Cim-based gis sound-vibration combined intelligent monitoring and fault prediction system and method

Abstract

The application relates to a GIS sound-vibration combined intelligent monitoring and fault prediction system and method based on CIM, and belongs to the technical field of power equipment monitoring. In the system, original signals including mechanical vibration and sound signals in GIS equipment are synchronously collected, multi-source sound-vibration data is output after Kalman filtering denoising and normalization processing through a built-in SRAM storage and calculation integrated computing structure; sound-vibration data received in real time is monitored to trigger a fault event, the time consistency of the multi-source sound-vibration data is ensured, and sound-vibration effective data is obtained through redundancy elimination and interference extraction; sound-vibration multi-domain features in the sound-vibration effective data are extracted for fault identification; a device health evaluation model is constructed to divide the health grades of the GIS equipment and output health prediction information; the fault identification result and the health prediction information are transmitted to a remote monitoring center, and remote regulation and control instructions are received. The application greatly improves the timeliness of data processing and can quickly cope with the identification demand of instantaneous faults.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, and in particular to a GIS-based acoustic-vibration joint intelligent monitoring and fault prediction system and method based on CIM. Background Technology

[0002] Gas-insulated switchgear (GIS) is a critical piece of equipment in the power grid, and its operating status directly affects the safe and stable operation of the power system. Mechanical vibration caused by mechanical defects and partial discharge caused by insulation defects are the two core causes of GIS equipment failures. Taking disconnect switches as an example, during frequent opening and closing operations, the impact and stress of mechanical closing forces can easily lead to poor contact of the contacts, resulting in continuous mechanical vibration, which may eventually cause serious equipment discharge failures. In addition, metal particles generated by contact wear are scattered throughout the gas chamber. Under the continuous action of electric field and mechanical vibration, they can jump to high field strength areas, causing partial discharge, which poses a potential threat to the insulation condition of the equipment. They may also spread to the surface of the insulator, causing surface flashover, accelerating insulation deterioration, and causing further deformation of the equipment due to thermal effects, inducing more serious mechanical defects. At the same time, different equipment manufacturers have differences in material selection and process level. In addition, various problems and defects are prone to occur during the design and installation process. After the equipment has been running for a long time, it is very easy to have typical faults such as abnormal contact of switch contacts or busbar joints, unbalanced housing connection, slight bending of guide rods, misalignment of guide rod connection, and loose parts.

[0003] In GIS monitoring, traditional methods typically combine vibration and sound, a method relied upon in traditional manual maintenance. However, traditional manual maintenance suffers from drawbacks such as difficulty in manually identifying faults and low maintenance efficiency. Even when using traditional vibration sensors and microphones, the field devices only have the function of collecting vibration data and cannot quickly identify and analyze faults. The collected data still requires extensive manual analysis, greatly reducing timeliness and failing to meet the need for rapid identification of instantaneous faults.

[0004] Although GIS possesses the structural characteristics of long distances and large scales, making it suitable for distributed monitoring applications, current defect detection methods are still primarily based on live-line testing due to limitations imposed by the traditional wired sensor cable layout requirements. This method not only consumes significant manpower and resources, but the time-consuming cable laying and probe installation process also severely limits the efficiency of the detection system. Furthermore, even when using online monitoring systems based on multiple sensors for critical GIS components, the cable laying method still needs to be adjusted according to the actual site conditions, and the collected data requires manual interpretation. Coupled with high daily maintenance costs, it is difficult to achieve the goal of economical operation. Therefore, developing more intelligent and efficient GIS monitoring technology has become an urgent need in the current power equipment monitoring field.

[0005] Traditional GIS monitoring technologies suffer from problems such as low fault identification efficiency, poor timeliness, high cost, and difficulty in achieving economical operation. Therefore, solutions are needed. Summary of the Invention

[0006] In view of the above analysis, the present invention aims to disclose a GIS-based acoustic-vibration joint intelligent monitoring and fault prediction system and method based on CIM, so as to solve the problems mentioned in the background art.

[0007] This invention discloses a CIM-based GIS sound-vibration joint intelligent monitoring and fault prediction system, including a system installed at the edge.

[0008] The in-memory computing unit is used to simultaneously acquire raw signals, including mechanical vibration and sound signals, from multiple locations in GIS equipment. Through the built-in SRAM in-memory computing structure, the raw signals are processed by Kalman filtering for noise reduction and normalization, and then multi-source sound and vibration data are output.

[0009] The fault triggering and synchronization processing module is used to monitor the received acoustic-vibration data in real time based on the fault triggering threshold to trigger fault events. It uses a timestamp synchronization method to ensure the time consistency of acoustic-vibration data from multiple sources, and obtains valid acoustic-vibration data through redundancy elimination and interference extraction.

[0010] The fault identification and prediction module is used to extract multi-domain features of acoustic-vibration from effective acoustic-vibration data for fault identification; and to build an equipment health assessment model to classify the health level of GIS equipment and output health prediction information.

[0011] The communication module is used to transmit fault identification results and health prediction information to the remote monitoring center, and to receive remote control commands, including adjusting the collected parameters and fault trigger thresholds.

[0012] Furthermore, the integrated storage and computing sound-vibration acquisition unit is deployed at least one of the disconnecting switch, busbar air chamber, and grounding switch in the GIS equipment;

[0013] Each in-memory acoustic-vibration acquisition unit includes a piezoelectric accelerometer, an electret microphone, and an SRAM in-memory computing structure;

[0014] Piezoelectric accelerometers collect mechanical vibration signals and electret microphones collect sound signals, which are then input into an SRAM in-memory computing architecture.

[0015] The SRAM in-memory computing architecture includes multiple 8TiC computing units and 6T-SRAM storage units;

[0016] The 8TiC computing unit and the 6T-SRAM storage unit are interconnected through word lines, bit lines, complementary bit lines and global bit lines. While collecting vibration and sound signals, the Kalman filter noise reduction and [0,1] interval normalization processing are performed in the storage unit.

[0017] Furthermore, the process of signal acquisition and processing by the in-memory integrated acoustic-vibration acquisition unit includes:

[0018] The raw signals collected by the piezoelectric accelerometer and electret microphone are converted into digital signals by the interface circuit and then written into the 6T-SRAM memory cell array through word line gating.

[0019] The 8TiC computing unit reads data from the storage unit in real time through the bit line, and performs Kalman filtering recursive operation in parallel within the storage array to eliminate random noise. The filtered data is normalized in the [0,1] interval by finding the maximum and minimum values ​​and linear scaling within the storage unit.

[0020] The processed valid data is output to the fault triggering and synchronization processing module via the global bit line.

[0021] Furthermore, the fault triggering and synchronization processing module includes:

[0022] The threshold triggering submodule is used to set the normal range threshold of vibration and sound signals according to the instructions of the remote monitoring center, and to monitor the received sound-vibration data in real time. When the data continuously exceeds the threshold, a fault event is triggered.

[0023] The data synchronization submodule is used to achieve consistent synchronization of multi-source data in the time dimension by adding precise timestamps to each set of vibration and sound data.

[0024] The interference cancellation submodule is used to analyze the synchronized data using wavelet transform, eliminate redundant information and interference components, extract effective acoustic-vibration data related to equipment faults, and output them to the edge fault identification and prediction module.

[0025] Furthermore, the fault identification and prediction module includes:

[0026] The fault identification module is used to perform multi-domain feature extraction and principal component analysis dimensionality reduction on effective acoustic-vibration data. It uses an SVM classifier and a ConvNeXt-T network in parallel to identify fault types and outputs fault type identification results.

[0027] The health prediction module is used to build a dynamic weighted health assessment model based on acoustic print, vibration, equipment, environment and operating condition data. It determines the weight of each influencing factor through cluster analysis and association reasoning, and uses an optimization algorithm to dynamically adjust the weights to classify the equipment health level and output health prediction information.

[0028] Furthermore, the fault identification module includes:

[0029] The feature extraction submodule is used to extract sound signal features using the Mel frequency cepstral coefficient (MFCC) method, and to extract vibration signal features using the Hilbert-Huang transform time-frequency analysis method and the high-frequency harmonic vibration amplitude ratio method, thus obtaining sound-vibration features.

[0030] The dimensionality reduction and optimization submodule is used to perform dimensionality reduction processing on the extracted acoustic-vibration features through principal component analysis;

[0031] The fault classification submodule is used to input the dimensionality-reduced features into the SVM classifier and the ConvNeXt-T network for parallel recognition, and then obtain and output the fault type after fusing the recognition results.

[0032] Furthermore, the fault classification submodule inputs the dimensionality-reduced acoustic-vibration feature vector into the SVM classifier, and simultaneously inputs the Mel-Gram angle difference field time spectrum of the sound signal extracted by the MFCC method into the ConvNeXt-T network for parallel identification. When the two identification results are consistent, the fault type is determined. When the results are inconsistent, features are re-extracted or combined with the historical operating data of the equipment for comprehensive judgment, and finally the fault type identification result is output.

[0033] Furthermore, the health prediction module includes:

[0034] The input submodule is used to receive acoustic data, vibration data, equipment data, environmental data, and operating condition information.

[0035] The data association submodule is used to associate the information input in the input submodule with device health factors through data matching and mapping.

[0036] The initial weight determination submodule is used to analyze the historical fault and normal operation data of GIS equipment using K-means clustering to determine the distribution range of each health factor under different health states, and to calculate the correlation degree between each health factor and the probability of equipment failure using the association reasoning method, thereby determining the initial weight coefficient of each health factor.

[0037] The dynamic adjustment submodule is used to dynamically adjust the weight coefficients based on the actual operating data and historical fault data of the equipment using the gradient descent algorithm;

[0038] The health assessment submodule is used to calculate the overall health score of the equipment based on the adjusted weighting coefficients, and to classify the equipment health level to output fault prediction information.

[0039] Furthermore, the communication module adopts 5G communication technology, and the module transmits the equipment operating status information, fault identification results and fault prediction information obtained by the fault identification and prediction module to the remote monitoring center in real time.

[0040] Simultaneously, it receives control commands from the remote monitoring center, including adjusting the acquisition frequency of the integrated storage and computing sound-vibration acquisition unit and modifying the fault event trigger threshold.

[0041] This invention also discloses a GIS-based intelligent monitoring and fault prediction method for acoustic-vibration joint intelligent monitoring and fault prediction using the CIM-based GIS-based intelligent monitoring and fault prediction system described above, comprising:

[0042] S1. Simultaneously collect raw signals, including mechanical vibration and sound signals, from multiple locations in the GIS equipment. Through the built-in SRAM storage and computing structure, the raw signals are processed by Kalman filtering for noise reduction and normalization, and then multi-source sound-vibration data is output.

[0043] S2. Real-time monitoring of received acoustic-vibration data based on fault triggering threshold to trigger fault events. Timestamp synchronization method is used to ensure time consistency of acoustic-vibration data from multiple sources. Valid acoustic-vibration data is obtained after redundancy elimination and interference extraction.

[0044] S3. Extract multi-domain features of acoustic-vibration from effective acoustic-vibration data for fault identification; and construct an equipment health assessment model to classify the health level of GIS equipment and output health prediction information.

[0045] S4. Transmit the fault identification results and health prediction information to the remote monitoring center, and receive remote control instructions, including adjusting the collection parameters and fault trigger thresholds.

[0046] This invention can achieve one of the following beneficial effects:

[0047] The present invention relates to a CIM-based GIS sound-vibration joint intelligent monitoring and fault prediction system and method. It adopts a CIM (Compute-In-Memory) architecture, which overcomes the bottlenecks in computing throughput and energy efficiency of the traditional von Neumann architecture with its separate storage and computing capabilities. It integrates sensing, storage and processing functions, and can efficiently and in real time collect and process GIS mechanical vibration and acoustic data, greatly improving the timeliness of data processing and enabling rapid response to the need for instantaneous fault identification.

[0048] This invention incorporates an SRAM-CIM structure within an integrated in-memory acoustic-vibration acquisition unit, enabling preliminary data processing while acquiring data. This reduces data transmission volume and subsequent processing pressure, and eliminates the need for complex cable routing, thereby lowering installation difficulty and cost. It is suitable for long-distance, large-scale distributed monitoring scenarios in GIS.

[0049] The fault identification and prediction algorithm of this invention, set at the edge, combines a variety of advanced signal processing and machine learning methods, such as MFCC feature extraction, HHT time-frequency analysis, principal component analysis, SVM classifier, and ConvNeXt-T network, which can accurately extract acoustic-vibration signal features, effectively identify typical GIS faults, and achieve high identification accuracy. At the same time, the equipment health assessment model can accurately assess the health status of equipment by dynamically adjusting the weight coefficients, realize fault prediction, and help to take maintenance measures in advance to reduce equipment failure rate.

[0050] This invention transmits monitoring data and analysis results to a remote monitoring center in real time, enabling remote real-time monitoring of the operating status of GIS equipment. This reduces the on-site workload of maintenance personnel, lowers maintenance costs, and ensures that staff can obtain equipment fault information in a timely manner, improving the timeliness and effectiveness of equipment maintenance and further enhancing the operational reliability of key power grid nodes. Attached Figure Description

[0051] 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.

[0052] Figure 1 This is a schematic diagram showing the components and connections of the GIS acoustic-vibration joint intelligent monitoring and fault prediction system in an embodiment of the present invention;

[0053] Figure 2a This is a diagram of the in-memory computing structure in an embodiment of the present invention;

[0054] Figure 2b This is a diagram of the traditional von Neumann architecture in an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of the specific structure of the in-memory computing integrated acoustic-vibration acquisition unit in an embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of the algorithm flow in the fault identification and prediction module of this invention.

[0057] Figure 5 This is a flowchart of the GIS-based intelligent monitoring and fault prediction method for acoustic-vibration combined in an embodiment of the present invention. Detailed Implementation

[0058] 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.

[0059] Example 1

[0060] One embodiment of the present invention discloses a CIM-based GIS-based joint intelligent monitoring and fault prediction system for acoustic-vibration, such as... Figure 1 As shown, including those set at the edge ends,

[0061] The in-memory computing unit is used to simultaneously acquire raw signals, including mechanical vibration and sound signals, from multiple locations in GIS equipment. Through the built-in SRAM in-memory computing structure, the raw signals are processed by Kalman filtering for noise reduction and normalization, and then multi-source sound and vibration data are output.

[0062] The fault triggering and synchronization processing module is used to monitor the received acoustic-vibration data in real time based on the fault triggering threshold to trigger fault events. It uses a timestamp synchronization method to ensure the time consistency of acoustic-vibration data from multiple sources, and obtains valid acoustic-vibration data through redundancy elimination and interference extraction.

[0063] The fault identification and prediction module is used to extract multi-domain features of acoustic-vibration from effective acoustic-vibration data for fault identification; and to build an equipment health assessment model to classify the health level of GIS equipment and output health prediction information.

[0064] The communication module is used to transmit fault identification results and health prediction information to the remote monitoring center, and to receive remote control commands, including adjusting the collected parameters and fault trigger thresholds.

[0065] Specifically, the integrated storage and computing sound-vibration acquisition unit is deployed at least one of the disconnecting switch, busbar air chamber, and grounding switch in the GIS equipment;

[0066] Each in-memory acoustic-vibration acquisition unit includes a piezoelectric accelerometer, an electret microphone, and an SRAM in-memory computing structure;

[0067] Among them, the piezoelectric accelerometer has high sensitivity and a wide frequency response range, which can accurately collect mechanical vibration signals of GIS equipment; the electret microphone has low noise and high signal-to-noise ratio characteristics, which can effectively collect abnormal noise signals during equipment operation.

[0068] Piezoelectric accelerometers collect mechanical vibration signals and electret microphones collect sound signals, which are then input into an SRAM in-memory computing architecture.

[0069] The SRAM in-memory computing architecture includes multiple 8TiC computing units and 6T-SRAM storage units;

[0070] Each 8TiC computing unit is an 8-transistor SRAM unit, which adds computing function transistors compared to the 6T-SRAM unit. It can perform binary multiplication operations and realize computing functions by reconstructing the read / write path.

[0071] The 8TiC computing unit and the 6T-SRAM storage unit are interconnected through word lines, bit lines, complementary bit lines and global bit lines. While collecting vibration and sound signals, the Kalman filter noise reduction and [0,1] interval normalization processing are performed in the storage unit.

[0072] In-memory computing architecture such as Figure 2a As shown, the traditional von Neumann architecture is as follows: Figure 2b As shown, compared with the traditional von Neumann 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] The in-memory computing integrated sound-vibration acquisition unit was developed based on the typical fault identification and prediction requirements of GIS, and its internal structure adopts a typical SRAM in-memory computing architecture.

[0074] like Figure 3 As shown, the integrated in-memory computing sound-vibration acquisition unit comprises multiple 8TiC units and 6T-SRAM units. Data input and output are achieved through WL (word line), BL (bit line), BLB (complementary bit line), and GBL (global bit line) circuits. The unit utilizes a piezoelectric accelerometer with a sensitivity of up to 100mV / g and a frequency response range of 0.1Hz-10kHz, accurately acquiring mechanical vibration signals generated by GIS equipment during operation. The electret microphone employed has a noise level below 15dB(A) and a signal-to-noise ratio greater than 60dB, effectively acquiring normal sound signals during equipment operation and abnormal noise signals under fault conditions. During acquisition, the integrated SRAM computing structure performs preliminary filtering and amplification of the acquired raw data, removing some high-frequency noise and interference signals, reducing the workload of subsequent data processing.

[0075] Specifically, the process of signal acquisition and processing by the integrated in-memory acoustic-vibration acquisition unit includes:

[0076] The raw signals collected by the piezoelectric accelerometer and electret microphone are converted into digital signals by the interface circuit and then written into the 6T-SRAM memory cell array through word line gating.

[0077] The 8TiC computing unit reads data from the storage unit in real time through the bit line, and performs Kalman filtering recursive operation in parallel within the storage array to eliminate random noise. The filtered data is normalized in the [0,1] interval by finding the maximum and minimum values ​​and linear scaling within the storage unit.

[0078] By pre-storing the Kalman gain matrix in an SRAM array, the state prediction and update calculations of the Kalman filter are performed in parallel using 8 TiC cells.

[0079] The normalization formula is:

[0080] x max x min Here, x represents the maximum and minimum values ​​of the data; x is the data before normalization. norm This is the normalized data.

[0081] The processed valid data is output to the fault triggering and synchronization processing module via the global bit line.

[0082] In a specific scheme, the recursive operation of Kalman filtering and normalization processing are implemented;

[0083] The SRAM array is divided into three regions.

[0084] State storage area: Stores the state or information of the Kalman filter;

[0085] Gain pre-stored area: Solidifies the K-matrix of the Kalman filter gain;

[0086] Intermediate computation area: stores the residuals of the Kalman filter;

[0087] The state vector of the Kalman filter can be set as the displacement, velocity, acceleration, sound pressure, trend, and entropy of the GIS switch;

[0088] When performing Kalman filter recursive operations in parallel, the 8TiC computing unit is divided into a multiply-accumulate unit and a state prediction unit according to its functions. The multiply-accumulate unit and the state prediction unit work together to complete each iteration in a pipelined timing sequence of reading data, parallel multiplication, accumulation operation, and writing back the result.

[0089] Reading phase: Read the state variables from the state storage area via the bit line BL, and read the gain matrix elements from the gain pre-store area;

[0090] Parallel multiplication calculation stage: The multiply-accumulate unit and the state prediction unit perform multiplication operations in parallel within the same clock cycle;

[0091] Accumulation phase: The multiply-accumulate unit performs multi-stage pipeline accumulation on the product result;

[0092] Write-back phase: The final calculation result is written back to the state storage area via the global bit line GBL, and the write-back operation and the read operation of the next iteration are triggered on different clock edges to avoid read-write conflicts.

[0093] During normalization implementation, maximum value lookup: 8TiC is reconstructed into a comparator, and 6 levels of comparisons are used to obtain x.max / x min (Without reading data) Linear scaling: Use 8TiC for fixed-point multiplication and addition, and use a pre-stored reciprocal table for division.

[0094] Specifically, the fault triggering and synchronization processing module includes:

[0095] The threshold triggering submodule is used to set the normal range threshold of vibration and sound signals according to the instructions of the remote monitoring center, and to monitor the received sound-vibration data in real time. When the data continuously exceeds the threshold, a fault event is triggered.

[0096] This module presets the vibration signal amplitude range (e.g., 0-5 m / s) under normal operating conditions of the GIS equipment. 2 The system monitors the sound signal intensity range (e.g., 30-60dB). When the monitored data exceeds the preset range, it is determined to be a "potential fault" and a fault event is immediately triggered; the time and location of the fault are recorded.

[0097] The data synchronization submodule is used to achieve consistent synchronization of multi-source data in the time dimension by adding precise timestamps to each set of vibration and sound data.

[0098] This module uses a timestamp synchronization method, preferably by adding a timestamp with an accuracy of 1μs to each set of collected data through a GPS module, to achieve time synchronization of multi-source data.

[0099] The interference cancellation submodule is used to analyze the synchronized data using wavelet transform, eliminate redundant information and interference components, extract effective acoustic-vibration data related to equipment faults (such as abnormal noises caused by partial discharge and specific frequency vibrations caused by mechanical defects), and output them to the edge fault identification and prediction module.

[0100] Specifically, the fault identification and prediction module includes:

[0101] The fault identification module is used to perform multi-domain feature extraction and principal component analysis dimensionality reduction on effective acoustic-vibration data. It uses an SVM classifier and a ConvNeXt-T network in parallel to identify fault types and outputs fault type identification results.

[0102] The health prediction module is used to build a dynamic weighted health assessment model based on acoustic print, vibration, equipment, environment and operating condition data. It determines the weight of each influencing factor through cluster analysis and association reasoning, and uses an optimization algorithm to dynamically adjust the weights to classify the equipment health level and output health prediction information.

[0103] The fault identification module includes:

[0104] The feature extraction submodule is used to extract sound signal features using the Mel frequency cepstral coefficient (MFCC) method, and to extract vibration signal features using the Hilbert-Huang transform time-frequency analysis method and the high-frequency harmonic vibration amplitude ratio method, thus obtaining sound-vibration features.

[0105] The Mel frequency cepstral coefficient (MFCC) method divides the sound data in the effective sound-vibration data into multiple frames, performs a Fourier transform on each frame to obtain the power spectrum, then uses a Mel filter bank to obtain the Mel power spectrum, takes the logarithm of the Mel power spectrum and performs a discrete cosine transform to finally obtain a 12-16 dimensional MFCC feature vector.

[0106] For the vibration data in the sound data of the effective sound-vibration data, on the one hand, the Hilbert-Huang Transform (HHT) is used to perform time-frequency analysis to decompose the vibration signal into multiple Intrinsic Mode Functions (IMFs) and extract characteristic parameters such as instantaneous frequency and amplitude of each IMF. On the other hand, the ratio of the amplitude of high-frequency harmonics (such as 200Hz and 300Hz) in the vibration signal to the amplitude of 100Hz vibration is calculated as the characteristic parameter of the vibration signal.

[0107] The dimensionality reduction and optimization submodule is used to perform dimensionality reduction processing on the extracted acoustic-vibration features through principal component analysis;

[0108] Principal component analysis (PCA) is used to optimize the dimensionality of the extracted sound and vibration feature parameters, thereby reducing the data dimensionality, computational load, and preserving the main information of the data.

[0109] The fault classification submodule is used to input the dimensionality-reduced features into the SVM classifier and the ConvNeXt-T network for parallel recognition, and then obtain and output the fault type after fusing the recognition results.

[0110] The fault classification submodule inputs the dimensionality-reduced acoustic-vibration feature vector into the SVM classifier, and simultaneously inputs the Mel-Gram angle difference field time spectrum of the sound signal extracted by the MFCC method into the ConvNeXt-T network for parallel identification. When the two identification results are consistent, the fault type is determined. When the results are inconsistent, features are re-extracted or combined with the historical operating data of the equipment for comprehensive judgment, and finally the fault type identification result is output.

[0111] Specifically, the SVM classifier uses a radial basis function kernel function and trains the classifier with training samples so that it can accurately identify typical faults of GIS equipment, such as poor contact of switch contacts, abnormal contact of busbar joints, unbalanced housing connection, bent guide rods, and loose components.

[0112] The ConvNeXt-T network extracts deep features from the temporal spectrogram through multi-layer convolution and pooling operations, and fuses them with the recognition results of the SVM classifier, ultimately achieving a fault recognition accuracy of over 99%.

[0113] When making a comprehensive judgment based on the historical operating data of the equipment, the following steps are performed in sequence: time-series trend verification based on short-term data, health status matching based on K-means clustering based on medium-term data, and confidence weighting of Apriori association rules based on long-term data throughout the entire life cycle. Finally, the historical accuracy of each model and the adaptive weight of the operating condition are combined through dynamic weight fusion decision, and the result with the highest fusion confidence and exceeding the threshold is taken as the final output.

[0114] The health prediction module includes:

[0115] The input submodule is used to receive acoustic data, vibration data, equipment data, environmental data, and operating condition information.

[0116] The acoustic fingerprint data is sound signal feature data extracted using the MFCC method; the vibration data is vibration signal feature data extracted using the Hilbert-Huang transform time-frequency analysis method and the high-frequency harmonic vibration amplitude ratio method; the equipment data includes equipment operating time and insulating gas pressure data; the environmental data includes ambient temperature and humidity data; the operating condition information includes load current, voltage imbalance, and number of switching actions data.

[0117] The data association submodule is used to associate the information input in the input submodule with device health factors through data matching and mapping.

[0118] Establish a five-dimensional health influencing factor space and map the input data into standardized factor values ​​F. j ∈[0,1]:

[0119] The five-dimensional health influencing factor space includes:

[0120] Voiceprint factor (F1): includes the mean of MFCC feature vector, sound pressure level anomaly index, and spectral steepness;

[0121] Vibration factor (F2): includes HHT time-frequency entropy, high-frequency harmonic energy ratio, and dominant frequency offset;

[0122] Equipment Factor (F3): Normalized value of runtime, gas pressure deviation rate, and normalized value of number of switching actions;

[0123] Environmental factors (F4): the degree of temperature deviation from the standard value and the percentage of cumulative time exceeding the humidity standard;

[0124] Operating condition factor (F5): load current fluctuation rate, voltage imbalance, and recent operating frequency.

[0125] A Gaussian membership function can be used to achieve a non-linear mapping from data to health factors.

[0126] The initial weight determination submodule is used to analyze the historical fault and normal operation data of GIS equipment using K-means clustering to determine the distribution range of each health factor under different health states, and to calculate the correlation degree between each health factor and the probability of equipment failure using the association reasoning method, thereby determining the initial weight coefficient of each health factor.

[0127] In one specific embodiment, during K-means clustering analysis, multiple sets of historical fault and normal operation data of GIS equipment are input, with each set containing 5-dimensional health factors;

[0128] Clustering settings: k = 5 classes, corresponding to excellent, good, average, warning, and fault; set the number of iterations and convergence threshold, using Mahalanobis distance.

[0129] Output results: 5 cluster centers; and calculate the covariance matrix, the standard deviation of each health factor in each cluster, and form a distribution range table as a benchmark;

[0130] The Apriori algorithm is used to calculate the correlation between each health factor and the probability of equipment failure. The inverse of the mean standard deviation of the health factors is used as the dispersion. Based on the set correlation and dispersion contribution values ​​(with a weight ratio of correlation:dispersion = 4:6), the initial weight coefficients of each normalized health factor are determined.

[0131] The dynamic adjustment submodule is used to dynamically adjust the weight coefficients based on the actual operating data and historical fault data of the equipment using the gradient descent algorithm;

[0132] The gradient of the batch data is calculated using a weighted mean squared error loss function that includes a fault sample penalty term and a weight smoothing term. The Adam optimizer is used to perform a batch weight update once every certain number of data sets with a set initial learning rate. At the same time, monotonicity constraints are applied to ensure that the weights of factors that are positively correlated with faults do not decrease. The weight range is restricted and normalized. When there are sudden changes in the environment or shocks in the operating conditions, the corresponding factor weights are temporarily increased immediately. If the new weights cause the validation accuracy to drop by more than a set threshold, the system will automatically roll back to the previous stable version and mark the abnormal data.

[0133] The health assessment submodule is used to calculate the overall health score of the equipment based on the adjusted weighting coefficients, and to classify the equipment health level to output fault prediction information.

[0134] Equipment health levels are divided into five categories: excellent (85-100 points), good (70-84 points), average (55-69 points), warning (40-54 points), and fault (0-39 points). When the equipment health level drops to the warning or fault level, a corresponding fault prediction warning message is issued.

[0135] Figure 4 This is a schematic diagram of the algorithm flow in the fault identification and prediction module of this invention.

[0136] Specifically, the communication module uses 5G communication technology. The module transmits the equipment operating status information (such as vibration amplitude, sound intensity, insulating gas pressure, etc.), fault identification results (such as fault type, fault location, etc.) and fault prediction information (such as health level, early warning prompts, etc.) obtained by the fault identification and prediction module to the remote monitoring center in real time.

[0137] Simultaneously, it receives control commands from the remote monitoring center, such as adjusting the acquisition frequency of the integrated storage and computing sound-vibration acquisition unit (from the original 1kHz to 2kHz) and modifying the fault event trigger threshold (from the upper limit of vibration signal amplitude to 5m / s). 2 Adjusted to 4.5 m / s 2 ).

[0138] In summary, the GIS acoustic-vibration joint intelligent monitoring and fault prediction system based on CIM disclosed in this embodiment adopts a CIM (Compute-In-Memory) architecture, which overcomes the bottlenecks in computing throughput and energy efficiency of the traditional von Neumann architecture with its separate storage and computing capabilities. It integrates sensing, storage, and processing functions, enabling efficient and real-time collection and processing of GIS mechanical vibration and acoustic data, significantly improving the timeliness of data processing and quickly responding to the need for instantaneous fault identification.

[0139] The integrated storage and computing sound-vibration acquisition unit incorporates a built-in SRAM-CIM structure, which can perform preliminary data processing while acquiring data, reducing the amount of data transmission and subsequent processing pressure. It also eliminates the need for complex cable layout, reducing installation difficulty and cost, and is suitable for long-distance, large-scale distributed monitoring scenarios in GIS.

[0140] The fault identification and prediction algorithm set at the edge combines a variety of advanced signal processing and machine learning methods, such as MFCC feature extraction, HHT time-frequency analysis, principal component analysis, SVM classifier, and ConvNeXt-T network, which can accurately extract acoustic-vibration signal features, effectively identify typical GIS faults, and achieve high identification accuracy. At the same time, the equipment health assessment model can accurately assess the health status of equipment by dynamically adjusting the weight coefficients, realize fault prediction, and help to take maintenance measures in advance to reduce equipment failure rate.

[0141] By transmitting monitoring data and analysis results to the remote monitoring center in real time, the operation status of GIS equipment can be monitored remotely in real time. This reduces the on-site workload of maintenance personnel, lowers maintenance costs, and ensures that staff can obtain equipment fault information in a timely manner, improving the timeliness and effectiveness of equipment maintenance and further enhancing the operational reliability of key nodes in the power grid.

[0142] Example 2

[0143] One embodiment of the present invention discloses a GIS-based intelligent monitoring and fault prediction method for acoustic-vibration joint intelligent monitoring using the CIM-based GIS acoustic-vibration joint intelligent monitoring and fault prediction system as described in Embodiment 1. Figure 5 As shown, it includes:

[0144] S1. Simultaneously collect raw signals, including mechanical vibration and sound signals, from multiple locations in the GIS equipment. Through the built-in SRAM storage and computing structure, the raw signals are processed by Kalman filtering for noise reduction and normalization, and then multi-source sound-vibration data is output.

[0145] The in-memory and in-storage sound and vibration acquisition unit is installed in key parts of the GIS equipment, such as disconnect switches, busbar air chambers, and grounding switches, to ensure that the acquisition unit can accurately collect the vibration and sound signals of the equipment. When the monitoring device is started, the in-memory and in-storage sound and vibration acquisition unit begins to collect mechanical vibration and sound signals during the operation of the GIS equipment. The built-in SRAM in-memory computing structure of the acquisition unit performs preliminary processing on the collected data to remove noise interference from the data.

[0146] S2. Real-time monitoring of received acoustic-vibration data based on fault triggering threshold to trigger fault events. Timestamp synchronization method is used to ensure time consistency of acoustic-vibration data from multiple sources. Valid acoustic-vibration data is obtained after redundancy elimination and interference extraction.

[0147] The fault event triggering and synchronization processing analysis module receives the pre-processed data transmitted by the acquisition unit, monitors the data in real time, and triggers a fault event when the data exceeds the preset normal range. At the same time, it uses a timestamp synchronization method to synchronize the vibration and sound data to ensure the time consistency of multi-source data, and further analyzes the synchronized data to extract effective information.

[0148] S3. Extract multi-domain features of acoustic-vibration from effective acoustic-vibration data for fault identification; and construct an equipment health assessment model to classify the health level of GIS equipment and output health prediction information.

[0149] The edge fault identification and prediction module calls the built-in fault identification and prediction algorithm to extract features from the effective data processed synchronously in S3. For sound signals, the MFCC method is used to extract features, while for vibration signals, HHT time-frequency analysis and the ratio of high-frequency harmonic vibration amplitude to 100Hz vibration amplitude are used to extract features. Then, principal component analysis is used to optimize the features through dimensionality reduction. An SVM classifier and a ConvNeXt-T network are combined to identify faults and determine the fault type. Finally, combined with the equipment health assessment model, based on acoustic data, vibration data, equipment data, environmental data, and operating condition information, the weight coefficients of each health influencing factor are determined through data matching mapping, cluster analysis, and association reasoning. After dynamically adjusting the weight coefficients, the equipment health level is classified to achieve fault prediction.

[0150] S4. Transmit the fault identification results and health prediction information to the remote monitoring center, and receive remote control instructions, including adjusting the collection parameters and fault trigger thresholds.

[0151] The communication module transmits the obtained equipment operating status information, fault identification results, and fault prediction information to the remote monitoring center. Staff can monitor the operation of the GIS equipment in real time through the remote monitoring center. When a fault warning or fault information is received, corresponding maintenance and repair measures can be taken in a timely manner. At the same time, the remote monitoring center can send control commands to the monitoring device through the communication module, such as adjusting the acquisition frequency and modifying the fault trigger threshold.

[0152] 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.

[0153] 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 CIM-based GIS-based intelligent monitoring and fault prediction system for acoustic-vibration joint monitoring, characterized in that, Includes settings at the edge. The in-memory computing unit is used to simultaneously acquire raw signals, including mechanical vibration and sound signals, from multiple locations in GIS equipment. Through the built-in SRAM in-memory computing structure, the raw signals are processed by Kalman filtering for noise reduction and normalization, and then multi-source sound and vibration data are output. The fault triggering and synchronization processing module is used to monitor the received acoustic-vibration data in real time based on the fault triggering threshold to trigger fault events. It uses a timestamp synchronization method to ensure the time consistency of acoustic-vibration data from multiple sources, and obtains valid acoustic-vibration data through redundancy elimination and interference extraction. The fault identification and prediction module is used to extract multi-domain acoustic-vibration features from effective acoustic-vibration data for fault identification. Furthermore, an equipment health assessment model is constructed to classify the health levels of GIS equipment and output health prediction information; The communication module is used to transmit fault identification results and health prediction information to the remote monitoring center, and to receive remote control commands, including adjusting the collected parameters and fault trigger thresholds.

2. The CIM-based GIS acoustic-vibration joint intelligent monitoring and fault prediction system according to claim 1, characterized in that, The integrated storage and computing acoustic-vibration acquisition unit is deployed at least one of the disconnecting switch, busbar air chamber, and grounding switch in the GIS equipment; Each in-memory acoustic-vibration acquisition unit includes a piezoelectric accelerometer, an electret microphone, and an SRAM in-memory computing structure; Piezoelectric accelerometers collect mechanical vibration signals and electret microphones collect sound signals, which are then input into an SRAM in-memory computing architecture. The SRAM in-memory computing architecture includes multiple 8TiC computing units and 6T-SRAM storage units; The 8TiC computing unit and the 6T-SRAM storage unit are interconnected through word lines, bit lines, complementary bit lines and global bit lines. While collecting vibration and sound signals, the Kalman filter noise reduction and [0,1] interval normalization processing are performed in the storage unit.

3. The CIM-based GIS acoustic-vibration joint intelligent monitoring and fault prediction system according to claim 2, characterized in that, The process of signal acquisition and processing by the in-memory integrated acoustic-vibration acquisition unit includes: The raw signals collected by the piezoelectric accelerometer and electret microphone are converted into digital signals by the interface circuit and then written into the 6T-SRAM memory cell array through word line gating. The 8TiC computing unit reads data from the storage unit in real time through the bit line, and performs Kalman filtering recursive operation in parallel within the storage array to eliminate random noise. The filtered data is normalized in the [0,1] interval by finding the maximum and minimum values ​​and linear scaling within the storage unit. The processed valid data is output to the fault triggering and synchronization processing module via the global bit line.

4. The CIM-based GIS-based acoustic-vibration joint intelligent monitoring and fault prediction system according to claim 1, characterized in that, The fault triggering and synchronization processing module includes: The threshold triggering submodule is used to set the normal range threshold of vibration and sound signals according to the instructions of the remote monitoring center, and to monitor the received sound-vibration data in real time. When the data continuously exceeds the threshold, a fault event is triggered. The data synchronization submodule is used to achieve consistent synchronization of multi-source data in the time dimension by adding precise timestamps to each set of vibration and sound data. The interference cancellation submodule is used to analyze the synchronized data using wavelet transform, eliminate redundant information and interference components, extract effective acoustic-vibration data related to equipment faults, and output them to the edge fault identification and prediction module.

5. The GIS-based acoustic-vibration joint intelligent monitoring and fault prediction system based on CIM according to claim 1, characterized in that, The fault identification and prediction module includes: The fault identification module is used to perform multi-domain feature extraction and principal component analysis dimensionality reduction on effective acoustic-vibration data. It uses an SVM classifier and a ConvNeXt-T network in parallel to identify fault types and outputs fault type identification results. The health prediction module is used to build a dynamic weighted health assessment model based on acoustic print, vibration, equipment, environment and operating condition data. It determines the weight of each influencing factor through cluster analysis and association reasoning, and uses an optimization algorithm to dynamically adjust the weights to classify the equipment health level and output health prediction information.

6. The CIM-based GIS acoustic-vibration joint intelligent monitoring and fault prediction system according to claim 5, characterized in that, The fault identification module includes: The feature extraction submodule is used to extract sound signal features using the Mel frequency cepstral coefficient (MFCC) method, and to extract vibration signal features using the Hilbert-Huang transform time-frequency analysis method and the high-frequency harmonic vibration amplitude ratio method, thus obtaining sound-vibration features. The dimensionality reduction and optimization submodule is used to perform dimensionality reduction processing on the extracted acoustic-vibration features through principal component analysis; The fault classification submodule is used to input the dimensionality-reduced features into the SVM classifier and the ConvNeXt-T network for parallel recognition, and then obtain and output the fault type after fusing the recognition results.

7. The CIM-based GIS acoustic-vibration joint intelligent monitoring and fault prediction system according to claim 6, characterized in that, The fault classification submodule inputs the dimensionality-reduced acoustic-vibration feature vector into the SVM classifier, and simultaneously inputs the Mel-Gram angle difference field time spectrum of the sound signal extracted by the MFCC method into the ConvNeXt-T network for parallel identification. When the two identification results are consistent, the fault type is determined. When the results are inconsistent, features are re-extracted or combined with the historical operating data of the equipment for comprehensive judgment, and finally the fault type identification result is output.

8. The CIM-based GIS acoustic-vibration joint intelligent monitoring and fault prediction system according to claim 5, characterized in that, The health prediction module includes: The input submodule is used to receive acoustic data, vibration data, equipment data, environmental data, and operating condition information. The data association submodule is used to associate the information input in the input submodule with device health factors through data matching and mapping. The initial weight determination submodule is used to analyze the historical fault and normal operation data of GIS equipment using K-means clustering to determine the distribution range of each health factor under different health states, and to calculate the correlation degree between each health factor and the probability of equipment failure using the association reasoning method, thereby determining the initial weight coefficient of each health factor. The dynamic adjustment submodule is used to dynamically adjust the weight coefficients based on the actual operating data and historical fault data of the equipment using the gradient descent algorithm; The health assessment submodule is used to calculate the overall health score of the equipment based on the adjusted weighting coefficients, and to classify the equipment health level to output fault prediction information.

9. The CIM-based GIS-based acoustic-vibration joint intelligent monitoring and fault prediction system according to claim 1, characterized in that, The communication module uses 5G communication technology. The module transmits the equipment operating status information, fault identification results and fault prediction information obtained by the fault identification and prediction module to the remote monitoring center in real time. Simultaneously, it receives control commands from the remote monitoring center, including adjusting the acquisition frequency of the integrated storage and computing sound-vibration acquisition unit and modifying the fault event trigger threshold.

10. A GIS-based intelligent monitoring and fault prediction method for acoustic-vibration joint intelligent monitoring and fault prediction using the CIM-based GIS-based intelligent monitoring and fault prediction system as described in any one of claims 1-9, characterized in that, include: S1. Simultaneously collect raw signals, including mechanical vibration and sound signals, from multiple locations in the GIS equipment. Through the built-in SRAM storage and computing structure, the raw signals are processed by Kalman filtering for noise reduction and normalization, and then multi-source sound-vibration data is output. S2. Real-time monitoring of received acoustic-vibration data based on fault triggering threshold to trigger fault events. Timestamp synchronization method is used to ensure time consistency of acoustic-vibration data from multiple sources. Valid acoustic-vibration data is obtained after redundancy elimination and interference extraction. S3. Extract multi-domain acoustic-vibration features from valid acoustic-vibration data for fault identification; Furthermore, an equipment health assessment model is constructed to classify the health levels of GIS equipment and output health prediction information; S4. Transmit the fault identification results and health prediction information to the remote monitoring center, and receive remote control instructions, including adjusting the collection parameters and fault trigger thresholds.

Images