchy Process (AHP) is used to dynamically assign weights to conduct a health assessment of the GIS equipment and classify fault levels.

[0061] Specifically, in S1, the sound sensor, vibration sensor, and GIS positioning module at the edge are deployed in the disconnecting switch, busbar air chamber, or grounding switch of the GIS equipment;

[0062] The sound sensor is a high-sensitivity microphone with a sampling frequency ≥44.1kHz, used to acquire broadband acoustic data of GIS equipment;

[0063] Preferably, the high-sensitivity microphone is an electret microphone, which has low noise and high signal-to-noise ratio characteristics and can effectively collect abnormal noise signals during equipment operation;

[0064] The vibration sensor is a piezoelectric accelerometer with a sampling frequency ≥10kHz, which has high sensitivity and a wide frequency response range, and is used to acquire time-varying vibration acceleration data of GIS equipment.

[0065] The GIS positioning module adopts GPS / BeiDou dual-mode positioning technology with a positioning accuracy of ≤5 meters. While providing accurate geographic location information, it uses GPS / BeiDou 1PPS pulses as a reference unified timestamp for the collection of acoustic and vibration data. It integrates acoustic data, vibration data, and geographic location information into a raw acoustic-vibration data stream with geographic tags.

[0066] Specifically, S2 includes:

[0067] S2-1. Write the original acoustic-vibration data stream into the in-memory SRAM chip;

[0068] S2-2. Sequentially complete the extraction of voiceprint MFCC (Mel frequency cepstral coefficients), vibration HHT (Hilbert-Huang transform) time-frequency features, and GIS location fusion within the in-memory SRAM chip to obtain a multi-parameter feature matrix containing geographic tags.

[0069] S2-3. A lightweight convolutional neural network is used to infer the feature matrix and output suspected fault markers. The in-memory SRAM chip contains multiple 8TiC computing units and 6T-SRAM storage units.

[0070] like Figure 2 As shown, each 8TiC computing unit is an 8-transistor SRAM unit, which adds computing function transistors compared to the 6T-SRAM unit, and can perform binary multiplication operations. The computing function is realized by reconstructing the read / write path.

[0071] The 8TiC computing unit and the 6T-SRAM memory unit are interconnected through word lines, bit lines, complementary bit lines and global bit lines to achieve physical integration of data storage and computing, reducing latency and energy consumption caused by data transfer.

[0072] like Figure 3 As shown, compared with the traditional von Neumann architecture (the left side of the figure shows the traditional von Neumann architecture, and the right side shows the in-memory computing architecture), the in-memory computing architecture integrates storage and computing functions, avoiding the time and energy loss caused by frequent data transfer between memory and arithmetic unit in the traditional architecture. This significantly improves computing throughput and energy efficiency, providing strong computing and control support for the stable operation of the entire monitoring device.

[0073] In the voiceprint MFCC feature extraction in S2-2, the 8TiC computing unit performs in-situ pipelined computation on the input voiceprint data stream, and the data does not need to be moved out of the storage area; 12-20 dimensional MFCC feature vectors are extracted through pre-emphasis / framing, FFT power spectrum, Mel filtering and DCT transformation.

[0074] In the pre-emphasis / framing process, coefficients are stored in the form of charge. The voiceprint data stream is serially input through the bit line BL, and multiplication and addition are performed in the current domain. Frame boundaries are controlled by address jumps through the word line WL. During FFT power spectrum calculation, 8 TiC computing units form a CORDIC rotating array. Complex rotation is performed in parallel in the charge domain, and the real / imaginary part results are stored in place. In Mel filtering, the triangular filter coefficients are fixed row by row, and the power spectrum is broadcast to each row through BL. Each channel is weighted and summed in parallel, and the result is normalized by WL current mirroring. In DCT transform, the DCT coefficient matrix is ​​stored in an array. The Mel spectrum is input through WL pulse width modulation, and matrix multiplication is output by BL charge accumulation, supporting dynamic dimension selection. Output: 12-20 dimensional MFCC feature vector.

[0075] In the vibration HHT time-frequency feature extraction in S2-2, the 8TiC computing unit performs in-situ Hilbert-Huang transform on the vibration acceleration data stream. The data does not need to be moved out of the storage array. The EMD empirical mode decomposition, Hilbert transform, and instantaneous feature calculation are completed sequentially, and the instantaneous frequency, instantaneous amplitude envelope, and Hilbert spectral statistics are output. The steps are as follows:

[0076] a) EMD decomposition: The 6T-SRAM array buffers vibration sampling sequences row by row, and the 8TiC cells operate in a four-stage pipeline of "extreme value detection - envelope fitting - mean value calculation - intrinsic mode (IMF) screening":

[0077] Extreme value detection is triggered by a sign transition between adjacent cells, and the maximum / minimum value positions are latched at WL level;

[0078] Envelope fitting uses distributed CORDIC to solve cubic spline coefficients. The coefficients are fixed in columns in the form of charge. The extreme value positions are broadcast by BL and spline interpolation is performed in parallel. The upper and lower envelopes are output in the current domain.

[0079] The mean value is obtained by using an 8TiC difference pair structure, subtracting the currents in the charge domain from the upper and lower envelopes to obtain the mean value sequence, which is then written back in situ.

[0080] IMF screening compares the mean energy with the threshold through a gated loop. It stops when SD < 0.25, and the remaining amount is used as the input for the next order. The loop continues until the residual is monotonic, and finally outputs the first N order IMFs (N = 3 to 6). The order index is encoded by WL pulse counting to achieve dynamic order configuration.

[0081] b) Hilbert Transform: Each IMF undergoes FFT → phase shift → IFFT to complete a 90° phase shift. The FFT / IFFT uses a CORDIC rotating array, and the rotation factor coefficients are fixed row by row. The real and imaginary parts are operated in parallel in the charge domain, and the phase shift result is stored in place. Subsequently, 8 TiC units realize "IMF in-phase - orthogonal" complex pairs through differential pairs to construct an analytical signal.

[0082] c) Instantaneous characteristic calculation: The amplitude and phase of the analytic signal are solved in parallel in the current domain:

[0083] Amplitude envelope: Using the CORDIC vector modulus mode, the point-by-point calculation results are directly cached in the corresponding 6T cell;

[0084] Instantaneous phase: The CORDIC rotating mode is used, and the phase difference is followed by unwrap processing in the charge domain;

[0085] Instantaneous frequency: It is achieved by the charge difference between adjacent cells and the current mirror integration, and Δt is calibrated by the global bit line GBL clock period;

[0086] d) Hilbert spectral statistics: Performed within the array using a two-dimensional frequency-amplitude histogram: the frequency axis is segmented by 1 / 3 octave bands, with fixed segment boundary coefficients; the amplitude axis is logarithmically quantized. The f(t) and A(t) values ​​for each IMF are broadcast via BL and accumulated in parallel to the corresponding cell. The cell counter is selected by WL, and the final output is:

[0087] Instantaneous frequency mean and variance (6 dimensions, first 3 orders of IMF)

[0088] Instantaneous amplitude envelope peak value, RMS value, kurtosis (6 dimensions, first 3 orders of IMF)

[0089] Hilbert spectral energy entropy, barycentric frequency (2D)

[0090] It has a total of 14-dimensional HHT time-frequency feature vectors, and no data needs to be moved out of the storage area throughout the process. The single-frame latency is <1ms and the power consumption is <0.8pJ / sample point.

[0091] In S2-2, the GIS location fusion involves a 6T-SRAM array that receives acoustic and vibration feature vectors while latching the latitude, longitude, and altitude of the GIS positioning module according to the corresponding sampling timestamp. The 8TiC unit encodes the location into a grid ID with an accuracy of ≤5m using WL pulses and splices it with the feature vector in the charge domain to form an integrated "acoustic-vibration-position" multi-parameter feature matrix. After being output in situ, it can be directly used by the subsequent lightweight neural network without the need for additional data transfer.

[0092] In the S2-3 lightweight convolutional neural network inference, the in-memory computing array treats the "sound-vibration-position" multi-parameter feature matrix as a 1×D input feature map, and performs MobileNet depthwise separable convolution in situ.

[0093] Deep convolution: 3×1 convolution kernel coefficients are fixed row by row, feature maps are broadcast via BL, each channel is multiplied and added in parallel in the charge domain, and ReLU6 activation is achieved by WL voltage clamping;

[0094] Point convolution: 1×1 convolution kernels complete channel fusion in the form of difference pairs, outputting 16-dimensional latent features. BatchNorm scaling bias is merged into the convolution kernel, and the mean and variance are fixed during the training phase.

[0095] Global average pooling: The current of each channel is integrated by mirroring and then discharged uniformly to directly output a scalar;

[0096] Softmax: Calculates the exponent and normalizes the CORDIC vector pattern. The category with the highest probability is the "suspected fault marker." The fault flag is set to "1" when the confidence level is ≥0.7; otherwise, it is set to "0." Typical fault types include loose contacts, metal particle discharge, and component misalignment.

[0097] Preferably, the edge device supports LoRa, NB-IoT and Ethernet communication methods and has a "fault data priority transmission" mechanism: when a suspected fault is identified and the fault flag position is "1", a high-priority transmission channel is triggered to ensure that the fault data is uploaded in real time before fault identification and health assessment are performed; the target of the edge device data upload can be a locally set processing terminal or the cloud.

[0098] With a total parameter count of less than 500,000, a computation latency of less than 2ms, and a power consumption of less than 1.2pJ / MAC, "real-time edge screening" can be completed without moving data out of the storage area.

[0099] Specifically, S3, the multi-parameter feature matrix with suspected fault markers is used to identify faults through a Support Vector Machine (SVM) model to obtain the final fault type and confidence level; including:

[0100] S3-1. Select samples marked as 1 to form a subset of suspected fault features, and standardize the subset so that the mean of each dimension is 0 and the variance is 1.

[0101] S3-2. Input the standardized feature subset into a pre-trained RBF kernel SVM classifier, which has determined the optimal penalty parameter C and kernel width γ through grid search.

[0102] S3-3. Employ a one-to-many voting strategy to output the label of the fault type with the highest probability, including three categories: loose contacts, metal particle discharge, and component misalignment.

[0103] S3-4. Based on Platt scaling, the SVM decision value is transformed into a posterior probability, which is used as the confidence level of the fault type and output along with the fault type.

[0104] Specifically, the AHP judgment matrix described in S4 is jointly constructed by domain experts and historical data, supports online weight updates, and adapts to different operating environments and equipment aging stages.

[0105] S4 includes the following sub-steps:

[0106] S4-1. Construct a set of health assessment indicators: use failure frequency, failure severity, equipment runtime, and maintenance record missingness as assessment indicators.

[0107] Among them, the fault frequency and fault severity are obtained by statistical analysis of the fault identification structure of the support vector machine (SVM) model in S3;

[0108] Based on the output of the fault identification model unit, the fault frequency and fault severity are statistically analyzed.

[0109] S4-2. Establish the AHP judgment matrix using the 1–9 scaling method, and dynamically calculate the weight vector W of each indicator by passing the consistency test CR<0.1.

[0110] S4-3. Quantize each indicator according to the normalized interval [0,1] and multiply it by the corresponding weight, and then sum the weighted values ​​to obtain the equipment health score S, S∈[0,100].

[0111] S4-4. Based on the scoring threshold, the equipment health level is divided into four levels: healthy (S≥85), sub-healthy (70≤S<85), fault warning (50≤S<70), and serious fault (S<50), and corresponding maintenance suggestions are output.

[0112] In the exemplary maintenance recommendations, the inspection cycle for healthy status is extended to 6 months and standard monitoring is maintained; for sub-health status, encrypted data collection is carried out once per hour, with targeted retesting of abnormal frequency bands and dedicated personnel for tracking; for fault warning status, continuous data collection and edge-cloud collaborative diagnosis are triggered, and based on the fault types identified by SVM, such as loose contacts and metal discharge, live-line testing or disassembly and repair are scheduled within 36-72 hours, while simultaneously reducing the load rate; for severe fault status, the highest level alarm is automatically pushed, the equipment is immediately taken out of operation, emergency response is carried out within 2 hours, and the entire gas chamber is forcibly disassembled or expanded for inspection based on the degree of missing maintenance records. Afterwards, fault restoration analysis is completed and the model is updated within 72 hours, forming a closed-loop intelligent operation and maintenance system of "monitoring-diagnosis-maintenance-feedback".

[0113] In this embodiment, an edge-cloud hierarchical storage architecture is adopted: real-time data is stored in the local storage of the edge computing architecture, and historical data and diagnostic results are synchronized to the distributed database in the cloud, realizing fast local access and cloud backup of data.

[0114] In summary, the solution of this embodiment has the following beneficial effects:

[0115] By using high-sampling-frequency sound sensors, vibration sensors, and high-precision GIS positioning modules, multi-dimensional data of "soundprint + vibration + geographical location" are collected simultaneously, breaking through the limitations of traditional single-parameter monitoring and providing a basis for accurate fault diagnosis.

[0116] The GIS positioning module binds acoustic and vibration data with geographic location, which can directly locate the specific equipment or component with GIS faults, solving the problem of "knowing the fault exists but not knowing its location" in traditional monitoring, and improving the efficiency of fault tracing.

[0117] It adopts an integrated SRAM architecture of "8TiC computing units + 6T-SRAM storage array" to eliminate the latency and energy consumption of traditional storage and computing separation, improve computing throughput and energy efficiency, and ensure the real-time performance of high-frequency data processing.

[0118] It completes acoustic and vibration feature extraction, multi-parameter fusion, and preliminary identification by a lightweight neural network at the edge, without relying on remote computing power, and can respond to GIS instantaneous faults in milliseconds.

[0119] The edge device supports multiple communication standards and has a "fault data priority transmission" mechanism, which can prioritize uploading fault data when the network is congested, thus saving time for timely maintenance.

[0120] By adopting a "edge-cloud" tiered storage approach, local storage ensures real-time data access speed while cloud backups guarantee historical data security and multi-device collaborative analysis.

[0121] The two-level algorithm architecture of "lightweight neural network for preliminary screening + parameter-optimized SVM model for accurate classification" significantly improves the accuracy of GIS fault identification.

[0122] The health assessment process incorporates the analytic hierarchy process (AHP) to quantify multiple indicators into health scores and levels, and generates maintenance recommendations, thereby promoting the transformation of GIS operation and maintenance from "passive emergency repair" to "proactive prevention".

[0123] In summary, the modular design of this embodiment is adapted to the distributed characteristics of GIS, the in-store computing architecture reduces the need for cable laying, and the automated analysis reduces the workload of manual labor, significantly reducing the operation and maintenance costs of large-scale GIS applications.

[0124] Example 2

[0125] One embodiment of the present invention discloses a system for implementing the intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data as described in Embodiment 1, such as... Figure 4 As shown, it includes:

[0126] Data acquisition unit, edge computing and storage integrated processing unit, and analysis, diagnosis and health assessment unit;

[0127] in,

[0128] The data acquisition unit uses sound sensors, vibration sensors, and GIS positioning modules set at the edge to collect acoustic data, vibration data, and geographic location information of the GIS device, and adds a unified timestamp to the three types of data to form a raw acoustic-vibration data stream with geographic labels.

[0129] The edge-end in-memory computing unit writes the raw acoustic-vibration data stream into the edge-end in-memory computing architecture, and sequentially completes the feature extraction and fusion of acoustic and vibration data within the same chip to obtain a multi-parameter feature matrix with geographic tags; and performs lightweight convolutional neural network inference on the multi-parameter feature matrix to mark suspected faults.

[0130] The analysis, diagnosis, and health assessment unit uses a support vector machine (SVM) model to identify faults and obtain the final fault type and confidence level. It also uses the analytic hierarchy process (AHP) to dynamically assign weights based on equipment runtime, maintenance records, fault frequency, and GIS location information to conduct a health assessment of the GIS equipment and classify fault levels.

[0131] Figure 4 As shown, the data acquisition unit 1 includes:

[0132] The sound sensor 11 is a high-sensitivity microphone with a sampling frequency ≥44.1kHz, used to collect broadband acoustic data during the operation of GIS equipment;

[0133] Vibration sensor 12 is a piezoelectric accelerometer with a sampling frequency ≥10kHz, used to collect time-varying data of vibration acceleration.

[0134] The GIS positioning module 13 adopts GPS / BeiDou dual-mode positioning technology with a positioning accuracy of ≤5 meters. It is used to bind acoustic and vibration data with geographic location information to realize the spatial dimension association of multi-parameter data.

[0135] Edge-end in-memory processing unit 2 includes:

[0136] The acoustic and vibration data storage subunit 21 adopts an integrated SRAM architecture, which includes 8 TiC computing units and a 6T-SRAM storage array, realizing the physical integration of data storage and computing, and reducing the latency and energy consumption caused by data transfer.

[0137] The real-time analysis and calculation subunit 22 extracts acoustic features from the acoustic data using Mel frequency cepstral coefficients (MFCC) to generate 12-20 dimensional MFCC feature vectors; it performs time-frequency analysis on the vibration data using Hilbert-Huang transform (HHT) to extract vibration features such as instantaneous frequency and amplitude envelope; it integrates GIS positioning information with acoustic and vibration features to construct a multi-parameter feature matrix; and it uses a lightweight convolutional neural network to perform real-time preliminary classification of the multi-parameter feature matrix to identify suspected fault data.

[0138] The edge communication subunit 23 supports multiple communication methods including LoRa, NB-IoT, and Ethernet, and features a "fault data priority transmission" mechanism. When a suspected fault is detected, a high-priority channel is automatically triggered to ensure that fault data is uploaded in real time. For example, during peak grid load periods, even if the communication network is congested, fault data can still be transmitted with priority, buying time for timely maintenance.

[0139] Analysis of diagnostic and health assessment unit 3 includes:

[0140] Feature extraction unit 31 is used to further optimize the transmitted feature matrix;

[0141] The fault identification model unit 32 is used by the SVM model to accurately identify fault types (such as loose contacts, metal particle discharge, etc.) based on the trained classifier and output confidence scores.

[0142] Health assessment model unit 33 combines fault frequency, severity and equipment historical data, uses the analytic hierarchy process to determine weights and calculate health scores and classify levels;

[0143] Data storage unit 34 is used to complete hierarchical storage and backup of data.

[0144] Display and output unit 35 is used to output evaluation results and maintenance recommendations.

[0145] In this embodiment, the more specific technical details and corresponding technical effects are the same as those in Embodiment 1. Please refer to them for details, and they will not be repeated here.

[0146] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent processing and health assessment of GIS multi-parameter acoustic-vibration data, characterized in that, include: S1. Use the sound sensor, vibration sensor and GIS positioning module set at the edge to collect the acoustic data, vibration data and geographical location information of the GIS device respectively, and add a unified timestamp to the three types of data to form a raw sound-vibration data stream with geographical tags; S2. Write the original acoustic-vibration data stream into the in-memory computing architecture at the edge, and sequentially complete the feature extraction and fusion of acoustic data and vibration data within the same chip to obtain a multi-parameter feature matrix with geographic tags; and perform lightweight convolutional neural network inference on the multi-parameter feature matrix to mark suspected faults. S3. The multi-parameter feature matrix with suspected fault markers is used to identify faults through a support vector machine (SVM) model to obtain the final fault type and confidence level. S4. Combining equipment runtime, maintenance records, fault frequency, and GIS location information, the Analytic Hierarchy Process (AHP) is used to dynamically assign weights to conduct a health assessment of the GIS equipment and classify fault levels.

2. The intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data according to claim 1, characterized in that, In S1, the sound sensor, vibration sensor and GIS positioning module set at the edge are deployed in the disconnecting switch, busbar air chamber or grounding switch of the GIS equipment; The sound sensor is a high-sensitivity microphone with a sampling frequency ≥44.1kHz, used to acquire broadband acoustic data of GIS equipment; The vibration sensor is a piezoelectric accelerometer with a sampling frequency of ≥10kHz, used to acquire time-varying vibration acceleration data of GIS equipment; The GIS positioning module adopts GPS / BeiDou dual-mode positioning technology, which provides accurate geographic location information and uses GPS / BeiDou 1PPS pulse as a reference unified timestamp for the collection of voiceprint data and vibration data. Voiceprint data, vibration data, and geographic location information are fused into a raw acoustic-vibration data stream with geographic tags.

3. The intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data according to claim 1, characterized in that, S2 includes: S2-1. Write the original acoustic-vibration data stream into the in-memory SRAM chip; S2-2. Sequentially complete the extraction of voiceprint MFCC features, vibration HHT time-frequency features, and GIS location fusion within the in-memory computing SRAM chip to obtain a multi-parameter feature matrix containing geographic tags. S2-3. Use a lightweight convolutional neural network to infer the feature matrix and output suspected fault markers.

4. The intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data according to claim 3, characterized in that, The in-memory SRAM chip contains multiple 8TiC computing units and 6T-SRAM storage units; In the voiceprint MFCC feature extraction in S2-2, the 8TiC computing unit performs in-situ pipelined computation on the input voiceprint data stream, and the data does not need to be moved out of the storage area; 12-20 dimensional MFCC feature vectors are extracted through pre-emphasis / framing, FFT power spectrum, Mel filtering and DCT transformation. In the vibration HHT time-frequency feature extraction in S2-2, the 8TiC computing unit performs in-situ Hilbert-Huang transform on the vibration acceleration data stream, and sequentially completes EMD empirical mode decomposition, Hilbert transform and instantaneous feature calculation, outputting instantaneous frequency, instantaneous amplitude envelope and Hilbert spectral statistics; In S2-2, the GIS location fusion uses a 6T-SRAM array to receive acoustic and vibration feature vectors while simultaneously latching the latitude, longitude, and altitude of the GIS positioning module according to the corresponding sampling timestamp. The 8TiC unit encodes the location as a grid ID through WL pulses and splices it with the feature vector in the charge domain to form an integrated "acoustic-vibration-position" multi-parameter feature matrix.

5. The intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data according to claim 4, characterized in that, In the S2-3 lightweight convolutional neural network inference, the in-memory computing array treats the multi-parameter feature matrix as the input feature map and performs MobileNet depthwise separable convolution in situ: suspected faults are marked; when a suspected fault is identified, the fault flag is set to "1", otherwise the fault flag is set to "0".

6. The intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data according to claim 5, characterized in that, S3. The multi-parameter feature matrix with suspected fault markers is used to perform fault identification through a Support Vector Machine (SVM) model, yielding the final fault type and confidence level; including: S3-1. Select samples marked as 1 to form a subset of suspected fault features, and standardize the subset so that the mean of each dimension is 0 and the variance is 1. S3-2. Input the standardized feature subset into a pre-trained RBF kernel SVM classifier, which has determined the optimal penalty parameter C and kernel width γ through grid search. S3-3. Employ a one-to-many voting strategy to output the label of the fault type with the highest probability, including three categories: loose contacts, metal particle discharge, and component misalignment. S3-4. Based on Platt scaling, the SVM decision value is transformed into a posterior probability, which is used as the confidence level of the fault type and output along with the fault type.

7. The intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data according to claim 6, characterized in that, S4 includes the following sub-steps: S4-1. Construct a set of health assessment indicators: use failure frequency, failure severity, equipment runtime, and maintenance record missingness as assessment indicators. S4-2. Establish the AHP judgment matrix using the 1–9 scaling method, and dynamically calculate the weight vector W of each indicator through consistency test; S4-3. Quantize each indicator according to the normalized interval [0,1] and multiply it by the corresponding weight, and then sum the weighted values ​​to obtain the equipment health score S, S∈[0,100]. S4-4. Based on the scoring threshold, the equipment health level is divided into four levels: healthy, sub-healthy, fault warning, and serious fault; and corresponding maintenance suggestions are output.

8. The intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data according to claim 5, characterized in that, The edge device supports LoRa, NB-IoT and Ethernet communication methods and has a "fault data priority transmission" mechanism: when a suspected fault is identified and the fault flag is "1", a high-priority transmission channel is triggered to ensure that the fault data is uploaded in real time before fault identification and health assessment are performed.

9. The intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data according to claim 8, characterized in that, The system adopts an "edge-cloud" hierarchical storage architecture: real-time data is stored in the local storage of the edge computing architecture, while historical data is synchronized to the distributed database in the cloud, enabling fast local access and cloud backup of data.

10. A system for implementing the intelligent processing and health assessment method for GIS multi-parameter acoustic-vibration data as described in any one of claims 1-9, characterized in that, include: The system comprises a data acquisition unit, an edge computing and storage processing unit, and an analysis, diagnosis, and health assessment unit; among which, The data acquisition unit uses sound sensors, vibration sensors, and GIS positioning modules set at the edge to collect acoustic data, vibration data, and geographic location information of the GIS device, and adds a unified timestamp to the three types of data to form a raw acoustic-vibration data stream with geographic labels. The edge-end in-memory computing unit writes the raw acoustic-vibration data stream into the edge-end in-memory computing architecture, and sequentially completes the feature extraction and fusion of acoustic and vibration data within the same chip to obtain a multi-parameter feature matrix with geographic tags; and performs lightweight convolutional neural network inference on the multi-parameter feature matrix to mark suspected faults. The analysis, diagnosis, and health assessment unit uses a support vector machine (SVM) model to identify faults and obtain the final fault type and confidence level. It also uses the analytic hierarchy process (AHP) to dynamically assign weights based on equipment runtime, maintenance records, fault frequency, and GIS location information to conduct a health assessment of the GIS equipment and classify fault levels.

Images