GIS latent defect diagnosis method and system based on multi-source signal fusion

Abstract

The application discloses a GIS latent defect diagnosis method and system based on multi-source signal fusion, and belongs to the field of power equipment fault diagnosis. The method comprises the following steps: monitoring a GIS internal PD signal by using a UWB sensor, combining a FPGA preprocessing chip to perform real-time filtering and spectrum analysis on the signal, and dynamically adjusting sampling sensitivity according to the PD signal strength; using a three-dimensional ultrasonic array to realize three-dimensional imaging of an ultrasonic source inside the GIS; processing the PD and ultrasonic signals based on wavelet transform, processing the PD and ultrasonic data based on CNN, and processing the signal time sequence based on LSTM; constructing a GIS fault knowledge graph, setting expert rules, and calculating the GIS fault occurrence probability by using Bayesian inference; predicting the GIS fault evolution path according to a Markov chain model, and dynamically adjusting in combination with reinforcement learning. The application establishes a fault state transition model of the GIS equipment based on the Markov chain, quantifies the evolution probability between different fault states, provides long-term steady-state fault risk assessment, and is helpful for predicting the operation health condition of the GIS equipment.

Description

Technical Field

[0001] This invention belongs to the field of power equipment fault diagnosis technology, based on multi-source signal fusion technology (IPC main classification number G01R, measurement of electrical variables in the physical domain), and particularly relates to a method and system for diagnosing latent defects in GIS based on multi-source signal fusion. It is used for defect diagnosis and analysis of GIS. Background Technology

[0002] Currently, fault detection in GIS equipment mainly relies on methods such as regular maintenance, partial discharge detection, infrared thermography, and ultrasonic testing. However, these methods have the following problems: High dependence on manual labor and long detection cycles: Traditional detection methods often require manual inspection and data analysis, resulting in high maintenance costs and the inability to achieve real-time monitoring. Single signal analysis is susceptible to interference: Using a single signal (such as ultrasonic or electromagnetic waves) for GIS fault diagnosis is easily affected by noise interference, leading to misdiagnosis or missed diagnosis.

[0003] Current GIS fault diagnosis mainly relies on expert experience or simple signal characteristic threshold judgments, lacking data-driven intelligent prediction models and making it difficult to adapt to complex operating environments. Fault evolution prediction capabilities are insufficient: existing methods often only detect the current state and cannot predict the future health status of GIS equipment, lacking modeling and analysis of fault evolution paths, making it difficult to provide early warnings of potential risks. Summary of the Invention

[0004] In view of the problems of the above-mentioned or existing methods and systems for diagnosing latent defects in GIS based on multi-source signal fusion, this invention is proposed.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] This invention provides a method for diagnosing latent defects in GIS based on multi-source signal fusion, comprising: using a UWB sensor to monitor PD signals inside the GIS, combining an FPGA preprocessing chip to perform real-time filtering and spectrum analysis on the signals, and using an AI adaptive gain algorithm to dynamically adjust the sampling sensitivity according to the PD signal strength; using a three-dimensional ultrasonic array to achieve three-dimensional imaging of the ultrasonic source inside the GIS, and combining a neural network noise reduction algorithm to filter environmental noise; processing PD and ultrasonic signals based on wavelet transform, processing PD and ultrasonic data based on CNN, and processing signal time series using LSTM to learn long-term evolution trends; constructing a GIS fault knowledge graph, using a graph neural network for data association reasoning, setting expert rules, and using Bayesian inference to calculate the probability of GIS fault occurrence; predicting the GIS fault evolution path based on a Markov chain model, and combining reinforcement learning for dynamic adjustment.

[0007] As a preferred embodiment of the GIS latent defect diagnosis method based on multi-source signal fusion described in this invention, the method includes: using a UWB sensor to monitor PD signals inside the GIS, combining an FPGA preprocessing chip to perform real-time filtering and spectrum analysis on the signals, and using an AI adaptive gain algorithm to dynamically adjust the sampling sensitivity according to the PD signal strength, including:

[0008] Set the center frequency f of the UWB sensor c Within the 3.1GHz-10.6GHz range, the sampling frequency is calculated using the Nyquist sampling theorem: ;

[0009] Among them, f max If the highest frequency component of the PD signal is represented by the minimum sampling frequency: ;

[0010] Select f s =1GH or higher ensures complete acquisition of PD signals;

[0011] An adaptive notch filter is used to remove 50Hz power frequency interference and its harmonics. ;

[0012] Where f0 is 50Hz or its harmonics, T is the sampling period, and r is the bandwidth of the control filter;

[0013] The pulse amplitude and signal-to-noise ratio of the PD signal are extracted, and the gain G(t) is adaptively adjusted using deep Q-network reinforcement learning. ;

[0014] Among them, R t The current reward is Q(s,a), and the gain policy value is Q(s,a).

[0015] If the signal-to-noise ratio is <10dB, the gain G increases by 5dB; if the signal-to-noise ratio is >30dB, the gain G decreases by 3dB.

[0016] As a preferred embodiment of the GIS latent defect diagnosis method based on multi-source signal fusion described in this invention, the method includes: employing a three-dimensional ultrasonic array to achieve three-dimensional imaging of ultrasonic sources inside the GIS, and combining this with a neural network noise reduction algorithm to filter environmental noise, including:

[0017] A spherical ultrasonic array is used, with the array diameter D set to 30cm-50cm. The sensor arrangement adopts a uniformly distributed 64×64×64 element to form a three-dimensional phased array ultrasonic sensor network. The ultrasonic signal generated by the PD is acquired by array distributed sampling.

[0018] Let the center coordinates of the sensor array be (x0, y0, z0), and the position of sensor i be (x0, y0, z0). i ,y i ,zi The time delay of the PD signal arriving from sensor i is calculated. By summing the time-delay corrected signals, a preliminary spatial distribution image of the PD ultrasound signal is formed. CNN extracts the time-frequency features of the ultrasound, and RNN performs temporal noise reduction.

[0019] ;

[0020] Among them, h t For the hidden layer state, W h This is the weight matrix;

[0021] By combining bidirectional LSTM, the noise reduction capability of ultrasonic timing signals is enhanced.

[0022] As a preferred embodiment of the GIS latent defect diagnosis method based on multi-source signal fusion described in this invention, the method includes: processing PD and ultrasonic signals based on wavelet transform, and processing PD and ultrasonic data based on CNN, comprising:

[0023] The PD and ultrasound signals were decomposed into four levels using the db4 wavelet: ;

[0024] Where A4 represents the low-frequency component and D4-D1 represent the high-frequency components;

[0025] The energy of wavelet coefficients is calculated for subsequent CNN training; CNN is used for PD and ultrasound signal feature extraction and classification, and PD type classification is performed through fully connected layers to output PD diagnosis results.

[0026] As a preferred embodiment of the GIS latent defect diagnosis method based on multi-source signal fusion described in this invention, wherein: LSTM processes the signal time series and learns the long-term evolution trend, including:

[0027] The sliding window technique is used to convert PD and ultrasound signals into time series: ;

[0028] Where n represents the window length, ensuring that LSTM captures both short-term and long-term changes;

[0029] An LSTM model is constructed, taking PD spectral features and ultrasonic features as input. A bidirectional LSTM is used to improve feature extraction capability, and multiple stacked LSTMs are used to extract deep temporal features. Then, the LSTM output is transformed through a fully connected layer to output PD discharge trend prediction or GIS health index. According to the dataset division, PD data from the past 100 moments are input, and LSTM predicts the discharge trend for the next 10 to 50 moments. This prediction is then fused with the ultrasonic signal to generate a GIS health trend curve.

[0030] As a preferred embodiment of the GIS latent defect diagnosis method based on multi-source signal fusion described in this invention, the method includes: constructing a GIS fault knowledge graph, using a graph neural network for data association reasoning, setting expert rules, and using Bayesian inference to calculate the probability of GIS fault occurrence, including:

[0031] The GIS knowledge graph is represented by triples, including GIS buses, GIS insulators, and PD signals. After the knowledge graph is constructed, a graph neural network (GNN) is used to mine GIS equipment fault patterns. The GNN formula is as follows:

[0032] ;

[0033] in, is the feature representation of node v at layer l+1, N(v) is the set of neighboring nodes of node v, W is the training parameter, b is the bias, and σ is the activation function;

[0034] Components, fault types, and sensor data from the GIS knowledge graph are selected as inputs to the graph neural network.

[0035] Graph feature extraction using GCN:

[0036] ;

[0037] in, It is an adjacency matrix. It is a degree matrix. These are the training weights for the GNN;

[0038] Train the GNN to predict equipment failure types; set expert rules: if the GIS PD signal strength > 50pC, the probability of insulation failure increases; if the GIS equipment temperature anomaly ΔT > 5℃, the probability of internal overheating failure increases; if PD + ultrasonic anomalies occur simultaneously, the probability of GIS equipment failure is greater than 80%.

[0039] If new sensor data arrives, the posterior probability is updated, and the GIS fault risk assessment is adjusted in real time.

[0040] As a preferred embodiment of the GIS latent defect diagnosis method based on multi-source signal fusion described in this invention, the method includes: predicting the GIS fault evolution path based on a Markov chain model and dynamically adjusting it using reinforcement learning, including:

[0041] Define the set of fault states for GIS equipment: ;

[0042] Where S0 represents a healthy state, S1 represents a minor anomaly, S2 represents a moderate anomaly, and S3 represents a severe failure, the state transitions of the Markov chain are described by the probability matrix P: ;

[0043] Among them, P ij Indicates that the system starts from state S i Transition to state S j The probability of;

[0044] Deep Q-learning is used for training. The Q-network is initialized, and GIS fault states and maintenance operations are set.

[0045] Perform actions to gain new status and rewards, and update your Q value.

[0046] A latent defect diagnosis system for GIS based on multi-source signal fusion includes: a discharge signal acquisition module, which uses a UWB sensor to monitor the PD signal inside the GIS, combines an FPGA preprocessing chip to perform real-time filtering and spectrum analysis on the signal, and uses an AI adaptive gain algorithm to dynamically adjust the sampling sensitivity according to the PD signal strength;

[0047] The ultrasonic signal acquisition module is used to realize three-dimensional imaging of ultrasonic sources inside GIS using a three-dimensional ultrasonic array, and to filter environmental noise by combining a neural network noise reduction algorithm.

[0048] The deep learning feature extraction module is used to process PD and ultrasonic signals based on wavelet transform, process PD and ultrasonic data based on CNN, process signal time series using LSTM, and learn long-term evolution trends.

[0049] The multi-source signal fusion module is used to construct a GIS fault knowledge graph, use graph neural networks for data association reasoning, set expert rules, and use Bayesian inference to calculate the probability of GIS fault occurrence.

[0050] The defect identification module is used to predict the failure evolution path of GIS based on the Markov chain model and to dynamically adjust it by combining reinforcement learning.

[0051] A computing device, the computing device comprising:

[0052] At least one processor, memory, and input / output unit;

[0053] The memory is used to store computer programs, and the processor is used to call the computer programs stored in the memory to execute the steps of the GIS latent defect diagnosis method based on multi-source signal fusion.

[0054] A computer-readable storage medium includes instructions that, when executed on a computer, cause the computer to perform steps of a GIS latent defect diagnosis method based on multi-source signal fusion.

[0055] The beneficial effects of this invention are as follows: Based on Markov chains, this invention establishes a fault state transition model for GIS equipment, which can quantify the evolution probability between different fault states, provide long-term steady-state fault risk assessment, and help predict the operational health status of GIS equipment in advance. Combined with reinforcement learning, an intelligent maintenance strategy for GIS equipment is constructed. The system can autonomously learn the optimal maintenance measures, reasonably balance maintenance costs and equipment health status, reduce maintenance frequency while extending the service life of GIS equipment. During continuous training, the reinforcement learning strategy can adaptively adjust the optimal maintenance plan according to changes in the GIS operating environment, making it applicable to various power grid operating conditions. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 A flowchart of a GIS latent defect diagnosis method based on multi-source signal fusion provided in an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram of the structure of a GIS latent defect diagnosis system based on multi-source signal fusion provided in an embodiment of the present invention.

[0059] Figure 3 A schematic diagram of the structure of a medium according to an embodiment of the present invention is shown.

[0060] Figure 4 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown.

[0061] In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts. Detailed Implementation

[0062] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0063] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0064] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0065] Example

[0066] The following is for reference. Figures 1-4 , Figure 1 This is a flowchart illustrating a method for diagnosing latent defects in GIS based on multi-source signal fusion, as provided in one embodiment of the present invention. It should be noted that the embodiments of the present invention can be applied to any applicable scenario.

[0067] Figure 1 The flowchart of a GIS latent defect diagnosis method based on multi-source signal fusion provided in an embodiment of the present invention includes:

[0068] S1: UWB sensors are used to monitor PD signals inside the GIS. Combined with FPGA preprocessing chips, the signals are filtered and analyzed in real time. AI adaptive gain algorithm is used to dynamically adjust the sampling sensitivity according to the PD signal strength.

[0069] Preferably, the center frequency f of the UWB sensor is set. c Within the 3.1GHz-10.6GHz range, the sampling frequency is calculated using the Nyquist sampling theorem: ;

[0070] Among them, f max If the highest frequency component of the PD signal is represented by the minimum sampling frequency: ;

[0071] Select f s =1GH or higher ensures complete acquisition of PD signals;

[0072] An adaptive notch filter is used to remove 50Hz power frequency interference and its harmonics. ;

[0073] Where f0 is 50Hz or its harmonics, T is the sampling period, and r is the bandwidth of the control filter;

[0074] The pulse amplitude and signal-to-noise ratio of the PD signal are extracted, and the gain G(t) is adaptively adjusted using deep Q-network reinforcement learning. ;

[0075] Among them, R t The current reward is Q(s,a), and the gain policy value is Q(s,a).

[0076] If the signal-to-noise ratio is <10dB, the gain G increases by 5dB; if the signal-to-noise ratio is >30dB, the gain G decreases by 3dB.

[0077] Furthermore, the center frequency range of the UWB sensor is set: f c =3.1GHz-10.6GHz

[0078] Set the highest frequency component of the PD signal: f max =500MHz

[0079] According to the Nyquist sampling theorem, substituting the value of fmax: f s ≥2×500MHz=1GHz

[0080] Set the sampling frequency f s =1.2GHz, ensuring complete acquisition of PD signals;

[0081] Design a fifth-order notch filter to filter out power frequency harmonics at f0=50Hz, 100Hz, 150Hz, 200Hz, and 250Hz. Use MATLAB / Simulink to simulate and verify the filtering effect.

[0082] After filtering, the pulse amplitude and signal-to-noise ratio of the PD signal are calculated. The time-frequency characteristics of the PD signal are analyzed by short-time Fourier transform, and the signal-to-noise ratio is calculated. The initial signal-to-noise ratio SNR0=15dB is set. The deep Q network method is used to adaptively adjust the gain G(t) according to the real-time SNR value of the PD signal to optimize the PD signal detection effect.

[0083] Define the state space S: S = {current signal-to-noise ratio (SNR)} t )}

[0084] Define the action space A: A = {gain adjustment ΔG}

[0085] Define the gain adjustment rule: if SNR < 10dB, increase the gain by 5dB; if SNR > 30dB, decrease the gain by 3dB; otherwise, keep the gain unchanged. Define the reward function R. t R t =-|SNR target-SNR t |;

[0086] The goal is to adjust the SNR to the optimal range of 15dB-25dB;

[0087] DQN was trained using Python and TensorFlow, with reinforcement learning training performed on a real PD dataset from GIS devices. Experience replay was used to train the DQN to enhance stability. After training, the PD signal detection accuracy was tested.

[0088] 1 15 20 3 19 2 12 18 5 17 3 30 20 -3 27 4 9 15 5 14.5

[0089] Through DQN reinforcement learning, the system can intelligently adjust the gain to keep the SNR of the PD signal within the optimal range (15dB-25dB), avoiding over-amplification or signal distortion. By employing a joint optimization method combining a UWB sensor, an adaptive notch filter, and reinforcement learning gain control, the false positive rate of PD signal detection is reduced by 20%, and the detection accuracy is increased to over 95%.

[0090] S2: A three-dimensional ultrasonic array is used to realize three-dimensional imaging of ultrasonic sources inside the GIS, and environmental noise is filtered by a neural network noise reduction algorithm.

[0091] Preferably, a spherical ultrasonic array is used, with the array diameter D set to 30cm-50cm. The sensor arrangement adopts a uniformly distributed 64×64×64 element to form a three-dimensional phased array ultrasonic sensor network, and the ultrasonic signal generated by the PD is acquired by array distributed sampling.

[0092] Let the center coordinates of the sensor array be (x0, y0, z0), and the position of sensor i be (x0, y0, z0). i ,y i ,z i The time delay of the PD signal arriving from sensor i is calculated. By summing the time-delay corrected signals, a preliminary spatial distribution image of the PD ultrasound signal is formed. CNN extracts the time-frequency features of the ultrasound, and RNN performs temporal noise reduction.

[0093] ;

[0094] Among them, h t For the hidden layer state, W h This is the weight matrix;

[0095] By combining bidirectional LSTM, the noise reduction capability of ultrasonic timing signals is enhanced.

[0096] Furthermore, short-time Fourier transform (STFT) was used to extract the time-frequency image of the PD ultrasound signal, and a convolutional neural network (CNN) was used to extract the time-frequency features.

[0097] First layer: 3×3 convolutional kernel, ReLU activation

[0098] Second layer: 5×5 convolutional kernel, ReLU activation

[0099] Pooling layer: 2×2 max pooling

[0100] Fully connected layer outputs feature vector

[0101] A CNN model was trained using Python TensorFlow + Keras, and CNN feature extraction was performed using the PD dataset. An RNN+LSTM model was trained using PyTorch + TensorFlow. The initial noise SNR was set to 10dB, and the target SNR was >20dB. Simulation results analysis:

[0102] CNN only 10 15 generally CNN+RNN 10 18 good CNN + Bidirectional LSTM 10 22 Optimal

[0103] CNN can effectively extract the time-frequency features of PD ultrasound signals, but its denoising ability is limited. RNN enhances the temporal denoising effect of ultrasound and improves signal quality. Bidirectional LSTM combined with CNN has the best denoising effect, improves SNR by 12dB, and increases PD recognition accuracy to 98%.

[0104] S3: Processes PD and ultrasonic signals based on wavelet transform, processes PD and ultrasonic data based on CNN, processes signal time series using LSTM, and learns long-term evolution trends.

[0105] Preferably, the PD and ultrasound signals are decomposed into four levels using the db4 wavelet: ;

[0106] Where A4 represents the low-frequency component and D4-D1 represent the high-frequency components;

[0107] The energy of wavelet coefficients is calculated for subsequent CNN training; CNN is used for PD and ultrasound signal feature extraction and classification, and PD type classification is performed through fully connected layers to output PD diagnosis results.

[0108] Preferably, a sliding window technique is used to convert the PD and ultrasound signals into time series: ;

[0109] Where n represents the window length, ensuring that LSTM captures both short-term and long-term changes;

[0110] An LSTM model is constructed, taking PD spectral features and ultrasonic features as input. A bidirectional LSTM is used to improve feature extraction capability, and multiple stacked LSTMs are used to extract deep temporal features. Then, the LSTM output is transformed through a fully connected layer to output PD discharge trend prediction or GIS health index. According to the dataset division, PD data from the past 100 moments are input, and LSTM predicts the discharge trend for the next 10 to 50 moments. This prediction is then fused with the ultrasonic signal to generate a GIS health trend curve.

[0111] Furthermore, a CNN was constructed for PD diagnostic classification.

[0112] (1) CNN network structure

[0113] Input layer: Employs wavelet coefficient energy features (A4, D4, D3, D2, D1)

[0114] Convolutional layer 1: Convolutional kernel (3×3), extracts local features;

[0115] Pooling layer 1: Max pooling (2×2), reducing computational cost;

[0116] Convolutional layer 2: Convolutional kernel (5×5), extracts deeper features;

[0117] Pooling layer 2: Max pooling (2×2);

[0118] Fully connected layer: PD type classification includes normal discharge, surface discharge, floating potential discharge, internal discharge, and local arc. Training was performed using 10,000 sets of PD signals + ultrasonic signals, with a learning rate η = 0.001, a batch size of 32, and 100 iterations. The cross-entropy loss function was used for training. The training results are as follows:

[0119] Traditional FFT+SVM 85.6% 150 db4 wavelet + CNN 97.2% 95

[0120] Wavelet decomposition preserves important time-frequency features of PD and is more suitable for non-stationary signals than FFT; CNN automatically learns features and has higher classification accuracy than SVM; training time is reduced by 36%, improving diagnostic efficiency.

[0121] Further, LSTM memory cell calculation: ;

[0122] Where, f t i t o t These are the forget gate, input gate, and output gate, respectively. (C) t h represents the state of a memory unit. t Here, W represents the hidden layer state, and b represents the weights and biases.

[0123] Bidirectional LSTM enhances time series features through forward and backward computation: ;

[0124] in, Processing forward time series, Processing inverted time series;

[0125] A CNN is used to extract the time-frequency features of the ultrasonic signal, and the ultrasonic feature Ut is fused into the hidden layer of an LSTM to generate a GIS health trend curve; the prediction accuracy is shown through simulation.

[0126] Traditional ARIMA 0.034 Sliding window + LSTM 0.008 LSTM+ultrasound signal fusion 0.005

[0127] Using LSTM reduces prediction error by 76.5%, and combining it with ultrasonic signals further reduces prediction error by 85.3%. In the normal phase of the GIS health trend curve, PD changes are stable with low ultrasonic signal intensity. In early faults, PD gradually increases with increasing ultrasonic signal intensity. As the fault worsens, PD signal spikes and ultrasonic signal becomes abnormal. In critical faults, PD spectrum changes significantly, requiring GIS equipment maintenance. A sliding window technique is used to construct time series data, capturing short-term and long-term changes in PD and ultrasonic signals. The LSTM model extracts discharge characteristics, and combining it with ultrasonic signals improves prediction accuracy by 85.3%. A bidirectional LSTM+CNN ultrasonic fusion is used to generate the GIS health trend curve.

[0128] S4: Construct a GIS fault knowledge graph, use graph neural networks for data association reasoning, set expert rules, and use Bayesian inference to calculate the probability of GIS fault occurrence.

[0129] Preferably, the GIS knowledge graph is represented by triples, including GIS busbars, GIS insulators, and PD signals. After the knowledge graph is constructed, a graph neural network (GNN) is used to mine GIS equipment fault modes. The GNN formula is:

[0130] ;

[0131] in, is the feature representation of node v at layer l+1, N(v) is the set of neighboring nodes of node v, W is the training parameter, b is the bias, and σ is the activation function;

[0132] Components, fault types, and sensor data from the GIS knowledge graph are selected as inputs to the graph neural network.

[0133] Graph feature extraction using GCN:

[0134] ;

[0135] in, It is an adjacency matrix. It is a degree matrix. These are the training weights for the GNN;

[0136] Train the GNN to predict equipment failure types; set expert rules: if the GIS PD signal strength > 50pC, the probability of insulation failure increases; if the GIS equipment temperature anomaly ΔT > 5℃, the probability of internal overheating failure increases; if PD + ultrasonic anomalies occur simultaneously, the probability of GIS equipment failure is greater than 80%.

[0137] If new sensor data arrives, the posterior probability is updated, and the GIS fault risk assessment is adjusted in real time.

[0138] Furthermore, Rule 1: Abnormal PD discharge

[0139] P(Insulation fault | PD > 50pC)

[0140] When the PD signal strength > 50pC, the probability of GIS insulation failure increases by 30%.

[0141] Let the initial probability be P0, and after adjustment:

[0142] P′=P0+0.3(1-P0)

[0143] Rule 2: GIS equipment overheating

[0144] P(Internal overheating fault |ΔT>5∘C)

[0145] If the GIS temperature anomaly ΔT > 5°C, the probability of internal overheating failure increases by 20%.

[0146] Rule 3: PD + ultrasound abnormalities

[0147] P(GIS failure | PD + ultrasound anomaly) > 80%

[0148] When PD discharge and abnormal ultrasonic signals occur simultaneously, the probability of GIS equipment failure exceeds 80%.

[0149] Data on GIS equipment structure and sensor data is collected to construct a triplet database. Neo4j is used to store the GIS knowledge graph. A GCN is trained by inputting a GIS adjacency matrix and equipment feature vectors to extract equipment fault modes.

[0150] Output GIS fault category prediction. With new sensor data input, the fault probability is dynamically updated: P′=P0+α(1-P0);

[0151] Wherein, α is calculated by expert rules: if PD > 50 pC, α = 0.3; if ΔT > 5℃, α = 0.2.

[0152] If the PD+ ultrasound is abnormal, P′>80%;

[0153] Failure risk assessment:

[0154] PD=30pC, ΔT=3℃ 0.2 0.2 PD=60pC, ΔT=4℃ 0.3 0.51 PD=80pC, ΔT=7℃ 0.5 0.82

[0155] Knowledge graphs provide a structured representation of GIS fault relationships, improving the training effect of GNNs; GCNs extract graph features, increasing prediction accuracy by 14.3%; expert rules dynamically update fault probabilities, enabling real-time GIS fault risk assessment.

[0156] S5: Predict the failure evolution path of GIS based on the Markov chain model, and make dynamic adjustments by combining reinforcement learning.

[0157] Preferably, a set of fault states for GIS equipment is defined: ;

[0158] Where S0 represents a healthy state, S1 represents a minor anomaly, S2 represents a moderate anomaly, and S3 represents a severe failure, the state transitions of the Markov chain are described by the probability matrix P: ;

[0159] Among them, P ij Indicates that the system starts from state S i Transition to state S j The probability of;

[0160] Deep Q-learning is used for training. The Q-network is initialized, and GIS fault states and maintenance operations are set.

[0161] Perform actions to gain new status and rewards, and update your Q value.

[0162] Furthermore, data from GIS equipment sensors (PD signals, temperature, and ultrasonic signals) are collected.

[0163] Calculate the GIS device state transition probability matrix P and initialize the Q-network. Set the GIS device state S. t Randomly select maintenance action A t ; Calculate the state S at the next time step t+1 Record reward R t DQN was trained using experience replay: ;

[0164] Train DQN for 20,000 iterations to obtain the optimal maintenance strategy, inputting the real-time GIS device status S. t DQN selects the optimal maintenance operation A t Perform maintenance operations and adjust the fault status of GIS equipment in real time; continuously optimize the Q network to improve decision-making capabilities;

[0165] After one year of operation of the GIS equipment, a comparison was made between the traditional maintenance strategy and the DQN maintenance strategy:

[0166] Traditional maintenance 12.4% 50 15 DQN maintenance 5.7% 30 20

[0167] The DQN maintenance strategy reduces GIS equipment failure rate by 54%, lowers maintenance costs by 40%, and extends GIS equipment lifespan by 33%.

[0168] After introducing the method of exemplary embodiments of the present invention, the following references are made. Figure 2 An exemplary embodiment of the present invention will be described, comprising a GIS latent defect diagnosis system based on multi-source signal fusion, the system comprising:

[0169] The discharge signal acquisition module is used to monitor the PD signal inside the GIS using a UWB sensor. Combined with the FPGA preprocessing chip, the signal is filtered and analyzed in real time. The sampling sensitivity is dynamically adjusted according to the PD signal strength using an AI adaptive gain algorithm.

[0170] The ultrasonic signal acquisition module is used to realize three-dimensional imaging of ultrasonic sources inside GIS using a three-dimensional ultrasonic array, and to filter environmental noise by combining a neural network noise reduction algorithm.

[0171] The deep learning feature extraction module is used to process PD and ultrasonic signals based on wavelet transform, process PD and ultrasonic data based on CNN, process signal time series using LSTM, and learn long-term evolution trends.

[0172] The multi-source signal fusion module is used to construct a GIS fault knowledge graph, use graph neural networks for data association reasoning, set expert rules, and use Bayesian inference to calculate the probability of GIS fault occurrence.

[0173] The defect identification module is used to predict the failure evolution path of GIS based on the Markov chain model and to dynamically adjust it by combining reinforcement learning.

[0174] The following is an explanation of some of the terms used in this invention:

[0175] Convolutional Neural Networks (CNN).

[0176] Ultra Wide Band (UWB): A wireless carrier communication technology.

[0177] Long Short-Term Memory (LSTM): A type of time-recurrent neural network.

[0178] PD (Partial Discharge).

[0179] After introducing the methods and apparatus of exemplary embodiments of the present invention, the following references are made. Figure 3 A computer-readable storage medium according to exemplary embodiments of the present invention will be described, please refer to... Figure 3The computer-readable storage medium shown is an optical disc 30, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it implements the steps described in the above-described method implementation. For example, it uses a UWB sensor to monitor the PD signal inside the GIS, combines an FPGA preprocessing chip to perform real-time filtering and spectrum analysis on the signal, and uses an AI adaptive gain algorithm to dynamically adjust the sampling sensitivity according to the PD signal strength; it uses a three-dimensional ultrasonic array to realize three-dimensional imaging of the ultrasonic source inside the GIS, and combines a neural network noise reduction algorithm to filter environmental noise; it processes PD and ultrasonic signals based on wavelet transform, processes PD and ultrasonic data based on CNN, processes signal time series using LSTM, and learns long-term evolution trends; it constructs a GIS fault knowledge graph, uses a graph neural network for data association reasoning, sets expert rules, and uses Bayesian reasoning to calculate the probability of GIS fault occurrence; it predicts the GIS fault evolution path based on a Markov chain model and combines reinforcement learning for dynamic adjustment; the specific implementation methods of each step will not be repeated here.

[0180] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical and magnetic storage media, which will not be elaborated here.

[0181] After introducing the methods, apparatus, and media of exemplary embodiments of the present invention, the following references are made. Figure 4 A computing device for diagnosing latent defects in GIS based on multi-source signal fusion, according to an exemplary embodiment of the present invention.

[0182] Figure 4 A block diagram is shown of an exemplary computing device 40 suitable for implementing embodiments of the present invention. The computing device 40 may be a computer system or a server. Figure 4 The computing device 40 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0183] like Figure 4 As shown, the components of computing device 40 may include, but are not limited to: one or more processors or processing units 401, system memory 402, and bus 403 connecting different system components (including system memory 402 and processing unit 401).

[0184] The computing device 40 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by the computing device 40, including volatile and non-volatile media, and removable and non-removable media.

[0185] System memory 402 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 4021 and / or cache memory 4022. Computing device 40 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, ROM 4023 may be used to read and write non-removable, non-volatile magnetic media (…). Figure 4 (Not shown in the image, usually referred to as "hard drive"). Although not shown in... Figure 4 The diagram illustrates that disk drives for reading and writing to removable non-volatile disks (e.g., "floppy disks") and optical disc drives for reading and writing to removable non-volatile optical discs (e.g., CD-ROMs, DVD-ROMs, or other optical media) can be provided. In these cases, each drive can be connected to bus 403 via one or more data media interfaces. System memory 402 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0186] A program / utility 4025 having a set (at least one) of program modules 4024 may be stored, for example, in system memory 402, and such program modules 4024 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment. Program modules 4024 typically perform the functions and / or methods described in the embodiments of the present invention.

[0187] The computing device 40 can also communicate with one or more external devices 404 (such as a keyboard, pointing device, display, etc.). This communication can be performed via the input / output (I / O) interface 405. Furthermore, the computing device 40 can also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter 406. Figure 4 As shown, network adapter 406 communicates with other modules of computing device 40 (such as processing unit 401) via bus 403. It should be understood that, although... Figure 4 As not shown, it can be used in conjunction with computing device 40 with other hardware and / or software modules.

[0188] The processing unit 401 executes various functional applications and data processing by running programs stored in the system memory 402. For example, it uses UWB sensors to monitor PD signals inside the GIS, combines FPGA preprocessing chips to perform real-time filtering and spectrum analysis of the signals, and uses AI adaptive gain algorithms to dynamically adjust the sampling sensitivity according to the PD signal strength; it uses a three-dimensional ultrasonic array to realize three-dimensional imaging of the ultrasonic source inside the GIS, and combines neural network noise reduction algorithms to filter environmental noise; it processes PD and ultrasonic signals based on wavelet transform, processes PD and ultrasonic data based on CNN, processes signal time series using LSTM, and learns long-term evolution trends; it constructs a GIS fault knowledge graph, uses graph neural networks for data association reasoning, sets expert rules, and uses Bayesian inference to calculate the probability of GIS fault occurrence; it predicts the GIS fault evolution path based on Markov chain models and combines reinforcement learning for dynamic adjustment.

[0189] The specific implementation methods of each step will not be repeated here. It should be noted that although several units / modules or sub-units / sub-modules of the synchronous escape wiring device based on multi-commodity flow have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0190] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0191] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0192] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0193] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0194] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0195] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0196] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0197] Furthermore, although the operations of the method of the present invention are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

Claims

1. A method for diagnosing latent defects in GIS based on multi-source signal fusion, characterized in that, include: UWB sensors are used to monitor PD signals inside the GIS. Combined with FPGA preprocessing chips, the signals are filtered and analyzed in real time. The sampling sensitivity is dynamically adjusted according to the PD signal strength using AI adaptive gain algorithm. A three-dimensional ultrasonic array is used to realize three-dimensional imaging of ultrasonic sources inside GIS, and environmental noise is filtered by combining neural network noise reduction algorithm; The system uses wavelet transform to process PD and ultrasonic signals, CNN to process PD and ultrasonic data, and LSTM to process signal time series to learn long-term evolution trends. Construct a GIS fault knowledge graph, use graph neural networks for data association reasoning, set expert rules, and use Bayesian inference to calculate the probability of GIS fault occurrence. The failure evolution path of GIS is predicted based on the Markov chain model, and dynamic adjustment is carried out by combining reinforcement learning.

2. The GIS latent defect diagnosis method based on multi-source signal fusion as described in claim 1, characterized in that, The method involves using a UWB sensor to monitor PD signals within the GIS, combining this with an FPGA preprocessing chip to perform real-time filtering and spectrum analysis of the signals, and utilizing an AI adaptive gain algorithm to dynamically adjust the sampling sensitivity based on the PD signal strength. This includes: Set the center frequency f of the UWB sensor c Within the 3.1GHz-10.6GHz range, the sampling frequency is calculated using the Nyquist sampling theorem: ; Among them, f max If the highest frequency component of the PD signal is represented by the minimum sampling frequency: ; Select f s =1GH or higher ensures complete acquisition of PD signals; An adaptive notch filter is used to remove 50Hz power frequency interference and its harmonics. ; Where f0 is 50Hz or its harmonics, T is the sampling period, and r is the bandwidth of the control filter; The pulse amplitude and signal-to-noise ratio of the PD signal are extracted, and the gain G(t) is adaptively adjusted using deep Q-network reinforcement learning. ; Among them, R t The current reward is Q(s,a), and the gain policy value is Q(s,a). If the signal-to-noise ratio is <10dB, the gain G increases by 5dB; if the signal-to-noise ratio is >30dB, the gain G decreases by 3dB.

3. The GIS latent defect diagnosis method based on multi-source signal fusion as described in claim 1, characterized in that, The method employs a three-dimensional ultrasonic array to achieve three-dimensional imaging of ultrasonic sources within the GIS, and combines this with a neural network noise reduction algorithm to filter environmental noise, including: A spherical ultrasonic array is used, with the array diameter D set to 30cm-50cm. The sensor arrangement adopts a uniformly distributed 64×64×64 element to form a three-dimensional phased array ultrasonic sensor network. The ultrasonic signal generated by the PD is acquired by array distributed sampling. The center coordinates of the sensor array are set to (x0, y0, z0), and the position of sensor i is (xi, yi, zi). The time delay of the PD signal arriving from sensor i is calculated. By summing the signals after time delay correction, a preliminary spatial distribution image of the PD ultrasound signal is formed. CNN extracts the time-frequency features of the ultrasound, and RNN performs temporal noise reduction. ； Among them, h t For the hidden layer state, W h This is the weight matrix; By combining bidirectional LSTM, the noise reduction capability of ultrasonic timing signals is enhanced.

4. The GIS latent defect diagnosis method based on multi-source signal fusion as described in claim 1, characterized in that, The wavelet transform-based processing of PD and ultrasonic signals, and the CNN-based processing of PD and ultrasonic data, include: The PD and ultrasound signals were decomposed into four levels using the db4 wavelet: ; Where A4 represents the low-frequency component and D4-D1 represent the high-frequency components; The energy of wavelet coefficients is calculated for subsequent CNN training; CNN is used for PD and ultrasound signal feature extraction and classification, and PD type classification is performed through fully connected layers to output PD diagnosis results.

5. The GIS latent defect diagnosis method based on multi-source signal fusion as described in claim 1, characterized in that, The LSTM processes signal time series to learn long-term evolution trends, including: The sliding window technique is used to convert PD and ultrasound signals into time series: ; Where n represents the window length, ensuring that LSTM captures both short-term and long-term changes; An LSTM model is constructed, taking PD spectral features and ultrasonic features as input. A bidirectional LSTM is used to improve feature extraction capability, and multiple stacked LSTMs are used to extract deep temporal features. Then, the LSTM output is transformed through a fully connected layer to output PD discharge trend prediction or GIS health index. According to the dataset division, PD data from the past 100 moments are input, and LSTM predicts the discharge trend for the next 10 to 50 moments. This prediction is then fused with the ultrasonic signal to generate a GIS health trend curve.

6. The GIS latent defect diagnosis method based on multi-source signal fusion as described in claim 1, characterized in that, The construction of the GIS fault knowledge graph, the use of graph neural networks for data association and reasoning, the setting of expert rules, and the use of Bayesian inference to calculate the probability of GIS fault occurrence include: The GIS knowledge graph is represented by triples, including GIS buses, GIS insulators, and PD signals. After the knowledge graph is constructed, a graph neural network (GNN) is used to mine GIS equipment fault patterns. The GNN formula is as follows: ； in, is the feature representation of node v at layer l+1, N(v) is the set of neighboring nodes of node v, W is the training parameter, b is the bias, and σ is the activation function; Components, fault types, and sensor data from the GIS knowledge graph are selected as inputs to the graph neural network. Graph feature extraction using GCN: ； in, It is an adjacency matrix. It is a degree matrix. These are the training weights for the GNN; Train the GNN to predict equipment failure types; set expert rules: if the GIS PD signal strength > 50pC, the probability of insulation failure increases; if the GIS equipment temperature anomaly ΔT > 5℃, the probability of internal overheating failure increases; if PD + ultrasonic anomalies occur simultaneously, the probability of GIS equipment failure is greater than 80%. If new sensor data arrives, the posterior probability is updated, and the GIS fault risk assessment is adjusted in real time.

7. The GIS latent defect diagnosis method based on multi-source signal fusion as described in claim 1, characterized in that, The method of predicting GIS failure evolution paths based on Markov chain models and dynamically adjusting them using reinforcement learning includes: Define the set of fault states for GIS equipment: ; Where S0 represents a healthy state, S1 represents a minor anomaly, S2 represents a moderate anomaly, and S3 represents a severe failure, the state transitions of the Markov chain are described by the probability matrix P: ; Among them, P ij Indicates that the system starts from state S i Transition to state S j The probability of; Deep Q-learning is used for training. The Q-network is initialized, and GIS fault states and maintenance operations are set. Perform actions to gain new status and rewards, and update your Q value.

8. A GIS latent defect diagnosis system based on multi-source signal fusion, characterized in that, include: The discharge signal acquisition module is used to monitor the PD signal inside the GIS using a UWB sensor. Combined with the FPGA preprocessing chip, the signal is filtered and analyzed in real time. The sampling sensitivity is dynamically adjusted according to the PD signal strength using an AI adaptive gain algorithm. The ultrasonic signal acquisition module is used to realize three-dimensional imaging of ultrasonic sources inside GIS using a three-dimensional ultrasonic array, and to filter environmental noise by combining a neural network noise reduction algorithm. The deep learning feature extraction module is used to process PD and ultrasonic signals based on wavelet transform, process PD and ultrasonic data based on CNN, process signal time series using LSTM, and learn long-term evolution trends. The multi-source signal fusion module is used to construct a GIS fault knowledge graph, use graph neural networks for data association reasoning, set expert rules, and use Bayesian inference to calculate the probability of GIS fault occurrence. The defect identification module is used to predict the failure evolution path of GIS based on the Markov chain model and to dynamically adjust it by combining reinforcement learning.

9. A computing device, the computing device comprising: At least one processor, memory, and input / output unit; The memory is used to store computer programs, and the processor is used to call the computer programs stored in the memory to execute the steps of the GIS latent defect diagnosis method based on multi-source signal fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium comprising instructions that, when executed on a computer, cause the computer to perform the steps of the GIS latent defect diagnosis method based on multi-source signal fusion as described in any one of claims 1 to 7.

Images