Generator partial discharge on-line monitoring method, system, device and medium

Abstract

The invention belongs to the field of power equipment state monitoring, and relates to a generator partial discharge on-line monitoring method, system, equipment and medium, and the method comprises the following steps: obtaining four types of original signals of a generator; extracting a partial discharge feature significant signal layer from the original signal by using an adaptive hierarchical denoising algorithm; constructing a feature quantity library based on the partial discharge feature significant signal layer, and extracting the frequency spectrum gravity center and wavelet energy entropy of each type of original signals; performing fuzzy logic weighting on the frequency spectrum gravity center and the wavelet energy entropy of each type of original signals to obtain a primary fusion result of each type of original signals; performing deep fusion on the primary fusion result of each type of original signals by adopting an LSTM neural network to obtain a partial discharge primary judgment result; and obtaining the score of the partial discharge preliminary judgment result based on the abnormal degree score of the mahalanobis distance, and if the score is greater than a decision threshold, determining that partial discharge occurs. The partial discharge of the generator can be quickly and effectively judged, and field operation and maintenance personnel can respond to the partial discharge conveniently.

Description

Technical Field

[0001] The present invention belongs to the field of power equipment status monitoring, and relates to a method, system, equipment and medium for online monitoring of partial discharge of a generator. Background Art

[0002] Generators, as core power equipment in power systems, have a direct impact on the stable operation of the power grid through their insulation condition. Approximately 42% of high-voltage generator failures are due to insulation degradation caused by partial discharge. Throughout the equipment's lifecycle, winding insulation material can gradually develop due to factors such as production process defects (such as bubbles in epoxy resin casting), mechanical vibration stress during operation, material aging caused by long-term temperature rise, and environmental moisture penetration. These factors can cause tip discharges caused by metal burrs on the stator bar surface, surface discharges along the insulation layer due to moisture, air gap discharges caused by infiltration of the cooling medium, and suspended discharges caused by loose fasteners. These partial discharges are initially minimal, but sustained discharge accelerates insulation carbonization, ultimately leading to breakdown. Currently, mainstream detection methods include ultrasonic, high-frequency, neutral point current, and transient ground voltage methods. Ultrasonic methods are widely used in online generator monitoring due to their strong resistance to electromagnetic interference. However, they have significant limitations: single signals lack reliability and are susceptible to mechanical vibration and noise. Offline detection efficiency is low: the current IEC 60034-27 standard recommends offline partial discharge testing, requiring shutdown and application of high voltage, which takes an average of 8-12 hours per unit, severely impacting power generation efficiency. Defect type identification is difficult: existing technologies often rely on PRPD (Phase Resolved Partial Discharge) spectrum analysis, requiring manual comparison of discharge pulse phase distribution, amplitude statistics, and other characteristics, relying heavily on the expertise of operators and maintenance personnel. Furthermore, dynamic interference suppression and feature fusion decision-making under complex operating conditions remain technical challenges. Load fluctuations and temperature changes during generator operation can cause the time-frequency characteristics of the discharge signal to drift, and existing threshold settings often use fixed values, which are unable to adapt to environmental changes. At the same time, a single sensor cannot fully characterize the discharge characteristics: ultrasonic signals reflect mechanical vibration effects, UHF (Ultra High Frequency) signals contain electromagnetic radiation information, and neutral point current can be associated with discharge energy. The lack of multi-dimensional data fusion will lead to missed detection of key features. Summary of the Invention

[0003] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a method, system, equipment and medium for online monitoring of partial discharge of generators, which can quickly and effectively determine partial discharge of generators, thereby facilitating on-site operation and maintenance personnel to respond to partial discharge.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions: A method for online monitoring of partial discharge of a generator includes the following steps: S1, obtain four types of original signals of the generator, including ultrasonic signals, neutral point current signals, voltage phase signals and ultra-high frequency current signals; S2, uses an adaptive hierarchical denoising algorithm to extract the significant signal layer of partial discharge characteristics from the original signal; S3, constructing a feature library based on the significant signal layer of partial discharge characteristics, and extracting the spectrum center of gravity and wavelet energy entropy of each type of original signal from the feature library; S4, performing fuzzy logic weighting on the spectral center of gravity and wavelet energy entropy of each type of original signal to obtain the primary fusion result of each type of original signal; S5, uses LSTM neural network to deeply fuse the primary fusion results of each type of original signal to obtain the preliminary judgment result of partial discharge; S6. A score for the preliminary partial discharge judgment result is obtained based on the abnormality score of the Mahalanobis distance, and the partial discharge judgment result is determined by comparing the score with a preset decision threshold. If the score is less than or equal to the decision threshold, no partial discharge has occurred; if the score is greater than the decision threshold, partial discharge has occurred.

[0005] Preferably, after acquiring the four types of original signals of the generator, the method further includes: performing time compensation on the four types of original signals using a clock synchronization compensation method to align the time of each type of original signal.

[0006] Preferably, the specific process of extracting the local discharge feature significant signal layer from the original signal using the adaptive layered denoising algorithm is as follows: filtering the original signal; adaptively selecting the local discharge feature significant signal layer from the filtered original signal based on the number of VMD decomposition layers of the kurtosis index; and performing wavelet threshold denoising on the local discharge feature significant signal layer through a threshold function.

[0007] Preferably, after constructing the feature database, the spectrum center of gravity and wavelet energy entropy of each type of original signal are extracted by analyzing the time domain characteristics, frequency domain characteristics and time-frequency characteristics of each type of original signal in the feature database.

[0008] Preferably, the specific process of fuzzy logic weighting of the spectral center of gravity and wavelet energy entropy of each type of original signal is: determine the three-dimensional variables of each type of original signal respectively; wherein the three-dimensional variables include signal confidence, feature stability and historical matching degree; according to the interference effect of the three-dimensional variables on the original signal, fuzzy logic weighting is performed on the spectral center of gravity and wavelet energy entropy of each type of original signal.

[0009] Preferably, the decision threshold is generated as follows: after the initial decision threshold is preset, a historical database including multiple operating parameters is established, and based on the operating parameters in the historical database, the initial decision threshold is dynamically corrected through the Bayesian network to obtain the final decision threshold.

[0010] Preferably, after step S6, the method further includes: comparing the actual partial discharge state with the final judgment result of the partial discharge; when the partial discharge judgment result does not correspond to the actual partial discharge state, marking the current partial discharge judgment result as an abnormal data sample; when any one of the preset conditions is met, optimizing the decision threshold based on the incremental learning algorithm; wherein the preset conditions include: no abnormal data samples are generated in a certain quarter, 500 groups of abnormal data samples are added cumulatively, and the number of manual corrections to the partial discharge judgment results is greater than 10 times per month.

[0011] A generator partial discharge online monitoring system, comprising: A signal acquisition module is used to obtain four types of original signals from the generator, including ultrasonic signals, neutral point current signals, voltage phase signals, and ultra-high frequency current signals; A signal layer extraction module is used to extract the local discharge feature-significant signal layer from the original signal using an adaptive layered denoising algorithm; A feature extraction module is used to construct a feature library based on the PD feature significant signal layer, and extract the spectral center of gravity and wavelet energy entropy of each type of original signal from the feature library; The fuzzy logic weighting module is used to perform fuzzy logic weighting on the spectral center of gravity and wavelet energy entropy of each type of original signal to obtain the primary fusion result of each type of original signal; The preliminary judgment module is used to deeply fuse the primary fusion results of each type of original signal using an LSTM neural network to obtain a preliminary judgment result of partial discharge; The final judgment module is used to obtain a score for the preliminary partial discharge judgment result based on the abnormality score of the Mahalanobis distance, and determine the partial discharge judgment result based on the relationship between the score and a preset decision threshold. If the score is less than or equal to the decision threshold, partial discharge has not occurred; if the score is greater than the decision threshold, partial discharge has occurred.

[0012] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the on-line monitoring method for partial discharge of a generator are implemented.

[0013] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the on-line monitoring method for partial discharge of a generator.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention uses an adaptive layered denoising algorithm to separate high-frequency noise from effective discharge pulses in the signal, addressing the issues of discontinuous hard thresholds and constant soft threshold deviations found in traditional methods. Initial fusion is achieved through fuzzy logic weighting, followed by deep fusion using an LSTM (Long Short-Term Memory) neural network. Finally, anomaly scoring using Mahalanobis distance is performed to produce the final judgment result. This allows on-site operators to quickly and effectively identify and respond to generator partial discharges. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of online monitoring of partial discharge of a generator according to embodiment 1 of the present invention. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application.

[0017] Example 1: like Figure 1 As shown, this embodiment provides a method for online monitoring of partial discharge of a generator, including the following processes: S1, acquires four types of original signals of the generator in real time. The four types of original signals include ultrasonic signals, neutral point current signals, voltage phase signals and ultra-high frequency current signals.

[0018] S2, an adaptive hierarchical denoising algorithm is used to extract the PD feature-signal layer from the original signal.

[0019] S3, constructing a feature library based on the PD feature significant signal layer, and extracting the spectrum center of gravity and wavelet energy entropy of each type of original signal from the feature library.

[0020] S4, the spectral center of gravity and wavelet energy entropy of each type of original signal are weighted by fuzzy logic to obtain the primary fusion result of each type of original signal.

[0021] In S5, the LSTM neural network is used to deeply fuse the primary fusion results of each type of original signal to obtain the preliminary judgment results of partial discharge.

[0022] S6. A score for the preliminary partial discharge judgment result is obtained based on the abnormality score of the Mahalanobis distance, and the partial discharge judgment result is determined by comparing the score with a preset decision threshold. If the score is less than or equal to the decision threshold, no partial discharge has occurred; if the score is greater than the decision threshold, partial discharge has occurred.

[0023] Example 2: This embodiment provides a method for online monitoring of partial discharge of a generator, including the following process: 1. Use a multi-dimensional online monitoring system for uninterrupted monitoring.

[0024] The multi-dimensional online monitoring system uses a GPS (Global Positioning System) clock to achieve μs-level time alignment and synchronously collects the following four types of signals as raw signals: 1. An ultrasonic sensor is used to collect ultrasonic signals generated by partial discharge, and the operating frequency band of the ultrasonic sensor is 40-300kHz; 2. Neutral point current signals are collected using a Rogowski coil with an accuracy of at least ±0.5%; 3. Voltage phase signals are collected using a capacitive voltage divider; 4. Ultra-high frequency current signals are collected using a UHF sensor, and the operating frequency band of the UHF sensor is 300MHz-3GHz. The multi-dimensional online monitoring system is powered by electromagnetic induction to achieve 24 / 7 uninterrupted monitoring.

[0025] When the multi-dimensional online monitoring system performs synchronous acquisition, a clock synchronization compensation method is used to compensate the acquired signals for time, ensuring the time alignment of multi-sensor data and accurately correlating the multi-dimensional features of the same discharge event. The specific method is as follows:

[0026] where t correct represents the timestamp after correction, Δ cable Indicates the difference in cable transmission delay (such as the wiring length difference between the UHF sensor and the current transformer, synchronization accuracy <100ns, ensuring accurate phase resolution), t local Represents the original timestamp recorded by the local sensor, t master Indicates the time base of the master clock, t slave Indicates the timestamp reported by the slave sensor. Indicates half-value compensation of the master-slave time difference.

[0027] 2. Adaptive layered denoising algorithm.

[0028] An adaptive hierarchical denoising algorithm is used to extract the significant signal layer of partial discharge characteristics from the original signal. The adaptive hierarchical denoising algorithm includes the use of improved wavelet packet transform combined with variational mode decomposition, including the following steps: Step 1: Use a bandpass filter bank to implement hardware filtering on the original signal; Step 2: Based on the kurtosis index, VMD (Variational Mode Decomposition) is used to adaptively select the PD feature-signal layer: the original signal Decomposed into K modal components with different center frequencies ωk , effectively separate high-frequency noise and effective discharge pulses, and compare them with the signal decomposition under normal conditions. If obvious pulse changes occur, it is considered that local discharge has occurred in the layer, and the signal of this layer is extracted; In step 2, the number of VMD decomposition layers is adaptively selected based on the kurtosis index as follows:

[0029]

[0030] The modal components is the kth modal component, representing the local discharge components of different frequencies in the signal (such as high-frequency pulses of tip discharge or low-frequency oscillations of surface discharge). k is the center frequency of the kth mode, which identifies the main frequency band of the discharge pulse and is used to distinguish the discharge type (for example, air gap discharge is concentrated in 100–300 kHz). is the original signal, K is the number of adaptive decomposition layers, which is dynamically adjusted according to the signal kurtosis index to ensure efficient separation of noise and effective signals under complex working conditions, K∈[3,8].

[0031] Step 3: Perform wavelet threshold denoising on the signal layer with significant local discharge characteristics by improving the threshold function to solve the problems of discontinuity of traditional hard threshold and constant deviation of soft threshold, so as to improve the signal-to-noise ratio, suppress noise through the threshold function with smooth transition, and retain the steep edge characteristics of the discharge pulse.

[0032] In step 3, the threshold function is improved to implement wavelet threshold denoising as follows:

[0033] in is the threshold function, x is the wavelet coefficient, α is the smoothing factor, which controls the slope of the transition zone to avoid signal distortion caused by traditional hard threshold or energy loss of soft threshold. The wavelet threshold smoothing factor α∈[3,7], and e is a natural constant.

[0034]

[0035] Where σ is the noise standard deviation, N is the signal length, and the threshold τ is adaptively adjusted according to the noise standard deviation σ and the signal length N to ensure effective denoising in strong noise environments. ln is the logarithm.

[0036] 3. Multi-source feature fusion decision model.

[0037] A multi-source feature fusion decision model is used to perform four types of signal fusion judgment on the PD feature significant signal layer to obtain the PD judgment result. The multi-source feature fusion decision model includes the following steps: Step 1: construct a feature library based on the significant signal layer of partial discharge characteristics, and extract the spectral center of gravity and wavelet energy entropy of the four types of signals by analyzing the time domain characteristics, frequency domain characteristics and time-frequency characteristics of the four types of signals in the feature library.

[0038] In step 1, the spectral center of gravity and wavelet energy entropy feature extraction are as follows:

[0039] Among them, SC (Spectral Center) represents the central frequency index of the signal energy distribution in the frequency domain, which is used to quantify the "center position" of the signal energy in the spectrum. represents the signal spectrum, f k It represents the kth frequency component. The spectrum centroid characterizes the central frequency of the energy distribution of the partial discharge signal. The discharge type is distinguished by the difference between the high-frequency centroid and the low-frequency centroid. For example, the tip discharge usually has a higher SC value due to the steep pulse, while the surface discharge usually has a lower SC value due to the oscillation attenuation.

[0040]

[0041] J represents the number of decomposition layers, E represents the wavelet energy entropy, and E j Indicates the frequency band energy corresponding to the j-th layer wavelet coefficient, E total It represents the total energy of the signal in all frequency bands, that is, the global energy benchmark, quantifies the complexity of the energy distribution of the signal in different frequency bands, reflects the randomness of the discharge pattern, and distinguishes different discharge types by high and low entropy values. For example, air gap discharge presents multi-band energy dispersion due to random bubble breakdown, and the entropy value is relatively high; suspended discharge has a low entropy value because the periodic arc is concentrated in a specific frequency band.

[0042] Step 2: Fuzzy logic weighting is performed on the spectral center of gravity and wavelet energy entropy of each type of original signal to obtain the primary fusion result of each type of original signal. The fuzzy logic weighting uses three-dimensional variables of each type of original signal. The three-dimensional variables include signal confidence (calculated based on the signal-to-noise ratio to reflect signal quality, such as the reduced confidence of UHF signals in high electromagnetic interference scenarios), feature stability (assessing the volatility of feature values ​​through sliding window variance analysis to eliminate transient interference), and historical matching (comparing typical discharge patterns in the historical database to identify abnormal features). The fuzzy logic weighting is performed on the spectral center of gravity and wavelet energy entropy of each type of original signal, considering the interference effects of signal confidence, feature stability, and historical matching on the original signal. This achieves dynamic weight allocation and optimizes decision-making. In step 2, the fuzzy logic weighted fusion is as follows:

[0043] W i is the fusion weight of the i-th type of signal, dynamically adjusting the contribution of each signal to the final decision. i is the signal confidence membership function, μS i is the characteristic stability membership function, μH i is the historical matching membership function, (x) represents the input variable, which is used to dynamically evaluate signal characteristics or system status.

[0044] In step 3, the LSTM neural network is used to deeply fuse the primary fusion results of different signals to obtain a preliminary judgment result of partial discharge. The LSTM neural network is used to capture the temporal correlation of partial discharge signals (such as the periodicity and burstiness of discharge pulses) and improve the ability to predict the evolution of insulation defects.

[0045] In step 3, the LSTM time series feature modeling unit state is updated as follows:

[0046]

[0047]

[0048]

[0049]

[0050]

[0051] f tW represents the forget gate, which controls the degree of retention of historical memory (such as ignoring background noise in steady-state operation). σ represents the Sigmoid activation function, which compresses the linear transformation result to [0, 1] to control the degree of forgetting. f Represents the forget gate weight matrix, for input x t and the previous hidden state h t−1 Perform feature fusion. t−1 , x t ] represents the concatenation operation of hidden state and input: the hidden state h at the previous moment is t−1 and the current input x t Splice by dimension to form a comprehensive feature vector. b f Represents the forget gate bias term, adjusts the baseline value of the linear transformation, and enhances the flexibility of the multi-source feature fusion decision model; i t Represents the input gate, which filters the effective discharge-related features in the current input (such as the rising edge slope of the ultra-high frequency pulse), h t−1 Represents the hidden state of the previous moment, carrying the time series accumulation characteristics (such as the amplitude decay pattern of the historical discharge pulse), x t Represents the current input feature vector, which may contain real-time sensor data (such as UHF signal amplitude, phase spectrum entropy value), W i represents the dynamic feature selection matrix, which learns the weight distribution of different features through training, b i represents the adaptive bias, which controls the gate activation threshold (e.g., setting the baseline noise level to enable information flow only when the pulse energy exceeds the threshold); represents the candidate memory at the current moment (the original update amount that has not been fused with the historical memory), tanh represents the numerical characteristics, that is, the output range [-1, 1], the central symmetric gradient, W C represents the weight matrix, b C Indicates adjusting the activation threshold by offset, x t Represents the sensor input feature at the current moment, h t−1 Represents the hidden state of the previous moment (short-term memory carrier), [h t−1 , x t ] represents the concatenation operation, combining historical information with the current input; C t Represents the current unit state, which is used to store the accumulated information from the beginning of the sequence to the current moment, C t−1 Represents the historical unit state, that is, the global time series characteristics up to time t-1; t represents the output gate, output timing modeling results, and characterizes the changing trend of discharge severity; W o represents the weight matrix, b o represents the bias vector, each neuron independently adjusts the activation threshold, h tRepresents the hidden layer state, ⊙ represents the Hadamard product, tanh (C t ) means that the unit state C t Compressed to the range of [-1, 1], the number of LSTM hidden layer nodes ∈ [32, 128].

[0052] Step 4: Perform a secondary partial discharge judgment based on the abnormality score of the Mahalanobis distance: The score of the initial partial discharge judgment result is obtained based on the abnormality score of the Mahalanobis distance. This score measures the degree (score) of the deviation of the current feature vector x (spectral center of gravity and wavelet energy entropy) from the historical normal state (mean μ and covariance Σ). When the degree of deviation exceeds the preset decision threshold, specifically: when the scores of 3 / 4 of the four types of signals exceed the decision threshold, partial discharge is considered to have occurred, and the partial discharge judgment result is output.

[0053] In step 4, the abnormality score based on the Mahalanobis distance is as follows:

[0054] Among them, D 2 Represents the generalized square distance from the data point x to the center μ of the data set in the multidimensional space, through the covariance matrix Σ −1 Normalize and decorrelate the features, x∈R 4 Represents the current feature vector, μ∈R 4 represents the sliding window mean (historical 24 hours), Σ∈R 4×4 Represents the covariance matrix, reflecting the statistical correlation between different signal features (such as ultrasonic amplitude and UHF band energy), to avoid misjudgments caused by fluctuations in a single feature. μ and Σ are updated every 24 hours to adapt to changes in generator load, temperature, and other operating conditions.

[0055] 4. Decision threshold adjustment mechanism.

[0056] The decision threshold adjustment mechanism is used to dynamically modify the decision threshold. The decision threshold adjustment mechanism includes the following steps: Step 1: preset the initial decision threshold and establish a historical database containing operating parameters such as ambient temperature, ambient humidity, and load rate.

[0057] Step 2: Based on the operating parameters in the historical database, the initial decision threshold is dynamically modified through the Bayesian network to obtain the final decision threshold.

[0058] In step 2, the Bayesian network dynamically modifies the decision threshold and adjusts the decision threshold through P(T|L, N) to reduce false positives caused by environmental changes, as follows:

[0059] P uniformly represents a probability function. Its specific meaning and type vary depending on the combination of variables in the brackets. T represents the decision threshold, L represents the load factor (%), N represents the noise baseline (dB), and E represents the ambient temperature and humidity.

[0060] 5. Self-learning model optimization.

[0061] A self-learning model is used to optimize the decision threshold. The self-learning model optimization includes establishing a closed-loop feedback mechanism and using new data to continuously optimize the parameters of the multi-source feature fusion decision model to adapt to generator insulation aging or new defect modes. The steps include: Step 1: compare the actual partial discharge state with the partial discharge judgment result. When a false alarm is manually confirmed, that is, the partial discharge judgment result does not correspond to the actual partial discharge state, the current partial discharge judgment result is automatically marked as an abnormal data sample.

[0062] Step 2: When any one of the following conditions is met: no abnormal data samples are generated in the first quarter, 500 sets of abnormal data are added cumulatively, and the number of manual corrections is greater than 10 times / month, the parameters (decision threshold) of the multi-source feature fusion decision model are optimized based on the incremental learning algorithm.

[0063] In step 2, the specific process of optimizing the parameters of the multi-source feature fusion decision model based on the incremental learning algorithm is as follows:

[0064] Where θ represents the parameters of the multi-source feature fusion decision model, θ t Represents the parameter vector of the multi-source feature fusion decision model at the current moment (step t), θ t+1 Represents the updated parameter vector, which is calculated based on the current gradient and historical parameter changes. Represents the loss function with respect to the parameter θ t The gradient vector of , η = 0.001, which represents the learning rate, λ = 0.5, which represents the momentum factor, prevents the model from deviating from the learned general features due to short-term data fluctuations; L represents the loss function.

[0065] This embodiment simultaneously collects four types of signals and performs time compensation, improving the accuracy of multi-channel synchronization errors and clock drift compensation. It employs a three-layer hardware-based approach combining improved wavelet packet transforms and VMD (Variational Mode Decomposition) to separate high-frequency noise and effective discharge pulses from the signal, resolving the issues of discontinuous hard thresholds and constant soft threshold deviations in traditional systems. Feature extraction is then weighted using fuzzy logic to fuse the extracted features across three dimensions: signal confidence, feature stability, and historical matching. This is then further integrated through LSTM (Long Short-Term Memory) time series modeling, ultimately achieving anomaly scoring using the Mahalanobis distance. The decision threshold is revised based on Bayesian threshold judgment, and an early warning is triggered when the abnormal scores of 3 / 4 signal channels exceed the decision threshold. On this basis, adaptive incremental optimization learning is completed. When a false alarm is manually confirmed, the abnormal data samples are automatically marked. When any of the following conditions are met: no abnormal data samples are generated in the first quarter, 500 sets of abnormal data are added cumulatively, and the number of manual corrections is greater than 10 times / month, the multi-source feature fusion decision model parameter optimization based on the incremental learning algorithm is triggered. This not only makes it convenient for on-site operation and maintenance personnel to quickly judge the partial discharge of the generator, but also can continuously optimize and achieve improved accuracy in partial discharge identification.

[0066] Example 3: In this embodiment, a generator partial discharge online monitoring system is provided. The generator partial discharge online monitoring system can be used to implement the above-mentioned generator partial discharge online monitoring method. Specifically, the generator partial discharge online monitoring system includes a signal acquisition module, a signal layer extraction module, a feature extraction module, a fuzzy logic weighting module, a preliminary judgment module, and a final judgment module.

[0067] Among them, the signal acquisition module is used to obtain four types of original signals of the generator, which include ultrasonic signals, neutral point current signals, voltage phase signals and ultra-high frequency current signals.

[0068] The signal layer extraction module is used to extract the local discharge feature-significant signal layer from the original signal using an adaptive hierarchical denoising algorithm.

[0069] The feature extraction module is used to construct a feature library based on the local discharge feature significant signal layer, and extract the spectrum center of gravity and wavelet energy entropy of each type of original signal from the feature library.

[0070] The fuzzy logic weighting module is used to perform fuzzy logic weighting on the spectral center of gravity and wavelet energy entropy of each type of original signal to obtain the primary fusion result of each type of original signal.

[0071] The preliminary judgment module is used to deeply fuse the primary fusion results of each type of original signal using the LSTM neural network to obtain the preliminary judgment results of partial discharge.

[0072] The final judgment module is used to obtain a score for the preliminary partial discharge judgment result based on the abnormality score of the Mahalanobis distance, and determine the partial discharge judgment result based on the relationship between the score and a preset decision threshold. If the score is less than or equal to the decision threshold, partial discharge has not occurred; if the score is greater than the decision threshold, partial discharge has occurred.

[0073] Example 4: In this embodiment, a terminal device is provided, which includes a processor and a memory, wherein the memory is used to store a computer program, wherein the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function. The processor described in the embodiment of the present invention can be used for the operation of the online monitoring method for partial discharge of a generator, including: S1, obtaining four types of original signals of the generator, the four types of original signals including ultrasonic signals, neutral point current signals, voltage phase signals, and ultra-high frequency current signals; S2, extracting the significant signal layer of partial discharge characteristics from the original signals using an adaptive hierarchical denoising algorithm; S3, based on A feature library is constructed at the local discharge feature-significant signal layer, and the spectral centroid and wavelet energy entropy of each type of original signal are extracted from the feature library; S4, the spectral centroid and wavelet energy entropy of each type of original signal are fuzzy-logic weighted to obtain the primary fusion results of each type of original signal; S5, the primary fusion results of each type of original signal are deeply fused using an LSTM neural network to obtain a preliminary judgment result of local discharge; S6, a score of the preliminary judgment result of local discharge is obtained based on the abnormality score of the Mahalanobis distance, and the local discharge judgment result is determined by the relationship between the score and the preset decision threshold; wherein, if the score is less than or equal to the decision threshold, local discharge has not occurred, and if the score is greater than the decision threshold, local discharge has occurred.

[0074] Example 5: In this embodiment, a computer-readable storage medium (Memory) is provided. The computer-readable storage medium is a memory device in a terminal device, used to store programs and data. It is understood that the computer-readable storage medium herein may include both built-in storage media in the terminal device and, of course, extended storage media supported by the terminal device. The computer-readable storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for being loaded and executed by a processor. These instructions may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium herein may be high-speed RAM (Random Access Memory) or non-volatile memory, such as at least one disk storage device.

[0075] The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the on-line monitoring method for partial discharge of a generator in the above embodiment; the processor loads one or more instructions in the computer-readable storage medium and executes the following steps: S1, obtaining four types of original signals of the generator, the four types of original signals including ultrasonic signals, neutral point current signals, voltage phase signals and ultra-high frequency current signals; S2, extracting a local discharge feature significant signal layer from the original signals using an adaptive hierarchical denoising algorithm; S3, constructing a feature quantity library based on the local discharge feature significant signal layer, and extracting a local discharge feature significant signal layer from the feature quantity library. Take the spectral centroid and wavelet energy entropy of each type of original signal; S4, perform fuzzy logic weighting on the spectral centroid and wavelet energy entropy of each type of original signal to obtain the primary fusion results of each type of original signal; S5, use the LSTM neural network to deeply fuse the primary fusion results of each type of original signal to obtain the preliminary judgment result of partial discharge; S6, obtain the score of the preliminary judgment result of partial discharge based on the abnormality score of Mahalanobis distance, and determine the partial discharge judgment result by the relationship between the score and the preset decision threshold; wherein, if the score is less than or equal to the decision threshold, no partial discharge has occurred, and if the score is greater than the decision threshold, partial discharge has occurred.

[0076] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, optical storage, etc.) containing computer-usable program code.

[0077] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0078] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0079] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0080] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0081] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

[0082] It should be understood that the above description is for illustration and not for limitation. Many embodiments and many applications beyond the examples provided will be apparent to those skilled in the art upon reading the above description.

Claims

1. A method for online monitoring of partial discharge of a generator, characterized in that: The following processes are included: S1, obtain four types of original signals of the generator, including ultrasonic signals, neutral point current signals, voltage phase signals and ultra-high frequency current signals; S2, uses an adaptive hierarchical denoising algorithm to extract the significant signal layer of partial discharge characteristics from the original signal; S3, constructing a feature library based on the significant signal layer of partial discharge characteristics, and extracting the spectrum center of gravity and wavelet energy entropy of each type of original signal from the feature library; S4, performing fuzzy logic weighting on the spectral center of gravity and wavelet energy entropy of each type of original signal to obtain the primary fusion result of each type of original signal; S5, uses LSTM neural network to deeply fuse the primary fusion results of each type of original signal to obtain the preliminary judgment result of partial discharge; S6. A score for the preliminary partial discharge judgment result is obtained based on the abnormality score of the Mahalanobis distance, and the partial discharge judgment result is determined by comparing the score with a preset decision threshold. If the score is less than or equal to the decision threshold, no partial discharge has occurred; if the score is greater than the decision threshold, partial discharge has occurred.

2. The method for online monitoring of partial discharge of a generator according to claim 1, characterized in that: After obtaining the four types of original signals of the generator, the method further includes: using a clock synchronization compensation method to perform time compensation on the four types of original signals so as to align the time of each type of original signal.

3. The method for online monitoring of partial discharge of a generator according to claim 1, characterized in that: The specific process of extracting the significant signal layer of partial discharge characteristics from the original signal using the adaptive layered denoising algorithm is as follows: The original signal is filtered; based on the number of VMD decomposition layers of the kurtosis index, the local discharge feature-signal layer is adaptively selected from the filtered original signal; and the local discharge feature-signal layer is subjected to wavelet threshold denoising through a threshold function.

4. The method for online monitoring of partial discharge of a generator according to claim 1, characterized in that: After constructing the feature database, the spectrum center of gravity and wavelet energy entropy of each type of original signal are extracted by analyzing the time domain characteristics, frequency domain characteristics and time-frequency characteristics of each type of original signal in the feature database.

5. The method for online monitoring of partial discharge of a generator according to claim 1, characterized in that: The specific process of fuzzy logic weighting of the spectrum center and wavelet energy entropy of each type of original signal is as follows: Determine the three-dimensional variables of each type of original signal respectively; wherein the three-dimensional variables include signal confidence, feature stability and historical matching; According to the interference effect of three-dimensional variables on the original signal, fuzzy logic weighting is performed on the spectrum center of gravity and wavelet energy entropy of each type of original signal.

6. The method for online monitoring of partial discharge of a generator according to claim 1, characterized in that: The decision threshold is generated as follows: An initial decision threshold is preset, and a historical database including various operating parameters is established. Based on the operating parameters in the historical database, the initial decision threshold is dynamically corrected through the Bayesian network to obtain the final decision threshold.

7. The method for online monitoring of partial discharge of a generator according to claim 1, characterized in that: After step S6, the method further includes: Comparing the actual partial discharge state with the partial discharge judgment result, and marking the current partial discharge judgment result as an abnormal data sample when the partial discharge judgment result does not correspond to the actual partial discharge state; When any of the preset conditions is met, the decision threshold is optimized based on the incremental learning algorithm; the preset conditions include: no abnormal data samples are generated in a quarter, 500 sets of abnormal data samples are added cumulatively, and the number of manual corrections to the partial discharge judgment results is greater than 10 times / month.

8. A generator partial discharge online monitoring system, characterized in that: include: A signal acquisition module is used to obtain four types of original signals from the generator, including ultrasonic signals, neutral point current signals, voltage phase signals, and ultra-high frequency current signals; A signal layer extraction module is used to extract the local discharge feature-significant signal layer from the original signal using an adaptive layered denoising algorithm; A feature extraction module is used to construct a feature library based on the significant signal layer of partial discharge characteristics, and extract the spectrum center of gravity and wavelet energy entropy of each type of original signal from the feature library; The fuzzy logic weighting module is used to perform fuzzy logic weighting on the spectral center of gravity and wavelet energy entropy of each type of original signal to obtain the primary fusion result of each type of original signal; The preliminary judgment module is used to deeply fuse the primary fusion results of each type of original signal using an LSTM neural network to obtain a preliminary judgment result of partial discharge; The final judgment module is used to obtain a score for the preliminary partial discharge judgment result based on the abnormality score of the Mahalanobis distance, and determine the partial discharge judgment result based on the relationship between the score and a preset decision threshold. If the score is less than or equal to the decision threshold, partial discharge has not occurred; if the score is greater than the decision threshold, partial discharge has occurred.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method for online monitoring of partial discharge of a generator as claimed in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for online monitoring of partial discharge of a generator according to any one of claims 1 to 7 are implemented.

Images