A method for monitoring power failures in intelligent mobile substations

Abstract

The present invention relates to the technical field of electric fault monitoring, and specifically to an intelligent mobile substation electric fault monitoring method. The method synchronously collects electrical monitoring signals, multi-dimensional environmental noise signals and dynamic operating parameters of the mobile substation through multi-source sensors; adopts multi-scale spectral modal decomposition to construct an environmental noise basis vector group, filters out environmental noise interference components from the electrical monitoring signal through orthogonal projection filtering, and outputs a baseline correction signal; constructs an operating condition disturbance response field matrix based on time domain and frequency domain correlation analysis of dynamic operating condition parameters, calculates gradient sensitivity coefficients, separates components strongly correlated with operating condition disturbances and residual components weakly correlated with equipment faults; reconstructs the residual components into pure fault feature vectors; and finally performs fault type diagnosis and risk warning based on the pure fault feature vectors. The method effectively solves the problem of fault feature annihilation caused by noise pollution in complex environments, and the problem of false alarms and missed alarms caused by confusion between operating condition disturbances and real fault signals.

Description

Technical Field

[0001] The present invention relates to the technical field of electric fault monitoring, and in particular to an electric fault monitoring method for an intelligent mobile substation. Background Art

[0002] Due to their high mobility, mobile substations are exposed to complex industrial environments and harsh outdoor conditions for long periods of time. They face multiple technical challenges. For example, dynamic and variable environmental interference causes non-stationary interference sources such as mechanical vibration and spatial electromagnetic noise to alias the equipment's electrical signal spectrum. Traditional fixed-bandwidth filters are unable to adapt to the non-stationary characteristics of the noise spectrum, resulting in significant interference components remaining in the filtered signal.

[0003] The strong coupling between operating disturbances and fault characteristics leads to sudden changes in on-load tap-changing states, instantaneous load fluctuations and other dynamic operating conditions, which can trigger synchronous distortion of electrical signals (such as harmonic increments and transient overvoltages). The frequency domain characteristics are highly similar to those of early insulation degradation, poor contact and other faults, making it difficult for existing methods to effectively distinguish them. In addition, the feature extraction dimension relies on time domain thresholds (such as current effective value exceeding the limit) or narrowband spectrum analysis, and cannot separate the operating disturbance component from the actual fault residual.

[0004] The above defects lead to frequent false alarms and missed detections in traditional monitoring systems in mobile substation scenarios, seriously restricting the precise control of equipment status and the timeliness of fault warnings. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent mobile substation electrical fault monitoring method to solve the problems raised in the above-mentioned background technology. The core problems to be solved include how to achieve high-precision separation of multi-source interference to solve the problem of fault signal annihilation caused by environmental noise and operating disturbances; and how to construct a decoupling model of dynamic operating conditions and electrical distortion to solve the problem that fault characteristics are difficult to separate and identify under fluctuating operating conditions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring electrical faults in an intelligent mobile substation, the method comprising the following steps:

[0007] S1. By deploying a multi-source sensor network at key equipment nodes and in the surrounding environment of the mobile station, an operating data set containing electrical monitoring signals, multi-dimensional environmental noise signals, and dynamic operating parameters is collected in real time and synchronously. The electrical monitoring signals specifically cover the voltage and current waveforms, harmonic components, and transient overvoltage signals of the main circuit and key components; the multi-dimensional environmental noise signals include the sound spectrum, mechanical vibration, and spatial electromagnetic noise radiation signals; and the dynamic operating parameters include the internal temperature field distribution of the equipment, on-load tap change status, cooling intensity, and instantaneous load rate parameters. This process involves deploying a multi-source sensor network at key equipment nodes and in the surrounding environment of the mobile station, collecting an operating data set containing specific electrical signals, multi-dimensional environmental noise, and dynamic operating parameters in real time and synchronously. This allows for comprehensive, synchronous data characterizing the equipment's operating status, environmental interference, and working conditions, laying the data foundation for subsequent noise stripping, operating condition decoupling, and fault diagnosis.

[0008] S2. Perform multi-scale spectral modal decomposition on the collected multi-dimensional environmental noise signal. This process uses a wavelet packet transform combined with an empirical mode decomposition algorithm to decompose the multi-dimensional environmental noise signal into sub-signals of different frequency bands through a wavelet packet tree structure. Then, empirical mode decomposition is performed on each frequency band sub-signal to extract the dominant intrinsic mode components whose energy proportion exceeds a preset threshold. The time domain waveform of the dominant intrinsic mode component is used as the basis function, and an independent environmental noise basis vector group is generated through orthogonalization processing. This process performs multi-scale spectral modal decomposition on the multi-dimensional environmental noise signal by using a wavelet packet transform combined with an empirical mode decomposition algorithm, so as to generate an independent environmental noise basis vector group, thereby accurately capturing and characterizing the core mode of the environmental noise in the time-frequency domain.

[0009] Based on the generated environmental noise basis vector group, a high-dimensional noise subspace is established. The electrical monitoring signal is orthogonally projected onto this high-dimensional noise subspace, and the signal's projection component on the noise basis is calculated. This noise projection component is then subtracted from the signal vector to remove interference components closely related to the environmental noise represented by the environmental noise basis vector group. After back-projection reconstruction, a baseline-corrected signal is output that retains the device's inherent electrical characteristics. This process establishes a high-dimensional noise subspace based on the environmental noise basis vector group and orthogonally projects the electrical monitoring signal onto this subspace (calculating the noise projection component, performing vector subtraction, and back-projection reconstruction). This removes interference components closely related to the environmental noise and outputs a baseline-corrected signal, effectively filtering out interference from complex, multi-dimensional, strong noise on the electrical signal. This significantly improves the signal-to-noise ratio and preserves the device's inherent electrical characteristics, resolving the problem of weak fault characteristics being contaminated or drowned out by noise.

[0010] The time domain and frequency domain correlation analysis of dynamic operating parameters is performed. The time-delayed cross-correlation values ​​between the fluctuations of each operating parameter and the electrical signal distortion are calculated to identify the key coupling time window between the two; at the same time, the coherence function of the operating parameter spectrum and the electrical distortion spectrum is calculated to quantify the frequency domain energy transfer relationship; based on this, the operating condition disturbance response field matrix is ​​constructed, in which the row vectors are the operating condition variables and the column vectors are the distortion characteristics. The matrix element values ​​accurately represent the coupling strength of the specific operating condition disturbance to the electrical signal distortion. This process identifies the key coupling time window by calculating the time-delayed cross-correlation values ​​between the fluctuations of each operating parameter and the electrical signal distortion, and calculates the coherence function between the operating parameter spectrum and the electrical distortion spectrum to quantify the frequency domain energy transfer relationship. This allows the construction of the operating condition disturbance response field matrix, thereby accurately quantifying the complex coupling relationship between dynamic operating condition disturbances and various electrical signal distortions in the time domain and frequency domain.

[0011] S3. Calculate the gradient sensitivity coefficient of the baseline correction signal along each dimension of the constructed operating condition disturbance response field matrix; perform partial derivative operations along each operating condition dimension in the operating condition disturbance response field matrix, fine-tune the target operating condition value while fixing other operating condition parameters, and simultaneously monitor the incremental changes in the key characteristic quantities of the electrical signal; calculate the partial derivative ratio between the rate of change of the characteristic quantity and the operating condition disturbance quantity, and generate a gradient sensitivity coefficient that characterizes the sensitivity of the electrical signal to specific operating condition changes. This process performs partial derivative operations along each operating condition dimension in the operating condition disturbance response field matrix, calculates the partial derivative ratio between the rate of change of the characteristic quantity and the operating condition disturbance quantity, and generates a gradient sensitivity coefficient, thereby quantitatively evaluating the sensitivity of the key characteristic quantities of the electrical signal to small changes in each operating condition parameter.

[0012] Signal components are separated based on the distribution characteristics of the gradient sensitivity coefficients. Coefficients with gradient sensitivity coefficient moduli exceeding the set sensitivity threshold are screened out, and the baseline-corrected signal segments within the corresponding operating condition fluctuation period are determined to be components strongly correlated with the operating condition disturbance. The remaining signal components with gradient sensitivity coefficient moduli below the set sensitivity threshold are determined to be residual components weakly correlated with equipment failures. By screening the gradient sensitivity coefficient moduli, this process successfully separates signal components strongly correlated with operating condition disturbances and pure residual components weakly correlated with equipment failures. This effectively decouples dynamic operating condition fluctuation signals from potential true fault signals, resolving the problem of fault characteristics being difficult to separate and identify under highly fluctuating operating conditions, leading to false alarms and missed diagnoses.

[0013] The isolated residual components, which are weakly correlated with the equipment fault, are deeply processed to generate a pure fault feature vector. First, a short-time Fourier transform is performed on the residual components to generate a time-frequency spectrum. A sparse autoencoder with sparsity constraints is constructed to perform deep feature extraction on the residual signal to compress redundant information. Simultaneously, envelope spectrum kurtosis analysis is performed on the residual signal or the features extracted by the sparse autoencoder. A Hanning window is used to calculate the kurtosis value within a specific frequency band to enhance the non-stationary impact characteristics. The frequency corresponding to the maximum kurtosis value is extracted as the potential resonant frequency band, and the autocorrelation function of the envelope signal is calculated to extract the impact period information. Finally, a low-dimensional vector containing time-domain statistics, frequency-domain energy entropy, and impact characteristic factors is constructed as a pure fault feature vector. The time-domain statistics include root mean square value, peak factor, kurtosis, etc.; the frequency-domain energy entropy includes wavelet packet energy entropy; and the impact characteristic factors include the extracted impact period and resonant frequency band kurtosis value. This process generates a time-frequency spectrum by performing short-time Fourier transform on the residual components weakly correlated with equipment faults, constructs a sparse autoencoder for deep feature extraction, and performs envelope spectrum kurtosis analysis. This reveals the time-frequency characteristics, compresses redundant information, enhances the impact characteristics, and integrates time domain statistics, frequency domain energy entropy, and impact characteristic factors to construct a highly streamlined, information-rich, and fault-type-sensitive low-dimensional pure fault feature vector.

[0014] S4. Input the generated pure fault feature vector into a pre-trained multi-class support vector machine classifier for fault type diagnosis. This classifier uses a radial basis kernel function and is trained based on a labeled historical fault feature sample dataset. The classifier outputs a probability distribution vector for each fault type for the pure fault feature vector. Based on this probability distribution vector, the fault type with the highest matching degree is determined as the final diagnosis result. This process inputs the pure fault feature vector into the pre-trained multi-class support vector machine classifier, outputting a probability distribution vector for each fault type for the feature vector and determining the fault type with the highest matching degree, thereby accurately identifying the specific fault type that has occurred in the equipment.

[0015] Risk warnings are implemented for the diagnosed fault type. Feature parameters strongly correlated with the fault type are selected from the pure fault feature vector. Based on the current values ​​and historical trends of these feature parameters, combined with the equipment failure mechanism model, a fault degradation index reflecting the severity of the current fault is calculated. A state transition model based on a hidden Markov model is established, using the current fault degradation index and real-time dynamic operating parameters as input to predict the evolution of the fault degradation index over a period of time. Multiple risk level thresholds corresponding to the predicted range of the fault degradation index are set, and warning signals of corresponding levels are triggered. This process predicts the evolution of the fault degradation index over a period of time by selecting strongly correlated parameters from the pure fault feature vector for the diagnosed fault type, calculating the fault degradation index based on their values ​​and historical trends combined with the equipment mechanism model, and establishing a state transition model based on the hidden Markov model. This predicts the evolution of the fault degradation index over a period of time, and triggers warning signals of corresponding levels based on preset risk level thresholds. This achieves accurate, graded risk warnings based on the current state and future evolution predictions, providing a key basis for proactive operation and maintenance decision-making.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] By combining multi-scale spectral modal decomposition with orthogonal projection filtering of the ambient noise floor, strong noise interference in complex environments is removed with high precision, improving the signal-to-noise ratio of weak fault characteristics. By constructing a disturbance response field matrix reflecting the dynamic coupling relationship between operating conditions and electrical distortion and analyzing the gradient sensitivity coefficient, the signal distortion components caused by severe operating condition fluctuations are effectively decoupled and separated, avoiding misjudgments and missed diagnoses.

[0018] Furthermore, through deep feature extraction and impact characteristic enhancement of pure residual signals, high-fidelity, low-dimensional feature vectors rich in fault-sensitive information are efficiently extracted, thereby improving the accuracy of fault type identification; finally, combined with intelligent classification and degradation prediction based on the hidden Markov model, accurate judgment of fault types and predictable assessment of future risks are achieved, providing key technical support for the safe operation and predictive maintenance of mobile substations in complex and harsh environments, and solving core operation and maintenance pain points such as false alarms and missed alarms caused by noise interference and confused working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the method steps of the present invention. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0021] See also Figure 1 The present invention provides a technical solution: a method for monitoring power failures in an intelligent mobile substation, comprising the following steps:

[0022] S1. Using multi-source sensors to collect real-time operating data sets of the mobile substation, the operating data sets include electrical monitoring signals, multi-dimensional environmental noise signals, and dynamic operating parameters;

[0023] The data collection process of the mobile substation is realized through a multi-source sensor network deployed at key nodes of the equipment (such as transformers, circuit breakers, busbars, bushings, etc.) and the surrounding environment. Various sensors collect the operating data set of the mobile substation in real time, including

[0024] Electrical monitoring signals: Use high-precision voltage transformers and current transformers to measure core electrical parameters such as voltage, current waveform, harmonic components, transient overvoltage, etc. of the main circuit and key components in real time, forming a continuous electrical signal time series as the electrical monitoring signal;

[0025] Multi-dimensional environmental noise signal: Acoustic sensors, vibration sensors, and electromagnetic interference sensors are deployed to continuously monitor the sound spectrum, mechanical vibration, and spatial electromagnetic noise radiation generated during substation operation. This forms a multi-dimensional time series signal representing the comprehensive environmental noise, which serves as the multi-dimensional environmental noise signal.

[0026] Dynamic operating parameters: Utilize temperature sensors, on-load tap-changer position sensors, cooling system status monitors, and load current monitoring units to record in real time dynamic parameters reflecting operating conditions, such as the internal temperature field distribution, on-load tap-changer status, cooling intensity, and instantaneous load rate, as dynamic operating parameters.

[0027] All collected mobile substation operational data sets are transmitted via high-speed industrial fieldbuses or wireless networks to an edge computing gateway or data acquisition unit. This unit performs anti-aliasing filtering, analog-to-digital conversion, and time-stamp synchronization. Ultimately, it integrates high-sampling-rate electrical waveforms, medium- and high-speed (kHz-level) acoustic and electromagnetic noise data, and low-speed (second-level) operating parameters into a structured, precisely time-stamped "mobile substation operational data set," providing the raw data foundation for subsequent noise filtering and fault analysis. The entire process ensures sensor accuracy, synchronization, anti-interference capabilities, and compliance with relevant safety standards.

[0028] S2. Perform multi-scale spectral modal decomposition on the multi-dimensional ambient noise signal, extract the basis functions of the dominant noise mode, and generate an ambient noise basis vector group, specifically including:

[0029] Multi-scale spectral modal decomposition of multidimensional environmental noise signals is performed using a combined algorithm of wavelet packet transform and empirical mode decomposition. First, the multidimensional environmental noise signal is decomposed into sub-signals of different frequency bands through the wavelet packet tree structure. Then, modal decomposition is performed on each sub-signal to extract the dominant intrinsic mode components whose energy proportion exceeds the threshold. The time domain waveform of the dominant intrinsic mode component is used as the basis function, and a group of independent environmental noise basis vectors is generated through Gram-Schmidt orthogonalization. This vector group can characterize the typical spectral structure characteristics of equipment operation noise.

[0030] For dynamic operating parameters such as on-load tap change gear, load current, and cooling intensity, a method combining time-domain cross-correlation analysis and frequency-domain transfer function calculation is used to construct an operating disturbance response field matrix that reflects the coupling relationship between operating fluctuations and signal distortion. Specifically, it includes:

[0031] First, the time-delay cross-correlation values ​​between the fluctuations of various operating parameters and the electrical signal distortion (such as harmonic distortion rate and phase offset) are calculated to identify the key coupling time window; then, the coherence function of the operating parameter spectrum and the electrical distortion spectrum is calculated through fast Fourier transform to quantify the frequency domain energy transfer relationship; finally, a matrix is ​​constructed with row vectors as operating condition variables and column vectors as distortion characteristics. The matrix element values ​​represent the coupling strength of specific operating condition disturbances on electrical signal distortion, forming an operating condition disturbance response field matrix.

[0032] The environmental noise base vector group is used to perform orthogonal projection filtering on the electrical monitoring signal to filter out the interference components related to the environmental noise and output a baseline correction signal. Specifically, it includes:

[0033] Based on the environmental noise base vector group, a high-dimensional noise subspace is established. The electrical monitoring signals (such as three-phase voltage and current waveforms) are projected onto this subspace, and the projection components of the signals on the noise base are calculated. The interference components related to the environmental noise are removed by subtracting the noise projection components from the signal vector. The baseline correction signal output after inverse projection reconstruction retains the electrical characteristics of the equipment itself, eliminates the contamination of the electrical waveform by environmental noise such as fan vibration and electromagnetic radiation, and enables subsequent fault analysis to focus on the actual electrical status.

[0034] S3. Input the baseline correction signal into the operating condition disturbance response field matrix and perform partial derivative operations along each operating condition dimension in the matrix (such as load factor, voltage regulation level, and cooling intensity). While keeping other operating condition parameters fixed, gradually fine-tune the target operating condition value (such as a ±1% step in load factor) while simultaneously monitoring the incremental changes in key characteristic quantities of the electrical signal (harmonic amplitude, phase offset, and waveform distortion rate). Calculate the ratio of the characteristic quantity change rate to the partial derivative of the operating condition disturbance quantity to generate a gradient sensitivity coefficient that characterizes the electrical signal's sensitivity to specific operating conditions. The magnitude of its modulus reflects the intensity of coupling between the signal and the operating condition disturbance. The gradual fine-tuning of the target operating condition value is performed in an offline simulation environment before the mobile substation is put into operation, so it will not directly impact the stability and security of the power grid.

[0035] Based on the distribution law of the gradient sensitivity coefficient, feature decoupling is performed. By screening the sensitivity coefficients whose modulus values ​​of the gradient sensitivity coefficients exceed the set sensitivity threshold (e.g., >0.85), the electrical signal segments within the corresponding working condition fluctuation period are defined as components that are strongly correlated with the working condition disturbance (e.g., periodic harmonic fluctuations caused by sudden load changes). The remaining signal components with gradient modulus values ​​below the sensitivity threshold (e.g., random pulses, asynchronous oscillations) that have no significant statistical correlation with the working condition parameters are determined to be residual components that are weakly correlated with the equipment fault, where:

[0036] Time-frequency domain fusion processing is performed on the residual components weakly correlated with equipment faults. A sparse autoencoder is used to extract deep features from the residual signal and compress redundant information. Envelope spectrum kurtosis analysis is used to enhance non-stationary shock characteristics and extract fault-sensitive indicators such as shock period and resonance frequency band. Finally, a low-dimensional vector containing time-domain statistics, frequency-domain energy entropy, and shock characteristic factor dimensions is constructed as a pure fault feature vector that eliminates working condition interference. The specific method process is as follows:

[0037] Perform short-time Fourier transform or wavelet transform on the residual components weakly correlated with equipment faults to generate a time-frequency spectrum; construct a sparse autoencoder with sparsity constraints to perform deep feature extraction on the time domain or time-frequency domain residual signal to compress redundant information; perform envelope spectrum kurtosis analysis on the features extracted by the autoencoder or the original residual signal (using a Hanning window to calculate the kurtosis value within a specific frequency band) to enhance the non-stationary impact characteristics, extract the frequency corresponding to the maximum kurtosis as the potential resonant frequency band, and calculate the autocorrelation function of the envelope signal to extract the impact period; finally, construct a low-dimensional vector containing time domain statistics (such as root mean square value, peak factor, kurtosis), frequency domain energy entropy (such as wavelet packet energy entropy), and impact characteristic factors (such as extracted impact period, resonant frequency band kurtosis value) as a pure fault feature vector.

[0038] S4. Input the pure fault feature vector into a pre-trained multi-class support vector machine classifier. This classifier uses a radial basis kernel function and is trained based on a sample data set of labeled historical fault features (covering typical faults such as winding deformation, partial discharge, multi-point grounding of the core, and abnormal bushing dielectric loss). It outputs a probability distribution vector for each type of fault for which the pure fault feature vector belongs. Based on the probability distribution vector, the fault type with the highest matching degree is determined as the diagnostic result (for example, "winding deformation: 92% probability, partial discharge: 5% probability").

[0039] For the diagnosed fault type, characteristic parameters that are strongly correlated with the fault type in the pure fault feature vector are selected; based on the selected characteristic parameter values ​​and their historical change trends, combined with the equipment mechanism model, a fault degradation index reflecting the severity of the current fault is calculated; the equipment mechanism model is dynamically selected according to the fault type. When the fault type is an electrical fault (such as partial discharge), the equivalent circuit equipment mechanism model is used, with the input being the pulse current amplitude / frequency, and the output being the insulation aging index; when the fault type is a thermal fault (such as winding overheating), the thermodynamic equipment mechanism model is used, with the input being the temperature gradient and load rate, and the output being the hotspot temperature prediction value; when the fault type is a mechanical fault (such as a loose core), the vibration transfer function equipment mechanism model is used, with the input being the vibration spectrum energy, and the output being the deformation displacement; the fault degradation index is the normalized weighted comprehensive evaluation value output by the above model, specifically including:

[0040] According to the type of fault diagnosed, three types of indicators are selected: time domain statistics, frequency domain energy entropy, and impact characteristic factor corresponding to the pure fault feature vector;

[0041] According to the formula Calculate the initial value of the degradation index, where is the normalized value of the time domain statistics (such as RMS value, crest factor); Represents the normalized value of frequency domain energy entropy (such as wavelet packet energy entropy); Indicates the normalized value of the impact characteristic factor (such as impact period, envelope spectrum kurtosis); 、 and The weights represent the normalized values ​​of the time-domain statistics, the frequency-domain energy entropy, and the impact characteristic factor, respectively. The weights are determined based on the principle of adapting the fault physical mechanism to the characteristic sensitivity. Specifically, different initial weight coefficients are assigned to the time-domain statistics, frequency-domain energy entropy, and impact characteristic factor according to the physical characteristics of the fault type (thermal fault / discharge fault / mechanical fault), ensuring that the sum of the weights is a fixed constant. Based on the fault type diagnosed in real time, the weights of feature categories that are strongly correlated with the fault are increased, while the weights of weakly correlated features are reduced year-on-year, ensuring that the fusion model always focuses on the dominant fault features. Using a historical fault sample library, the weight allocation strategy is iteratively optimized through feature validity verification algorithms (such as sensitivity-specificity analysis) to ensure that the weight setting maximizes fault identification accuracy.

[0042] Each indicator is based on the upper limit of the equipment safe operation threshold As a benchmark, press Calculate the normalized value where Indicates the actual measured value of the indicator. Set according to the device's technical specifications.

[0043] Gradient analysis is the core technology path to achieve high-precision decoupling of operating disturbance components and fault residuals. For example:

[0044] In strong noise environments, while orthogonal projection filtering removes ambient noise, signal distortion caused by operating condition fluctuations (such as a sudden increase in harmonics due to load changes) remains highly intertwined with true fault signatures (such as partial discharge pulses) in the time-frequency domain. The gradient sensitivity coefficient, based on the operating condition disturbance response field matrix, precisely quantifies the coupling strength between each signal segment and the operating condition disturbance through mathematical differentiation (calculating the partial derivative of the electrical characteristic with respect to the operating condition variable), thereby identifying the residual component as the device's own fault. This dynamic decoupling mechanism directly distinguishes between pseudo-fault signatures driven by external operating conditions and true fault signatures driven by internal device degradation, ultimately enabling the precise extraction of weak fault signals in strong noise environments and the proactive mitigation of operating condition interference.

[0045] A state transition model based on the hidden Markov model is established. The current fault degradation index and dynamic operating condition parameters are used as input to predict the evolution trajectory of the fault degradation index in the future. Multiple risk level thresholds corresponding to the predicted value range of the fault degradation index are set. When the predicted fault degradation index exceeds the set high-risk level threshold, a warning signal of the corresponding level is triggered, for example:

[0046] Level 1 warning: If the characteristic fluctuation exceeds the limit but does not endanger the operation, the operation and maintenance suggestions will be sent to the monitoring center;

[0047] Level 2 warning: Failure probability > 80% and degradation is accelerating, and the equipment load reduction protection interlock is activated;

[0048] Level 3 warning: When approaching the insulation failure threshold, the circuit breaker is emergency tripped and the positioning information is pushed to the emergency repair terminal.

[0049] The setting of risk level thresholds is based on a comprehensive quantitative process of equipment safety operation standards, statistical analysis of historical failure data, and assessment of economic losses caused by failure consequences. First, the theoretical safety limits of each failure type are determined based on the equipment manufacturer's technical specifications and industry standards. Second, the mapping relationship between the failure degradation index and the consequences of equipment failure in the historical failure case library is analyzed to establish a regression model of "degradation index-failure probability-downtime". Finally, combined with the economic loss assessment under different risk levels, the risk thresholds are weighted and revised. For example, the first-level warning threshold is set at 0.6, the second-level warning threshold is set at 0.8, and the third-level warning threshold is set at 0.95, thereby achieving operability and repeatability of threshold setting.

[0050] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring power failures in an intelligent mobile substation, characterized in that: The method steps are as follows: S1. Using multi-source sensors to collect real-time operating data sets of the mobile substation, the operating data sets include electrical monitoring signals, multi-dimensional environmental noise signals, and dynamic operating parameters; S2. Perform multi-scale spectral modal decomposition on the multidimensional environmental noise signal, extract the basis functions of the dominant noise mode, and generate an environmental noise basis vector group. Perform time-domain and frequency-domain correlation analysis on the dynamic operating parameters to construct an operating disturbance response field matrix that reflects the coupling relationship between operating fluctuations and signal distortion. Use the environmental noise basis vector group to perform orthogonal projection filtering on the electrical monitoring signal to filter out interference components related to the environmental noise and output a baseline correction signal. S3. Calculate the gradient sensitivity coefficient of the baseline correction signal in each dimension of the working condition disturbance response field matrix; According to the distribution characteristics of the gradient sensitivity coefficient, the components strongly related to the working condition disturbance and the residual components weakly related to the equipment failure are separated, and the residual components are reconstructed into a pure fault feature vector; where: The calculation process of the gradient sensitivity coefficient includes: Perform partial derivative operations along each operating dimension in the operating condition disturbance response field matrix, fine-tune the target operating condition value while keeping other operating condition parameters fixed, and simultaneously monitor the incremental changes of key characteristic quantities of the electrical signal. Calculate the ratio of the characteristic quantity change rate to the partial derivative of the operating condition disturbance quantity to generate a gradient sensitivity coefficient that characterizes the electrical signal's sensitivity to a specific operating condition. The process of separating the distribution characteristics of the gradient sensitivity coefficient into components specifically includes: The gradient sensitivity coefficients whose modulus exceeds the set sensitivity threshold are screened, and the baseline correction signal segments within the corresponding operating condition fluctuation period are defined as components with strong correlation to the operating condition disturbance; the signal components with the remaining gradient modulus below the set sensitivity threshold are determined as residual components with weak correlation to the equipment fault; The generation of pure fault feature vectors specifically includes: Performing short-time Fourier transform on the residual components weakly related to the equipment fault to generate a time-frequency spectrum; constructing a sparse autoencoder with sparsity constraints to perform deep feature extraction on the time-frequency domain residual signal to compress redundant information; The envelope spectrum kurtosis analysis is performed on the features extracted by the autoencoder or the original residual signal to enhance the non-stationary shock characteristics, the frequency corresponding to the maximum kurtosis is extracted as the potential resonance frequency band, and the autocorrelation function of the envelope signal is calculated to extract the shock period; Finally, a low-dimensional vector including time domain statistics, frequency domain energy entropy and impact characteristic factor dimensions is constructed as the fault feature vector; S4. Perform fault type diagnosis and risk warning based on the pattern recognition results of the pure fault feature vector.

2. The intelligent mobile substation power fault monitoring method according to claim 1, characterized in that: The electrical monitoring signals include the voltage and current waveforms, harmonic components and transient overvoltage signals of the main circuit and key components; the multi-dimensional environmental noise signals include sound spectrum, mechanical vibration and spatial electromagnetic noise radiation signals; the dynamic operating parameters include the internal temperature field distribution of the equipment, on-load tap-changing status, cooling intensity, and instantaneous load rate parameters.

3. The intelligent mobile substation power fault monitoring method according to claim 1, characterized in that: The process of multi-scale spectral modal decomposition specifically includes: A wavelet packet transform combined with empirical mode decomposition (EMD) is used to decompose the multidimensional environmental noise signal into sub-signals of different frequency bands through the wavelet packet tree structure. Empirical mode decomposition is then performed on each sub-signal to extract the dominant eigenmode components whose energy proportion exceeds a preset threshold. The time domain waveform of the dominant eigenmode component is used as the basis function, and an independent set of environmental noise basis vectors is generated through orthogonalization processing.

4. The intelligent mobile substation power fault monitoring method according to claim 1, characterized in that: The construction process of the working condition disturbance response field matrix specifically includes: The time-delayed cross-correlation values ​​between the fluctuations of each operating parameter and the electrical signal distortion are calculated to identify the key coupling time window; the coherence function between the operating parameter spectrum and the electrical distortion spectrum is calculated to quantify the frequency domain energy transfer relationship; and finally, an operating disturbance response field matrix is ​​constructed, in which the row vectors are the operating variables, the column vectors are the distortion characteristics, and the element values ​​represent the coupling strength of the specific operating disturbance on the electrical signal distortion.

5. The intelligent mobile substation power fault monitoring method according to claim 1, characterized in that: The generation process of the baseline correction signal specifically includes: A high-dimensional noise subspace is established based on the environmental noise base vector group; the electrical monitoring signal is projected onto the subspace, and the projection component of the signal on the noise base is calculated; the interference components related to the environmental noise are removed by subtracting the noise projection component from the signal vector; and after back-projection reconstruction, a baseline correction signal that retains the electrical characteristics of the device itself is output.

6. The intelligent mobile substation power fault monitoring method according to claim 1, characterized in that: In step S4, the process of performing fault type diagnosis based on the pattern recognition result of the pure fault feature vector specifically includes: The pure fault feature vector is input into a pre-trained multi-class support vector machine classifier; the classifier uses a radial basis kernel function and is trained based on a labeled historical fault feature sample data set to output a probability distribution vector of each type of fault for which the pure fault feature vector belongs; based on the probability distribution vector, the fault type with the highest matching degree is determined as the diagnosis result.

7. The intelligent mobile substation power fault monitoring method according to claim 1, characterized in that: The risk warning process specifically includes: For the diagnosed fault type, characteristic parameters that are strongly correlated with the fault type in the pure fault feature vector are selected; based on the selected characteristic parameter values ​​and their historical change trends, combined with the equipment mechanism model, the fault degradation index reflecting the severity of the current fault is calculated; a state transition model based on the hidden Markov model is established, and the current fault degradation index and dynamic operating condition parameters are used as input to predict the evolution trajectory of the fault degradation index in the future period; multiple risk level thresholds corresponding to the predicted value range of the fault degradation index are set, and warning signals of corresponding levels are triggered.

Images