A Deep Learning-Based Online Monitoring Method for Leakage Current of Zinc Oxide Surge Arresters

Abstract

This invention discloses a deep learning-based online monitoring method for leakage current of zinc oxide surge arresters. The method generates a preprocessed signal sequence with robust operating conditions; combines channel coupling features with dual-channel input signals to form a dual-channel input tensor; constructs a dual-channel autoencoder network framework and divides the dual-channel input tensor into training and validation datasets; inputs a new dual-channel input signal and outputs channel A reconstructed signal, channel B reconstructed signal, and corresponding latent variable vector; calculates channel reconstruction residuals, latent variable offsets, and channel consistency metrics, and combines them to generate an anomaly scoring vector; obtains a parameter posterior set; forms an online optimized parameter set; uses the online optimized parameter set to perform threshold determination on the anomaly scoring vector; outputs anomaly identification results with confidence intervals and triggers hierarchical alarm commands. This invention effectively reduces the false anomaly rate caused by non-abnormal interference from power grid transients, EMI, and operating condition disturbances.

Description

Technical Field

[0001] This invention relates to the field of zinc oxide surge arrester technology, and in particular to a method for online monitoring of leakage current of zinc oxide surge arresters based on deep learning. Background Technology

[0002] With the continuous improvement of the intelligence and automation level of power systems, zinc oxide surge arresters, as important overvoltage protection devices for substations and transmission lines, require online monitoring and anomaly diagnosis to ensure the safe and stable operation of the power grid. Currently, most mainstream zinc oxide surge arrester condition assessment methods are based on single or simplified indicators such as leakage current amplitude, power frequency fundamental component, or harmonic characteristics, and judge abnormal states by setting fixed thresholds. Existing methods generally have the following problems:

[0003] Traditional anomaly criteria based on single features or fixed thresholds are easily affected by changes in environmental conditions, making it difficult to stably separate resistive leakage components from harmonic distortion characteristics. The drift of capacitive current and harmonic components can easily mask the weak resistive leakage signals in the early degradation stage of zinc oxide surge arresters, leading to long-term mismatch of anomaly thresholds, prominent false alarm and false alarm problems, and difficulty in interpretably characterizing degradation modes under complex operating conditions.

[0004] Currently, the model's adaptive and parameter calibration capabilities are limited under conditions of small sample size, weak labeling, and cross-site migration. Most existing technologies rely on experience to set model hyperparameters and alarm thresholds, lacking automatic adjustment and compensation mechanisms for different seasons, different sites, and different bus voltage levels. They cannot effectively quantify and manage the uncertainty of identification results, making it difficult to adapt to long-term dynamic operating conditions and data distribution drift, resulting in insufficient model generalization ability and practical stability in actual operation and maintenance. Summary of the Invention

[0005] One objective of this invention is to propose an online monitoring method for leakage current of zinc oxide surge arresters based on deep learning. This invention effectively reduces the false anomaly rate caused by non-abnormal interferences such as power grid transients, EMI, and operating condition disturbances.

[0006] A method for online monitoring of leakage current of zinc oxide surge arresters based on deep learning according to an embodiment of the present invention includes:

[0007] Synchronous measurement data of the operating voltage and base current of the zinc oxide surge arrester are collected and preprocessed to generate a preprocessed signal sequence with robust operating conditions.

[0008] Based on the resistive leakage mechanism and harmonic distortion mechanism, the preprocessed signal sequence is mapped into a dual-channel input signal, and the channel coupling characteristics are combined with the dual-channel input signal to form a dual-channel input tensor;

[0009] A dual-channel autoencoder network framework is constructed, and the dual-channel input tensor is divided into a training dataset and a validation dataset.

[0010] Based on the health condition samples corresponding to the training dataset, a dual-channel autoencoder network is trained. The trained dual-channel autoencoder network is then deployed to the online recognition terminal. A new dual-channel input signal is input, and the channel A reconstructed signal, the channel B reconstructed signal, and the corresponding latent variable vector are output.

[0011] Based on the reconstructed signals of channel A and channel B and the latent variable vector, calculate the channel reconstruction residual, latent variable offset and channel consistency measure, and combine them to generate an anomaly score vector;

[0012] A parameter prior set is established, and probabilistic models are performed on the dual-channel autoencoder network structure parameters, loss weight parameters, environmental compensation parameters, and anomaly alarm threshold parameters. The anomaly score vector and the parameter prior set are input into the variational Bayesian optimization module to obtain the parameter posterior set.

[0013] Based on the posterior set of parameters and the expected utility function, the structural parameters of the dual-channel autoencoder network are updated to form an online optimized parameter set. The online optimized parameter set is used to determine the threshold of the anomaly scoring vector, output the anomaly identification result with confidence interval, and trigger the hierarchical alarm instruction.

[0014] Optionally, generating a preprocessed signal sequence with robust operating conditions includes:

[0015] The time series data of the operating voltage of the zinc oxide surge arrester, the time series data of the base current, the time series data of the ambient temperature, the time series data of the ambient humidity, and the time series data of the bus voltage level are acquired and synchronized according to a unified sampling timestamp to form a raw synchronized data set containing the main measured quantities and auxiliary operating condition information.

[0016] The base current in the original synchronization data set is processed using a bandpass filter to obtain the bandpass-processed current signal.

[0017] A power frequency reference phase signal is constructed using the time series data of the operating voltage as a reference. The current signal after bandpass processing is phase-aligned with the power frequency reference phase signal to obtain a current signal with a unified phase reference.

[0018] Based on Fourier transform, spectral analysis is performed on the current signal with a unified phase reference to form the odd harmonic energy characteristic vector.

[0019] Impulse transient detection is performed on a current signal with a unified phase reference to identify all short-term non-stationary abrupt changes in the signal, detect all abrupt change points, and form a set of transient pulse markers;

[0020] The current signal with a unified phase reference is normalized and compensated to obtain the normalized and compensated current signal.

[0021] The current signal after bandpass processing, the current signal with unified phase reference, the odd harmonic energy feature vector, the transient pulse mark set, and the normalized compensated current signal are combined to form a preprocessed signal sequence with operating condition robustness.

[0022] Optionally, the step of combining the channel coupling features with the dual-channel input signal to form a dual-channel input tensor includes:

[0023] The channel A input signal is constructed, which includes the current signal after bandpass processing, the current signal with unified phase reference, and the current signal after normalization compensation, and all of them are organized according to a unified sampling timestamp.

[0024] Construct the channel B input signal, which includes the odd harmonic energy eigenvectors and the transient pulse marker set;

[0025] The instantaneous phase difference between the phase-referenced current signal in the input signal of channel A and the harmonic mapping signal in the input signal of channel B is calculated, and the cosine value of the instantaneous phase difference is used as the channel coupling characteristic at each moment.

[0026] The dual-channel input tensor is obtained by temporally cascading the input signals from channel A and channel B with the channel coupling feature sequences at a unified time step.

[0027] Optionally, the dual-channel autoencoder network framework includes:

[0028] Construct a dual-channel autoencoder network framework that includes a channel A-specific encoder, a channel B-specific encoder, a shared latent variable pool, and a collaborative decoder;

[0029] The encoder specific to channel A encodes and maps the input signal of channel A to generate the latent variable representation of channel A;

[0030] The encoder specific to channel B encodes and maps the input signal of channel B to generate the latent variable representation of channel B;

[0031] The latent variable representations of channel A and channel B are concatenated to obtain a shared latent variable pool.

[0032] The collaborative decoder takes a shared latent variable pool as input and generates the reconstructed signals for channel A and channel B respectively.

[0033] The constructed dual-channel autoencoder network framework is applied to the dual-channel input tensor, and the dual-channel input tensor is divided into training dataset and validation dataset according to time order.

[0034] Optionally, training a dual-channel autoencoder network based on health condition samples corresponding to the training dataset includes:

[0035] The overall training loss function is obtained by weighting the reconstruction error of channel A, the reconstruction error of channel B, and the cross-channel consistency loss.

[0036] Based on the training dataset and the overall training loss function value, the backpropagation algorithm is used to update the parameters of all modules in the dual-channel autoencoder network until the overall training loss function value converges, thus obtaining the set of model parameters after training.

[0037] The trained dual-channel autoencoder network is deployed to the online recognition terminal. The newly acquired dual-channel input tensor is input, and the output is the reconstructed signal of channel A, the reconstructed signal of channel B, the latent variable representation of channel A, and the latent variable representation of channel B.

[0038] Optionally, training a dual-channel autoencoder network based on health condition samples corresponding to the training dataset includes:

[0039] Calculate the channel A reconstruction residual, which is used to measure the mean square reconstruction error between the channel A input signal and the channel A reconstructed signal;

[0040] Calculate the latent variable offset, which measures the overall deviation between the latent variable distribution under the current operating condition and the reference vector of the latent variable mean under the healthy operating condition;

[0041] Calculate a channel consistency metric to measure the structural consistency between the latent variable representations of channel A and channel B;

[0042] The anomaly score vector is obtained by weighted summing of the channel A reconstruction residual, the channel B reconstruction residual, the latent variable offset, and the channel consistency metric.

[0043] Optionally, the step of inputting the anomaly score vector and the parameter prior set into the variational Bayesian optimization module includes:

[0044] Establish a parameter prior set, which includes dual-channel autoencoder network structure parameters, loss weight parameters, environmental compensation parameters, and anomaly alarm threshold parameters;

[0045] For each parameter in the prior set of parameters, a prior distribution model is established. The network structure parameters are modeled using a discrete uniform distribution, the loss weight parameters are modeled using a weight normalization distribution, the environmental compensation parameters are modeled using a multivariate normal distribution, and the abnormal alarm threshold parameters are modeled using an interval probability distribution.

[0046] The abnormal score vector and the parameter prior set are used as inputs to the variational Bayesian optimization module to establish a variational inference problem with the maximum likelihood of the posterior distribution as the objective.

[0047] The target is iteratively optimized using the variational Bayesian optimization module to obtain the posterior set of parameters.

[0048] Optionally, the output of the anomaly identification result with confidence interval and the triggering of the hierarchical alarm instruction includes:

[0049] Based on the parameter posterior set, the mean of the optimal posterior distribution of network structure parameters, the mean of the optimal posterior distribution of loss weight parameters, the mean of the optimal posterior distribution of environmental compensation parameters, and the mean of the optimal posterior distribution of abnormal alarm threshold parameters are extracted respectively, and used as the online optimization parameter set for the current round.

[0050] The threshold of the abnormal score vector is determined by using an online optimized parameter set. The mean of the optimal posterior distribution of the abnormal alarm threshold parameters is extracted as the current alarm threshold. The upper and lower limits of the confidence interval of the alarm threshold are determined by combining the variance of the optimal posterior distribution of the abnormal alarm threshold parameters.

[0051] Alarm classification is based on the relationship between the anomaly scoring vector and the confidence interval of the alarm threshold:

[0052] When the anomaly score vector is less than the lower limit of the confidence interval of the alarm threshold, a prompt-level alarm is output, indicating that there is no obvious abnormality in the current working condition and routine inspection is recommended;

[0053] When the abnormal scoring vector is greater than or equal to the lower limit of the confidence interval of the alarm threshold and less than the upper limit of the confidence interval of the alarm threshold, a warning-level alarm is output, indicating that there is a deviation in the current working condition but it is not serious. It is recommended to strengthen monitoring or conduct on-site re-inspection.

[0054] When the abnormal score vector is greater than or equal to the upper limit of the confidence interval of the alarm threshold, an alarm-level alarm is output, indicating that the current operating condition is highly abnormal, and it is recommended to immediately perform infrared temperature measurement or power outage inspection and handling measures.

[0055] The beneficial effects of this invention are:

[0056] This invention improves the detectability of early degradation and the resolution of abnormal signals under complex operating conditions through dual-channel structural decoupling under mechanistic constraints. It adopts a dual-channel self-encoding network structure based on resistive leakage mechanism and harmonic distortion mechanism. Channel A is constructed to characterize resistive leakage and power frequency fundamental frequency dominant characteristics, and channel B is constructed to characterize harmonic distortion and transient impulse dominant characteristics. Channel coupling characteristics are introduced into the latent space to structurally separate and jointly model the manifold structure of multiple physical quantity current signals under healthy operating conditions. This significantly improves the sensitivity of anomaly detection under weak resistive leakage and mixed harmonic interference, accurately captures the degradation symptoms of zinc oxide surge arresters, and effectively reduces the false anomaly rate caused by non-abnormal interference such as power grid transients, EMI, and operating condition disturbances.

[0057] This invention introduces a variational Bayesian optimization module to probabilistically model network structure parameters, loss weight parameters, environmental compensation parameters, and anomaly alarm threshold parameters. It dynamically inputs the anomaly scoring vector and the parameter prior set to obtain the parameter posterior distribution and automatically updates the parameters online based on the expected utility criterion. This enables the parameters to self-calibrate with seasonal changes, site conditions, bus voltage, and external environmental conditions, ensuring the robustness of the method and low false alarms and low false negatives during cross-seasonal, cross-regional, and long-term operation. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a flowchart of an online monitoring method for leakage current of zinc oxide surge arresters based on deep learning, as proposed in this invention. Detailed Implementation

[0060] Example 1: Reference Figure 1 A deep learning-based online monitoring method for leakage current of zinc oxide surge arresters includes:

[0061] Synchronous measurement data of the operating voltage and base current of the zinc oxide surge arrester are collected and preprocessed to generate a preprocessed signal sequence with robust operating conditions.

[0062] In this embodiment, a preprocessed signal sequence with robust operating conditions is generated, including:

[0063] The time series data of the operating voltage of the zinc oxide surge arrester, the time series data of the base current, the time series data of the ambient temperature, the time series data of the ambient humidity, and the time series data of the bus voltage level are acquired and synchronized according to a unified sampling timestamp to form a raw synchronized data set containing the main measured quantities and auxiliary operating condition information.

[0064] The total number of sampling times is N. The operating voltage of the zinc oxide surge arrester is the phase voltage value at each sampling time. The base current of the zinc oxide surge arrester is the base current value at each sampling time. The ambient temperature of the zinc oxide surge arrester is the ambient temperature value at each sampling time. The ambient humidity of the zinc oxide surge arrester is the ambient humidity value at each sampling time. The bus voltage level of the zinc oxide surge arrester is the bus voltage level discrete label at each sampling time.

[0065] The base current in the original synchronization data set is processed using a bandpass filter to obtain the bandpass-processed current signal.

[0066] The passband range of a bandpass filter is defined by a set of lower and upper frequency limits. Signal processing involves filtering out all current signal components that are below the set lower bandpass filter frequency limit or above the set upper bandpass filter frequency limit.

[0067] A power frequency reference phase signal is constructed using the time series data of the operating voltage as a reference. The current signal after bandpass processing is phase-aligned with the power frequency reference phase signal to obtain a current signal with a unified phase reference.

[0068] Based on Fourier transform, spectral analysis is performed on the current signal with a unified phase reference to form the odd harmonic energy characteristic vector.

[0069] The energy of all odd harmonic components in the signal is extracted. The energy of the odd harmonic component is defined as the current signal energy at each odd harmonic frequency. The energy of all odd harmonic components is calculated as odd multiples of the power frequency. The highest odd harmonic order is set to K to form an odd harmonic energy feature vector. The odd harmonic energy feature vector is used to measure the energy distribution of each order of odd harmonics in the base current signal.

[0070] Impulse transient detection is performed on a current signal with a unified phase reference to identify all short-term non-stationary abrupt changes in the signal, detect all abrupt change points, and form a set of transient pulse markers;

[0071] The transient pulse marker set consists of all detected moments of impulse abrupt changes.

[0072] The current signal with a unified phase reference is normalized and compensated to obtain the normalized and compensated current signal.

[0073] The current signal after bandpass processing, the current signal with unified phase reference, the odd harmonic energy feature vector, the transient pulse mark set, and the normalized compensated current signal are combined to form a preprocessed signal sequence with operating condition robustness.

[0074] Based on the resistive leakage mechanism and harmonic distortion mechanism, the preprocessed signal sequence is mapped into a dual-channel input signal, and the channel coupling characteristics are combined with the dual-channel input signal to form a dual-channel input tensor;

[0075] In this embodiment, the channel coupling feature is combined with the dual-channel input signal to form a dual-channel input tensor, including:

[0076] The channel A input signal is constructed, which includes the current signal after bandpass processing, the current signal with unified phase reference, and the current signal after normalization compensation, and all of them are organized according to a unified sampling timestamp.

[0077] The input signal of channel A represents the resistive leakage and power frequency fundamental frequency dominance characteristics of the zinc oxide surge arrester, and each group of input signals has a uniform current unit.

[0078] Construct the channel B input signal, which includes the odd harmonic energy eigenvectors and the transient pulse marker set;

[0079] Each dimension of the odd harmonic energy feature vector represents the current signal energy corresponding to a specific odd harmonic frequency. The amplitude is obtained by taking the square root of the energy at the frequency. The harmonic frequency is an odd multiple of the power frequency. Both the amplitude and frequency have the same physical meaning. The transient pulse marker set marks each detected impulse change moment as 1 and other moments as 0 according to a unified sampling timestamp. The channel B input signal represents the harmonic distortion and transient impulse dominant characteristics of the zinc oxide arrester base current. The amplitude characteristics and logic characteristics are combined according to a unified time step.

[0080] The instantaneous phase difference between the phase-referenced current signal in the input signal of channel A and the harmonic mapping signal in the input signal of channel B is calculated, and the cosine value of the instantaneous phase difference is used as the channel coupling characteristic at each moment.

[0081] The channel coupling characteristic value ranges from negative one to one, representing the synchronization relationship between the two types of signals at the corresponding time.

[0082] The dual-channel input tensor is obtained by temporally cascading the input signals from channel A and channel B with the channel coupling feature sequences at a unified time step.

[0083] The dual-channel input tensor can simultaneously express multiple features such as resistive leakage, power frequency fundamental wave, harmonic distortion and transient impact, and uses a unified data structure as the standard input for the dual-channel autoencoder network. All input signals and features maintain consistency in sampling time, unit and physical semantics.

[0084] A dual-channel autoencoder network framework is constructed, and the dual-channel input tensor is divided into a training dataset and a validation dataset.

[0085] In this embodiment, the dual-channel autoencoder network framework includes:

[0086] Construct a dual-channel autoencoder network framework that includes a channel A-specific encoder, a channel B-specific encoder, a shared latent variable pool, and a collaborative decoder;

[0087] The encoder specific to channel A encodes and maps the input signal of channel A to generate the latent variable representation of channel A;

[0088] In Example 1, the channel A-specific encoder is a deep neural network composed of multiple nonlinear mappings. Specifically, it takes the channel A input signal as input data and sequentially extracts the resistive leakage and power frequency fundamental wave dominant features from the channel A input signal through a linear transformation layer, an activation function layer, and a batch normalization layer. The latent variable representation of channel A is output in the last layer of the deep neural network. All network parameters of the channel A-specific encoder are obtained by end-to-end backpropagation optimization based on the training dataset. The optimization objective is to minimize the reconstruction error between the channel A input signal and the channel A reconstructed signal, while taking into account the cross-channel consistency loss. The optimized channel A-specific encoder is used to generate the latent variable representation of channel A. The latent variable representation is the latent space feature expression at each sampling time and corresponds to the normalized expression of the current signal in physical semantics.

[0089] The encoder specific to channel B encodes and maps the input signal of channel B to generate the latent variable representation of channel B;

[0090] In Example 1, the channel B-specific encoder is a deep neural network structure dedicated to processing frequency domain features and impulse logic features. Specifically, it takes the channel B input signal as input data and sequentially passes it through a pyramidal dilated convolutional layer, a multi-scale feature extraction layer, a nonlinear activation layer, and a feature fusion layer to progressively extract the dominant harmonic distortion features and transient impulse features in the channel B input signal. The latent variable representation of channel B is output at the last layer of the network. All network parameters of the channel B-specific encoder are obtained through end-to-end backpropagation optimization based on the training dataset. The optimization objective is to minimize the reconstruction error between the channel B input signal and the channel B reconstructed signal, while also taking into account cross-channel consistency loss. The optimized channel B-specific encoder is used to generate the latent variable representation of channel B. The latent variable representation is the latent space feature expression at each sampling time and corresponds to the joint expression of frequency domain energy features and impulse markers in physical semantics.

[0091] The latent variable representations of channel A and channel B are concatenated to obtain a shared latent variable pool.

[0092] The shared latent variable pool serves as the input representation of the collaborative decoder, simultaneously expressing the structural commonalities and differences in the latent space between two signals: resistive leakage and power frequency fundamental wave dominant features, and harmonic distortion and transient impulse dominant features.

[0093] The collaborative decoder takes a shared latent variable pool as input and generates the reconstructed signals for channel A and channel B respectively.

[0094] The reconstructed signal of channel A is a time series of resistive leakage and power frequency fundamental wave dominant characteristics recovered by the collaborative decoder based on the shared hidden variable pool. The reconstructed signal of channel B is a time series of harmonic distortion and transient impulse dominant characteristics recovered by the collaborative decoder based on the shared hidden variable pool. The two sets of reconstructed signals correspond one-to-one with the input signals of channel A and channel B at each sampling time, and the unit of the reconstructed signal is consistent with that of the original input signal.

[0095] The constructed dual-channel autoencoder network framework is applied to the dual-channel input tensor, and the dual-channel input tensor is divided into training dataset and validation dataset according to time order.

[0096] The training dataset includes data from the first few time steps, and the validation dataset includes data from the subsequent few time steps. The training dataset is used to train network parameters, and the validation dataset is used to evaluate network performance. All data after the split maintains consistency in input structure, units, and physical semantics.

[0097] Based on the health condition samples corresponding to the training dataset, a dual-channel autoencoder network is trained. The trained dual-channel autoencoder network is then deployed to the online recognition terminal. A new dual-channel input signal is input, and the channel A reconstructed signal, the channel B reconstructed signal, and the corresponding latent variable vector are output.

[0098] In this embodiment, a dual-channel autoencoder network is trained based on the health condition samples corresponding to the training dataset, including:

[0099] The overall training loss function is obtained by weighting the reconstruction error of channel A, the reconstruction error of channel B, and the cross-channel consistency loss.

[0100] The channel A reconstruction error is used to measure the mean square reconstruction error between the channel A input signal and the channel A reconstructed signal. The channel B reconstruction error is used to measure the mean square reconstruction error between the channel B input signal and the channel B reconstructed signal. The cross-channel consistency loss is used to measure the degree of difference between channel A and channel B in the latent variable space representation. The three are multiplied by the weighting coefficient of the loss function and then added together to obtain the overall training loss function value.

[0101] The channel A reconstruction error is obtained by squaring the difference between the channel A input signal and the channel A reconstructed signal at each sampling time, summing the results, and then dividing by the total number of sampling times. The channel B reconstruction error is obtained by squaring the difference between the channel B input signal and the channel B reconstructed signal at each sampling time, summing the results, and then dividing by the total number of sampling times. The cross-channel consistency loss is obtained by squaring the difference between the channel A latent variable representation and the channel B latent variable representation at each sampling time, summing the results, and then dividing by the total number of sampling times.

[0102] Based on the training dataset and the overall training loss function value, the backpropagation algorithm is used to update the parameters of all modules in the dual-channel autoencoder network until the overall training loss function value converges, thus obtaining the set of model parameters after training.

[0103] The model parameter set corresponds to the optimal parameter set for the channel A encoder, channel B encoder, shared latent variable pool, channel A decoder, and channel B decoder, respectively.

[0104] The trained dual-channel autoencoder network is deployed to the online recognition terminal. The newly acquired dual-channel input tensor is input, and the output is the reconstructed signal of channel A, the reconstructed signal of channel B, the latent variable representation of channel A, and the latent variable representation of channel B.

[0105] The channel A input signal is mapped to a channel A latent variable representation by the channel A encoder, and the channel B input signal is mapped to a channel B latent variable representation by the channel B encoder. The channel A latent variable representation and the channel B latent variable representation are concatenated at each sampling time to form a shared latent variable pool. The shared latent variable pool is input to the collaborative decoder, which generates the channel A reconstructed signal and the channel B reconstructed signal, respectively. The channel A reconstructed signal maintains the same sampling step size and unit as the channel A input signal, and the channel B reconstructed signal maintains the same dimension and physical semantics as the channel B input signal. The channel A latent variable representation and the channel B latent variable representation are used for subsequent latent space drift detection.

[0106] Based on the reconstructed signals of channel A and channel B and the latent variable vector, calculate the channel reconstruction residual, latent variable offset and channel consistency measure, and combine them to generate an anomaly score vector;

[0107] In this embodiment, a dual-channel autoencoder network is trained based on the health condition samples corresponding to the training dataset, including:

[0108] Calculate the channel A reconstruction residual, which is used to measure the mean square reconstruction error between the channel A input signal and the channel A reconstructed signal;

[0109] The channel A reconstruction residual is calculated as follows: for each sampling time, the difference between the channel A input signal and the channel A reconstruction signal is calculated and squared. The squared differences of all sampling times are accumulated, and the accumulated value is divided by the total number of sampling times to obtain the channel A reconstruction residual.

[0110] The channel B reconstruction residual is calculated as follows: for each sampling time, first calculate the difference between the channel B input signal and the channel B reconstructed signal and square it. Then, accumulate the squared differences of all sampling times and finally divide the accumulated value by the total number of sampling times to obtain the channel B reconstruction residual.

[0111] Calculate the latent variable offset, which measures the overall deviation between the latent variable distribution under the current operating condition and the reference vector of the latent variable mean under the healthy operating condition;

[0112] The latent variable offset is calculated as follows: for each sampling time, the difference between the current shared latent variable pool representation and the reference vector of the latent variable mean under healthy operating conditions is calculated and squared. The squared differences of all sampling times are accumulated, and the accumulated value is divided by the total number of sampling times to obtain the latent variable offset.

[0113] Calculate a channel consistency metric to measure the structural consistency between the latent variable representations of channel A and channel B;

[0114] The channel consistency metric is calculated as follows: For each sampling time, first calculate the difference between the latent variable representation of channel A and the latent variable representation of channel B, then square the difference, sum the squared differences of all sampling times, and finally divide the sum by the total number of sampling times. The result is the channel consistency metric.

[0115] The anomaly score vector is obtained by weighted summing of the channel A reconstruction residual, the channel B reconstruction residual, the latent variable offset, and the channel consistency metric.

[0116] The larger the abnormality score vector value, the greater the deviation between the current operating condition and the healthy operating condition.

[0117] A parameter prior set is established, and probabilistic models are performed on the dual-channel autoencoder network structure parameters, loss weight parameters, environmental compensation parameters, and anomaly alarm threshold parameters. The anomaly score vector and the parameter prior set are input into the variational Bayesian optimization module to obtain the parameter posterior set.

[0118] In this embodiment, the anomaly score vector and the parameter prior set are input into the variational Bayesian optimization module, including:

[0119] Establish a parameter prior set, which includes dual-channel autoencoder network structure parameters, loss weight parameters, environmental compensation parameters, and anomaly alarm threshold parameters;

[0120] Network structure parameters represent the number of layers, the number of neurons per layer, and the bottleneck dimension. The parameter values ​​range between the given minimum and maximum layers, and the unit is the number of nodes. The loss weight parameter is used to balance the contributions of channel A reconstruction error, channel B reconstruction error, and cross-channel consistency loss. The environmental compensation parameter describes the impact of temperature, humidity, and bus voltage on the normalization compensation of the current signal. Each environmental compensation parameter is the adjustment rate of the impact of each environmental variable on the current gain. The abnormal alarm threshold parameter represents the final scoring threshold used for alarms.

[0121] For each parameter in the prior set of parameters, a prior distribution model is established. The network structure parameters are modeled using a discrete uniform distribution, the loss weight parameters are modeled using a weight normalization distribution, the environmental compensation parameters are modeled using a multivariate normal distribution, and the abnormal alarm threshold parameters are modeled using an interval probability distribution.

[0122] The abnormal score vector and the parameter prior set are used as inputs to the variational Bayesian optimization module to establish a variational inference problem with the maximum likelihood of the posterior distribution as the objective.

[0123] The objective is to minimize the information distance between the prior and posterior distributions, while maximizing the probability of generating anomaly rating vectors under the given parameters. The posterior distribution is the probability distribution of all parameters, and the prior distribution is the distribution model of each parameter in the prior set of parameters. The information distance between the prior and posterior distributions is measured by an information distance metric function, and the probability of generating anomaly rating vectors is measured by the logarithm of the conditional probability.

[0124] The target is iteratively optimized using the variational Bayesian optimization module to obtain the posterior set of parameters.

[0125] The parameter posterior set includes the optimal posterior distributions of network structure parameters, loss weight parameters, environmental compensation parameters, and abnormal alarm threshold parameters. Each parameter posterior distribution contains the mean and variance.

[0126] Based on the posterior set of parameters and the expected utility function, the structural parameters of the dual-channel autoencoder network are updated to form an online optimized parameter set. The online optimized parameter set is used to determine the threshold of the anomaly scoring vector, output the anomaly identification result with confidence interval, and trigger the hierarchical alarm instruction.

[0127] In this embodiment, the anomaly identification result with confidence interval is output and a tiered alarm command is triggered, including:

[0128] Based on the parameter posterior set, the mean of the optimal posterior distribution of network structure parameters, the mean of the optimal posterior distribution of loss weight parameters, the mean of the optimal posterior distribution of environmental compensation parameters, and the mean of the optimal posterior distribution of abnormal alarm threshold parameters are extracted respectively, and used as the online optimization parameter set for the current round.

[0129] The threshold of the abnormal score vector is determined by using an online optimized parameter set. The mean of the optimal posterior distribution of the abnormal alarm threshold parameters is extracted as the current alarm threshold. The upper and lower limits of the confidence interval of the alarm threshold are determined by combining the variance of the optimal posterior distribution of the abnormal alarm threshold parameters.

[0130] The upper and lower limits of the confidence interval for the alarm threshold are respectively the mean of the optimal posterior distribution minus the product of the confidence level factor and the variance of the optimal posterior distribution, and the mean of the optimal posterior distribution plus the product of the confidence level factor and the variance of the optimal posterior distribution. The confidence level factor is used to adjust the confidence strength of the interval.

[0131] Alarm classification is based on the relationship between the anomaly scoring vector and the confidence interval of the alarm threshold:

[0132] When the anomaly score vector is less than the lower limit of the confidence interval of the alarm threshold, a prompt-level alarm is output, indicating that there is no obvious abnormality in the current working condition and routine inspection is recommended;

[0133] When the abnormal scoring vector is greater than or equal to the lower limit of the confidence interval of the alarm threshold and less than the upper limit of the confidence interval of the alarm threshold, a warning-level alarm is output, indicating that there is a deviation in the current working condition but it is not serious. It is recommended to strengthen monitoring or conduct on-site re-inspection.

[0134] When the abnormal score vector is greater than or equal to the upper limit of the confidence interval of the alarm threshold, an alarm-level alarm is output, indicating that the current operating condition is highly abnormal, and it is recommended to immediately perform infrared temperature measurement or power outage inspection and handling measures.

[0135] Example 2: In a real-world operation and maintenance scenario for online monitoring of zinc oxide surge arresters, the system performs high-frequency synchronous data acquisition on the operating voltage, base current, ambient temperature, ambient humidity, and bus voltage level of multiple surge arresters around the clock. A set of signal segments is automatically generated every hour, each segment containing 60 seconds of high-sampling-rate multi-source raw data. During a certain equipment inspection, the operation and maintenance system detected abnormal fluctuations in the base current signal of a certain surge arrester over several consecutive cycles. The abnormal score of the signal segment repeatedly approached and exceeded the system's adaptive threshold.

[0136] In a set of data collected under high humidity and high temperature conditions, the reconstruction residual of channel A of the surge arrester base current gradually increased from 0.007 under normal conditions to 0.021, while the reconstruction residual of channel B increased from 0.008 to 0.013. Simultaneously, the latent variable offset increased from 0.006 to 0.020, the channel consistency metric also increased from 0.011 to 0.029, and the final anomaly score increased from 0.07 to 0.28. At this point, the system dynamically determines, based on the online optimization parameter set, that the anomaly score has exceeded the alarm threshold of the current confidence interval, automatically generates a warning-level alarm command, and pushes an anomaly report through the monitoring backend.

[0137] The anomaly report details all signal characteristics during this data acquisition period, including instantaneous changes in base current, power frequency phase-locked loop amplitude, odd-order harmonic energy (e.g., 3rd harmonic energy 0.027 A², 5th harmonic energy 0.014 A²), the specific time of sudden impact, and all operating parameters such as temperature (e.g., 27℃), humidity (e.g., 83%), and bus voltage level (e.g., level 2). The report also lists the confidence interval determination process, anomaly scoring sub-indicators, and displays the alarm level as "early warning," recommending on-site infrared thermography and insulation retesting.

[0138] After receiving the alarm, the maintenance engineer conducted an on-site inspection and found localized moisture damage to the equipment surface, resulting in reduced insulation resistance. Subsequent comparison of manual maintenance logs and offline inspection reports confirmed that the alarm was a genuine early degradation signal. Insulation treatment was performed on-site according to the recommended measures, and the equipment returned to healthy operation. During the same period, using the traditional amplitude threshold method and frequency domain energy method, the threshold was passively raised due to capacitive current and harmonic drift under high humidity conditions, failing to output alarms in a timely manner, resulting in missed alarms.

[0139] During another equipment monitoring operation, a surge arrester experienced a strong transient disturbance in its base current signal during a grid switching operation. The reconstruction residual of channel B instantly increased from 0.009 to 0.024, while the reconstruction residual of channel A remained below 0.008, with the highest anomaly score reaching 0.19. Since the dynamic threshold of this invention was adjusted to 0.22 (confidence interval upper limit 0.25) based on online Bayesian optimization, the system automatically output a prompt-level alarm. No warning or alarm was triggered, indicating normal grid operation interference. Traditional methods, unable to distinguish between power frequency dominance and harmonic disturbances, sometimes misinterpreted abnormal fluctuations as alarms, even after actual verification confirmed no equipment anomaly.

[0140] To further verify the robustness and effectiveness of the present invention, detailed statistics were performed on three consecutive months of on-site monitoring data, and hundreds of signal segments were randomly selected for manual annotation and verification. Taking 100 signal segments manually annotated as early signs of moisture degradation as an example, the method of the present invention could accurately detect 98 segments, with 1 false alarm and 1 missed alarm. In the same batch, the traditional amplitude threshold method detected only 89 segments, with 5 false alarms and 6 missed alarms, while the traditional frequency domain energy method detected 92 segments, with 3 false alarms and 5 missed alarms.

[0141] Taking the segment in Example 2 as an example, the signal samples collected for the healthy segment are as follows: base current 0.36A, channel A reconstruction residual 0.007, channel B reconstruction residual 0.009, latent variable offset 0.005, and anomaly score 0.06. The abnormal segment samples are: base current 0.41A, channel A reconstruction residual 0.021, channel B reconstruction residual 0.015, latent variable offset 0.019, and anomaly score 0.31. The system automatically determines that the healthy segment is at the prompt level and the abnormal segment is at the alarm level.

[0142] Meanwhile, this invention can dynamically adjust the compensation coefficients and alarm thresholds based on the variational Bayesian optimization mechanism under complex seasonal changes, bus voltage switching, and various disturbance scenarios. In Example 2, when the ambient temperature suddenly rises from 15℃ to 32℃, the model automatically adjusts the temperature compensation parameter to 0.006, maintaining the discriminative effectiveness of the reconstruction residuals and anomaly scores of each channel, and avoiding false alarms and threshold drift problems under high-temperature conditions. The adaptive adjustment process of all parameters is automatically completed with each new data acquisition, and the confidence interval and suggested maintenance level are output simultaneously, greatly improving the engineering practicality and traceability of anomaly detection.

[0143] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based zinc oxide surge arrester leakage current online monitoring method, characterized in that, The method comprises the following steps: Synchronization measurement data of operating voltage and base current of zinc oxide surge arrester are collected and preprocessed to generate preprocessed signal sequences with robust working conditions; According to the resistive leakage mechanism and the harmonic distortion mechanism, the preprocessed signal sequences are mapped into double-channel input signals, and the channel coupling features and the double-channel input signals are combined to form a double-channel input tensor; A double-channel auto-encoding network framework is constructed, and the double-channel input tensor is divided into a training data set and a validation data set; Based on the healthy working condition samples corresponding to the training data set, the double-channel auto-encoding network is trained, and the trained double-channel auto-encoding network is deployed to an online identification end to input new double-channel input signals, output channel A reconstructed signals, channel B reconstructed signals and corresponding latent variable vectors; According to the channel A reconstructed signals, the channel B reconstructed signals and the latent variable vectors, channel reconstruction residuals, latent variable offsets and channel consistency measures are calculated to generate an abnormal score vector; A parameter prior set is established to probabilistically model the double-channel auto-encoding network structure parameters, loss weight parameters, environment compensation parameters and abnormal alarm threshold parameters, and the abnormal score vector and the parameter prior set are input into a variational Bayesian optimization module to obtain a parameter posterior set; According to the parameter posterior set and an expected utility function, the double-channel auto-encoding network structure parameters are updated to form an online optimization parameter set, the abnormal score vector is thresholded using the online optimization parameter set, and an abnormal identification result with a confidence interval is output and a hierarchical alarm instruction is triggered.

2. The method according to claim 1, wherein, The preprocessed signal sequences with robust working conditions are generated by the following steps: Time series data of operating voltage, time series data of base current, time series data of environmental temperature, time series data of environmental humidity and time series data of bus voltage grade are obtained, and are synchronized according to a unified sampling timestamp to form an original synchronization data set containing main measurement quantities and auxiliary working condition information; The base current in the original synchronization data set is processed by a band-pass filter to obtain a band-pass processed current signal; A power frequency reference phase signal is constructed with the time series data of operating voltage as a reference, and the band-pass processed current signal and the power frequency reference phase signal are phase-aligned to obtain a phase reference unified current signal; The phase reference unified current signal is subjected to frequency spectrum analysis based on Fourier transform to form an odd harmonic energy feature vector; The phase reference unified current signal is subjected to impulse transient detection to identify all short-time non-stationary abrupt components in the signal, detect all abrupt points and form a transient pulse marker set; The phase reference unified current signal is normalized and compensated to obtain a normalized and compensated current signal; The band-pass processed current signal, the phase reference unified current signal, the odd harmonic energy feature vector, the transient pulse marker set and the normalized and compensated current signal are combined to form the preprocessed signal sequences with robust working conditions.

3. The method according to claim 2, wherein the method is characterized by, The double-channel input tensor is formed by combining the channel coupling features and the double-channel input signals. The channel A input signal is constructed to include the band-pass processed current signal, the phase reference unified current signal and the normalized compensated current signal, and is organized according to a unified sampling time stamp; The channel B input signal is constructed to include the odd harmonic energy feature vector and the transient pulse label set; The instantaneous phase difference between the phase reference unified current signal in the channel A input signal and the harmonic mapping signal in the channel B input signal is calculated, and the cosine value of the instantaneous phase difference is taken as the channel coupling feature at each time point; The double-channel input tensor is obtained by sequentially cascading the channel A input signal, the channel B input signal and the channel coupling feature sequence according to a unified time step.

4. The method according to claim 3, wherein, The double-channel auto-encoding network framework includes: The double-channel auto-encoding network framework is constructed to include the channel A specific encoder, the channel B specific encoder, the shared hidden variable pool and the collaborative decoder; The channel A specific encoder encodes and maps the channel A input signal to generate the latent variable representation of the channel A; The channel B specific encoder encodes and maps the channel B input signal to generate the latent variable representation of the channel B; The latent variable representation of the channel A and the latent variable representation of the channel B are spliced to obtain the shared hidden variable pool; The collaborative decoder takes the shared hidden variable pool as input and generates the reconstruction signal of the channel A and the reconstruction signal of the channel B, respectively; The constructed double-channel auto-encoding network framework is applied to the double-channel input tensor, and the double-channel input tensor is divided into a training data set and a validation data set in time sequence.

5. The method according to claim 1, wherein, The double-channel auto-encoding network is trained based on the healthy working condition samples corresponding to the training data set, including: The overall training loss function is obtained by weighting the channel A reconstruction error, the channel B reconstruction error and the cross-channel consistency loss; Based on the training data set and the overall training loss function value, the parameters of all modules in the double-channel auto-encoding network are updated by the back propagation algorithm until the overall training loss function value converges, and the model parameter set after training is obtained; The trained double-channel auto-encoding network is deployed to the online identification end, and the newly obtained double-channel input tensor is input to output the channel A reconstruction signal, the channel B reconstruction signal, the channel A latent variable representation and the channel B latent variable representation.

6. The method according to claim 5, wherein the method is characterized by, The double-channel auto-encoding network is trained based on the healthy working condition samples corresponding to the training data set, including: The channel A reconstruction residual is calculated to measure the mean square reconstruction error between the channel A input signal and the channel A reconstruction signal; The latent variable offset is calculated to measure the overall deviation between the latent variable distribution under the current working condition and the latent variable mean reference vector under the healthy working condition; The channel consistency measure value is calculated to measure the structural consistency between the channel A latent variable representation and the channel B latent variable representation; The channel A reconstruction residual, the channel B reconstruction residual, the latent variable offset and the channel consistency measure value are weighted and summed to obtain an abnormal score vector.

7. The method according to claim 6, wherein the method is characterized by, The abnormal score vector and the parameter prior set are input into the variational Bayesian optimization module, including: The parameter prior set includes a double-channel auto-encoding network structure parameter, a loss weight parameter, an environment compensation parameter, and an abnormal alarm threshold parameter. A prior distribution model is established for each parameter in the parameter prior set. The network structure parameter is modeled using a discrete uniform distribution, the loss weight parameter is modeled using a weight uniformity distribution, the environment compensation parameter is modeled using a multivariate normal distribution, and the abnormal alarm threshold parameter is modeled using an interval probability distribution. The abnormal score vector and the parameter prior set are input into the variational Bayesian optimization module to establish a variational inference problem with the maximum likelihood of the posterior distribution as the target. The parameter posterior set is obtained by iteratively optimizing the target using the variational Bayesian optimization module.

8. The method according to claim 7, wherein the method is characterized by, The output of the abnormal recognition result with a confidence interval and the triggering of the hierarchical alarm instruction include: According to the parameter posterior set, the optimal posterior distribution mean of the network structure parameter, the optimal posterior distribution mean of the loss weight parameter, the optimal posterior distribution mean of the environment compensation parameter, and the optimal posterior distribution mean of the abnormal alarm threshold parameter are extracted as the online optimization parameter set of the current round. The online optimization parameter set is used to perform threshold determination on the abnormal score vector, and the optimal posterior distribution mean of the abnormal alarm threshold parameter is extracted as the current alarm threshold. The confidence interval upper and lower limits of the alarm threshold are determined by combining the optimal posterior distribution variance of the abnormal alarm threshold parameter. According to the relationship between the abnormal score vector and the confidence interval of the alarm threshold, hierarchical alarm determination is performed: When the abnormal score vector is less than the confidence interval lower limit of the alarm threshold, a prompt-level alarm is output, indicating that the current working condition has no obvious abnormality, and routine inspection is recommended. When the abnormal score vector is greater than or equal to the confidence interval lower limit of the alarm threshold and less than the confidence interval upper limit of the alarm threshold, a pre-warning level alarm is output, indicating that the current working condition deviates but is not serious, and enhanced monitoring or on-site re-inspection is recommended. When the abnormal score vector is greater than or equal to the confidence interval upper limit of the alarm threshold, an alarm-level alarm is output, indicating that the current working condition is highly abnormal, and immediate infrared temperature measurement or power-off inspection and disposal measures are recommended.

Images