A switch cabinet multi-parameter composite online monitoring method

Abstract

The application discloses a kind of switch cabinet multi-parameter composite online monitoring method, belong to switch cabinet state monitoring technical field.The method includes: synchronous acquisition switch cabinet bus three-phase current signal, obtains intrinsic mode function component by variational mode decomposition, obtains instantaneous frequency sequence and instantaneous amplitude sequence by hilbert-huang transform;Extract partial discharge characteristic parameter and mechanical vibration characteristic parameter, construct high-dimensional monitoring characteristic vector;Adaptive dimension reduction mapping is carried out to high-dimensional monitoring characteristic vector, generate low-dimensional state representation vector, input pre-trained degradation discriminant model, obtain switch cabinet insulation degradation grade and mechanical looseness index.The application can realize the synchronous evaluation of insulation defect and mechanical looseness state by using only non-invasive current signal, and provides multi-parameter quantitative basis for switch cabinet state maintenance.

Description

Technical Field

[0001] This invention relates to the field of switchgear condition monitoring technology, specifically to a multi-parameter composite online monitoring method for switchgear. Background Technology

[0002] Switchgear, as a critical power distribution device in power systems, is primarily susceptible to operational failures due to insulation degradation and mechanical loosening. Existing monitoring solutions typically rely on multiple discrete sensors, such as ultrasonic sensors and transient ground voltage sensors deployed inside the cabinet to capture partial discharge signals, along with vibration sensors to collect vibration parameters caused by mechanical loosening. The installation locations of these sensors are limited, and construction is complex, leading to increased hardware costs and maintenance workload. Furthermore, sensors based on different physical principles face obstacles in time synchronization and signal fusion, making it difficult to correlate insulation and mechanical conditions for analysis, resulting in fragmented evaluation results.

[0003] At the signal processing level, partial discharge monitoring often extracts high-frequency pulses from ultrasonic or transient ground voltage signals, while mechanical condition monitoring analyzes the vibration acceleration spectrum. When attempting to achieve composite sensing using only non-invasive bus current signals, existing technologies often employ Fourier transform or wavelet transform for feature separation. However, since partial discharge pulses and mechanical disturbances manifest as weak components with overlapping frequencies and aliasing in the current signal, conventional methods struggle to effectively decouple them. Under strong background noise, the stability and discriminative power of feature extraction are insufficient, leading to a high false alarm rate.

[0004] In the condition assessment phase, on-site alarms are typically triggered by a single feature threshold, or a binary judgment of normal or abnormal is given by inputting feature vectors into a classifier. These methods cannot provide a graded description of insulation degradation or a continuous quantitative indicator of mechanical loosening, making it difficult for maintenance personnel to formulate maintenance strategies based on precise degradation trends. High-dimensional feature vectors contain information redundancy and multicollinearity; directly using them for model evaluation reduces generalization performance. Furthermore, common dimensionality reduction methods are often designed independently of the evaluation model, lacking an adaptive processing mechanism for joint degradation assessment.

[0005] Therefore, it is necessary to solve the problem of reliably separating and simultaneously extracting partial discharge characteristic parameters and mechanical vibration characteristic parameters from the bus current signal of the switchgear, and how to achieve the coordinated quantification of insulation degradation level and mechanical loosening index based on high-dimensional multi-parameter characteristics, so as to overcome the shortcomings of existing monitoring methods, such as multi-sensor dependence, single evaluation results and inability to quantify. Summary of the Invention

[0006] This invention provides a multi-parameter composite online monitoring method for switchgear, which aims to simultaneously extract partial discharge and mechanical vibration characteristics through a single non-invasive current signal, and output insulation degradation level and mechanical loosening index, replacing the multi-sensor deployment scheme, and realizing online quantitative assessment of the dual state of insulation and mechanical properties of switchgear.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a multi-parameter composite online monitoring method for switchgear. This method synchronously acquires the three-phase current signals of the switchgear busbar to obtain a first current timing signal, a second current timing signal, and a third current timing signal. The three-phase current signals are then subjected to variational mode decomposition (VMD) to generate multiple intrinsic mode function (IMF) components. Hilbert-Huang transform is then applied to these IMF components to obtain instantaneous frequency and amplitude sequences. Utilizing the time-frequency analysis capabilities of VMD and Hilbert-Huang transform, the method effectively adapts to the non-stationary characteristics of current signals, accurately captures weak transient characteristics caused by partial discharge and mechanical loosening, and separates modal components reflecting different physical processes.

[0008] Partial discharge characteristic parameters and mechanical vibration characteristic parameters are extracted from the instantaneous frequency sequence and instantaneous amplitude sequence to construct a high-dimensional monitoring feature vector. As a technical solution of this invention, pulse amplitude peak value, pulse repetition rate, and phase distribution characteristics are extracted from the instantaneous amplitude sequence as partial discharge characteristic parameters; fundamental frequency offset, harmonic energy ratio, and frequency fluctuation intensity are extracted from the instantaneous frequency sequence as mechanical vibration characteristic parameters; the partial discharge characteristic parameters and mechanical vibration characteristic parameters are concatenated to obtain a high-dimensional monitoring feature vector, thereby integrating multi-dimensional information from electrical insulation and mechanical fastening to improve the sensitivity of composite fault correlation identification.

[0009] An adaptive dimensionality reduction mapping process is performed on the high-dimensional monitoring feature vector to generate a low-dimensional state representation vector. Preferably, the covariance matrix between each dimension of the high-dimensional monitoring feature vector is calculated, the eigenvalues ​​and eigenvectors are solved, the number of principal components is selected based on the cumulative contribution rate of the eigenvalues, the high-dimensional monitoring feature vector is projected onto the subspace spanned by the corresponding eigenvectors, and the coefficient vector obtained after projection is used as the low-dimensional state representation vector. This process retains the main information about state changes while eliminating feature redundancy and reducing the computational complexity of subsequent discrimination models.

[0010] The low-dimensional state representation vector is input into a pre-trained degradation discrimination model for state assessment, obtaining the insulation degradation level and mechanical loosening index of the switchgear. This degradation discrimination model simultaneously outputs discrete label values ​​corresponding to the insulation degradation level and continuous values ​​corresponding to the mechanical loosening index, enabling a single model to classify the severity of insulation defects and quantitatively assess the degree of mechanical loosening, effectively improving the real-time performance and comprehensive judgment efficiency of online monitoring.

[0011] As a further improvement of the present invention, the process of performing variational mode decomposition (VMD) on the three-phase current signal to generate multiple intrinsic mode function (IMF) components includes: setting the number of VMD decomposition levels to a preset value, and initializing the center frequency and bandwidth constraint parameters of each IMF; during the iterative solution process, dividing the frequency domain of each phase current signal in the three-phase current signal, updating the frequency domain representation of each IMF through an alternating direction multiplier algorithm until the convergence condition is met; and performing inverse Fourier transform on each frequency domain representation obtained after meeting the convergence condition to obtain the time domain waveform corresponding to each IMF component. This process can adaptively divide the signal frequency band, suppress mode aliasing, and ensure that the decomposition results have good physical interpretability.

[0012] When performing Hilbert-Huang transform on multiple intrinsic mode function components, an analytic signal is constructed by performing Hilbert transform on each intrinsic mode function component. The instantaneous amplitude envelope and instantaneous phase are extracted from the analytic signal. The instantaneous phase is differentiated over time to obtain the instantaneous frequency sequence. The instantaneous amplitude sequence and the instantaneous frequency sequence are aligned according to the time index to generate the Hilbert time spectrum, which provides high-resolution time-frequency joint distribution information for feature extraction.

[0013] Preferably, when the degradation discrimination model processes the low-dimensional state representation vector, it maps the low-dimensional state representation vector to a first latent feature through a first fully connected layer and a ReLU nonlinear activation function. The attention mechanism layer performs global average pooling, dimensionality reduction, and dimensionality increase on the first latent feature, generates weight coefficients for each feature dimension through a Sigmoid activation function, and multiplies them element-wise with the first latent feature to obtain a weighted latent feature. This automatically focuses on key features that are highly correlated with the degree of degradation and suppresses the influence of irrelevant disturbances. The weighted latent feature is input into a second fully connected layer, which outputs a continuous value of the mechanical loosening index through a linear mapping of the regression branch. At the same time, it outputs the probability of each category through a classification branch and a Softmax activation function, and takes the category label corresponding to the maximum probability as the insulation degradation level to achieve multi-task joint judgment.

[0014] As another preferred embodiment of the present invention, the method further includes: simultaneously acquiring ambient temperature and humidity data inside the switchgear during the acquisition of the first current timing signal, the second current timing signal, and the third current timing signal; using the ambient temperature and humidity data as auxiliary input parameters, concatenating them with the low-dimensional state representation vector to form an augmented state representation vector; inputting the augmented state representation vector into a pre-trained degradation discrimination model for state assessment to obtain the corrected switchgear insulation degradation level and mechanical loosening index. Introducing ambient temperature and humidity information can compensate for the influence of temperature and humidity on partial discharge characteristics and current characteristics, improving the accuracy and robustness of state assessment under complex environmental conditions.

[0015] Using the method of this invention, there is no need to install additional partial discharge or vibration sensors inside the switchgear. Based solely on the three-phase bus current signal and environmental parameters, composite features are extracted through variational mode decomposition and Hilbert-Huang transform. Combined with adaptive dimensionality reduction and attention mechanism degradation discrimination model, online monitoring and quantitative assessment of insulation degradation and mechanical loosening status can be completed simultaneously. This improves the sensitivity, reliability, and engineering practicality of monitoring, helps to detect early fault hazards in a timely manner, and reduces the operational risks of switchgear.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0017] Variational mode decomposition (VMD) is employed to process the three-phase current signal of the switchgear busbar. This adaptively divides the current waveform into multiple intrinsic mode function (EMF) components. Compared to fixed basis function transformation, this method can perform frequency band segmentation based on the signal's inherent fluctuation characteristics, effectively separating partial discharge pulse components and mechanical modulation components with similar frequencies. Subsequently, Hilbert-Huang transform is applied to each EMF component to extract instantaneous amplitude and frequency sequences. Partial discharge characteristic parameters such as pulse amplitude peak value, pulse repetition rate, and phase distribution are obtained from the instantaneous amplitude envelope. Mechanical vibration characteristic parameters such as fundamental frequency offset, harmonic energy ratio, and frequency fluctuation intensity are analyzed from the instantaneous frequency sequence. This process utilizes only non-intrusive busbar current signals, avoiding the need for additional partial discharge and vibration sensors inside the switchgear. It eliminates synchronization errors and mutual interference from multiple sensors, simplifying the system structure while obtaining high-dimensional monitoring feature vectors responding to both local insulation defects and mechanical loosening. Variational mode decomposition optimizes the mode separation capability under frequency domain aliasing conditions. Hilbert-Huang transform directly provides time-varying frequency response information in a single-component demodulation manner, making the composite fault characteristics more clearly identifiable in the amplitude and frequency dimensions, and exhibiting good robustness to background noise and load fluctuations.

[0018] An adaptive dimensionality reduction mapping based on covariance matrix eigenvalue decomposition is performed on the high-dimensional monitoring feature vector. Principal components are selected according to the cumulative contribution rate of eigenvalues, and the high-dimensional vector is projected onto the low-dimensional subspace to generate a low-dimensional state representation vector. This process removes feature redundancy and collinearity, while retaining the main variation directions sensitive to degradation states. The low-dimensional state representation vector is input into a pre-trained degradation discrimination model. This model generates latent features through a first fully connected layer, and then introduces an attention mechanism layer to perform global average pooling, dimensionality reduction, and dimensionality increase mapping on the latent features. After Sigmoid activation, dimensionality weight coefficients are obtained, and the latent features are then weighted element-wise to strengthen the feature components that play a key role in degradation discrimination and suppress secondary interference information. The weighted latent features are fed into the regression and classification branches of the second fully connected layer, outputting continuous values ​​of the mechanical loosening index and discrete labels of the insulation degradation level, respectively. This processing incorporates feature dimensionality reduction and attention weighting into a unified evaluation framework, enabling the model to adaptively focus on high-contribution dimensions. The quantitative loosening index provides a continuous reference for predicting the attenuation of mechanical fastening torque, while the qualitative degradation level provides staged alarms for insulation deterioration. The binary output jointly characterizes the combined state of the switchgear, reducing the ambiguity of single threshold alarms and providing multi-dimensional and quantitative decision information for differentiated maintenance. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0020] Figure 1 This is a flowchart of a multi-parameter composite online monitoring method for switchgear;

[0021] Figure 2 This is a flowchart of Hilbert's time spectrum generation process;

[0022] Figure 3 This is a flowchart of the dimensionality reduction process for high-dimensional monitoring feature vectors;

[0023] Figure 4 This is a schematic diagram of the instantaneous amplitude envelope and pulse amplitude peak value of the three-phase current;

[0024] Figure 5 This is a scatter plot of principal component analysis of the insulation degradation level of the switchgear;

[0025] Figure 6 This is a curve comparing the online monitoring results of the insulation degradation level and mechanical loosening index of the switchgear. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] See Figure 1 This invention provides a multi-parameter composite online monitoring method for switchgear. The method synchronously acquires three-phase current signals from the switchgear busbar to obtain a first current time-series signal, a second current time-series signal, and a third current time-series signal. The three-phase current signals are then subjected to variational mode decomposition (VMD) to generate multiple intrinsic mode function (IMF) components. Hilbert-Huang transform is then applied to these IMF components to obtain instantaneous frequency and amplitude sequences. Partial discharge and mechanical vibration characteristic parameters are extracted from the instantaneous frequency and amplitude sequences to construct a high-dimensional monitoring feature vector. The high-dimensional monitoring feature vector is then subjected to adaptive dimensionality reduction mapping to generate a low-dimensional state representation vector. This low-dimensional state representation vector is input into a pre-trained degradation discrimination model for state evaluation to obtain the switchgear insulation degradation level and mechanical loosening index.

[0028] Example 1:

[0029] In specific implementation, please refer to Figure 2The number of mode decomposition levels for variational mode decomposition is set to a preset value. Based on the sampling frequency of the switchgear bus current signal and the distribution range of the high-frequency modulation components caused by partial discharge and mechanical vibration in the current signal, the number of mode decomposition levels is set to 5. The center frequencies of each intrinsic mode function are initialized, and five initial center frequency values ​​are selected at equal intervals between 0 and π radians (corresponding to the Nyquist frequency); the bandwidth constraint parameter is initialized to 2000. The bandwidth constraint parameter is used to adjust the weight of the squared penalty term of the bandwidth of each intrinsic mode function in the optimization objective of variational mode decomposition. The bandwidth constraint parameter is set to 2000 to balance the ability to preserve signal detail components with the ability to suppress aliasing of adjacent mode spectra. During the iterative solution process, the frequency domain of each phase current signal in the first, second, and third current time series signals is divided, and the frequency domain representation of each intrinsic mode function is updated by the alternating direction multiplier algorithm. Each iteration of the alternating direction multiplier algorithm includes: fixing the frequency domain representations of all eigenmode functions except the current eigenmode function and the Lagrange multipliers; updating the frequency domain representation of the current eigenmode function with the objective of minimizing the augmented Lagrange function of the variational mode decomposition; re-estimating the center frequency of the current eigenmode function based on the updated frequency domain representation; and updating the Lagrange multipliers. This update process is repeated until the convergence condition is met. The convergence condition is set to a relative mean square error of less than 10⁻⁻⁶ between two consecutive iterations of the frequency domain representations of all eigenmode functions. 6 Alternatively, the iteration count can reach 500. After the convergence condition is met, the frequency domain representations of the final intrinsic mode functions are subjected to inverse Fourier transform to obtain the time domain waveforms corresponding to each intrinsic mode function component. The inverse Fourier transform converts the frequency domain representations of the intrinsic mode functions from the angular frequency domain to the time domain, recovering the time domain signals of each intrinsic mode function component.

[0030] A Hilbert transform is performed on each intrinsic mode function (EMF) component to construct an analytic signal. The Hilbert transform is implemented using a linear time-invariant system with an impulse response of 1 / (πt). The EMF components are used as input signals, and the output signal is the imaginary part orthogonal to the EMF components. The EMF components are used as the real parts, and the imaginary part output by the Hilbert transform is used as the imaginary part to construct the analytic signal. The instantaneous amplitude envelope and instantaneous phase are extracted from the analytic signal. The instantaneous amplitude envelope is given by the magnitude of the analytic signal, and the instantaneous phase is calculated using the arctangent function of the ratio of the imaginary to the real parts of the analytic signal. The arctangent calculation result is then subjected to phase dewinding. By detecting phase jumps between adjacent sampling points and accumulating ±2π compensation, a continuously changing instantaneous phase sequence is obtained. The instantaneous phase is differentiated over time to obtain the instantaneous frequency sequence, which is calculated using the following formula:

[0031]

[0032] in, Indicates at discrete sampling time points The instantaneous frequency value at that point, A function representing the continuous change of the instantaneous phase sequence after phase unwinding over time. Indicates the first Each sampling time point For time index number, Indicates the instantaneous phase function at The first derivative value at time step. In the discrete implementation, the first derivative of the instantaneous phase function is calculated using the central difference approximation, with the central difference spanning two sampling intervals. The instantaneous amplitude sequence and the instantaneous frequency sequence are assigned the same time index. Alignment is performed to generate a Hilbert time spectrum, which consists of a time index sequence, an instantaneous frequency sequence, and an instantaneous amplitude sequence, represented in the form of a two-dimensional matrix. The rows of the matrix correspond to the discrete frequency points of the instantaneous frequency sequence, the columns of the matrix correspond to the sampling points of the time index sequence, and the matrix element values ​​are the instantaneous amplitudes at the corresponding time and frequency.

[0033] Example 2:

[0034] In practice, the method for extracting pulse amplitude peaks from the instantaneous amplitude sequence is as follows: A pulse detection threshold is set, which is the root mean square value of the instantaneous amplitude sequence multiplied by a coefficient of 3. The coefficient 3 is chosen to ensure that the pulse detection threshold is higher than the background noise level of the instantaneous amplitude sequence while being lower than the minimum pulse amplitude caused by partial discharge. Each sampling point in the instantaneous amplitude sequence is iterated. When the instantaneous amplitude at a certain sampling point exceeds the pulse detection threshold, a pulse event is recorded. Centered on this sampling point, the maximum value of the instantaneous amplitude sequence is searched within a time window with a width of 100 sampling points. The found maximum value is taken as the pulse amplitude peak of that pulse event. The arithmetic mean of the pulse amplitude peaks of all detected pulse events is calculated, and this arithmetic mean is used as the pulse amplitude peak characteristic parameter.

[0035] The pulse repetition rate is extracted as follows: The total number of pulse events detected within a preset monitoring duration is counted, and the total number of pulse events is divided by the preset monitoring duration to obtain the pulse repetition rate characteristic parameter. The preset monitoring duration is the time length corresponding to 10 power frequency cycles.

[0036] The method for extracting phase distribution features is as follows: Zero-crossing detection is performed on the first current time-series signal to obtain the power frequency reference phase sequence. Each sampling point of the power frequency reference phase sequence corresponds to a phase value that varies from 0° to 360°. For each pulse event, the power frequency reference phase value corresponding to the peak sampling point of the pulse event is determined. The power frequency reference phase values ​​of all pulse events are divided into 12 phase intervals, each interval being 30°. The proportion of pulse events falling into each phase interval is counted to form a 12-dimensional phase distribution feature vector, which serves as the phase distribution feature parameter.

[0037] The method for extracting the fundamental frequency offset from the instantaneous frequency sequence is as follows: calculate the average value of the instantaneous frequency sequence within a preset monitoring period, subtract the nominal power frequency of 50Hz from the average value of the instantaneous frequency sequence, and use the difference as the fundamental frequency offset characteristic parameter.

[0038] The method for extracting the harmonic energy proportion is as follows: Subtract the average value of the instantaneous frequency sequence from the instantaneous frequency sequence, then perform a Discrete Fourier Transform to obtain the amplitude spectrum of the instantaneous frequency sequence. The harmonic energy proportion is calculated using the following formula:

[0039]

[0040] in, The parameter representing the proportion of harmonic energy is... Represents the first digit after the discrete Fourier transform. Discrete angular frequency points, Represents the discrete angular frequency points Complex spectrum value at that location, The amplitude spectrum value, and These represent the lower and upper bounds of the discrete angular frequency point index corresponding to the fundamental frequency component of the power frequency, respectively. The fundamental frequency component of the power frequency is 50Hz, and the lower bound of the index is... The corresponding frequency is 49.5Hz, and the upper bound of the index is... The corresponding frequency is 50.5Hz. This represents the total number of effective discrete angular frequency points in the amplitude spectrum after the discrete Fourier transform. Indicates the first Discrete angular frequency points, For amplitude spectrum in The value at the numerator. The energy of the fundamental frequency band, denominator term The total energy of the instantaneous frequency sequence.

[0041] The method for extracting frequency fluctuation intensity is as follows: Calculate the standard deviation of the instantaneous frequency sequence and use it as a characteristic parameter of frequency fluctuation intensity. The obtained pulse amplitude peak value characteristic parameter, pulse repetition rate characteristic parameter, 12-dimensional phase distribution characteristic vector, fundamental frequency offset characteristic parameter, harmonic energy proportion characteristic parameter, and frequency fluctuation intensity characteristic parameter are then concatenated in a preset order: pulse amplitude peak value is in the first position of the vector, pulse repetition rate is in the second position, the 12-dimensional phase distribution characteristic vector is in the third to fourteenth positions, the fundamental frequency offset is in the fifteenth position, the harmonic energy proportion is in the sixteenth position, and the frequency fluctuation intensity is in the seventeenth position. After concatenation, a 17-dimensional high-dimensional monitoring characteristic vector is obtained.

[0042] See Figure 4 In the figure, the horizontal axis represents time in milliseconds (ms), and the vertical axis represents instantaneous amplitude in volts (V). The figure shows the instantaneous amplitude envelope curves extracted from the three-phase current signals using Hilbert transform in Example 2, along with the corresponding pulse peak distribution. The blue solid line, orange dashed line, and green dotted line represent the instantaneous amplitude envelopes of the first, second, and third currents, respectively. The three curves generally exhibit low-amplitude background fluctuations, with amplitudes roughly concentrated in the range of 0.4V to 0.7V, reflecting a normal background noise level.

[0043] The red dotted line in the figure represents the set pulse detection threshold, which is approximately 1.5V, significantly higher than the background level of the instantaneous amplitude of the three-phase current. This meets the requirement in Example 2 that the threshold is set to the root mean square value of the instantaneous amplitude sequence multiplied by 3. The dots marked with different symbols in the figure represent pulse peak events exceeding the threshold in the instantaneous amplitude sequence of the three-phase current. The blue circles, orange triangles, and green squares correspond to the pulse peaks of the first, second, and third currents, respectively. The pulse peaks are scattered, with peak amplitudes generally exceeding the threshold of 1.5V, mostly concentrated in the 2.5V to 3.5V range, and some peaks reaching 4.5V to 5.1V, indicating that partial discharge pulse events have a significant response in each phase current.

[0044] From a timeline perspective, the pulse events occur intermittently. The number and amplitude of peaks fluctuate relatively steadily throughout the 200ms monitoring period, without a significant upward or downward trend, reflecting the repetitive and transient characteristics of partial discharge activity. Each pulse event, within a width of approximately 100 sampling points (corresponding to the time period), exhibits a clear peak in its instantaneous amplitude envelope curve, facilitating accurate extraction and statistical analysis of subsequent pulse amplitude peaks.

[0045] Example 3:

[0046] In specific implementation, please refer to Figure 3To obtain high-dimensional monitoring feature vectors, multiple sets of historical high-dimensional monitoring feature vectors are generated based on current signals from multiple historical samplings, forming a sample dataset. The sample dataset contains N samples, each a 17-dimensional high-dimensional monitoring feature vector. The covariance matrix between the dimensions of the high-dimensional monitoring feature vectors is calculated as follows: the sample dataset is organized into an N x 17 matrix X. The element in the a-th row and b-th column of matrix X represents the value of the b-th feature dimension of the a-th sample, where a ranges from 1 to N, and b ranges from 1 to 17. The mean of each column of matrix X is calculated to obtain a 17-dimensional mean vector μ. The j-th component of the mean vector μ... Let $\mathbf{j}$ be the arithmetic mean of all elements in the $j$-th column of matrix $X$, where $j$ ranges from 1 to 17. To remove the mean from matrix $X$, subtract the average of the corresponding column from each element in each row of matrix $X$. This yields a mean-reduced matrix. The covariance matrix C is calculated using the following formula:

[0047]

[0048] Where C represents the 17-row, 17-column covariance matrix, N is the total number of samples in the sample dataset, and N-1 is the mean-reduced matrix. The unbiased estimate of the sample covariance denominator, Represents the mean matrix The transpose of the matrix. The element in the u-th row and v-th column of the covariance matrix C. Let represent the covariance between the u-th and v-th dimensions of the high-dimensional monitoring feature vector, where u and v both range from 1 to 17.

[0049] The eigenvalues ​​and eigenvectors of the covariance matrix are obtained by performing eigenvalue decomposition on the covariance matrix C using the Jacobi iteration method. The Jacobi iteration method progressively diagonalizes the covariance matrix C through a series of planar rotation transformations. Each planar rotation transformation selects the position with the largest absolute value of the off-diagonal elements in the covariance matrix C, and constructs a rotation matrix to eliminate the off-diagonal elements at that position. This iterative process continues until the absolute values ​​of all off-diagonal elements are less than [a certain value]. This yields a set of eigenvalues ​​and their corresponding eigenvectors. All 17 eigenvalues ​​obtained are then sorted in descending order of numerical value, denoted as [e.g., eigenvalues ​​... ,in The largest eigenvalue, The smallest eigenvalue is found. The eigenvectors corresponding to each eigenvalue are rearranged in descending order of eigenvalue to obtain the eigenvector sequence. ,in For eigenvalues The corresponding 17-dimensional column vector, and all eigenvectors are unit vectors, satisfy the following condition: The value of k ranges from 1 to 17.

[0050] The method for selecting the number of principal components based on the cumulative contribution rate of eigenvalues ​​is as follows: A cumulative contribution rate threshold of 0.98 is set. The rationale for setting this threshold to 0.98 is to ensure that the reduced-dimensional state representation vector retains more than 98% of the variance information of the original high-dimensional monitoring feature vector. The cumulative contribution rate of the first p eigenvalues ​​is then calculated. :

[0051]

[0052] in, This represents the cumulative contribution rate of the first p eigenvalues, where p is the number of principal components to be selected, and p is an integer ranging from 1 to 17. The numerator term is the k-th eigenvalue after descending order. The term represents the sum of the first p eigenvalues, and the denominator term is... This represents the sum of all 17 eigenvalues. The cumulative contribution rate is calculated incrementing from p=1. This will be the first time that the requirement is met. The p-value is determined as the number of principal components to be selected. The eigenvectors corresponding to the selected p eigenvalues ​​are then... The columns are combined to form a 17-row, p-column projection matrix W, and the k-th column vector of the projection matrix W is... The value of k ranges from 1 to p.

[0053] The method for projecting high-dimensional monitoring feature vectors onto the subspace spanned by the corresponding feature vectors is as follows: For any 17-dimensional high-dimensional monitoring feature vector y to be reduced in dimensionality, first subtract the mean vector μ from the high-dimensional monitoring feature vector y to perform mean-reduction processing, resulting in the mean-reduced vector. The vector after removing the mean. Multiply by the projection matrix W to calculate the projection coefficient vector. The coefficient vector z obtained after projection is a p-dimensional column vector, and the k-th component of the coefficient vector z is... The value of k ranges from 1 to p. The coefficient vector z obtained after projection is used as the low-dimensional state representation vector, and the dimension of the low-dimensional state representation vector is equal to the number of principal components p selected.

[0054] See Figure 5This figure shows the scatter plot of the low-dimensional state representation vector in the two-dimensional principal component space after adaptive dimensionality reduction mapping based on the high-dimensional monitoring feature vector in Example 3. The horizontal axis represents the first principal component value, and the vertical axis represents the second principal component value. Both are projection space coordinates constructed by calculating the covariance matrix of the high-dimensional monitoring feature vector sample set and solving for the eigenvalues ​​and eigenvectors using the Jacobian iteration method, selecting the first two principal components. In the figure, the scatter points of different colors and shapes represent the sample distribution of the switchgear under different insulation degradation levels: green dots represent normal state samples, yellow squares represent slightly degraded samples, orange triangles represent moderately degraded samples, and red crosses represent severely degraded samples.

[0055] As shown in the figure, normal state samples are mainly concentrated in the range of approximately 0 to 5 for the first principal component and approximately 0 to 5 for the second principal component, forming relatively dense and clearly distinguishable clustered areas; mildly degraded samples are mainly distributed in the range of approximately -2 to 2 for the first principal component and approximately -2 to 2 for the second principal component, showing a clear clustering trend but overlapping with normal state samples; moderately degraded samples are mainly clustered in the range of approximately -4 to 0 for the first principal component and approximately 0 to 3 for the second principal component, located below normal and mildly degraded samples, showing relatively independent distribution characteristics; severely degraded samples are mainly distributed in the range of approximately -5 to -1 for the first principal component and approximately -4 to 0 for the second principal component, forming a clear cluster in the lower left of the scatter plot, which is well distinguishable from other state samples.

[0056] Example 4:

[0057] In its implementation, the degradation discrimination model comprises a first fully connected layer, a ReLU nonlinear activation function, an attention mechanism layer, and a second fully connected layer, connected sequentially. The second fully connected layer contains parallel regression and classification branches. The pre-training process of the degradation discrimination model is as follows: Multiple sets of historical low-dimensional state representation vectors of switchgear under known operating conditions are collected using the method described in Example 3, and the insulation degradation level label and mechanical loosening index label corresponding to each set of historical low-dimensional state representation vectors are recorded. The insulation degradation level is divided into four categories: normal, mild degradation, moderate degradation, and severe degradation. The discrete label value for normal is 0, for mild degradation it is 1, for moderate degradation it is 2, and for severe degradation it is 3. The mechanical loosening index label is a continuous value normalized to 0 to 1, where 0 indicates no mechanical loosening and 1 indicates complete loosening.

[0058] The low-dimensional state representation vector is input into the first fully connected layer of the degradation discrimination model. The dimension of the low-dimensional state representation vector is denoted as p, which is determined by adaptive dimensionality reduction mapping. The weight matrix W1 of the first fully connected layer has a dimension of 128 rows and p columns, and the bias vector b1 has a dimension of 128. The low-dimensional state representation vector is multiplied by the weight matrix of the first fully connected layer, and the bias vector is added to obtain the linear transformation result. :

[0059]

[0060] in, This represents a low-dimensional state representation vector of the input, which is a p-dimensional column vector. This represents the weight matrix of the first fully connected layer. Each element in the matrix has a mean of 0 and a standard deviation of 0. The truncated normal distribution is randomly initialized. This represents the bias vector of the first fully connected layer, with each element in the bias vector initialized to 0. The linear transformation result is the output of the first fully connected layer, a 128-dimensional column vector. This linear transformation result is then input into a ReLU activation function for non-linear mapping to obtain the first hidden feature. : . This means taking the larger of two values ​​for each element.

[0061] The first latent feature is input into the attention mechanism layer of the degradation discrimination model. In the attention mechanism layer, global average pooling is performed on the first latent feature. The first latent feature is a 128-dimensional vector, treated as a feature representation with 128 channels, each channel having a spatial size of 1. Global average pooling is performed on each channel, i.e., taking the numerical value of each channel itself, to obtain the first global feature vector with a dimension of 128. The first global feature vector undergoes a first dimensionality reduction process, implemented through a dimensionality reduction fully connected layer. The weight matrix of the dimensionality reduction fully connected layer... The dimension is r rows and 128 columns, and the bias vector is... The dimension is r, where r is the dimensionality reduction and compression coefficient, and r takes a value of 16. The first global feature vector is then combined with the weight matrix. Multiply, plus the bias vector The first dimensionality-reduced feature, with a dimension of 16, is obtained by applying the ReLU activation function. This first dimensionality-reduced feature is then subjected to a second dimensionality-up process, implemented through a fully connected layer. The weight matrix of this fully connected layer... The dimension is 128 rows and 16 columns, and the bias vector is... The dimension is 128. The first dimensionality-reduced feature is then combined with the weight matrix. Multiply, plus the bias vector This yields a second global feature vector with 128 dimensions. The second global feature vector is then mapped to the 0-1 interval using a Sigmoid activation function, generating weight coefficient vectors for each feature dimension. The dimension is 128, and the weight coefficient vector is... The Each component The value range of is (0,1). The value range is from 1 to 128. The weight coefficient vector... With the first implicit feature Element-wise multiplication yields weighted latent features. : . This indicates element-wise multiplication.

[0062] The weighted latent features are input into the second fully connected layer of the degradation discrimination model. The regression branch of the second fully connected layer consists of a linear mapping layer, and the weight matrix of the regression branch... The dimension is 1 row and 128 columns, and the bias vector is... The dimension is 1-dimensional. The weighted latent features are combined with the weight matrix. Multiply, plus the bias vector The system outputs a continuous numerical value corresponding to the mechanical loosening index. This continuous value is constrained to the 0-1 range by a Sigmoid activation function and then used as the final mechanical loosening index output value. The classification branch of the second fully connected layer consists of a Softmax classification layer, and the weight matrix of the classification branch... The dimension is 4 rows and 128 columns, and the bias vector is... The dimension is 4-dimensional. The weighted latent features are combined with the weight matrix. Multiply, plus the bias vector The system obtains a 4-dimensional category score vector, inputs the category score vector into the Softmax activation function, outputs the probability of each category, and outputs the category label corresponding to the maximum probability as the insulation degradation level.

[0063] The degradation discrimination model is optimized using a joint loss function. This joint loss function is a weighted sum of the cross-entropy loss of the classification branch and the mean squared error loss of the regression branch. The cross-entropy loss of the classification branch is calculated as follows: for a single training sample, the cross-entropy between the predicted probability vectors of each category and the one-hot encoded vector of the insulation degradation level label is taken. The mean squared error loss of the regression branch is calculated as the squared difference between the mechanical loosening index label and the continuous values ​​of the mechanical loosening index output by the model. The classification loss weight in the joint loss function is set to 1.0, and the regression loss weight is set to 0.5. The Adam optimizer is used during training, with a learning rate of 0.001, a first-order moment decay coefficient of 0.9, and a second-order moment decay coefficient of 0.999. Mini-batch gradient descent is used, with each mini-batch containing 64 samples, and the number of training iterations is set to 200 epochs. After each iteration, the classification accuracy and regression mean absolute error are monitored using a validation set. When the joint loss on the validation set no longer decreases, the model parameters are saved as the pre-trained degradation discrimination model parameters. After training, the pre-trained degradation discrimination model is used to evaluate the state of the low-dimensional state representation vectors that have been collected and processed in real time.

[0064] See Figure 6 In the graph, the horizontal axis represents the monitoring time points, ranging from 0 to 1000. The left side of the vertical axis represents the insulation degradation level of the switchgear, ranging from 0 to 3, and the right side represents the mechanical loosening index, ranging from 0 to 1. The blue stepped curve represents the discrete change in the insulation degradation level, which increases sequentially from 0 to 3. Specifically, level 0 persists until about the 250th time point before jumping to level 1, then to level 2 at about the 500th time point, and finally to level 3 at about the 770th time point. The red continuous curve represents the trend of the mechanical loosening index, showing a gradual upward trend overall. In the initial stage (time points 0 to about 250), the mechanical loosening index slowly rises from near 0 to about 0.25; then it continues to rise steadily to about 0.55 at about the 500th time point; then it further increases to about 0.75 at about the 750th time point; in the final stage (time points 750 to 1000), the mechanical loosening index tends to stabilize, fluctuating around 0.9.

[0065] Example 5:

[0066] In practice, during the acquisition of the first, second, and third current timing signals, ambient temperature and humidity data inside the switchgear are simultaneously acquired. Ambient temperature data is obtained using a digital temperature sensor installed in the switchgear busbar compartment. The digital temperature sensor has a measurement range of -40℃ to 125℃, a measurement accuracy of ±0.3℃, and outputs a digital temperature value after analog-to-digital conversion. The sampling frequency is synchronized with the sampling frequency of the current signal. Ambient humidity data is obtained using a digital humidity sensor installed at the same location. The digital humidity sensor has a relative humidity measurement range of 0%RH to 100%RH, a measurement accuracy of ±3%RH, and outputs a digital relative humidity value after analog-to-digital conversion.

[0067] The collected raw ambient temperature data is recorded as follows: The unit is Celsius; the collected raw environmental humidity data is recorded as... The unit is relative humidity as a percentage. The raw ambient temperature and humidity data were normalized separately to obtain normalized ambient temperature values. and normalized ambient humidity value The normalized ambient temperature value is obtained using the following formula:

[0068]

[0069] in, This is the original ambient temperature data. This is the lowest permissible temperature value for the switchgear operating environment. The temperature is set to -25℃, based on the lower limit of the ambient temperature under extreme low-temperature conditions at the switchgear installation location. This represents the highest permissible temperature value for the switchgear's operating environment. The temperature is set at 85℃, based on the upper limit of the possible temperature inside the busbar room when the switchgear is operating at full load and the ambient heat dissipation is at its worst. The value ranges from 0 to 1. The normalized ambient humidity value is obtained by the following formula:

[0070]

[0071] in, This is the raw ambient humidity data. This represents the maximum percentage value of relative humidity. The value range is from 0 to 1.

[0072] After obtaining the low-dimensional state representation vector, the ambient temperature value is normalized. and normalized ambient humidity value As auxiliary input parameters, they are concatenated with the low-dimensional state representation vector to form an augmented state representation vector. The concatenation operation combines the low-dimensional state representation vector, the normalized ambient temperature value, and the normalized ambient humidity value into a new column vector. The concatenation order is such that all p components of the low-dimensional state representation vector are at the beginning, followed by the normalized ambient temperature value. Located at position p+1, normalized ambient humidity value Located at position p+2, the augmented state representation vector has a dimension of p+2.

[0073] When the augmented state representation vector is input into the pre-trained degradation discrimination model for state evaluation, the dimension of the input layer of the degradation discrimination model matches the dimension of the augmented state representation vector. The structure of the degradation discrimination model includes a first fully connected layer, a ReLU nonlinear activation function, an attention mechanism layer, and a second fully connected layer connected in sequence. The second fully connected layer contains regression and classification branches set in parallel. When using the augmented state representation vector, the pre-training process of the degradation discrimination model is as follows: collect historical low-dimensional state representation vectors, corresponding historical normalized ambient temperature values, and historical normalized ambient humidity values ​​of multiple sets of switchgear synchronously obtained under known operating conditions. Generate historical augmented state representation vectors in the same splicing method, and label each set of historical augmented state representation vectors with the corresponding insulation degradation level label and mechanical loosening index label. The dimension of the weight matrix of the first fully connected layer is adjusted to 128 rows (p+2) columns, the bias vector remains 128 dimensions, and the structure of the remaining layers remains unchanged. The degradation discrimination model is trained using historical augmented state representation vectors and corresponding labels. The training employs the joint loss function, Adam optimizer parameter settings, and training hyperparameters described in Example 4. After training, a pre-trained degradation discrimination model suitable for augmented state representation vectors is obtained. During the real-time monitoring phase, the real-time generated low-dimensional state representation vector is concatenated with synchronously collected and normalized ambient temperature and humidity data to form an augmented state representation vector, which is then input into the pre-trained degradation discrimination model. The model outputs the corrected switchgear insulation degradation level and mechanical loosening index.

[0074] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for multi-parameter composite online monitoring of switchgear, characterized in that, The method includes: The three-phase current signals of the switchgear bus are synchronously acquired to obtain the first current timing signal, the second current timing signal and the third current timing signal. The three-phase current signal is subjected to variational mode decomposition to generate multiple intrinsic mode function components, and Hilbert-Huang transform is performed on the multiple intrinsic mode function components to obtain the instantaneous frequency sequence and instantaneous amplitude sequence. The partial discharge characteristic parameters and mechanical vibration characteristic parameters of the instantaneous frequency sequence and instantaneous amplitude sequence are extracted to construct a high-dimensional monitoring feature vector; The high-dimensional monitoring feature vector is subjected to adaptive dimensionality reduction mapping to generate a low-dimensional state representation vector; The low-dimensional state representation vector is input into the pre-trained degradation discrimination model for state evaluation to obtain the insulation degradation level and mechanical loosening index of the switchgear.

2. The method for multi-parameter composite online monitoring of switchgear as described in claim 1, characterized in that, The step of performing variational mode decomposition on the three-phase current signal to generate multiple intrinsic mode function components includes: Set the number of mode decomposition layers of variational mode decomposition to a preset value, and initialize the center frequency and bandwidth constraint parameters of each intrinsic mode function; During the iterative solution process, the frequency domain of each phase current signal in the three-phase current signal is divided, and the frequency domain representation of each intrinsic mode function is updated by the alternating direction multiplier algorithm until the convergence condition is met. After satisfying the convergence condition, each frequency domain representation is subjected to inverse Fourier transform to obtain the time domain waveform corresponding to each intrinsic mode function component.

3. The method for multi-parameter composite online monitoring of switchgear as described in claim 2, characterized in that, The step of performing Hilbert-Huang transform on the plurality of intrinsic mode function components to obtain the instantaneous frequency sequence and instantaneous amplitude sequence includes: Perform a Hilbert transform on each intrinsic mode function component to construct an analytic signal; The instantaneous amplitude envelope and instantaneous phase are extracted from the analytical signal, and the instantaneous phase is differentiated in time to obtain the instantaneous frequency sequence; Align the instantaneous amplitude sequence and the instantaneous frequency sequence by time index to generate the Hilbert time spectrum.

4. The method for multi-parameter composite online monitoring of switchgear as described in claim 1, characterized in that, The step of extracting the partial discharge characteristic parameters and mechanical vibration characteristic parameters of the instantaneous frequency sequence and instantaneous amplitude sequence to construct a high-dimensional monitoring feature vector includes: The pulse amplitude peak value, pulse repetition rate, and phase distribution characteristics are extracted from the instantaneous amplitude sequence as partial discharge characteristic parameters; The fundamental frequency offset, harmonic energy ratio, and frequency fluctuation intensity are extracted from the instantaneous frequency sequence as mechanical vibration characteristic parameters. The partial discharge characteristic parameters are concatenated with the mechanical vibration characteristic parameters to obtain the high-dimensional monitoring feature vector.

5. The method for multi-parameter composite online monitoring of switchgear as described in claim 1, characterized in that, The adaptive dimensionality reduction mapping process for the high-dimensional monitoring feature vector to generate a low-dimensional state representation vector includes: Calculate the covariance matrix between each dimension of the high-dimensional monitoring feature vector, and solve for the eigenvalues ​​and eigenvectors of the covariance matrix; The number of principal components is selected based on the cumulative contribution rate of the eigenvalues, and the high-dimensional monitoring feature vector is projected onto the subspace spanned by the corresponding feature vectors. The coefficient vector obtained after projection is used as the low-dimensional state representation vector.

6. The method for multi-parameter composite online monitoring of switchgear as described in claim 1, characterized in that, The step of inputting the low-dimensional state representation vector into a pre-trained degradation discrimination model for state evaluation to obtain the insulation degradation level and mechanical loosening index of the switchgear includes: The low-dimensional state representation vector is input into the first fully connected layer of the degradation discrimination model, and the first latent feature is generated by mapping through a nonlinear activation function. The first latent feature is input into the attention mechanism layer of the degradation discrimination model, the weight distribution of each feature dimension is calculated, and the first latent feature is weighted based on the weight distribution to generate a weighted latent feature. The weighted latent features are input into the second fully connected layer of the degradation discrimination model for regression and classification, and the discrete label values ​​corresponding to the insulation degradation level and the continuous values ​​corresponding to the mechanical loosening index are output.

7. The method for multi-parameter composite online monitoring of switchgear as described in claim 6, characterized in that, The step of inputting the low-dimensional state representation vector into the first fully connected layer of the degradation discrimination model and generating the first latent feature through mapping using a nonlinear activation function includes: Multiply the low-dimensional state representation vector by the weight matrix of the first fully connected layer, and add the bias vector to obtain the linear transformation result; The linear transformation result is input into the ReLU activation function for nonlinear mapping to obtain the first hidden feature.

8. The method for multi-parameter composite online monitoring of switchgear as described in claim 6, characterized in that, The step of inputting the first latent feature into the attention mechanism layer of the degradation discrimination model and calculating the weight distribution of each feature dimension includes: The first latent feature is subjected to global average pooling to obtain a first global feature vector, and the first global feature vector is subjected to a first dimensionality reduction process to obtain a first dimensionality-reduced feature. The first dimensionality reduction feature is subjected to a second dimensionality increase process to obtain a second global feature vector, and the second global feature vector is mapped to the interval between 0 and 1 through the Sigmoid activation function to generate the weight coefficients of each feature dimension. The weighted latent feature is obtained by multiplying the weight coefficients element by element with the first latent feature.

9. The method for multi-parameter composite online monitoring of switchgear as described in claim 6, characterized in that, The step of inputting the weighted latent features into the second fully connected layer of the degradation discrimination model for regression and classification, and outputting the discrete label value corresponding to the insulation degradation level and the continuous value corresponding to the mechanical loosening index, includes: The weighted latent features are input into the regression branch of the second fully connected layer, and the continuous value of the mechanical loosening index is output through linear mapping; The weighted latent features are input into the classification branch of the second fully connected layer, and the probability of each category is output through the Softmax activation function. The category label corresponding to the maximum probability is taken as the insulation degradation level.

10. The method for multi-parameter composite online monitoring of switchgear as described in claim 1, characterized in that, The method further includes: During the acquisition of the first current timing signal, the second current timing signal, and the third current timing signal, the ambient temperature and humidity data inside the switch cabinet are acquired simultaneously. The ambient temperature and humidity data are used as auxiliary input parameters and concatenated with the low-dimensional state representation vector to form an augmented state representation vector. The augmented state representation vector is input into the pre-trained degradation discrimination model for state evaluation to obtain the corrected switchgear insulation degradation level and mechanical loosening index.

Images