Cable hazard assessment and prediction method based on natural time domain and fuzzy rough set

Abstract

The application discloses a wiring hidden danger evaluation and prediction method based on natural time domain and fuzzy rough set, and belongs to the technical field of power grid equipment operation and maintenance and overhaul. The method is based on collected multi-source monitoring signals such as leakage current, induced current and temperature, adopts a method combining empirical mode decomposition (EMD) and sample entropy to perform adaptive signal preprocessing, effectively extracts effective modal components and suppresses noise interference; natural time domain analysis is introduced on the basis of traditional time domain, event sequences are constructed, dynamic characteristics are extracted, and a hidden danger characteristic data set with time sequence evolution information is formed; further, the fuzzy rough set theory is used to perform unsupervised attribute reduction on high-dimensional characteristics, remove redundant information and reserve key discriminant characteristics; finally, a support vector machine (SVM) classifier is used to realize identification and trend prediction of wiring hidden danger types. The method improves the accuracy and robustness of hidden danger identification, and provides effective technical support for intelligent operation and maintenance of power wiring.

Description

Technical Field

[0001] This invention belongs to the field of power system equipment operation and maintenance and intelligent diagnosis technology. Specifically, it relates to a method for assessing and predicting wiring hazards based on natural time domain and fuzzy rough set. It is particularly suitable for application scenarios of signal processing, feature extraction, feature reduction and intelligent classification and identification of early fault hazards in power wiring under complex operating environments. Background Technology

[0002] As a critical link in power transmission within a power system, the operational status of power distribution lines directly impacts the security and stability of the power grid. With the development of smart grids, higher demands are placed on the monitoring of the operational status and fault early warning capabilities of distribution equipment. Traditional methods for detecting potential distribution line hazards mainly rely on manual inspections and simple threshold-based alarm mechanisms, which are insufficient for accurately identifying early-stage faults and predicting trends. In recent years, with the advancements in sensor technology, signal processing, and artificial intelligence algorithms, intelligent assessment methods based on multi-source monitoring data have gradually become a research hotspot.

[0003] In existing technologies, researchers have attempted to use wavelet transform, Fourier transform, and other methods to denoise and extract features from acquired signals such as current, voltage, and temperature, and to combine these with machine learning algorithms for fault classification and identification. However, the following key technical challenges remain in practical applications:

[0004] First, the operating environment of power distribution lines is complex, and the acquired signals are often subject to strong noise interference, which can easily mask effective hidden danger signals. Existing signal processing methods generally rely on preset basis functions or fixed parameters, making it difficult to adaptively process nonlinear and non-stationary signals, thus limiting the denoising effect and affecting the accuracy of subsequent feature extraction.

[0005] Secondly, the development of wiring hazards exhibits significant dynamism and time-varying characteristics, often undergoing a long accumulation process from early subtle anomalies to eventual failure. Traditional time-domain feature extraction methods struggle to fully characterize the evolution trend of hazards and lack sensitivity to changes in the system's critical state, resulting in insufficient early hazard identification capabilities.

[0006] Furthermore, multi-source monitoring data is characterized by high dimensionality, redundancy, and heterogeneity, which can easily lead to the "curse of dimensionality" during feature engineering, increasing model complexity and causing overfitting, thus affecting the generalization ability of the prediction model. Existing feature selection methods mostly rely on supervised learning or prior knowledge, making it difficult to adapt to the actual scenarios of unlabeled or weakly labeled data in power systems.

[0007] In summary, existing technologies still have significant shortcomings in signal preprocessing, dynamic feature extraction, and high-dimensional feature reduction. There is an urgent need for a wiring hazard assessment and prediction method that can adaptively process non-stationary signals, fuse multi-domain features, and possess robustness and generalization capabilities. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes a method for assessing and predicting wiring hazards based on the natural time domain and fuzzy rough sets. This method utilizes various monitoring data, including partial discharge, temperature, vibration, voltage, and current. It employs the EMD non-stationary signal processing algorithm to preprocess the signals, suppressing interference and noise based on sample entropy. Combining the traditional time domain and the natural time domain, it extracts a feature dataset of wiring hazards. Fuzzy rough sets are then used to reduce the attributes of the high-dimensional feature dataset. Finally, a classifier is used to predict the unknown relationships between entities.

[0009] The technical solution adopted in this invention is as follows:

[0010] A method for assessing and predicting wiring hazards based on natural time domain and fuzzy rough sets includes the following steps:

[0011] Step 1, Multi-source monitoring data acquisition: Collect various monitoring signals during the operation of power distribution lines. These signals include partial discharge, temperature, vibration, voltage, and current.

[0012] Step 2, Signal preprocessing based on EMD and sample entropy: For the acquired non-stationary monitoring signals, an adaptive denoising process is performed by combining empirical mode decomposition (EMD) and sample entropy analysis to extract the effective modal components in each signal and suppress noise interference.

[0013] Step 3, Feature extraction by integrating traditional time domain and natural time domain: Based on the effective signal modal components obtained from preprocessing, extract statistical features of traditional time domain, construct event sequences of natural time domain and calculate their dynamic features, and finally integrate the two types of features to form a wiring hazard feature dataset with dynamic representation capabilities.

[0014] Step 4: High-dimensional feature attribute reduction based on fuzzy rough set: For the constructed high-dimensional hidden danger feature dataset that integrates traditional time domain and natural time domain, fuzzy rough set theory is used to perform unsupervised attribute reduction, remove redundant features, and retain key features with high discriminative ability for hidden danger identification.

[0015] Step 5: Hazard classification and prediction based on support vector machine: Based on the low-dimensional feature set after attribute reduction, a support vector machine (SVM) classification model is constructed, and combined with parameter optimization strategies, to accurately classify and predict the trends of different types of hazards in power wiring.

[0016] Furthermore, step 2 includes the following steps:

[0017] Step 2.1, Signal segmentation: The non-stationary time-series signal is reasonably segmented according to the sampling rate. Each segment contains at least n typical hidden danger signal cycles and the number of sampling points does not exceed the limit.

[0018] Step 2.2, EMD mode decomposition: Empirical mode decomposition (EMD) is used to adaptively decompose the segmented signal, extract the intrinsic mode function components that satisfy the IMF conditions in sequence, and finally obtain the residual terms to complete the multi-level characterization of the non-stationary signal.

[0019] Step 2.3, Noise Suppression Based on Sample Entropy: By calculating the sample entropy of each IMF component and constructing a separation exponential function, combined with signal energy and complexity information, noise components are adaptively identified and suppressed, while retaining effective signal modes with potential hazards.

[0020] Furthermore, step 2.2, the EMD mode decomposition process, includes the following steps:

[0021] First, let's define the signal margin. The algorithm iterates through all local extrema of the residual signal, then uses cubic spline interpolation to fit the local maxima and minima, constructs the upper and lower envelopes, and calculates the average of the envelope values. ;

[0022] Then, calculate If it satisfies the intrinsic mode function (IMF) condition, then let If the condition is not met, repeat the above steps until the IMF condition is met; then calculate the residual component. Repeat the above steps until the residual component is reached. It is a monotonic function or a constant;

[0023] This is the original non-stationary time-series signal. Then it can be expressed as:

[0024] ;

[0025] In the formula, The total number of extracted IMF components is determined by the complexity of the original non-stationary time series signal, and is usually 5 to 10 layers. The sequence number of the IMF component, from Initially, high-frequency components are extracted layer by layer, from low-frequency components to high-frequency components.

[0026] Furthermore, in step 2.3, during the noise suppression process based on sample entropy, the steps for calculating the sample entropy of each IMF component for each intrinsic mode component are as follows:

[0027] Vector reconstruction: converting the original time series Reconstructed Dimensional vector:

[0028] ;

[0029] In the formula, for The dimensional reconstruction vector represents the reconstructed vector from... The beginning of the continuous A vector composed of points; The original time series represents the first... The signal values ​​at each sampling point; , jointly constructed indivual dimensional vector; The time series length represents the total number of sampling points.

[0030] Calculate the distance between vectors: for any two vectors and , Calculate the maximum absolute difference between them:

[0031] ;

[0032] In the formula, Let be the distance between vector pairs, representing the distance between the first and second vector pairs. The and the first indivual The maximum absolute difference between dimensional vectors; Let be the dimension of the vector, representing the dimension of the reconstructed vector;

[0033] Logarithm of similar vectors: given a similarity threshold Statistically count all that satisfy Count the number of vector pairs and calculate:

[0034] ;

[0035] In the formula, For dimension The proportion of similar vector pairs in a time series measures its performance. Regularity under dimensions;

[0036] Add dimensions to Expand the dimension of the vector to Calculate the distance between new vector pairs and statistically analyze the results. By counting the number of vector pairs, we get:

[0037] ;

[0038] In the formula, For dimension The proportion of similar vector pairs measures the regularity of time series changes in higher dimensions;

[0039] Calculate sample entropy:

[0040] ;

[0041] In the formula, The sample entropy measures the complexity of a time series; the larger the value, the more complex the series. For dimension The proportion of similar vector pairs measures the regularity of time series changes in higher dimensions.

[0042] Furthermore, in the noise suppression process based on sample entropy, a modal component separation exponential function is constructed. The steps are as follows:

[0043] Based on sample entropy and combined with signal energy, a comprehensive index is constructed to distinguish effective signals from noise; the separation index is proportional to the product of signal energy and sample entropy, expressed as:

[0044] ;

[0045] In the formula, For the j-th layer IMF The time separation index, which combines signal energy and entropy information, is used to identify valid signals and noise. For the first Layer IMF in the 1st The energy at a given moment represents a measure of the signal energy, which is the square of the signal amplitude. For the first Layer IMF in the 1st Sample entropy at time t;

[0046] Noise suppression logic: Set a threshold;

[0047] like If the threshold is reached, then the point is a valid signal;

[0048] like If the value is less than or equal to the threshold, then the point is considered noise and should be suppressed.

[0049] Furthermore, in step 3, during the feature extraction process that integrates the traditional time domain and the natural time domain, the traditional time domain feature extraction process is as follows:

[0050] For each decomposed modal component signal, time-domain features are extracted according to the segmented signal, including maximum value, minimum value, mean, absolute mean, standard deviation, variance, and root mean square; at the same time, peak factor, impulse factor, margin factor, and waveform factor are calculated.

[0051] The formula for calculating the peak factor is:

[0052] ;

[0053] The formula for calculating the impulse factor is:

[0054] ;

[0055] The formula for calculating the margin factor is:

[0056] ;

[0057] The formula for calculating the waveform factor is:

[0058] ;

[0059] In the formula, The original time series represents the first... The signal values ​​at each sampling point; This represents the total number of sampling points; The maximum value of the signal, the maximum amplitude value in this segment of the signal. ; The minimum value of the signal, the minimum amplitude in this segment of the signal. .

[0060] Furthermore, in step 3, during the feature extraction process of integrating the traditional time domain and the natural time domain, the process of constructing the natural time domain event sequence is as follows:

[0061] Construction of natural time-domain event sequences:

[0062] Define event energy, which includes: the pulse generated by partial discharge, the magnitude of temperature rise, the magnitude of voltage drop / rise, the change in signal harmonic THD, and the instantaneous frequency change of the signal;

[0063] After normalization, the event sequence is constructed: , , ..., ;

[0064] In the formula, For the first The time of occurrence of each event; No. The energy value of an event; Event number , This represents the total number of events.

[0065] Assign sequence numbers to the events and calculate the normalized energy:

[0066] Starting with the first event, assign an incrementing integer number to each event. Thus, the dataset is obtained. ;

[0067] Then the first An event is defined in the natural time domain as:

[0068] ;

[0069] Calculate the normalized energy:

[0070] ;

[0071] In the formula, The event sequence number; For the first The natural time value of an event; For the first The energy value of an event; No. Normalized energy of an event.

[0072] Furthermore, in step 3, during the feature extraction process that integrates the traditional time domain and the natural time domain, the calculation of its dynamic features is as follows:

[0073] Natural time-domain feature acquisition:

[0074] Extracting dynamic features, including: entropy, a measure of system disorder. Inverse entropy, reflecting time symmetry Variance describing the degree of dispersion of energy distribution ;

[0075] variance The calculation formula is as follows:

[0076] ;

[0077] In the formula, Variance represents the degree of dispersion in the energy distribution of events and reflects the volatility of the system state. Let be the first moment of natural time, representing the center position of the event sequence; It is a weighted average of the squares of natural time, which measures the spatial diffusion of the energy distribution of an event;

[0078] entropy The calculation formula is as follows:

[0079] ;

[0080] In the formula, Entropy in the natural time domain is used to measure the uncertainty and complexity of an event sequence. The weighted natural time logarithm mean describes the complexity of the product of event energy and time.

[0081] Entropy The calculation formula is as follows:

[0082] ;

[0083] ;

[0084] In the formula, The normalized energy after time reversal is the energy of the original event sequence rearranged from back to front; The event number is the sequence number after time reversal, indicating the original sequence number when the events are arranged from back to front. The new location of the event; The natural time value after time reversal is the event timeline mapped in reverse to form a time-symmetric sequence; The natural time value after time reversal represents the value of the first natural time series after the original natural time series is reversed. The natural time value at each location.

[0085] Further, step 4: High-dimensional feature attribute reduction based on fuzzy rough sets includes the following steps:

[0086] Step 4.1, Information System:

[0087] Treat all data as a collection of objects It takes all features of traditional time and natural time as a set of attributes. , build matrix;

[0088] Step 4.2, data normalization and fuzzification:

[0089] The min-max normalization method is used to normalize all attributes. The normalization calculation formula is as follows:

[0090] ;

[0091] In the formula, Normalized eigenvalues; These are the original eigenvalues; The minimum value for all objects under this property; The maximum value of all objects under this property;

[0092] Constructing a fuzzy similarity function:

[0093] ;

[0094] In the formula, For object and In attributes Fuzzy similarity on; , Represents any two objects in the dataset; For two objects in properties The distance below; is a scale parameter that controls the rate at which the similarity function decays; it is a positive real number.

[0095] Finally, the fuzzy similarity matrix corresponding to each attribute is obtained;

[0096] Step 4.3, define fuzzy dependency:

[0097] First, define the fuzzy neighborhood. :

[0098] ;

[0099] In the formula, Representation Object With object In attribute subset Minimum fuzzy similarity; For object and In attributes Fuzzy similarity;

[0100] Then, calculate the fuzzy dependency between attribute subsets. :

[0101] ;

[0102] In the formula, Represents a subset of attributes For attribute subsets The degree of dependence; the larger the value, the stronger the dependence. For object In attribute subset The fuzzy neighborhood below; For object In attribute subset The fuzzy neighborhood below;

[0103] Step 4.4, Attribute Reduction Algorithm:

[0104] First, initialize the reduced set. ;

[0105] Then, for each attribute a∈A\R that is not in the reduced attribute set R, calculate the contribution of adding it to R to the fuzzy dependency.

[0106] Then, select the attribute a* that contributes the most and add it to the reduced attribute set R until the dependency of the remaining attributes is lower than the set threshold.

[0107] Finally, the reduced property set is output.

[0108] Further, step 5: Hazard classification and prediction based on support vector machines, including the following steps:

[0109] Step 5.1, Normalize and reduce the attribute set:

[0110] Mini-maximum normalization is applied to the reduced features;

[0111] Step 5.2, One-to-many SVM training strategy:

[0112] A binary classification SVM is trained for each hazard category. During training, each SVM labels all samples of a specific category as positive and all samples of the other categories as negative. The final prediction mechanism is as follows:

[0113] ;

[0114] In the formula, The input feature vector; This indicates that the category with the highest score is selected as the prediction result; For the first The decision function of an SVM classifier represents the input feature vector. Belongs to the Confidence level or score of the type of potential hazard;

[0115] Step 5.3, Parameter Optimization:

[0116] Optimize classification parameters using soft margin:

[0117] ;

[0118] In the formula, Indicates the direction of the classification hyperplane; The penalty coefficient controls the trade-off between the cost of misclassification and the classification margin; These are slack variables; For the first The label of each sample, This indicates whether the sample belongs to the positive or negative class; This is a feature mapping function that maps the original input space to a high-dimensional feature space. The bias term controls the position of the classification hyperplane; Indicates the index for all samples That is, the constraints apply to every sample in the training set;

[0119] Radial basis function (RBF) kernels are used to handle nonlinear separable problems:

[0120] ;

[0121] In the formula, The kernel function output value represents the similarity between two samples in a high-dimensional space; These are kernel function parameters that control the range of influence of a single sample on the decision boundary; , Given two input samples, representing the first... and the One training sample; It is the square of the Euclidean distance between the feature vectors of two samples;

[0122] Then, the parameters were tuned using a combination of random search and K-fold cross-validation to find the optimal penalty coefficient. and kernel function parameters combination;

[0123] Finally, a predictive evaluation model is used to process the test data.

[0124] The beneficial effects of this invention are:

[0125] The present invention proposes a method for assessing and predicting potential hazards in power distribution lines based on natural time domain and fuzzy rough sets. This method has significant technical advantages and application value in monitoring the operating status and identifying potential hazards in power distribution lines, specifically in the following aspects:

[0126] 1. Strong signal preprocessing capability and significant denoising effect: The method of combining empirical mode decomposition (EMD) with sample entropy is used to adaptively decompose and suppress noise in the acquired non-stationary signals, effectively extracting effective mode components with hidden danger characteristics, improving the signal-to-noise ratio and the accuracy of subsequent feature extraction.

[0127] 2. Comprehensive extraction of hidden danger features and strong dynamic characterization ability: Based on the traditional time domain, natural time domain analysis methods are introduced to construct event sequences and extract dynamic features such as entropy, inversion entropy, and variance. This enables more sensitive capture of the critical process of power distribution hidden dangers evolving from a stable state to a fault state, and enhances the ability to identify early weak anomalies.

[0128] 3. Reasonable feature dimensionality compression and accurate and efficient attribute selection: The fuzzy rough set theory is used to perform unsupervised attribute reduction on high-dimensional heterogeneous features. While retaining key discriminative information, redundant features are effectively removed, which improves the robustness of the feature set and the generalization ability of the classification model.

[0129] 4. Excellent classification and prediction performance and strong applicability: Based on the support vector machine (SVM) to build a multi-classification prediction model and combined with parameter optimization strategies, it can accurately identify and predict the trend of various wiring hazards, and has good engineering applicability and promotion prospects.

[0130] In summary, this invention, by integrating natural time domain analysis and fuzzy rough set theory, constructs a complete system for assessing and predicting potential hazards in power distribution lines. This provides a scientific basis and technical support for the intelligent operation and maintenance and condition-based inspection of power equipment, and has significant theoretical and practical application value. Attached Figure Description

[0131] 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.

[0132] Figure 1 This is a flowchart of the wiring hazard assessment and prediction method based on natural time domain and fuzzy rough set of the present invention;

[0133] Figure 2 Flowchart of EMD signal decomposition;

[0134] Figure 3 A feature parameter map within the natural time domain;

[0135] Figure 4 This is a flowchart of the hazard assessment process based on SVM. Detailed Implementation

[0136] 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, and 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.

[0137] To address the challenges of efficient assessment and accurate prediction of potential hazards in power distribution lines, and to overcome issues such as severe signal and noise interference, incomplete hazard feature extraction, and feature redundancy affecting classification accuracy in existing technologies, this embodiment proposes a method for assessing and predicting power distribution line hazards based on natural time domain and fuzzy rough sets. This method utilizes various monitoring data, including partial discharge, temperature, vibration, voltage, and current. It employs the EMD non-stationary signal processing algorithm for signal preprocessing, suppressing interference and noise based on sample entropy. Combining traditional and natural time domain methods, it extracts a feature dataset of power distribution line hazards. Fuzzy rough sets are used to reduce the attributes of the high-dimensional feature dataset. Finally, a classifier is used to predict unknown relationships between entities.

[0138] Specifically, such as Figure 1 As shown, the wiring hazard assessment and prediction method based on natural time domain and fuzzy rough sets includes the following steps:

[0139] Step 1: Multi-source monitoring data acquisition:

[0140] This step aims to collect various monitoring signals during the operation of power distribution lines, serving as the raw data source for subsequent signal processing, feature extraction, and hazard assessment.

[0141] The collected signals include, but are not limited to, the following types:

[0142] Partial discharge signal, temperature signal, vibration signal, voltage signal, current signal.

[0143] The aforementioned signals reflect the electrical characteristics and physical changes of the wiring under different operating conditions, exhibiting typical features such as nonlinearity, non-stationarity, and multi-source heterogeneity. To ensure data integrity and subsequent algorithm processing efficiency, the acquired time-series signals are reasonably segmented according to the system sampling rate. Each signal segment should contain at least 10 complete cycles of typical wiring hazard signals to ensure sufficient preservation of characteristic information; simultaneously, to balance processing speed and computing resources, the number of sampling points in each signal segment does not exceed 2048 points.

[0144] This step provides a high-quality data foundation for subsequent signal preprocessing based on EMD and sample entropy, multi-domain feature extraction, and hazard classification and prediction.

[0145] Step 2, Signal preprocessing based on EMD and sample entropy:

[0146] This step aims to perform adaptive denoising on the non-stationary monitoring signals acquired in Step 1, extracting the effective modal components from each signal and suppressing noise interference, thereby improving signal quality and providing a high signal-to-noise ratio data foundation for subsequent feature extraction. For example... Figure 2 As shown, this step includes the following three sub-steps: signal segmentation processing, EMD mode decomposition, and noise suppression based on sample entropy.

[0147] Step 2.1, Signal segmentation processing:

[0148] The non-stationary monitoring signals collected in Step 1, such as leakage current, induced current, and temperature, reflect the electrical characteristics and physical changes of the wiring under different operating conditions, exhibiting typical characteristics such as nonlinearity, non-stationarity, and multi-source heterogeneity. To ensure data integrity and subsequent algorithm processing efficiency, the collected time-series signals are reasonably segmented according to the system sampling rate. Each signal segment should contain at least 10 complete cycles of typical wiring hazard signals to ensure sufficient preservation of characteristic information; at the same time, to balance processing speed and computing resources, the number of sampling points in each signal segment does not exceed 2048 points.

[0149] Step 2.2, EMD mode decomposition:

[0150] Empirical Mode Decomposition (EMD) is used to adaptively decompose the segmented signal, sequentially extracting modal components that satisfy the Intrinsic Mode Function (IMF) conditions, and finally obtaining the residual term, thereby achieving a multi-level characterization of non-stationary signals.

[0151] The specific process is as follows:

[0152] First, let's define the signal margin. The algorithm iterates through all local extrema of the residual signal and then uses cubic spline interpolation to fit the local maxima and minima, constructing the upper envelope. and lower envelope ;

[0153] Calculate the envelope mean :

[0154] ;

[0155] Then, the envelope mean is subtracted from the original non-stationary time series signal to obtain the signal h. i (t):

[0156] ;

[0157] If it satisfies the intrinsic mode function (IMF) condition, then let If the condition is not met, repeat the above steps until the IMF condition is met; then calculate the residual component. Repeat the above steps until the residual component is reached. It is a monotonic function or a constant;

[0158] This is the original non-stationary time-series signal. Then it can be expressed as:

[0159] ;

[0160] In the formula, The total number of extracted IMF components is determined by the complexity of the original non-stationary time series signal, and is usually 5 to 10 layers. The sequence number of the IMF component, from Initially, high-frequency components are extracted layer by layer, from low-frequency components to high-frequency components.

[0161] The EMD method does not require preset basis functions and is entirely data-driven, making it particularly suitable for handling the nonlinear and non-stationary characteristics of power line fault signals.

[0162] Step 2.3, Noise suppression based on sample entropy:

[0163] Based on the IMF components obtained from EMD decomposition, the signal complexity is further measured using sample entropy, and a separation exponential function is constructed by combining signal energy to achieve adaptive identification and suppression of noise components. The specific steps are as follows:

[0164] Vector reconstruction: converting the time series of each IMF component Reconstructed Dimensional vector:

[0165] ;

[0166] In the formula, for The dimensional reconstruction vector represents the reconstructed vector from... The beginning of the continuous A vector composed of points; the original non-stationary time-series signal The signal is a continuous-time signal. After EMD decomposition, multiple IMF components and residual terms are obtained. Each IMF component is a locally oscillating mode with physical meaning, and it is also a time series in itself. In this step, each IMF component is discretely sampled to obtain a discrete-time series. , This is the original time series, representing the first... The signal values ​​at each sampling point; , jointly constructed indivual dimensional vector; The time series length represents the total number of sampling points. To reconstruct the dimension, the value is usually 2 or 3.

[0167] Calculate the distance between vectors: for any two dimensional vector and , Calculate the maximum absolute difference between them:

[0168] ;

[0169] In the formula, Let be the distance between vector pairs, representing the distance between the first and second vector pairs. The and the first indivual The maximum absolute difference between dimensional vectors.

[0170] Logarithm of similar vectors: given a similarity threshold Statistically count all that satisfy Count the number of vector pairs and calculate:

[0171] ;

[0172] In the formula, For dimension The proportion of similar vector pairs in a time series measures its performance. Regularity under dimensions.

[0173] Expand to Dimensionality and similarity logarithm:

[0174] ;

[0175] In the formula, For dimension The proportion of similar vector pairs measures the regularity of time series changes in higher dimensions.

[0176] Calculate sample entropy:

[0177] ;

[0178] In the formula, The sample entropy measures the complexity of a time series; the larger the value, the more complex the series. For dimension The proportion of similar vector pairs measures the regularity of time series changes in higher dimensions.

[0179] Constructing the modal component separation exponential function :

[0180] Based on sample entropy and combined with signal energy, a comprehensive index is constructed to distinguish effective signals from noise; the separation index is proportional to the product of signal energy and sample entropy, expressed as:

[0181] ;

[0182] In the formula, For the j-th layer IMF The time separation index, which combines signal energy and entropy information, is used to identify valid signals and noise. For the first Layer IMF in the 1st The energy at a given moment represents a measure of the signal energy, which is the square of the signal amplitude. For the first Layer IMF in the 1st Sample entropy at time t;

[0183] Noise suppression logic: Set a threshold;

[0184] like If the threshold is reached, then the point is a valid signal;

[0185] like If the value is less than or equal to the threshold, then the point is considered noise and should be suppressed.

[0186] Through the above steps, noise components in non-stationary signals can be efficiently identified and suppressed, effective modal components with potential hazards can be preserved, and signal quality can be significantly improved.

[0187] Step 3: Feature extraction by integrating traditional time domain and natural time domain:

[0188] This step aims to combine traditional time-domain and natural time-domain analysis methods to extract representative hidden danger features from the modal components of the preprocessed effective signal, and construct a feature dataset that integrates information from both domains. This step includes the following three main parts: traditional time-domain feature extraction, natural time-domain event sequence construction, and natural time-domain dynamic feature extraction.

[0189] Step 3.1, Traditional time-domain feature extraction:

[0190] For the obtained decomposed modal component signals, time-domain features are extracted according to the segmented signals, including maximum value, minimum value, mean, absolute mean, standard deviation, variance, root mean square, etc. At the same time, the peak factor, impulse factor, margin factor, and waveform factor are calculated to sensitively reflect the characteristics of waveform changes.

[0191] Crest factor: The crest factor can be used to evaluate the morphological characteristics of a waveform. A high crest factor indicates that the waveform has obvious peaks; conversely, a low crest factor indicates that the waveform may be more stable without obvious peaks. The calculation formula is as follows:

[0192]

[0193] Pulse Factor: The pulse factor can be used to evaluate the shape of pulse signals in a waveform. A higher pulse factor indicates a more abrupt change in waveform amplitude; conversely, a lower pulse factor indicates a more gradual change in waveform amplitude. The calculation formula is as follows:

[0194]

[0195] Margin Factor: The margin factor can be used to assess the stability of a waveform. A higher margin factor indicates larger amplitude fluctuations and instability; conversely, a lower margin factor indicates smaller amplitude fluctuations and relative stability. The calculation formula is as follows:

[0196]

[0197] Waveform factor: The waveform factor can be used to evaluate the quality of a waveform. A higher waveform factor value indicates a higher signal-to-noise ratio (SNR); conversely, a lower waveform factor value indicates a lower SNR. The calculation formula is as follows:

[0198]

[0199] In the formula, The original time series represents the first... The signal values ​​at each sampling point; This represents the total number of sampling points; The maximum value of the signal, the maximum amplitude value in this segment of the signal. ; The minimum value of the signal, the minimum amplitude in this segment of the signal. .

[0200] Step 3.2, construct the natural time domain event sequence:

[0201] The natural time domain is suitable for describing the dynamic evolution of aperiodic and nonlinear systems. This step characterizes the evolution of wiring hazards by constructing an event sequence. The specific steps are as follows:

[0202] Event energy is defined as follows: Event energy includes partial discharge pulse amplitude, temperature change, voltage surge amplitude, harmonic distortion rate (THD) change, and instantaneous frequency change.

[0203] Normalization: Normalize the event energy and construct the event sequence.

[0204] , , ..., ;

[0205] In the formula, For the first The time of occurrence of each event; No. The energy value of an event; Event number , This represents the total number of events.

[0206] Assign sequence numbers to the events and calculate the normalized energy:

[0207] Starting with the first event, assign an incrementing integer number to each event. Thus, the dataset is obtained. ;

[0208] Then the first An event is defined in the natural time domain as:

[0209] ;

[0210] Calculate the normalized energy:

[0211] ;

[0212] In the formula, The event sequence number; For the first The natural time value of an event; For the first The energy value of an event; No. Normalized energy of an event.

[0213] Through the above steps, the original time series is mapped to the natural time domain, constructing an event sequence with time evolution characteristics, laying the foundation for dynamic feature extraction.

[0214] Step 3.3, then extract dynamic features in the time domain.

[0215] However, in the time domain, entropy Reverse entropy and variance Characteristic values ​​and their changes can reveal the internal change patterns of a system and the system's disorder or uncertainty; such as Figure 3 As shown, entropy Reverse entropy and variance Equivalent characteristic value combination curves can characterize the critical time region where the system state undergoes drastic changes.

[0216] Calculate based on the previous step. variance The calculation is as follows:

[0217]

[0218] In the formula, Variance represents the degree of dispersion in the energy distribution of events and reflects the volatility of the system state. Let be the first moment of natural time, representing the center position of the event sequence; It is a weighted average of the squares of natural time, measuring the spatial diffusion of the energy distribution of an event.

[0219] Entropy defined within the natural time domain It is dynamic entropy, which differs from the static entropy defined in the traditional time domain. It can capture the dynamic characteristics of the system. Entropy is calculated as follows:

[0220]

[0221] In the formula, Entropy in the natural time domain is used to measure the uncertainty and complexity of an event sequence. The weighted natural time logarithmic mean describes the complexity of the product of event energy and time.

[0222] Time Reversal Entropy The calculation formula is as shown above, but the formula for calculating it is as follows: and From time reversal and The calculation formula is as follows:

[0223] ;

[0224] ;

[0225] In the formula, The normalized energy after time reversal is the energy of the original event sequence rearranged from back to front; The event number is the sequence number after time reversal, indicating the original sequence number when the events are arranged from back to front. The new location of the event; The natural time value after time reversal is the event timeline mapped in reverse to form a time-symmetric sequence; The natural time value after time reversal represents the value of the first natural time series after the original natural time series is reversed. The natural time value at each location.

[0226] Reversed and Substituting into the entropy formula, we obtain the inverse entropy. .

[0227] By combining features such as entropy, inversion entropy, and variance, abrupt changes and critical phenomena in the system state can be effectively identified, enhancing the ability to perceive early potential problems.

[0228] Step 3.4, Integration

[0229] This dataset fuses statistical features extracted from the traditional time domain with dynamic features extracted from the natural time domain to form a unified dataset of potential hazards. This feature set has the following characteristics:

[0230] Rich in dimensions: covering waveform morphology characteristics and system dynamic evolution characteristics;

[0231] Information compression: Reducing data dimensionality while retaining key information;

[0232] Strong dynamic characterization capability: It can capture the key process of wiring hazards evolving from a stable state to a fault state.

[0233] This step integrates traditional time-domain and natural time-domain feature extraction methods to construct a feature dataset that comprehensively reflects the dynamic evolution of wiring hazards, providing high-quality, highly discriminative input features for subsequent attribute reduction and classification prediction.

[0234] Step 4: High-dimensional feature attribute reduction based on fuzzy rough sets:

[0235] This step aims to reduce the attributes of the high-dimensional potential hazard feature dataset constructed in Step 3, which integrates traditional and natural time domains. Redundant or irrelevant features are removed, while key features with high discriminative power are retained, thereby reducing feature dimensionality and improving the training efficiency and prediction performance of the classification model. This step employs fuzzy rough set theory to achieve unsupervised feature selection for continuous and uncertain data.

[0236] Step 4.1, Construct a decision-making information system:

[0237] All collected sample data are treated as a collection of objects. collection of objects This is called the domain of discourse, which uses the traditional time domain and natural time domain features extracted in step 3 as the attribute set. , build a The feature-object matrix. Each row in this matrix corresponds to a sample object, and each column corresponds to a feature attribute.

[0238] Step 4.2, Data Normalization and Fuzzification:

[0239] Since different features have different dimensions and value ranges, to avoid bias in similarity calculation, the min-max normalization method is first used to normalize all features. The calculation formula is as follows:

[0240] ;

[0241] In the formula, Normalized eigenvalues; These are the original eigenvalues; The minimum value for all objects under this property; The maximum value of all objects under this property.

[0242] Subsequently, a fuzzy similarity function is constructed to quantify the similarity between any two objects under a certain attribute:

[0243] ;

[0244] In the formula, For object and In attributes Fuzzy similarity on; , Represents any two objects in the dataset; For two objects in properties The distance below; The scaling parameter controls the rate at which the similarity function decays; it is a positive real number.

[0245] For each attribute in the dataset Calculate the fuzzy similarity between all objects and construct a dimension of Symmetric fuzzy similarity matrix ,in This represents the total number of objects. Ultimately, we obtain a fuzzy similarity set:

[0246] ;

[0247] in, The total number of attributes; each attribute in the collection is a... The symmetric fuzzy similarity matrix.

[0248] Step 4.3, Define and calculate fuzzy dependency:

[0249] Fuzzy dependency is used to measure the correlation between subsets of attributes and is an important basis for attribute reduction.

[0250] Define fuzzy neighborhood:

[0251] For any object and attribute subset Define its in Fuzzy neighborhood , representing an object and In attribute subset Minimum fuzzy similarity:

[0252] ;

[0253] In the formula, Representation Object With object In attribute subset Minimum fuzzy similarity; For object and In attributes Fuzzy similarity.

[0254] Calculate fuzzy dependency:

[0255] Measuring a subset of two attributes The dependency relationship between them, the fuzzy dependency degree is defined as:

[0256] ;

[0257] In the formula, Represents a subset of attributes For attribute subsets The degree of dependence; the larger the value, the stronger the dependence. For object In attribute subset The fuzzy neighborhood below; For object In attribute subset The fuzzy neighborhood below.

[0258] Fuzzy dependency reflects the discriminative ability between attribute subsets and is an important basis for attribute selection.

[0259] Step 4.4, Attribute Reduction Algorithm:

[0260] Based on fuzzy dependency, a greedy strategy is used for attribute reduction. The specific steps are as follows:

[0261] Initialize the reduced set: Let the initial reduced set be... This indicates that the currently selected set of attributes is empty.

[0262] Iterative selection of the optimal attribute: For each unselected attribute a∈A\R, calculate and add it to the list. The subsequent increase in fuzzy dependency;

[0263] Select the attribute a* that maximizes the fuzzy dependency gain and add it. ;

[0264] renew .

[0265] Termination condition judgment: When the improvement of fuzzy dependency after adding the remaining unselected attributes is less than the preset threshold, the iteration stops.

[0266] Output the final reduced attribute set: Obtain the final attribute reduction result. This set significantly reduces the feature dimension while retaining key discriminative information.

[0267] Compared with traditional feature selection methods, this step has the following advantages: fuzzy rough set introduces fuzzy similarity relationships, which is suitable for noise and uncertainty in power line monitoring data; it does not rely on sample labels, making it suitable for practical engineering scenarios lacking labeled data; it filters out the attributes that contribute the most to classification ability through fuzzy dependency, thereby improving the model's generalization ability; and it removes redundant features, significantly reducing classifier training time and resource consumption.

[0268] This step uses fuzzy rough set theory to perform unsupervised attribute reduction on high-dimensional heterogeneous features, which effectively reduces the feature dimension while retaining key discriminative information, improving the efficiency and robustness of subsequent classification models and providing optimized feature inputs for the intelligent identification of hidden dangers in power distribution lines.

[0269] Step 5: Hazard classification and prediction based on support vector machine:

[0270] like Figure 4 As shown, this step aims to construct a Support Vector Machine (SVM) classification model based on the low-dimensional feature set after attribute reduction in step 4, and combine it with parameter optimization strategies to achieve accurate classification and trend prediction of different types of hidden dangers in power distribution lines. This step includes the following three main parts: feature normalization, SVM classifier training strategy, and model parameter optimization and evaluation.

[0271] Step 5.1, Feature normalization processing:

[0272] Because SVM is sensitive to feature dimensions when calculating the similarity between sample points, attributes with large feature value ranges may dominate distance calculations, causing the model to favor these features. Therefore, the reduced low-dimensional feature set is subjected to min-max normalization again, mapping all feature values ​​uniformly to... Interval.

[0273] This step effectively improves the training efficiency and classification stability of the SVM model.

[0274] Step 5.2, One-to-Many SVM Classifier Training Strategy:

[0275] To address the multi-category classification problem of potential hazards in power distribution lines, this step employs a one-to-many strategy to train multiple binary classification SVM models. The specific process is as follows:

[0276] Train a separate SVM classifier for each hazard category;

[0277] During training, samples of this class are labeled as positive (+1), and all other samples are labeled as negative (-1).

[0278] All classifiers together constitute a multi-classification model;

[0279] In the prediction phase, a feature vector of a sample to be tested is input. Calculate its decision function value under all classifiers. And select the category with the highest score as the prediction result:

[0280] ;

[0281] In the formula, The input feature vector; This indicates that the category with the highest score is selected as the prediction result; For the first The decision function of an SVM classifier represents the input feature vector. Belongs to the The confidence level or score of the type of hidden danger.

[0282] This strategy effectively improves the model's ability to identify various types of hidden dangers and is applicable to the classification and identification of various typical hidden dangers in power distribution lines, such as partial discharge, insulation degradation, and poor contact.

[0283] Step 5.3, SVM model parameter optimization:

[0284] To improve the classification performance of the model, this step uses soft-margin SVM and combines kernel tricks and parameter optimization strategies for modeling and tuning.

[0285] Soft-margin SVM optimization objective function: To improve the model's robustness to noise and outliers, a soft-margin SVM with relaxation variables is used. Its optimization objective function is as follows:

[0286] ;

[0287] In the formula, Indicates the direction of the classification hyperplane; The penalty coefficient controls the trade-off between the cost of misclassification and the classification margin; These are slack variables; For the first The label of each sample, This indicates whether the sample belongs to the positive or negative class; This is a feature mapping function that maps the original input space to a high-dimensional feature space. The bias term controls the position of the classification hyperplane; Indicates the index for all samples This means that the constraints apply to every sample in the training set.

[0288] Nonlinear separable problem handling: To address the nonlinear characteristics of power distribution hidden danger data, a radial basis function (RBF) kernel is used to map the low-dimensional feature space to a high-dimensional space, thereby achieving nonlinear classification.

[0289] The kernel function is defined as follows:

[0290] ;

[0291] In the formula, The kernel function output value represents the similarity between two samples in a high-dimensional space; These are kernel function parameters that control the range of influence of a single sample on the decision boundary; , Given two input samples, representing the first... and the One training sample; It is the square of the Euclidean distance between the feature vectors of the two samples.

[0292] Parameter optimization method: To improve the model's generalization ability, a random search method combined with... Cross-validation is used to evaluate the key parameters of SVM. and Joint optimization will be performed. The specific process is as follows:

[0293] Randomly select C and within the preset parameter range. Combinations; perform parameter analysis on each group. Cross-validation is used to evaluate classification accuracy; the parameter combination that performs best on the validation set is selected for final model training.

[0294] Model evaluation and prediction:

[0295] After parameter optimization, the model is applied to the test set for evaluation, and the classification performance is measured by indicators such as accuracy, recall, and F1 score. Finally, the trained SVM model is deployed to the hazard prediction system to achieve automatic classification and hazard identification of newly collected data.

[0296] This step constructs a multi-classification prediction model based on support vector machines (SVM) using a low-dimensional feature set after attribute reduction. Through feature normalization, a one-to-many training strategy, and parameter optimization, it achieves efficient identification and accurate prediction of different types of potential hazards in power distribution lines. This method possesses good generalization ability and classification accuracy, providing reliable technical support for the intelligent operation and maintenance of power distribution lines.

[0297] In summary, this method for assessing and predicting wiring hazards based on natural time domain and fuzzy rough sets systematically solves key technical problems in power wiring hazard identification, such as severe signal noise interference, incomplete hazard feature extraction, and high-dimensional feature redundancy affecting model performance, starting from the entire process of signal preprocessing, feature extraction, feature reduction to classification and prediction. It achieves efficient assessment and accurate prediction of the operating status of power wiring.

[0298] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for evaluating and predicting wiring hazards based on natural time domain and fuzzy rough set, characterized in that, Includes the following steps: Step 1, Multi-source monitoring data acquisition: Collect various monitoring signals during the operation of power distribution lines. These signals include partial discharge, temperature, vibration, voltage, and current. Step 2, Signal preprocessing based on EMD and sample entropy: For the acquired non-stationary monitoring signals, an adaptive denoising process is performed by combining empirical mode decomposition (EMD) and sample entropy analysis to extract the effective modal components in each signal and suppress noise interference. Step 3, Feature extraction by integrating traditional time domain and natural time domain: Based on the effective signal modal components obtained from preprocessing, extract statistical features of traditional time domain, construct event sequences of natural time domain and calculate their dynamic features, and finally integrate the two types of features to form a wiring hazard feature dataset with dynamic representation capabilities. Step 4: High-dimensional feature attribute reduction based on fuzzy rough set: For the constructed high-dimensional hidden danger feature dataset that integrates traditional time domain and natural time domain, fuzzy rough set theory is used to perform unsupervised attribute reduction, remove redundant features, and retain key features with high discriminative ability for hidden danger identification. Step 5: Hazard classification and prediction based on support vector machine: Based on the low-dimensional feature set after attribute reduction, a support vector machine (SVM) classification model is constructed, and combined with parameter optimization strategy, the model is used to accurately classify and predict the trend of different types of hazards in power wiring. Step 3, in the process of integrating the feature extraction of the traditional time domain and the natural time domain, the process of constructing the event sequence in the natural time domain is as follows: Construction of natural time-domain event sequences: Define event energy, which includes: the pulse generated by partial discharge, the magnitude of temperature rise, the magnitude of voltage drop / rise, the change in signal harmonic THD, and the instantaneous frequency change of the signal; After normalization, the event sequence is constructed: , , …, ; In the formula, is the time of occurrence of the th event; is the energy value of the th event; is the event number , is the total number of events; Assign sequence numbers to the events and calculate the normalized energy: From the first event, assign an incremental integer sequence number to each event , thereby obtaining a dataset ; Then the first An event is defined in the natural time domain as: ； Calculate the normalized energy: ； In the formula, The event sequence number; For the first The natural time value of an event; For the first The energy value of an event; No. Normalized energy of an event; Step 3, in the process of integrating the traditional time domain and the natural time domain feature extraction, the calculation of its dynamic features is as follows: Natural time-domain feature acquisition: Extracting dynamic features, including: entropy, a measure of system disorder. Inverse entropy, reflecting time symmetry Variance describing the degree of dispersion of energy distribution ; variance The calculation formula is as follows: ； In the formula, Variance represents the degree of dispersion in the energy distribution of events and reflects the volatility of the system state. Let be the first moment of natural time, representing the center position of the event sequence; It is a weighted average of the squares of natural time, which measures the spatial diffusion of the energy distribution of an event; entropy The calculation formula is as follows: ； In the formula, Entropy in the natural time domain is used to measure the uncertainty and complexity of an event sequence. The weighted natural time logarithm mean describes the complexity of the product of event energy and time. Entropy The calculation formula is as follows: ； ； In the formula, The normalized energy after time reversal is the energy of the original event sequence rearranged from back to front; The event number is the sequence number after time reversal, indicating the original sequence number when the events are arranged from back to front. The new location of the event; The natural time value after time reversal is the event timeline mapped in reverse to form a time-symmetric sequence; The natural time value after time reversal represents the value of the first natural time series after the original natural time series is reversed. The natural time value at each location.

2. The wiring hazard assessment and prediction method based on natural time domain and fuzzy rough set as described in claim 1, characterized in that, Step 2 includes the following steps: Step 2.1, Signal segmentation: The non-stationary time-series signal is reasonably segmented according to the sampling rate. Each segment contains at least n typical hidden danger signal cycles and the number of sampling points does not exceed the limit. Step 2.2, EMD mode decomposition: Empirical mode decomposition (EMD) is used to adaptively decompose the segmented signal, extract the intrinsic mode function components that satisfy the IMF conditions in sequence, and finally obtain the residual terms to complete the multi-level characterization of the non-stationary signal. Step 2.3, Noise Suppression Based on Sample Entropy: By calculating the sample entropy of each IMF component and constructing a separation exponential function, combined with signal energy and complexity information, noise components are adaptively identified and suppressed, while retaining effective signal modes with potential hazards.

3. The wiring hazard assessment and prediction method based on natural time domain and fuzzy rough set as described in claim 2, characterized in that, Step 2.2, EMD mode decomposition process, includes the following steps: First, let's define the signal margin. ， Then, it iterates through all local extrema of the signal margin, and uses cubic spline interpolation to fit the local maxima and minima, constructing the upper and lower envelopes, and averaging them to obtain the envelope mean. ; Then, calculate If it satisfies the intrinsic mode function (IMF) condition, then let ， It is an IMF component; if it does not meet the requirements, repeat the above steps until the IMF conditions are met; then calculate the residual component. Repeat the above steps until the residual component is reached. It is a monotonic function or a constant; This is the original non-stationary time-series signal. Then it can be expressed as: ； In the formula, The total number of extracted IMF components is determined by the complexity of the original non-stationary time series signal, and is 5 to 10 layers. The sequence number of the IMF component, from Initially, high-frequency components are extracted layer by layer, from low-frequency to high-frequency components.

4. The wiring hazard assessment and prediction method based on natural time domain and fuzzy rough set as described in claim 2, characterized in that, Step 2.3, in the noise suppression process based on sample entropy, the steps for calculating the sample entropy of each IMF component for each intrinsic mode component are as follows: Vector reconstruction: converting the original time series Reconstructed Dimensional vector: ； In the formula, for The dimensional reconstruction vector represents the reconstructed vector from... The beginning of the continuous A vector composed of points; The original time series represents the first... The signal values ​​at each sampling point; , jointly constructed indivual dimensional vector; The time series length represents the total number of sampling points. Calculate the distance between vectors: for any two vectors and , Calculate the maximum absolute difference between them: ； In the formula, Let be the distance between vector pairs, representing the distance between the first and second vector pairs. The and the first indivual The maximum absolute difference between dimensional vectors; The dimension of the vector represents the dimensions of the reconstructed vector, typically 2 or 3. Logarithm of similar vectors: given a similarity threshold Statistically count all that satisfy Count the number of vector pairs and calculate: ； In the formula, For dimension The proportion of similar vector pairs in a time series measures its performance. Regularity under dimensions; This represents the number of vector logarithms that satisfy the condition. Add dimensions to Expand the dimension of the vector to Calculate the distance between new vector pairs and statistically determine if they satisfy the condition. By counting the number of vector pairs, we get: ； In the formula, For dimension The proportion of similar vector pairs measures the regularity of time series changes in higher dimensions; Calculate sample entropy: ； In the formula, The sample entropy measures the complexity of a time series; the larger the value, the more complex the series. For dimension The proportion of similar vector pairs measures the regularity of time series changes in higher dimensions.

5. The wiring hazard assessment and prediction method based on natural time domain and fuzzy rough set as described in claim 4, characterized in that, In the noise suppression process based on sample entropy, a modal component separation exponential function is constructed. The steps are as follows: Based on sample entropy and combined with signal energy, a comprehensive index is constructed to distinguish effective signals from noise; the separation index is proportional to the product of signal energy and sample entropy, expressed as: ； In the formula, For the first j Layer IMF The time separation index, which combines signal energy and entropy information, is used to identify valid signals and noise. For the first Layer IMF in the 1st The energy at a given moment represents a measure of the signal energy, which is the square of the signal amplitude. For the first Layer IMF in the 1st Sample entropy at time t; Noise suppression logic: Set a threshold; like If the threshold is reached, then the point is a valid signal; like If the value is less than or equal to the threshold, then that point is considered noise and should be suppressed.

6. The wiring hazard assessment and prediction method based on natural time domain and fuzzy rough set as described in claim 1, characterized in that, Step 3, in the process of integrating the traditional time domain and the natural time domain feature extraction, the traditional time domain feature extraction process is as follows: For each decomposed modal component signal, time-domain features are extracted according to the segmented signal, including maximum value, minimum value, mean, absolute mean, standard deviation, variance, and root mean square; at the same time, peak factor, impulse factor, margin factor, and waveform factor are calculated. The formula for calculating the peak factor is: ； The formula for calculating the impulse factor is: ； The formula for calculating the margin factor is: ； The formula for calculating the waveform factor is: ； In the formula, The original time series represents the first... The signal values ​​at each sampling point; This represents the total number of sampling points; The maximum value of the signal, the maximum amplitude value in this segment of the signal. ; The minimum value of the signal, the minimum amplitude in this segment of the signal. .

7. The wiring hazard assessment and prediction method based on natural time domain and fuzzy rough set as described in claim 1, characterized in that, Step 4: High-dimensional feature attribute reduction based on fuzzy rough sets, including the following steps: Step 4.1, Information System: Treat all data as a collection of objects It takes all features of traditional time and natural time as a set of attributes. , build matrix; Step 4.2, data normalization and fuzzification: The min-max normalization method is used to normalize all attributes. The normalization calculation formula is as follows: ； In the formula, Normalized eigenvalues; These are the original eigenvalues; The minimum value for all objects under this property; The maximum value of all objects under this property; Constructing a fuzzy similarity function: ； In the formula, For object and In attributes Fuzzy similarity on; , Represents any two objects in the dataset; For two objects in properties The distance below; is a scale parameter that controls the rate at which the similarity function decays; it is a positive real number. Finally, the fuzzy similarity matrix corresponding to each attribute is obtained; Step 4.3, define fuzzy dependency: First, define the fuzzy neighborhood. : ； In the formula, Representation Object With object In attribute subset Minimum fuzzy similarity; For object and In attributes Fuzzy similarity; Then, calculate the fuzzy dependency between attribute subsets. : ； In the formula, Represents a subset of attributes For attribute subsets The degree of dependence; the larger the value, the stronger the dependence. For object In attribute subset The fuzzy neighborhood below; For object In attribute subset The fuzzy neighborhood below; Step 4.4, Attribute Reduction Algorithm: First, initialize the reduced set. ; Then, for each property not in the reduced property set R Attributes in a ∈ A ∖ R Calculate and add it to R The subsequent contribution to fuzzy dependency; Then, select the attribute that contributes the most. a* Add to the set of reduced properties R The process continues until the remaining attribute dependency is below a set threshold. Finally, the reduced property set is output.

8. The wiring hazard assessment and prediction method based on natural time domain and fuzzy rough set as described in claim 1, characterized in that, Step 5: Hazard classification and prediction based on support vector machines, including the following steps: Step 5.1, Normalize and reduce the attribute set: Mini-maximum normalization is applied to the reduced features; Step 5.2, One-to-many SVM training strategy: A binary classification SVM is trained for each hazard category. During training, each SVM labels all samples of a specific category as positive and all samples of the other categories as negative. The final prediction mechanism is as follows: ； In the formula, The input feature vector; This indicates that the category with the highest score is selected as the prediction result; For the first The decision function of an SVM classifier represents the input feature vector. Belongs to the Confidence level or score of the type of potential hazard; Step 5.3, Parameter Optimization: Optimize classification parameters using soft margin: ； In the formula, Indicates the direction of the classification hyperplane; The penalty coefficient controls the trade-off between the cost of misclassification and the classification margin; These are slack variables; For the first The label of each sample, This indicates whether the sample belongs to the positive or negative class; This is a feature mapping function that maps the original input space to a high-dimensional feature space. The bias term controls the position of the classification hyperplane; Indicates the index for all samples That is, the constraints apply to every sample in the training set; Radial basis function (RBF) kernels are used to handle nonlinear separable problems: ； In the formula, The kernel function output value represents the similarity between two samples in a high-dimensional space; These are kernel function parameters that control the range of influence of a single sample on the decision boundary; , Given two input samples, representing the first... and the One training sample; It is the square of the Euclidean distance between the feature vectors of two samples; Then, the parameters were tuned using a combination of random search and K-fold cross-validation to find the optimal penalty coefficient. C and kernel function parameters γ combination; Finally, a predictive evaluation model is used to process the test data.

Images