Intelligent monitoring system for mechanical characteristics of high-voltage disconnector operating mechanism

Abstract

The application relates to the field of intelligent control of high-voltage isolator operating mechanisms, and discloses an intelligent monitoring system for mechanical characteristics of a high-voltage isolator operating mechanism. The system comprises: a multi-source signal synchronous acquisition end which synchronously acquires a driving motor current signal and a transmission shaft vibration signal; a time-frequency characteristic coupling extraction end which adopts adaptive noise auxiliary ensemble empirical mode decomposition to process the vibration signal to extract marginal spectrum characteristics, adopts dynamic time warping calculation to extract dynamic deformation characteristics of the current signal, orthogonally fuses the marginal spectrum characteristics and the dynamic deformation characteristics to construct a time-frequency domain coupling characteristic vector; and a state evolution prediction end which inputs the time-frequency domain coupling characteristic vector into a hidden Markov model based on dynamic correction of a state transition probability matrix, drives a forward-backward calculation process based on a current observation sequence to calculate a hidden state probability distribution, extracts a maximum probability hidden state evolution slope to output a trend prediction index. The application realizes the conversion of mechanical characteristics from post-alarm to trend prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent control of high-voltage disconnector operating mechanisms, and discloses an intelligent monitoring system for the mechanical characteristics of high-voltage disconnector operating mechanisms. Background Technology

[0002] High-voltage disconnect switches are critical equipment in power systems, and the mechanical characteristics of their operating mechanisms directly affect the reliability of switch operation and the safety of power grid operation. In existing technologies, monitoring the mechanical characteristics of the operating mechanism typically employs a single-signal monitoring method, primarily relying on current transformers installed in the drive motor circuit or vibration sensors mounted on the operating mechanism's transmission shaft. In practice, the system collects current waveforms during motor operation or vibration waveforms during mechanism operation, extracting time-domain characteristic indicators such as peak current, operating time, vibration amplitude, and duration. The system compares these extracted time-domain characteristic indicators with pre-set fixed thresholds. If any time-domain characteristic indicator exceeds the corresponding fixed threshold range, the system determines that the operating mechanism has a mechanical abnormality and issues an alarm signal. This method, based on comparing a single signal with a fixed threshold, constitutes the operational foundation of current mainstream monitoring systems.

[0003] Regarding the aforementioned collected signals, existing technologies also have limitations in data processing and state assessment. For vibration signal processing, conventional techniques often use Fast Fourier Transform to obtain spectral information or directly extract the statistical features of the vibration envelope. For current signals, existing solutions typically only focus on the start and end times of the overall current waveform, calculating the closing or opening time. In the state assessment stage, existing technologies only perform transient discrimination based on the signal features extracted from the current single action, lacking correlation calculation and in-depth mining of historical action data features. The system can only trigger the alarm mechanism when the mechanical defect has developed to a relatively serious extent, causing the time-domain or frequency-domain characteristic indicators to significantly exceed the limits. This is a typical post-event alarm logic and cannot characterize the continuous changes in mechanical characteristics during multiple actions.

[0004] High-voltage substations are characterized by complex electromagnetic interference and mechanical shock noise. The signal-to-noise ratio of a single sensor's measurement signal is low, and simple frequency domain transformation and envelope statistics cannot extract features reflecting early, minor wear from a strong noise background. Furthermore, the ability of a single signal feature to characterize complex mechanical degradation states is limited. Simultaneously, the degradation of the mechanical characteristics of the operating mechanism is a continuous, non-linear evolution process. Relying solely on threshold discrimination of features at the current moment severs the temporal correlation of feature evolution with increasing number of operations. This leads to the core technical problem of existing technologies: difficulty in extracting early degradation features of the operating mechanism's mechanical characteristics under complex electromagnetic and mechanical noise interference, and inability to predict the state development trend. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent monitoring system for the mechanical characteristics of the operating mechanism of a high-voltage disconnector, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] An intelligent monitoring system for the mechanical characteristics of a high-voltage disconnector operating mechanism includes: a multi-source signal synchronous acquisition end for synchronously acquiring the current signal of the drive motor and the vibration signal of the operating mechanism's transmission shaft; a time-frequency feature coupling extraction end for using adaptive noise-assisted ensemble empirical mode decomposition to decompose the vibration signal into intrinsic mode functions and extract marginal spectral features, using dynamic time warping to extract the dynamic deformation features of the current signal, and orthogonally fusing the marginal spectral features and the dynamic deformation features in the feature space to construct a time-frequency domain coupled feature vector; and a state evolution prediction end for inputting the time-frequency domain coupled feature vector into a hidden Markov model based on dynamic correction of the state transition probability matrix, using the current observation sequence to drive the forward-backward calculation process to calculate the probability distribution of each hidden state, extracting the slope of the maximum probability hidden state evolving over time, and outputting a trend prediction index for the mechanical characteristics.

[0008] Preferably, the method of decomposing the vibration signal into intrinsic mode functions and extracting marginal spectral features using adaptive noise-assisted ensemble empirical mode decomposition includes: adding white noise sequences of different amplitudes to the vibration signal multiple times; performing empirical mode decomposition on the vibration signal after adding white noise sequences to obtain a set of candidate intrinsic mode functions; calculating the correlation coefficient between adjacent candidate intrinsic mode functions in the set of candidate intrinsic mode functions; identifying candidate intrinsic mode functions with correlation coefficients less than a preset correlation threshold as noise components and removing them; determining the remaining candidate intrinsic mode functions as the intrinsic mode functions; performing Hilbert transform on the intrinsic mode functions to obtain an instantaneous frequency sequence; constructing a time spectrum based on the instantaneous frequency sequence and the amplitude of the intrinsic mode functions; and performing an integral operation along the time axis on the time spectrum to obtain the marginal spectral features.

[0009] Preferably, the dynamic deformation features of the current signal extracted using dynamic time warping calculation include: constructing a standard reference current waveform template for the current signal; using the current signal and the standard reference current waveform template as a test sequence and a reference sequence, respectively, calculating the Euclidean distance between each sampling point in the test sequence and each sampling point in the reference sequence, and constructing a distance cost matrix; searching for the optimal warping path with the minimum cumulative distance between the test sequence and the reference sequence based on the distance cost matrix, extracting the local slope and local curvature at each sampling point along the optimal warping path, and arranging the local slope and local curvature in chronological order to form the dynamic deformation features.

[0010] Preferably, constructing a time-frequency domain coupled feature vector by orthogonally fusing the marginal spectral features and the dynamic deformation features in the feature space includes: performing zero-mean processing on the marginal spectral features and the dynamic deformation features respectively to obtain a centered marginal spectral feature vector and a centered dynamic deformation feature vector; calculating the covariance matrix between the centered marginal spectral feature vector and the centered dynamic deformation feature vector, performing eigenvalue decomposition on the covariance matrix to obtain an orthogonal projection matrix; using the orthogonal projection matrix to map the centered marginal spectral feature vector and the centered dynamic deformation feature vector to the same orthogonal feature subspace, and concatenating the mapped vectors in the orthogonal feature subspace to generate the time-frequency domain coupled feature vector.

[0011] Preferably, the Hidden Markov Model based on dynamic correction of the state transition probability matrix includes: initializing the initial state probability vector, initial state transition probability matrix, and observation probability matrix of the Hidden Markov Model; inputting the time-frequency domain coupled feature vector as the observation sequence into the Hidden Markov Model, and calculating the posterior probability distribution of each hidden state under the observation sequence using the forward-backward computation process; based on the time label of the posterior probability distribution and the time-frequency domain coupled feature vector, counting the hidden state transition frequency within adjacent time steps, recursively updating and replacing the initial state transition probability matrix using the hidden state transition frequency, and substituting the updated state transition probability matrix into the forward-backward computation process iteratively until convergence.

[0012] Preferably, the method of constructing a time spectrum based on the instantaneous frequency sequence and the amplitude of the intrinsic mode function, and performing an integral operation on the time spectrum along the time axis to obtain the marginal spectral features includes: setting a frequency band division interval, performing frequency band clustering on the instantaneous frequency sequence, merging the instantaneous frequency sequences whose cluster center frequencies fall within the same frequency band division interval and their corresponding intrinsic mode function amplitudes to obtain the frequency band reconstructed time spectrum; calculating the energy entropy value corresponding to each frequency band component in the frequency band reconstructed time spectrum, determining the frequency band components with energy entropy values ​​less than a preset entropy threshold as aliasing interference frequency bands and filtering them out; integrating the frequency band reconstructed time spectrum after filtering out aliasing interference frequency bands along the time axis, extracting the marginal spectral peak value, marginal spectral variance, and marginal spectral skewness of the retained frequency bands, and combining the marginal spectral peak value, marginal spectral variance, and marginal spectral skewness to constitute the marginal spectral features.

[0013] Preferably, extracting the local slope and local curvature at each sampling point along the optimal regular path, and arranging the local slope and local curvature in chronological order to form the dynamic deformation features includes: segmenting the optimal regular path according to a fixed time window to obtain multiple local regular path segments; calculating the mean and standard deviation of the local slope and local curvature in each local regular path segment to construct a local path statistical feature vector; combining all local path statistical feature vectors in chronological order to generate an initial dynamic deformation feature matrix; performing principal component analysis on the initial dynamic deformation feature matrix; selecting feature vectors with a cumulative contribution rate exceeding a preset threshold to form a dimension-reduced projection matrix; and multiplying the initial dynamic deformation feature matrix with the dimension-reduced projection matrix to obtain the dynamic deformation features.

[0014] Preferably, concatenating the mapped vectors in the orthogonal feature subspace to generate the time-frequency domain coupled feature vector includes: calculating the Pearson correlation coefficient between the marginal spectral feature vector mapped to the orthogonal feature subspace and the dynamic deformation feature vector; when the Pearson correlation coefficient exceeds a preset correlation threshold, performing weighted fusion on the mapped marginal spectral feature vector and the dynamic deformation feature vector, and determining the weighted fused single vector as the time-frequency domain coupled feature vector; when the Pearson correlation coefficient does not exceed the preset correlation threshold, performing concatenating the mapped marginal spectral feature vector and the dynamic deformation feature vector, and performing manifold regularization processing on the concatenated vector, and determining the manifold regularized vector as the time-frequency domain coupled feature vector.

[0015] Preferably, the system further includes an online model calibration terminal, which is used to: record the measured mechanical response time of the operating mechanism when it performs subsequent actions after each output of the trend prediction index of the mechanical characteristics; calculate the residual value between the measured mechanical response time and the predicted response time corresponding to the trend prediction index; when the residual value exceeds a preset residual tolerance interval, extract the historical time-frequency domain coupled feature vector and the historical measured mechanical response time within a preset time window before the current moment to construct an incremental sample set; use the incremental sample set to perform stochastic gradient descent fine-tuning on the observation probability matrix of the Hidden Markov Model, and freeze the hidden state topology of the Hidden Markov Model during the fine-tuning process, and update the parameters of the observation probability matrix until the residual value falls into the preset residual tolerance interval.

[0016] Preferably, it further includes a multi-source data spatiotemporal alignment end, which is used to: acquire a first original timestamp of the current signal of the drive motor and a second original timestamp of the vibration signal of the transmission shaft of the operating mechanism; calculate a clock offset based on the first original timestamp and the second original timestamp, and use the clock offset to resample and interpolate the current signal and the vibration signal so that the current signal and the vibration signal are strictly aligned on the time axis; detect abrupt spike interference in the aligned current signal and vibration signal, use a sliding median filter to smoothly replace the abrupt spike interference, and transmit the smoothed current signal and vibration signal to the time-frequency feature coupling extraction end.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] 1. This invention synchronously acquires the current signal of the drive motor and the vibration signal of the transmission shaft of the operating mechanism. It employs adaptive noise-assisted ensemble empirical mode decomposition to decompose the vibration signal into intrinsic mode functions and extract marginal spectral features. Dynamic time warping is used to extract the dynamic deformation features of the current signal. The marginal spectral features and dynamic deformation features are orthogonally fused in the feature space to construct a time-frequency domain coupled feature vector. This overcomes the masking of weak fault features by complex electromagnetic and mechanical environmental noise, extracting high-dimensional coupled features reflecting early wear of mechanical transmission components, and solving the problem of insufficient characterization of complex degradation states by single signal features. The time-frequency domain coupled feature vector is input into a hidden Markov model based on dynamic correction of the state transition probability matrix. The forward-backward calculation process is driven by the current observation sequence to calculate the probability distribution of each hidden state. The slope of the maximum probability hidden state evolving over time is extracted to output a trend prediction index of mechanical characteristics. This realizes the transformation of the mechanical characteristics of the high-voltage disconnector operating mechanism from post-threshold alarm to early degradation trend quantitative prediction, avoiding sudden jamming or failure to operate of the operating mechanism during the operating cycle.

[0019] 2. This invention adds white noise to the vibration signal sequence for empirical mode decomposition and removes noise components based on the correlation coefficient between adjacent candidate intrinsic mode functions, suppressing mode aliasing and retaining intrinsic mode functions that reflect real physical motion. The optimal warped path obtained from dynamic time warping is cut by time windows, and the mean and standard deviation of local slope and local curvature are calculated and principal component analysis is performed, reducing the redundant dimensions of dynamic deformation features. During orthogonal fusion of feature spaces, weighted fusion or manifold regularization splicing is selected based on the Pearson correlation coefficient, eliminating the overlap between marginal spectral features and dynamic deformation features. The linear and nonlinear redundancy interference between the two states is eliminated. Based on the posterior probability distribution and time label statistics of the hidden state transition frequency, the initial state transition probability matrix is ​​recursively updated. After outputting the trend prediction index, the observation probability matrix is ​​fine-tuned by stochastic gradient descent based on the residual value of the measured mechanical response time and the predicted response time, and the hidden state topology is frozen. The above process adapts to the nonlinear decay law of mechanical characteristics of the operating mechanism throughout its entire life cycle, eliminates the model initial parameter setting error and model drift caused by long-term operation, and improves the tracking accuracy and robustness of the state prediction model in the long-term operation process. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall workflow of the intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector operating mechanism of the present invention.

[0021] Figure 2 This is a flowchart of the multi-source signal acquisition and spatiotemporal alignment processing of the present invention;

[0022] Figure 3 This is a flowchart of the vibration signal marginal spectrum feature extraction process of the present invention;

[0023] Figure 4 This is a flowchart of the current signal dynamic deformation feature extraction process of the present invention;

[0024] Figure 5 This is a flowchart of the orthogonal fusion process of time-frequency domain coupling features in this invention;

[0025] Figure 6 This is a flowchart of the state evolution prediction and online model calibration process of the present invention. Detailed Implementation

[0026] refer to Figure 1In one embodiment, the intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector switch operating mechanism includes a multi-source signal synchronous acquisition terminal, a time-frequency feature coupling extraction terminal, and a state evolution prediction terminal. The multi-source signal synchronous acquisition terminal establishes electrical and mechanical connections with the current loop of the drive motor and the transmission shaft of the operating mechanism, respectively, synchronously acquiring the current signal of the drive motor during the opening and closing operations and the vibration signal generated by the transmission shaft of the operating mechanism. The time-frequency feature coupling extraction terminal receives the current signal and vibration signal output from the multi-source signal synchronous acquisition terminal, and uses adaptive noise-assisted ensemble empirical mode decomposition to decompose the vibration signal, obtaining multiple intrinsic mode functions and extracting marginal spectral features from them; simultaneously, it uses a dynamic time warping calculation method to process the current signal, extracting the dynamic deformation features of the current signal; and then orthogonally fuses the marginal spectral features and dynamic deformation features in the feature space to construct a time-frequency domain coupled feature vector. The state evolution prediction end receives the time-frequency domain coupled feature vector output by the time-frequency feature coupling extraction end, inputs it into the hidden Markov model based on dynamic correction of the state transition probability matrix, drives the forward-backward calculation process with the observation sequence at the current moment, calculates the probability distribution of each hidden state in the model, extracts the slope of the maximum probability hidden state as time evolves, and outputs the trend prediction index of mechanical characteristics.

[0027] Specifically, the multi-source signal synchronous acquisition terminal includes a current acquisition module and a vibration acquisition module. The current acquisition module is connected in series to any one phase of the three-phase power supply circuit of the drive motor, acquiring the instantaneous current value of the drive motor during operation. The vibration acquisition module is fixedly installed on the bearing seat of the operating mechanism's drive shaft via a rigid connector, acquiring the mechanical vibration acceleration signal generated by the drive shaft during rotation and impact. The current acquisition module and the vibration acquisition module share the same clock source, ensuring that the acquired current signal and vibration signal have the same time reference. Simultaneously with the drive motor receiving a tripping or closing command, the multi-source signal synchronous acquisition terminal triggers the acquisition process, synchronously sampling the current signal and vibration signal according to a preset sampling frequency. The sampling process continues until the drive motor stops running and the operating mechanism's drive shaft is completely stationary.

[0028] The time-frequency feature coupling extraction end performs adaptive noise-assisted ensemble empirical mode decomposition (EMD) processing on the received vibration signal. This process first adds white noise sequences of different amplitudes to the original vibration signal multiple times. After each addition of white noise, EMD is performed on the noise-added vibration signal to obtain a set of candidate intrinsic mode functions (EMFs). After multiple additions of white noise and EMD, multiple sets of candidate EMFs are obtained, forming a candidate EMF set. The correlation coefficient between adjacent candidate EMFs in the candidate EMF set is calculated. Candidate EMFs with correlation coefficients less than a preset correlation threshold are identified as noise components and removed. The remaining candidate EMFs are the EMFs that reflect the true physical characteristics of the vibration signal. A Hilbert transform is performed on each EMF to obtain the corresponding instantaneous frequency sequence. A time-spectrum is constructed based on the instantaneous frequency sequence and the amplitude of the EMF. The time-spectrum is integrated along the time axis to obtain the marginal spectral features.

[0029] The time-frequency feature coupling extraction end performs dynamic time warping calculation on the received current signal. This process first constructs a standard reference current waveform template, derived from the average value of current signals collected during multiple opening and closing operations of the high-voltage disconnector during its factory commissioning phase, when the operating mechanism is in a brand-new state. Using the real-time acquired current signal as the test sequence and the standard reference current waveform template as the reference sequence, the Euclidean distance between each sampling point in the test sequence and each sampling point in the reference sequence is calculated, constructing a distance cost matrix. Based on the distance cost matrix, the optimal warping path that minimizes the cumulative distance between the test sequence and the reference sequence is searched. Local slopes and local curvatures at each sampling point are extracted along the optimal warping path, and these local slopes and curvatures are arranged in chronological order to form dynamic deformation features.

[0030] The time-frequency feature coupling extraction stage orthogonally fuses the extracted marginal spectral features and dynamic deformation features in the feature space. This fusion process first performs zero-mean processing on both the marginal spectral features and the dynamic deformation features, obtaining a centralized marginal spectral feature vector and a centralized dynamic deformation feature vector. The covariance matrix between the centralized marginal spectral feature vector and the centralized dynamic deformation feature vector is calculated, and eigenvalue decomposition is performed on the covariance matrix to obtain the orthogonal projection matrix. Using the orthogonal projection matrix, the centralized marginal spectral feature vector and the centralized dynamic deformation feature vector are mapped to the same orthogonal feature subspace. In this orthogonal feature subspace, the two mapped vectors are concatenated to generate a time-frequency domain coupled feature vector.

[0031] The state evolution prediction end inputs the time-frequency domain coupled feature vector into a Hidden Markov Model (HMM) dynamically modified based on the state transition probability matrix. This model pre-initializes the initial state probability vector, the initial state transition probability matrix, and the observation probability matrix. After inputting the time-frequency domain coupled feature vector as the observation sequence into the model, a forward-backward computation process is used to calculate the posterior probability distribution of each hidden state under this observation sequence. Based on the posterior probability distribution and the time label of the time-frequency domain coupled feature vector, the transition frequency of the hidden states within adjacent time steps is statistically analyzed. The initial state transition probability matrix is ​​recursively updated and replaced using the hidden state transition frequency. The updated state transition probability matrix is ​​then substituted into the forward-backward computation process for iteration until the calculation results converge. The slope of the maximum probability hidden state evolving over time after convergence is extracted, and this slope is output as a trend prediction index for mechanical characteristics.

[0032] In this embodiment, the multi-source signal synchronous acquisition end synchronously acquires current and vibration signals. The time-frequency feature coupling extraction end extracts the marginal spectral features of the vibration signal and the dynamic deformation features of the current signal, respectively, and performs orthogonal fusion. The state evolution prediction end outputs mechanical characteristic trend prediction indicators based on a dynamically modified hidden Markov model. This system can extract feature information reflecting the early mechanical deterioration of the operating mechanism from a strong noise background, and realize the quantitative prediction of the mechanical characteristic evolution trend.

[0033] Table 1 Correspondence between mechanical states and characteristics of high-voltage disconnector operating mechanisms

[0034] Normal state The marginal spectral peaks are concentrated in the low-frequency range, with relatively small variance and skewness. The local slope deviates little from the standard template, and the curvature is uniform. State 1 Slight wear The marginal spectral peaks shift towards the mid-frequency band, and the variance increases. The local slope fluctuates slightly, and the curvature increases locally. State 2 Moderate wear The marginal spectrum shows multiple peaks, and the skewness is significantly increased. Local slope fluctuations intensify, and curvature distribution becomes uneven. State 3 Severe wear The marginal spectral peaks diffuse into the high-frequency band, and the variance and skewness increase sharply. The local slope deviates significantly from the standard template, and the curvature changes abruptly. State 4

[0035] In this embodiment, Table 1 shows the correspondence between the marginal spectrum characteristics, dynamic deformation characteristics, and maximum probability hidden states of the high-voltage disconnector operating mechanism under different mechanical states. Under normal conditions, the transmission components of the operating mechanism are well-matched, the vibration signal energy is mainly concentrated in the low-frequency range, the current waveform is highly consistent with the standard template, and the maximum probability hidden state output by the model is state 1. As the degree of mechanical wear increases, the energy distribution of the vibration signal gradually spreads to the mid-to-high frequency range, the current waveform exhibits local deformation, and the maximum probability hidden states output by the model successively change to states 2, 3, and 4. By monitoring the changes in the maximum probability hidden states, the degree of mechanical degradation of the operating mechanism can be accurately determined.

[0036] refer to Figure 3In one optional embodiment, the process of decomposing a vibration signal into intrinsic mode functions and extracting marginal spectral features using adaptive noise-assisted ensemble empirical mode decomposition includes the following steps: White noise sequences of different amplitudes are added to the vibration signal multiple times. The amplitude of each added white noise sequence is 0.01 to 0.1 times the standard deviation of the original vibration signal, and the number of additions is set to 100 to 200 times. Empirical mode decomposition is performed on the vibration signal after each addition of white noise sequences to obtain a set of candidate intrinsic mode functions. The specific process of empirical mode decomposition is as follows: Identify all local maxima and local minima in the vibration signal, connect all local maxima with cubic spline curves to form an upper envelope, and connect all local minima to form a lower envelope; calculate the average of the upper and lower envelopes to obtain the mean envelope; subtract the mean envelope from the original vibration signal to obtain the first component; determine whether this component satisfies the two conditions of intrinsic mode functions, namely, the difference between the number of extrema and the number of zero-crossing points does not exceed 1, and the average of the upper and lower envelopes is zero at any time; if satisfied, this component is the first candidate intrinsic mode function; if not satisfied, this component is used as a new original signal, and the above process is repeated until a candidate intrinsic mode function that meets the conditions is obtained; subtract the first candidate intrinsic mode function from the original signal to obtain the remaining signal, and repeat the above process on the remaining signal until the remaining signal is a monotonic function or a constant, thus completing one empirical mode decomposition.

[0037] After repeatedly adding white noise and performing empirical mode decomposition, multiple sets of candidate intrinsic mode functions (EMFs) are obtained, forming a candidate EMF set. The correlation coefficient between adjacent candidate EMFs in the candidate EMF set is calculated using the following formula:

[0038]

[0039] in, Let represent the correlation coefficient between the i-th candidate intrinsic mode function and the (i+1)-th candidate intrinsic mode function. Let represent the magnitude of the i-th candidate intrinsic mode function at time t. This represents the average magnitude of the i-th candidate intrinsic mode function. Let represent the magnitude of the (i+1)th candidate intrinsic mode function at time t. Let N represent the average amplitude of the (i+1)th candidate intrinsic mode function, and N represent the total number of sampling points.

[0040] Candidate intrinsic mode functions (IMFs) with correlation coefficients less than a preset correlation threshold are identified as noise components and eliminated. The preset correlation threshold is set to 0.1 to 0.3. The remaining candidate IMFs are the IMFs that reflect the true physical characteristics of the vibration signal. A Hilbert transform is performed on each IMF to obtain the corresponding instantaneous frequency sequence. The formula for calculating the Hilbert transform is:

[0041]

[0042] in, Represents the intrinsic mode function The result of the Hilbert transform.

[0043] Based on the Hilbert transform results, construct the analytic signal:

[0044]

[0045] in, Represents the intrinsic mode function instantaneous amplitude, Represents the intrinsic mode function The instantaneous phase.

[0046] The formulas for calculating instantaneous amplitude and instantaneous phase are as follows:

[0047]

[0048]

[0049] Differentiating the instantaneous phase yields the instantaneous frequency sequence:

[0050]

[0051] The time spectrum is constructed based on the instantaneous frequency sequence and instantaneous amplitude. The expression for the time spectrum is:

[0052]

[0053] Where M represents the number of intrinsic mode functions, Dirac function.

[0054] Integrating the time spectrum along the time axis yields the marginal spectrum:

[0055]

[0056] Where T represents the total duration of the signal.

[0057] Marginal spectrum features are derived by extracting the marginal spectrum peak value, marginal spectrum variance, and marginal spectrum skewness. These features are then combined to form the marginal spectrum characteristics. The marginal spectrum peak value is the maximum value in the marginal spectrum, the marginal spectrum variance is the variance of the marginal spectrum amplitude, and the marginal spectrum skewness is the ratio of the third central moment of the marginal spectrum amplitude to the cube of the standard deviation.

[0058] refer to Figure 4 In one optional embodiment, the process of extracting the dynamic deformation characteristics of the current signal using dynamic time warping calculation includes the following steps: Constructing a standard reference current waveform template for the current signal. The construction process of the standard reference current waveform template is as follows: Collecting current signals from 50 to 100 opening and closing operations of the high-voltage disconnecting switch during the factory commissioning phase, when the operating mechanism is in a brand-new state; performing time alignment and amplitude normalization processing on all collected current signals; calculating the average amplitude of all current signals at each sampling point to obtain the standard reference current waveform template.

[0059] The real-time acquired current signal is used as the test sequence X, and the standard reference current waveform template is used as the reference sequence Y. The length of the test sequence X is N, and the length of the reference sequence Y is M. The Euclidean distance between each sampling point in the test sequence and each sampling point in the reference sequence is calculated, and the distance cost matrix D is constructed. The formula for calculating the element D(i,j) in the distance cost matrix is:

[0060]

[0061] in, This represents the amplitude of the i-th sampling point in the test sequence X. This represents the amplitude of the j-th sampling point in the reference sequence Y.

[0062] The optimal normalized path that minimizes the cumulative distance between the test sequence and the reference sequence is searched based on the distance cost matrix. The formula for calculating the cumulative distance matrix C is:

[0063]

[0064] Where C(1,1)=D(1,1), when i=1 and j>1, C(1,j)=D(1,j)+C(1,j-1); when j=1 and i>1, C(i,1)=D(i,1)+C(i-1,1).

[0065] From the bottom right element of the cumulative distance matrix Start by searching backwards to the top-left element C(1,1). Each time, select the position with the smallest cumulative distance to the left, above, or top-left of the current position as the next path point, until the top-left corner is reached, thus obtaining the optimal regular path P. The optimal regular path P satisfies the following conditions: the path starts from (1,1) and ends at (N,M); each point (i,j) in the path satisfies i≤N and j≤M; adjacent points in the path can only be (i+1,j), (i,j+1), or (i+1,j+1).

[0066] The local slope and local curvature at each sampling point are extracted along the optimal regular path. The formula for calculating the local slope is:

[0067]

[0068] Where k(t) represents the local slope at time t. and This represents two adjacent points on the optimal regularized path.

[0069] The formula for calculating local curvature is:

[0070]

[0071] in, This represents the local curvature at time t. and Let represent the first and second derivatives of the optimal regularized path at time t, respectively. and Let represent the first and second derivatives of the optimal regularized path at time t, respectively.

[0072] Arrange the local slope and local curvature in chronological order to form dynamic deformation characteristics.

[0073] refer to Figure 5 In one optional embodiment, the process of orthogonally fusing marginal spectral features and dynamic deformation features in the feature space to construct a time-frequency coupled feature vector includes the following steps: Zero-mean processing is applied to both the marginal spectral features and the dynamic deformation features to obtain a centered marginal spectral feature vector and a centered dynamic deformation feature vector. The formula for zero-mean processing is:

[0074]

[0075]

[0076] Where X represents the original marginal spectrum eigenvector. This represents the mean of the marginal spectral eigenvectors. Y represents the centered marginal spectral eigenvector; Y represents the original dynamic deformation eigenvector. This represents the mean of the dynamic deformation eigenvectors. This represents the centered dynamic deformation eigenvector.

[0077] Calculate the covariance matrix between the eigenvectors of the centered marginal spectrum and the eigenvectors of the centered dynamic deformation:

[0078]

[0079] Where N represents the number of samples, Let represent the covariance matrix.

[0080] Perform eigenvalue decomposition on the covariance matrix:

[0081]

[0082] Where U represents the orthogonal eigenvector matrix. Let represent a diagonal matrix, where the elements on the diagonal are the eigenvalues ​​of the covariance matrix.

[0083] Arrange the eigenvalues ​​in descending order, and select the eigenvectors corresponding to the K largest eigenvalues ​​to form an orthogonal projection matrix P. K is the number of eigenvalues ​​whose cumulative contribution rate exceeds 90%. Use the orthogonal projection matrix to map the centered marginal spectrum eigenvectors and the centered dynamic deformation eigenvectors to the same orthogonal eigenspace.

[0084]

[0085]

[0086] in, This represents the mapped marginal spectral eigenvector. This represents the dynamic deformation feature vector after mapping.

[0087] In the orthogonal feature subspace, the two mapped vectors are concatenated to generate a time-frequency domain coupled feature vector:

[0088]

[0089] Where Z represents the time-frequency domain coupled eigenvector.

[0090] This embodiment details the specific implementation process of marginal spectrum feature extraction of vibration signals, dynamic deformation feature extraction of current signals, and orthogonal feature fusion. By adding white noise to assist empirical mode decomposition, mode aliasing is suppressed; by dynamic time warping calculation, local deformation features of the current waveform are extracted; and by orthogonal fusion, redundant information between features is eliminated.

[0091] Table 2 Performance Comparison of Feature Extraction Methods under Different Noise Levels

[0092] 20 0.92 0.95 0.98 10 0.78 0.86 0.94 0 0.56 0.72 0.88 -10 0.32 0.51 0.76

[0093] In this embodiment, Table 2 shows a performance comparison of three feature extraction methods—Fast Fourier Transform, Empirical Mode Decomposition (EMD), and Adaptive Noise-Assisted Ensemble Empirical Mode Decomposition (ADE)—under different noise levels. The performance metric used is feature extraction accuracy, which is the proportion of extracted features that can correctly distinguish between normal and abnormal states. Table 2 shows that the feature extraction accuracy of all three methods decreases with decreasing signal-to-noise ratio (SNR), but the ADE method shows the smallest decrease. Even in a strong noise environment with an SNR of -10 dB, the ADE method still achieves a feature extraction accuracy of 0.76, significantly higher than the other two methods. This indicates that the method has strong noise resistance and can effectively extract feature information reflecting the mechanical characteristics of the operating mechanism in complex electromagnetic and mechanical noise environments.

[0094] refer to Figure 6 In one optional embodiment, the construction and operation process of the Hidden Markov Model based on dynamic correction of the state transition probability matrix includes the following steps: Initializing the initial state probability vector of the Hidden Markov Model. The initial state transition probability matrix A and the observation probability matrix B. Initial state probability vector. Let be an N-dimensional vector, where N represents the number of hidden states. Let represent the probability of being in the i-th hidden state at the initial time step, satisfying . The initial state transition probability matrix A is an N×N matrix, where... This represents the probability of transitioning from the i-th hidden state to the j-th hidden state, satisfying... The observation probability matrix B is an N×M matrix, where M represents the number of observations. This represents the probability of observing the j-th observation in the i-th hidden state, satisfying... .

[0095] The time-frequency coupled feature vector is used as the input to the hidden Markov model as the observation sequence O, and the length of the observation sequence O is T. The posterior probability distribution of each hidden state under the observation sequence O is calculated using a forward-backward computation process. Forward variables... This represents the probability that at time t, the user is in the i-th hidden state and has observed the first t observations. The formula for calculating this probability is:

[0096]

[0097]

[0098] in, This represents the observed value at time t.

[0099] Backward variables Let represent the probability that a given time t is in the i-th hidden state and has observed values ​​from t+1 to T. The formula for calculating this probability is:

[0100]

[0101]

[0102] Based on the forward and backward variables, calculate the posterior probability of being in the i-th hidden state at time t. :

[0103]

[0104] Calculate the posterior probability of being in the i-th hidden state at time t and in the j-th hidden state at time t+1. :

[0105]

[0106] Based on the time label of the posterior probability distribution coupled with the time-frequency domain feature vector, the transition frequency of hidden states within adjacent time steps is counted. From time 1 to time T-1, the total frequency of transitions from the i-th hidden state to the j-th hidden state is counted. :

[0107]

[0108] The initial state transition probability matrix is ​​recursively updated and replaced using the hidden state transition frequency. The updated state transition probability matrix is... medium elements The calculation formula is:

[0109]

[0110] The updated state transition probability matrix is ​​substituted into the forward-backward calculation process for iteration until the difference between the state transition probability matrices obtained from two adjacent iterations is less than the preset convergence threshold, thus completing the dynamic correction of the model.

[0111] Extract the slope of the maximum probability hidden state as it evolves over time after convergence. First, determine the maximum probability hidden state at each time t. :

[0112]

[0113] Constructing the maximum probability hidden state sequence Linear regression is performed on the sequence to obtain the slope of the fitted line, and this slope is output as a trend prediction index of mechanical characteristics.

[0114] In one optional embodiment, the process of constructing a time-frequency spectrum based on the instantaneous frequency sequence and the amplitude of the intrinsic mode function (EMF), and obtaining marginal spectral features by integrating the time-frequency spectrum along the time axis, includes the following steps: First, define frequency band division intervals, dividing the frequency range from 0 to half the sampling frequency into multiple equal-width frequency band division intervals, with the width of each interval set to 10Hz to 50Hz. Second, perform frequency band clustering on the instantaneous frequency sequence, using the K-means clustering algorithm to cluster the instantaneous frequency sequence into multiple clusters, each cluster corresponding to a cluster center frequency. Third, merge the instantaneous frequency sequences whose cluster center frequencies fall within the same frequency band division interval and their corresponding EMF amplitudes to obtain the reconstructed time-frequency spectrum.

[0115] The energy entropy value corresponding to each frequency band component in the spectrum is calculated during frequency band reconstruction. The formula for calculating the energy entropy value is as follows:

[0116]

[0117] in, This represents the energy entropy value of the k-th frequency band component. This represents the proportion of the energy of the k-th frequency band component at time t to the total energy.

[0118] Frequency band components with energy entropy values ​​less than a preset entropy threshold are identified as aliasing interference bands and filtered out. The preset entropy threshold is set to 0.5 to 0.8. When reconstructing the frequency bands after filtering out aliasing interference bands, the spectrum is integrated along the time axis, and the marginal spectral peak, marginal spectral variance, and marginal spectral skewness of the retained frequency bands are extracted. These features are combined to form marginal spectral features.

[0119] In one optional embodiment, the process of extracting the local slope and local curvature at each sampling point along the optimal regular path, and arranging the local slope and local curvature in chronological order to form dynamic deformation features, includes the following steps: The optimal regular path is segmented according to a fixed time window, with the time window set to 10ms to 50ms, to obtain multiple segments of the local regular path. The mean and standard deviation of the local slope and local curvature within each segment are calculated to construct a local path statistical feature vector. The local path statistical feature vector contains four elements: the mean local slope, the standard deviation of the local slope, the mean local curvature, and the standard deviation of the local curvature.

[0120] An initial dynamic deformation feature matrix is ​​generated by combining all local path statistical feature vectors in chronological order. The number of rows in the initial dynamic deformation feature matrix corresponds to the number of locally regularized path segments, and the number of columns is 4. Principal component analysis is performed on the initial dynamic deformation feature matrix to calculate its covariance matrix. Eigenvalue decomposition is then performed on the covariance matrix, and the eigenvalues ​​are arranged in descending order. Feature vectors with a cumulative contribution rate exceeding a preset threshold are selected to form a dimension-reduced projection matrix. The preset threshold is set to 0.85 to 0.95. The initial dynamic deformation feature matrix is ​​multiplied by the dimension-reduced projection matrix to obtain the dimension-reduced dynamic deformation features.

[0121] In an optional embodiment, the process of concatenating the mapped vectors in the orthogonal feature subspace to generate time-frequency domain coupled feature vectors includes the following steps: Calculating the Pearson correlation coefficient between the marginal spectral feature vector mapped to the orthogonal feature subspace and the dynamic deformation feature vector:

[0122]

[0123] Where r represents the Pearson correlation coefficient, and K represents the dimension of the orthogonal eigenspace. This represents the i-th element of the mapped marginal spectral eigenvector. This represents the mean of the mapped marginal spectral eigenvectors. This represents the i-th element of the mapped dynamic deformation eigenvector. This represents the mean of the dynamically deformed eigenvectors after mapping.

[0124] When the Pearson correlation coefficient exceeds a preset correlation threshold, a weighted fusion is performed on the mapped marginal spectral feature vector and the dynamic deformation feature vector. The single vector resulting from the weighted fusion is then determined as the time-frequency domain coupled feature vector. The formula for calculating the weighted fusion is as follows:

[0125]

[0126] in, This represents the weighting coefficient, with values ​​ranging from 0.4 to 0.6.

[0127] When the Pearson correlation coefficient does not exceed a preset correlation threshold, the mapped marginal spectral feature vector and the dynamic deformation feature vector are concatenated and then subjected to manifold regularization. The manifold-regularized vector is then determined as the time-frequency domain coupled feature vector. The objective function for manifold regularization is:

[0128]

[0129] in, This represents the original vector after concatenation and splicing. Represents the regularization parameter. Let represent the elements in the adjacency matrix, where the i-th sample and the j-th sample are nearest neighbors. ,otherwise .

[0130] This embodiment details the dynamic correction process of the Hidden Markov Model, the frequency band reconstruction and energy entropy screening process of marginal spectral features, the dimensionality reduction process of dynamic deformation features, and the adaptive selection process of feature fusion. Through frequency band reconstruction and energy entropy screening, aliasing interference is further filtered out; through principal component analysis dimensionality reduction, the redundant dimensions of dynamic deformation features are reduced; and through adaptive selection of the fusion method, the effectiveness of feature fusion is improved.

[0131] Table 3 Comparison of state recognition accuracy using different feature fusion methods

[0132] direct splicing 0.95 0.72 0.81 0.96 0.86 Weighted fusion 0.94 0.78 0.85 0.95 0.88 Orthogonal fusion 0.96 0.83 0.89 0.97 0.91 Adaptive fusion 0.97 0.88 0.92 0.98 0.94

[0133] In this embodiment, Table 3 shows a comparison of the state recognition accuracy of four feature fusion methods: direct splicing, weighted fusion, orthogonal fusion, and adaptive fusion. As can be seen from Table 3, the adaptive fusion method achieves higher recognition accuracy than the other three methods in all mechanical states, with an average accuracy of 0.94. Particularly in the case of slight wear, the adaptive fusion method's recognition accuracy is 0.16 higher than the direct splicing method, indicating that the adaptive fusion method can more effectively extract feature information reflecting early mechanical degradation and improve the system's ability to identify early faults.

[0134] In one optional embodiment, the intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector switch operating mechanism further includes an online model calibration terminal. After each output of the trend prediction index of the mechanical characteristics, the online model calibration terminal records the measured mechanical response time of the operating mechanism during subsequent actions. The measured mechanical response time is the time elapsed from when the drive motor receives the opening or closing command until the transmission shaft of the operating mechanism completely stops rotating.

[0135] Calculate the residual between the measured mechanical response time and the predicted response time corresponding to the trend prediction index:

[0136]

[0137] Where e represents the residual value, Indicates the measured mechanical response time. This indicates the predicted response time.

[0138] When the residual value exceeds the preset residual tolerance range, an incremental sample set is constructed by extracting the historical time-frequency domain coupled feature vector within the preset time window before the current time and the historical measured mechanical response time. The preset residual tolerance range is set to [-0.1s, 0.1s], and the preset time window is set to include data from the most recent 10 to 20 actions.

[0139] The observation probability matrix of the Hidden Markov Model (HMM) is fine-tuned using stochastic gradient descent with an incremental sample set. During fine-tuning, the hidden state topology of the HMM is frozen, meaning the initial state probability vector and state transition probability matrix remain unchanged; only the parameters of the observation probability matrix are updated. The update formula for stochastic gradient descent is:

[0140]

[0141] in, This represents the observation probability matrix before the update. This represents the updated observation probability matrix. Indicates the learning rate. This represents the gradient of the loss function with respect to the observation probability matrix.

[0142] The loss function used is the mean squared error loss function:

[0143]

[0144] Where M represents the number of samples in the incremental sample set. Let B represent the predicted response time of the i-th sample obtained based on the observation probability matrix B. This represents the measured mechanical response time of the i-th sample.

[0145] Update the observation probability matrix parameters until the residual values ​​fall within the preset residual tolerance range, thus completing the online calibration of the model.

[0146] refer to Figure 2 In one optional embodiment, the intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector switch operating mechanism further includes a multi-source data spatiotemporal alignment terminal. The multi-source data spatiotemporal alignment terminal acquires a first original timestamp of the current signal of the drive motor and a second original timestamp of the vibration signal of the operating mechanism's transmission shaft. The first and second original timestamps are added by the current acquisition module and the vibration acquisition module, respectively, during data acquisition.

[0147] The clock offset is calculated based on the first and second original timestamps. The formula for calculating the clock offset is:

[0148]

[0149] in, This represents the clock offset, and K represents the number of synchronization events. This represents the timestamp of the vibration signal corresponding to the k-th synchronization event. This represents the timestamp of the current signal corresponding to the k-th synchronization event. The synchronization event is the moment the drive motor starts, which is manifested in the current signal as a sudden increase in current and in the vibration signal as a sudden increase in vibration.

[0150] The current and vibration signals are resampled and interpolated using clock offset to ensure strict alignment between them on the time axis. The resampling interpolation employs cubic spline interpolation to uniformly resample the current and vibration signals to the same sampling frequency.

[0151] The method for detecting abrupt spike interference in the aligned current and vibration signals is as follows: calculate the local standard deviation of the signal; when the difference between the amplitude of a certain sampling point and the amplitude of the adjacent sampling points exceeds three times the local standard deviation, the sampling point is determined to be abrupt spike interference.

[0152] A sliding median filter is used to smoothly replace abrupt spike interference. The window length of the sliding median filter is set to 5 to 9 sampling points, and the median value within the window is used as the replacement value for the abrupt spike interference sampling point. The smoothed replacement current signal and vibration signal are then transmitted to the time-frequency feature coupling extraction terminal.

[0153] This embodiment details the implementation process of the online model calibration end and the multi-source data spatiotemporal alignment end. Through multi-source data spatiotemporal alignment, strict synchronization of the current signal and vibration signal on the time axis is ensured; through online model calibration, drift caused by long-term model operation is eliminated, improving the model's prediction accuracy.

[0154] Table 4 Comparison of prediction errors before and after online model calibration

[0155] 1 0.03 0.02 3 0.06 0.03 6 0.12 0.04 12 0.21 0.05 24 0.35 0.06

[0156] In this embodiment, Table 4 shows the comparison of the average prediction error before and after online model calibration at different running times. As can be seen from Table 4, with the increase in running time, the average prediction error of the model without online calibration gradually increases, reaching 0.35s after 24 months of operation. In contrast, the average prediction error of the model with online calibration increases slowly, reaching only 0.06s after 24 months of operation. This indicates that online calibration can effectively suppress model drift and maintain the prediction accuracy of the model during long-term operation.

Claims

1. An intelligent monitoring system for the mechanical characteristics of the operating mechanism of a high-voltage disconnector, characterized in that, include: The multi-source signal synchronous acquisition terminal is used to synchronously acquire the current signal of the drive motor and the vibration signal of the transmission shaft of the operating mechanism; The time-frequency feature coupling extraction end is used to decompose the vibration signal into intrinsic mode functions and extract marginal spectral features by using adaptive noise-assisted ensemble empirical mode decomposition, extract the dynamic deformation features of the current signal by using dynamic time warping calculation, and orthogonally fuse the marginal spectral features and the dynamic deformation features in the feature space to construct a time-frequency domain coupled feature vector. The state evolution prediction end is used to input the time-frequency domain coupled feature vector into the hidden Markov model based on the dynamic correction of the state transition probability matrix, drive the forward-backward calculation process with the observation sequence at the current moment to calculate the probability distribution of each hidden state, extract the slope of the maximum probability hidden state as time evolves, and output the trend prediction index of mechanical characteristics.

2. The intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector operating mechanism according to claim 1, characterized in that, The vibration signal is decomposed into intrinsic mode functions and marginal spectral features are extracted using adaptive noise-assisted ensemble empirical mode decomposition, including: The vibration signal is repeatedly added with white noise sequences of different amplitudes, and the vibration signal after adding white noise sequences is subjected to empirical mode decomposition to obtain a set of candidate intrinsic mode functions. Calculate the correlation coefficient between adjacent candidate intrinsic mode functions in the candidate intrinsic mode function set, determine the candidate intrinsic mode functions with the correlation coefficient less than a preset correlation threshold as noise components and remove them, and determine the remaining candidate intrinsic mode functions as the intrinsic mode functions; The intrinsic mode function is subjected to Hilbert transform to obtain an instantaneous frequency sequence. A time spectrum is constructed based on the instantaneous frequency sequence and the amplitude of the intrinsic mode function. The marginal spectral features are obtained by integrating the time spectrum along the time axis.

3. The intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector operating mechanism according to claim 1, characterized in that, The dynamic deformation characteristics of the current signal extracted using dynamic time warping calculations include: Construct a standard reference current waveform template for the current signal; The current signal and the standard reference current waveform template are used as the test sequence and the reference sequence, respectively. The Euclidean distance between each sampling point in the test sequence and each sampling point in the reference sequence is calculated, and a distance cost matrix is ​​constructed. Based on the distance cost matrix, search for the optimal regularized path that minimizes the cumulative distance between the test sequence and the reference sequence. Extract the local slope and local curvature at each sampling point along the optimal regularized path. Arrange the local slope and local curvature in chronological order to form the dynamic deformation feature.

4. The intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector operating mechanism according to claim 1, characterized in that, The process of orthogonally fusing the marginal spectral features and the dynamic deformation features in the feature space to construct a time-frequency domain coupled feature vector includes: The marginal spectral features and the dynamic deformation features are respectively subjected to zero-mean processing to obtain the centered marginal spectral feature vector and the centered dynamic deformation feature vector; Calculate the covariance matrix between the centered marginal spectrum eigenvector and the centered dynamic deformation eigenvector, and perform eigenvalue decomposition on the covariance matrix to obtain the orthogonal projection matrix; The orthogonal projection matrix is ​​used to map the centered marginal spectral feature vector and the centered dynamic deformation feature vector to the same orthogonal feature subspace. The mapped vectors are then concatenated in the orthogonal feature subspace to generate the time-frequency domain coupled feature vector.

5. The intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector operating mechanism according to claim 1, characterized in that, Hidden Markov models based on dynamic correction of the state transition probability matrix include: Initialize the initial state probability vector, initial state transition probability matrix, and observation probability matrix of the Hidden Markov Model; The time-frequency domain coupled feature vector is used as the observation sequence and input into the hidden Markov model. The forward-backward computation process is used to calculate the posterior probability distribution of each hidden state under the observation sequence. Based on the time label of the posterior probability distribution and the time-frequency domain coupled feature vector, the frequency of hidden state transitions within adjacent time steps is counted. The initial state transition probability matrix is ​​recursively updated and replaced using the hidden state transition frequency. The updated state transition probability matrix is ​​substituted into the forward-backward calculation process and iterated until convergence.

6. The intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector operating mechanism according to claim 2, characterized in that, Constructing a time-spectrum based on the instantaneous frequency sequence and the amplitude of the intrinsic mode function, and obtaining the marginal spectral features by integrating the time-spectrum along the time axis includes: Define frequency band division intervals, perform frequency band clustering on the instantaneous frequency sequence, merge the instantaneous frequency sequences whose cluster center frequencies fall within the same frequency band division interval and their corresponding intrinsic mode function amplitudes, and obtain the frequency band reconstruction spectrum; Calculate the energy entropy value corresponding to each frequency band component in the spectrum during the frequency band reconstruction, and determine the frequency band components whose energy entropy value is less than a preset entropy threshold as aliasing interference frequency bands and filter them out. When reconstructing the frequency band after filtering out aliasing interference, the spectrum is integrated along the time axis to extract the marginal spectral peak, marginal spectral variance, and marginal spectral skewness of the retained frequency band. The marginal spectral peak, marginal spectral variance, and marginal spectral skewness are combined to form the marginal spectral features.

7. The intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector operating mechanism according to claim 3, characterized in that, Extracting the local slope and local curvature at each sampling point along the optimal regularized path, and arranging the local slope and local curvature in chronological order to form the dynamic deformation feature includes: The optimal regular path is segmented according to a fixed time window to obtain multiple locally regular path segments. For each segment of the locally regular path, calculate the mean and standard deviation of the local slope and local curvature, and construct a local path statistical feature vector. All local path statistical feature vectors are combined in chronological order to generate an initial dynamic deformation feature matrix. Principal component analysis is performed on the initial dynamic deformation feature matrix, and feature vectors with a cumulative contribution rate exceeding a preset threshold are selected to form a dimension-reduced projection matrix. The initial dynamic deformation feature matrix is ​​multiplied by the dimension-reduced projection matrix to obtain the dynamic deformation features.

8. The intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector operating mechanism according to claim 4, characterized in that, In the orthogonal feature subspace, the mapped vectors are concatenated to generate the time-frequency domain coupled feature vector, which includes: Calculate the Pearson correlation coefficient between the marginal spectral eigenvector and the dynamic deformation eigenvector after mapping to the orthogonal eigenspace; When the Pearson correlation coefficient exceeds the preset correlation threshold, a weighted fusion is performed on the mapped marginal spectral feature vector and the dynamic deformation feature vector, and the single vector after weighted fusion is determined as the time-frequency domain coupling feature vector. When the Pearson correlation coefficient does not exceed the preset correlation threshold, the mapped marginal spectral feature vector and the dynamic deformation feature vector are concatenated and spliced ​​together, and the spliced ​​vector is subjected to manifold regularization processing. The vector after manifold regularization processing is determined as the time-frequency domain coupled feature vector.

9. The intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector operating mechanism according to claim 1, characterized in that, It also includes an online model calibration terminal, which is used for: After each output of the trend prediction index of the mechanical characteristics, the measured mechanical response time of the operating mechanism when performing subsequent actions is recorded. Calculate the residual value between the measured mechanical response time and the predicted response time corresponding to the trend prediction index. When the residual value exceeds the preset residual tolerance range, extract the historical time-frequency domain coupled feature vector within the preset time window before the current moment and the historical measured mechanical response time to construct an incremental sample set. The observation probability matrix of the Hidden Markov Model is fine-tuned using stochastic gradient descent with the incremental sample set. During the fine-tuning process, the hidden state topology of the Hidden Markov Model is frozen, and the parameters of the observation probability matrix are updated until the residual value falls into the preset residual tolerance interval.

10. The intelligent monitoring system for the mechanical characteristics of the high-voltage disconnector operating mechanism according to claim 1, characterized in that, It also includes a multi-source data spatiotemporal alignment endpoint, which is used for: The first original timestamp of the current signal of the drive motor and the second original timestamp of the vibration signal of the transmission shaft of the operating mechanism are obtained. The clock offset is calculated based on the first original timestamp and the second original timestamp. The clock offset is used to resample and interpolate the current signal and the vibration signal so that the current signal and the vibration signal are strictly aligned on the time axis. The sudden spike interference in the aligned current signal and vibration signal is detected, and the sudden spike interference is smoothly replaced by a sliding median filter. The smoothed current signal and vibration signal are then transmitted to the time-frequency feature coupling extraction end.

Images