Power transmission line operation risk prediction method based on big data mining

Abstract

The invention provides a power transmission line operation risk prediction method based on big data mining, and the method comprises the steps: collecting the high-frequency vibration data and environment parameters of a bolt connection part of a power transmission line tower in real time, and generating a time sequence data set containing the vibration frequency, the amplitude and the temperature; a Fourier transform algorithm is adopted for the time sequence data set, a dominant frequency component of the vibration frequency is extracted, linear correction is conducted on the vibration frequency according to the thermal expansion coefficient and the temperature of the bolt material, and a frequency feature set is generated; constructing a bolt state evaluation model according to the preliminary feature set and the stress distribution feature set, evaluating the bolt state, and outputting the bolt pre-tightening force attenuation degree; and according to the state score of each bolt output by the optimized state evaluation model, calculating the overall load sharing and vibration frequency drift degree of the bolt joint, and generating a load sharing feature set.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for predicting the operational risks of power transmission lines based on big data mining. Background Technology

[0002] As a crucial component of the power system, the structural safety of transmission line towers directly impacts the reliability of power grid operation and the normal functioning of socio-economic activities. The stability of the tower's bolted connections is critical to ensuring the overall structural stiffness and load-bearing capacity; bolt loosening can lead to structural failure or even catastrophic accidents. However, existing bolt condition assessment methods primarily rely on periodic inspections and static mechanical analysis, making it difficult to capture the dynamic changes in bolt preload decay in real time and insufficiently addressing vibration response characteristics under complex environments. This results in limited early warning capabilities for bolt loosening and significant uncertainty in structural safety margin assessments. Bolt preload decay is a core challenge to tower structural safety. Reduced preload increases the non-uniformity of stress distribution around the bolts, leading to localized stress concentration and accelerating material fatigue. This non-uniform stress distribution further imbalances the load-sharing ratio among bolts, causing some bolts to bear excessive loads and increasing the risk of fracture. This load-sharing imbalance alters the structure's dynamic response characteristics, manifesting as a drift in local vibration frequencies, affecting the accurate interpretation of high-frequency vibration monitoring data. These interrelated factors collectively exacerbate the complexity of dynamically adjusting bolt condition assessment models based on vibration data and historical inspection records. Therefore, the key issue of this study is how to construct a dynamically adjustable bolt condition assessment model by integrating high-frequency vibration monitoring data and historical inspection records, and how to quantify its impact on the safety margin assessment of tower structures. Summary of the Invention

[0003] This invention provides a method for predicting the operational risks of power transmission lines based on big data mining, mainly including:

[0004] (1) Real-time acquisition of high-frequency vibration data and environmental parameters at the bolted connection of transmission line towers to generate a time series dataset containing vibration frequency, amplitude and temperature;

[0005] (2) Extract the main frequency component of the vibration frequency from the time series dataset using the Fourier transform algorithm, and perform linear correction of the vibration frequency based on the thermal expansion coefficient and temperature of the bolt material to generate a frequency feature set;

[0006] (3) If the main frequency component in the frequency feature set deviates from the preset frequency range, the original high-frequency vibration signal is decomposed by wavelet transform, the local vibration energy distribution features of different frequency bands are extracted, and the local vibration energy distribution features are processed to obtain a preliminary feature set reflecting the degree of bolt preload attenuation.

[0007] (4) Perform feature selection and dimensionality reduction on the preliminary feature set, input it into the pre-established bolt finite element model, simulate the stress distribution of the bolt connection under different preload attenuation levels, extract stress distribution features, and generate a stress distribution feature set;

[0008] (5) Construct a bolt condition assessment model based on the preliminary feature set and stress distribution feature set, assess the bolt condition, and output the degree of bolt preload attenuation.

[0009] (6) Combine the bolt preload attenuation with the obtained bolt historical inspection record data, and use Bayesian inference method to update the parameters of the bolt condition assessment model to obtain the optimized bolt condition assessment model.

[0010] (7) Based on the state scores of each bolt output by the optimized state assessment model, calculate the overall load sharing and vibration frequency drift of the bolt connection and generate a load sharing feature set.

[0011] (8) Establish a Monte Carlo simulation model based on the load sharing feature set, simulate the failure probability of bolts under service conditions based on the current load sharing and vibration frequency drift, and generate a comprehensive evaluation result of bolt status based on the failure probability.

[0012] Preferably, step (1) specifically includes:

[0013] The vibration signal output by the sensor array is collected, and the vibration signal is filtered by a digital bandpass filter to obtain a filtered vibration signal.

[0014] The amplitudes of the lateral, longitudinal, and vertical vibrations in the filtered vibration signal are used to determine whether they exceed a preset threshold. If they do, the sampling frequency is adjusted. Triaxial accelerometers are installed at each bolt connection of the transmission tower with a sensor array spacing of at least 30 centimeters. When collecting vibration signals, the original vibration waveform is obtained at a sampling frequency of 2 kHz. The collected signal is then filtered to suppress power frequency interference and eliminate high-frequency noise through a digital bandpass filter.

[0015] Fourier transform is performed on the filtered vibration signal to obtain spectral features. The dominant vibration frequency and its amplitude are extracted from these spectral features, and ambient temperature data is recorded to obtain the original monitoring dataset. Amplitude detection is performed on the filtered vibration signal. When the lateral, longitudinal, or vertical vibration amplitude exceeds the 0.5 gravitational acceleration threshold, the sampling frequency is automatically adjusted to 4 kHz, and the original sampling frequency is restored after the vibration amplitude drops below the threshold. Fourier transform is performed on the triaxial vibration signal to obtain spectral features, extracting the dominant vibration frequency and its amplitude within the range of 0 to 1 kHz, and ambient temperature data is recorded simultaneously to form an original monitoring dataset containing vibration frequency, amplitude, and temperature.

[0016] A support vector regression algorithm was used to establish a mapping relationship between vibration feature vectors and temperature changes in the original monitoring dataset. The original monitoring dataset was segmented using a sliding window method, with a window length of one thousand data points. Outliers were removed using median filtering to obtain a preprocessed monitoring dataset. Wavelet decomposition was performed on the vibration frequency and amplitude data in the preprocessed monitoring dataset to extract the low-frequency fundamental component and high-frequency harmonic components. These components were then combined with temperature data to construct a multi-dimensional feature vector. The support vector regression algorithm was then used to establish a mapping relationship between the vibration feature vectors and temperature changes, generating a time-series dataset containing the vibration frequency, amplitude, temperature parameters, and their associated features for each monitoring point.

[0017] Preferably, step (2) specifically includes:

[0018] The Hanning window function is used to process vibration time series data. The windowed signal is subjected to spectrum analysis by fast Fourier transform to obtain the dominant frequency feature vector. The collected vibration time series data is windowed using the Hanning window function with the window length set to the length of the vibration data sequence. The spectrum analysis of the windowed signal is performed by fast Fourier transform, and the three frequency peaks with the largest amplitude are extracted from the spectrum as the dominant frequency feature vector.

[0019] Based on the displacement of the frequency peak value in the main frequency feature vector as a function of temperature, and combined with the thermal expansion coefficient of the bolt material, the temperature-frequency mapping relationship is obtained, and the frequency drift vector is obtained.

[0020] The frequency drift vector is curve-fitted by the least squares method to obtain a temperature compensation function containing a slope term and an intercept term. The temperature compensation function is then applied to the main frequency feature vector to eliminate the influence of temperature on the frequency and obtain the corrected frequency feature vector.

[0021] Wavelet decomposition is performed on the corrected frequency feature vector. Wavelet basis functions are selected to perform three-level decomposition of the signal and extract the energy coefficients of each frequency component. The energy coefficients are then classified using the support vector machine method to obtain a frequency feature set containing the frequency value of the main frequency component, the energy value, and the temperature compensation coefficient.

[0022] Preferably, step (3) specifically includes:

[0023] The original vibration signal is decomposed using wavelet transform to obtain the multi-scale decomposition coefficients of the vibration signal; the deviation between the main frequency component in the frequency feature set and the preset reference frequency is calculated using the spectrum analysis method. The reference frequency range is set in the range of 200 Hz to 800 Hz. A signal segment with a length of 4096 points is extracted from the original vibration signal, and the signal segment is decomposed into four levels using wavelet transform to obtain the multi-scale decomposition coefficients of the high-frequency vibration signal.

[0024] The frequency band energy values ​​of the low-frequency band, mid-frequency band, and high-frequency band are calculated based on the multi-scale decomposition coefficients, and the frequency band energy values ​​are compared with preset energy thresholds. The frequency bands are divided into low-frequency band (0 to 200 Hz), mid-frequency band (200 to 500 Hz), and high-frequency band (500 to 1000 Hz), and the energy threshold for each frequency band is set to 1.5 times the average energy of that frequency band.

[0025] For frequency bands exceeding the preset energy threshold, the corresponding wavelet coefficients are extracted, and the frequency band energy feature vector is obtained by calculating the mean, variance, and kurtosis factor of the wavelet coefficients.

[0026] Based on the frequency band energy feature vector, calculate the energy distribution characteristics including the energy centroid frequency value, frequency band energy ratio, and energy dispersion. Use a support vector regression model to establish the mapping relationship between the frequency offset and the energy distribution characteristics to obtain a preliminary feature set reflecting the degree of bolt preload attenuation.

[0027] The least squares method is used to fit the frequency band energy characteristic vector, and the energy ratio between each frequency band is calculated to obtain the frequency band energy distribution characteristics. Based on the frequency band energy distribution characteristics, an energy concentration index is calculated, including the energy centroid frequency, the frequency band energy proportion, and the energy dispersion.

[0028] A support vector regression model is constructed based on the energy concentration index to establish a mapping relationship between frequency offset and energy distribution characteristics, and to generate a feature dataset containing frequency offset, energy distribution characteristics, and preload attenuation index.

[0029] Preferably, step (4) specifically includes:

[0030] A hexahedral mesh model is constructed based on the bolt diameter and thread pitch, with the mesh size set to one-tenth of the bolt diameter. A preload boundary condition is applied to the bolt connection surface, with the preload value decreasing stepwise from the nominal preload. The force model is obtained by applying the preload boundary condition to the bolt connection surface of the mesh model.

[0031] The elastic modulus, Poisson's ratio, and yield strength parameters of the bolt material are set. Threaded pair contact constraints are applied to the mesh model. The nodal displacement field is calculated using the finite element method. The nodal displacement field is then converted into stress field distribution data using Hooke's law. The data includes the normal stress and shear stress values ​​at the contact surface between the bolt shank and the nut.

[0032] The stress field distribution data is resampled by grid, and the coordinates of the stress concentration region are obtained by extracting grid cells whose stress values ​​exceed the yield strength.

[0033] A deep autoencoder is used to extract features from the coordinates of the stress concentration region. By setting the number of hidden layer nodes of the encoder to half of the input dimension, a stress distribution feature set is obtained. The stress distribution feature set mainly includes stress distribution, deformation, and stress concentration location.

[0034] Preferably, step (5) specifically includes:

[0035] Outlier detection is performed on the preliminary feature set and stress distribution feature set to remove data points that exceed three times the standard deviation. The maximum and minimum value normalization method is used to map the feature values ​​to the zero-to-one interval to generate a standardized feature matrix.

[0036] Variance screening is performed based on the standardized feature matrix to remove feature columns with variance values ​​less than a preset threshold. Pearson correlation coefficient analysis is then used to obtain a feature subset; that is, redundant features with correlation values ​​higher than 0.9 are removed using the Pearson correlation coefficient to obtain a feature subset.

[0037] The feature subset is graded and labeled according to the preload attenuation amount, and the preload attenuation amount is divided into grade intervals. The training set and the validation set are obtained by stratified sampling; preferably, the training set and the validation set are divided in an 8:2 ratio.

[0038] Based on the training set, the parameters of the support vector machine are selected using cross-validation, and the parameters of the classification hyperplane are obtained through grid search to form a bolt state evaluation model.

[0039] Specifically, the radial basis function parameters of the support vector machine are selected using the five-fold cross-validation method, the parameter search grid range is set, and the classification accuracy corresponding to each set of parameters is calculated. The penalty coefficient is determined by the grid search method, and the parameter combination with the highest classification accuracy is selected. The training data is then fitted to obtain the classification hyperplane parameters in the feature space, thus forming the bolt state evaluation model.

[0040] For the validation set, a prediction is made, and the mean square error is calculated based on the predicted value and the true label. The mean square error is used to construct a confidence interval, and an evaluation result containing the preload attenuation degree and the confidence interval is generated; preferably, the confidence interval is 95%.

[0041] Preferably, step (6) specifically includes:

[0042] Based on the bolt inspection records, the preload measurement value and environmental parameter records are obtained. The missing data in the inspection records are supplemented by cubic spline interpolation to obtain a time series dataset.

[0043] The time series dataset is segmented according to the monitoring time, and the mean and variance of the preload decay rate in the dataset are obtained by statistical calculation method, so as to obtain the normal distribution model of the preload decay parameter.

[0044] The prior probability density function of the preload attenuation parameter is calculated based on the normal distribution model. The mean and covariance matrix of the function are obtained through the probability density function and can be used as prior information for Bayesian inference. The prior probability density function is iteratively updated using the Gibbs sampling method. In each iteration, a preset number of sample points are extracted, and the likelihood function value corresponding to the sample points is calculated. The posterior probability distribution is obtained by calculating the likelihood function value and the prior probability. Weight coefficients are set for historical data according to time distance, and the time weight decreases exponentially with increasing distance. The bolt condition assessment model parameters are updated according to the posterior probability distribution, including the preload attenuation threshold, confidence interval range, and warning level limit, to obtain the optimized bolt condition assessment model.

[0045] Preferably, step (7) specifically includes:

[0046] The stress of the connection structure is calculated based on the bolt condition assessment score. The axial and circumferential stresses of each bolt are calculated. The overall stress distribution at the connection is obtained by using the stress superposition principle, and the stress distribution matrix at the connection is obtained.

[0047] The stress concentration factor is calculated using the maximum stress value and the average stress value in the stress distribution matrix, and the load distribution non-uniformity is obtained based on the stress concentration factor.

[0048] Fourier transform is performed on the vibration signal at the connection point. Spectral analysis is then conducted using the Fourier transform to extract the amplitude and phase spectra of the vibration signal, and the deviation between the vibration frequency of each bolt and the initial assembly frequency is calculated. The stiffness attenuation vector is calculated based on the frequency response curve of the vibration signal. The amplitude response at the natural frequency point is extracted, and the change in structural stiffness is calculated to generate the stiffness attenuation vector. A neural network containing an input layer, a hidden layer, and an output layer is constructed. The input layer receives the load sharing unevenness and the stiffness attenuation vector. The load sharing feature set is obtained through iterative computation of the neural network, and this feature set includes the stress distribution ratio, frequency deviation, and stiffness change of each bolt.

[0049] Preferably, it further includes: obtaining material fatigue parameters and preload deviation parameters from the bolt condition assessment score to establish a dynamic stress distribution matrix, the dynamic stress distribution matrix containing contact pressure gradient data of each node in the bolt group, extracting contact surface deformation characteristics from the dynamic stress distribution matrix, inputting the contact surface deformation characteristics and vibration frequency drift into the load distribution calculation model, outputting the dynamic load transfer coefficient of each bolt node, and generating a load sharing feature set based on the dynamic load transfer coefficient, specifically including:

[0050] A gridded nodal stress field is constructed based on the material fatigue parameters and preload deviation parameters in the bolt condition score. The initial stress distribution matrix is ​​obtained by spatial discretization of the bolt connection using hexahedral elements. Discretization involves spatial discretization of the bolt connection using hexahedral elements and calculation of the normal and tangential stresses at each grid point.

[0051] The initial stress distribution matrix is ​​subjected to mesh refinement processing, and the stress values ​​of the refined mesh points are calculated using a stress interpolation algorithm to obtain refined stress distribution data;

[0052] The refined stress distribution data is fitted with a cubic spline interpolation method to fit the deformation surface of the contact surface, and the deformation of key feature points is extracted from the deformation surface.

[0053] A recurrent neural network is trained based on the deformation and frequency drift data. The network parameters are optimized using stochastic gradient descent to obtain the load transfer coefficient between nodes. A load sharing feature set is generated based on the load transfer coefficient and stress gradient data. The recurrent neural network is a five-layer recurrent neural network. The input layer contains deformation features and frequency drift data. The number of nodes in the hidden layer is halved with each layer. The network parameters are optimized using stochastic gradient descent. The load transfer coefficient between nodes is calculated based on the trained neural network. The coefficient value represents the load distribution ratio between adjacent nodes.

[0054] A load distribution matrix is ​​constructed based on the load transfer coefficient, and the stress distribution weight of each node is calculated by combining the stress gradient data to generate a feature dataset containing node stress distribution, deformation, and transfer coefficient.

[0055] Cross-validation is used to verify the accuracy of the feature dataset, and the mean square error between the predicted and measured values ​​is calculated to ensure the reliability of the dataset.

[0056] Preferably, step (8) specifically includes:

[0057] The maximum likelihood estimation method is used to process the load distribution and frequency drift data to obtain the joint probability density function, and the initial state probability distribution data is generated based on the joint probability density function.

[0058] Random sampling is performed based on the initial state probability distribution data to obtain a set of sample points. The stress amplitude and cycle number are calculated for the set of sample points. During random sampling, the Monte Carlo sampling number is set, and sample points are randomly selected from the initial state probability distribution. The stress amplitude and cycle number corresponding to each sample point are calculated, and the fatigue damage value is calculated according to the linear cumulative damage criterion.

[0059] A prediction model is constructed using a deep neural network, which contains three convolutional layers. The input dimension of each convolutional layer is the same as the number of features. The number of kernels in each convolutional layer is set to half that of the previous layer, and batch normalization is used to prevent gradient vanishing.

[0060] The deep neural network is trained on the sample point set to obtain a trained neural network model. The failure probability under different working conditions is calculated based on the trained neural network model to generate a comprehensive evaluation result of the bolt condition. Failure probability curves are generated by calculating the failure probability under different working conditions based on the trained neural network. The accuracy of the prediction results is evaluated using a five-fold cross-validation method, and the relative error between the predicted and measured values ​​is calculated.

[0061] Based on the failure probability curve, the bolt condition is divided into three levels: intact, warning, and dangerous. The confidence level for each level is calculated. A comprehensive evaluation dataset containing the condition level, failure probability value, and confidence interval is generated. The dataset format includes three fields: timestamp, evaluation index, and prediction result.

[0062] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0063] This invention discloses a method for predicting the operational risks of transmission lines based on big data mining. It collects high-frequency vibration data and environmental parameters at bolted connections in real time, extracts vibration features using algorithms such as Fourier transform and wavelet transform, and simulates stress distribution under different preload attenuation levels using a finite element model. A support vector machine algorithm is employed to fuse multi-source features, constructing a bolt condition assessment model, and Bayesian inference is used to optimize model parameters based on historical inspection records. Finally, a Monte Carlo simulation model is established based on load-sharing characteristics to assess the probability of bolt failure. This invention achieves accurate assessment and prediction of the bolt condition of transmission line towers, helping to promptly detect potential faults and improve the safety and reliability of transmission lines. Attached Figure Description

[0064] Figure 1 This is a flowchart of a method for predicting the operational risks of power transmission lines based on big data mining, according to the present invention.

[0065] Figure 2 This is a schematic diagram of a method for predicting the operation risk of power transmission lines based on big data mining, according to the present invention. Detailed Implementation

[0066] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0067] To facilitate understanding of the embodiments of the present invention, the relevant technical terms will first be briefly explained:

[0068] Bolt preload: refers to the initial clamping force generated when tightening bolts in a bolted connection, used to ensure the rigidity and stability of the connection. Preload decay can lead to loosening of the connection, affecting structural safety.

[0069] High-frequency vibration data: Vibration signals at bolted connections collected by sensors, including parameters such as frequency and amplitude, reflecting the dynamic state of the bolted connection, typically in the range of 20Hz to 1000Hz.

[0070] Time series dataset: A continuous data set containing parameters such as vibration frequency, amplitude, and ambient temperature, used to analyze the dynamic changes in bolt condition.

[0071] In this embodiment of the invention, the bolt condition assessment of transmission line towers involves high-frequency vibration monitoring and data mining techniques. Vibration data is collected by sensors and combined with environmental parameters to generate a time-series dataset. Features are then extracted using algorithms such as Fourier transform and wavelet decomposition. Subsequently, a dynamic assessment model is constructed using a finite element model and machine learning algorithms to generate a comprehensive bolt condition assessment result. The technical solution is detailed below through specific embodiments.

[0072] Figure 1 The implementation flow of the transmission line operation risk prediction method provided in this embodiment of the invention is shown, specifically including:

[0073] The S101 uses sensors to collect high-frequency vibration data and environmental parameters at the bolted connections of transmission line towers in real time, generating a time-series dataset containing vibration frequency, amplitude, and temperature.

[0074] In this embodiment of the invention, a sensor array is deployed at the bolted connection of the transmission line tower to collect high-frequency vibration data and environmental parameters in real time, thereby generating a time-series dataset. The specific acquisition and processing procedures are optimized according to the stress characteristics of the bolted nodes and environmental conditions to ensure high accuracy and reliability of the data.

[0075] Based on the stress characteristics of the bolted connection, a triaxial accelerometer was deployed with a spacing of 300 mm and an initial sampling frequency of 2000 Hz to collect vibration signals including lateral, longitudinal, and vertical directions.

[0076] In this embodiment of the invention, the triaxial accelerometer is mounted at the bolted connection with a spacing of 300 mm to effectively avoid signal interference. The initial sampling frequency of 2000 Hz satisfies the Nyquist sampling theorem, ensuring the capture of the main vibration frequencies in the range of 20 Hz to 800 Hz. The acquired vibration signals contain waveform data in all three axes, providing a basis for subsequent processing.

[0077] The collected vibration signal is processed by a digital bandpass filter to suppress power frequency interference and eliminate high-frequency noise, generating a filtered vibration signal.

[0078] In this embodiment of the invention, a digital bandpass filter with a passband range of 10Hz to 1000Hz is used to preprocess the vibration signal, eliminate 50Hz power frequency interference and high-frequency noise, ensure the purity of the filtered vibration signal, and provide high-quality data for frequency domain analysis.

[0079] The amplitude of the filtered vibration signal is detected. If the horizontal, longitudinal, or vertical amplitude exceeds the 0.5g threshold, the sampling frequency is adjusted to 4000Hz and then restored to 2000Hz after the amplitude returns to normal.

[0080] In this embodiment of the invention, the vibration state at the bolt connection is determined by amplitude detection. When the vibration amplitude in any direction exceeds 0.5g, it indicates a possible anomaly, and the sampling frequency is automatically increased to 4000Hz to capture transient characteristics. After the amplitude returns to normal, the sampling frequency is restored to 2000Hz to reduce power consumption and data processing burden.

[0081] The filtered vibration signal was subjected to Fourier transform to extract the main frequency and its amplitude in the range of 0Hz to 1000Hz. The ambient temperature was recorded simultaneously to generate the original monitoring dataset.

[0082] In this embodiment of the invention, a Fourier transform with a 1024-point window is used to perform frequency domain analysis on the vibration signal, extracting the main frequency peak and its amplitude in the range of 100Hz to 300Hz, while recording the ambient temperature to form an original monitoring dataset containing vibration frequency, amplitude and temperature, laying the foundation for feature extraction.

[0083] S1015 uses a sliding window method to segment the original monitoring dataset, with a window length of 1000 points. It combines median filtering to remove outliers and generate a preprocessed monitoring dataset.

[0084] In this embodiment of the invention, the original monitoring dataset is segmented using a 1000-point sliding window and a 50% overlap rate, and a 5-point median filter is used to remove sudden interference signals, ensuring the continuity and stability of the data. The preprocessed monitoring dataset provides reliable input for subsequent feature extraction.

[0085] S1016 performs wavelet decomposition on the preprocessed monitoring dataset to extract low-frequency fundamental and high-frequency harmonic components, and generates a multi-dimensional feature vector by combining it with temperature data.

[0086] In this embodiment of the invention, the db4 wavelet basis function is selected to perform four-level wavelet decomposition on the vibration signal, extracting the low-frequency fundamental component to reflect the overall vibration characteristics, and the high-frequency harmonic components to capture local loosening features. Combined with temperature data, a multi-dimensional feature vector containing vibration frequency, amplitude, and temperature is constructed to provide rich features for model training.

[0087] S1017 uses the support vector regression algorithm to establish the mapping relationship between vibration feature vectors and temperature changes, generating a time series dataset containing vibration parameters and associated features.

[0088] In this embodiment of the invention, a support vector regression algorithm based on radial basis function kernel function is used to train sample data including normal tightening and multi-level loosening states, establishing a mapping relationship between vibration characteristics and temperature changes. The generated time-series dataset contains the vibration frequency, amplitude, temperature, and related characteristics of each monitoring point, supporting dynamic evaluation of bolt condition.

[0089] In this embodiment of the invention, high-frequency vibration monitoring at bolted connections is applicable to various power transmission line scenarios. Real-time vibration data acquisition via a triaxial sensor accurately reflects the bolt tightness. The sensor arrangement considers the stress characteristics of angle steel connecting bolts, with a 300mm spacing to avoid signal interference. A 2000Hz sampling frequency meets the acquisition requirements of the main vibration frequencies, and a bandpass filter effectively retains valid signals from 10Hz to 1000Hz. When the amplitude exceeds a threshold, the sampling frequency is increased to 4000Hz to ensure the capture of transient features. Fourier transform provides high-resolution frequency analysis, with characteristic peaks from 100Hz to 300Hz reflecting bolt loosening. Temperature changes significantly affect vibration amplitude; in summer at 35℃, the amplitude may increase by 15% to 20%. Sliding window and median filtering ensure data quality, wavelet decomposition extracts multi-scale features, and a support vector regression model achieves accurate mapping between vibration and temperature. When the vibration amplitude increases by 30% and the main frequency decreases by 20%, it indicates significant bolt loosening, requiring timely warning. The embodiments of the present invention can improve the real-time performance and accuracy of bolt condition monitoring, providing technical support for the safe operation of power transmission lines.

[0090] S102 uses frequency domain analysis to extract the main frequency characteristics of bolt vibration signals from the time series dataset. It then performs frequency correction based on the thermal expansion characteristics of the bolt material and the ambient temperature, generating a feature dataset that includes frequency, energy, and temperature compensation coefficients.

[0091] In this embodiment of the invention, a time-series dataset collected from bolted connections of transmission line towers is analyzed using frequency domain transformation and feature extraction techniques to assess the frequency characteristics of vibration signals. Correction is then applied based on temperature effects to generate a frequency feature dataset for bolt condition evaluation. The processing includes windowing, spectral analysis, temperature compensation, and wavelet decomposition steps to ensure the accuracy and reliability of the feature data.

[0092] S1021 preprocesses vibration signals in time series datasets using a windowing function, extracts the main frequency peak and its amplitude in the range of 0Hz to 1000Hz using fast Fourier transform, and generates an initial frequency feature vector.

[0093] In this embodiment of the invention, the vibration signal is windowed using a 4096-point Hanning window function to reduce spectral leakage and improve frequency resolution. A fast Fourier transform is used to generate the spectrum, and the three frequency peaks with the largest amplitudes, typically located in the range of 100Hz to 400Hz, are extracted as the initial frequency feature vector, reflecting the inherent vibration characteristics of the bolted connection.

[0094] S1022 constructs a mapping relationship between temperature and frequency based on the offset of the frequency peak value in the initial frequency feature vector as a function of temperature, combined with the thermal expansion coefficient of the bolt material, and generates a frequency drift vector.

[0095] In this embodiment of the invention, the variation of peak frequency with temperature is analyzed. Taking Q345 steel bolts as an example, their coefficient of thermal expansion is 11.7 micrometers per meter per degree Celsius. When the temperature rises from 20°C to 40°C, the dominant frequency decreases by approximately 2Hz to 3Hz. Based on this, a temperature-frequency mapping relationship is constructed, the frequency drift vector is calculated, and the influence of temperature on vibration characteristics is quantified.

[0096] S1023 uses the least squares method to linearly fit the frequency drift vector, generates a temperature compensation function containing slope and intercept, and applies it to the initial frequency feature vector to generate the corrected frequency feature vector.

[0097] In this embodiment of the invention, the frequency drift vector is fitted using the least squares method, with temperature as the independent variable and frequency shift as the dependent variable, to generate a temperature compensation function. The fitting results show that within the temperature range of 20℃ to 60℃, frequency and temperature exhibit a linear relationship with a correlation coefficient of 0.96. The compensation function is then applied to correct the initial frequency eigenvector, eliminating the influence of temperature.

[0098] S1024 performs wavelet decomposition on the corrected frequency feature vector to extract the energy distribution features of each frequency band and generate a multidimensional feature vector containing frequency and energy.

[0099] In this embodiment of the invention, the db4 wavelet basis function is used to perform a three-level decomposition on the corrected frequency feature vector to extract low-frequency approximate components and mid-to-high-frequency detail components. Under normal conditions, low-frequency energy accounts for approximately 70%, and high-frequency energy accounts for 5%; when the bolts are loose, the proportion of high-frequency energy increases to over 15%. The multidimensional feature vector contains the frequency values ​​and energy coefficients of each frequency band.

[0100] S1025 uses the support vector machine algorithm to classify the energy coefficients of multidimensional feature vectors, generating a frequency feature dataset containing the main frequency value, energy value, and temperature compensation coefficient.

[0101] In this embodiment of the invention, a support vector machine with a radial basis function kernel is used to classify the energy coefficients. The kernel parameter is set to 0.8, and the penalty factor is 100. Cross-validation ensures that the classification accuracy reaches over 95%. The generated frequency feature dataset includes the dominant frequency value, energy value, and temperature compensation coefficient, supporting subsequent bolt condition assessment.

[0102] In this embodiment of the invention, bolt vibration frequency analysis is a core component of transmission line operation risk prediction. Windowing processing, using a Hanning window to reduce spectral leakage, ensures a frequency resolution of 0.5 Hz with a 4096-point window length, meeting the feature extraction requirements within the 100 Hz to 400 Hz range. The dominant frequency peak extracted by Fast Fourier Transform reflects the inherent vibration characteristics of the bolt, and frequency drift caused by temperature changes is effectively corrected through linear fitting. A compensation function improves the stability of the feature data. Wavelet decomposition captures multi-scale vibration features; low-frequency components reflect the overall tightness, while high-frequency components indicate local loosening. Support Vector Machine classification further enhances the discriminative power of the features. In practical applications, the dominant frequency of a normal bolt is between 200 Hz and 400 Hz, with an energy value below 0.5 g RMS, and the temperature compensation coefficient fluctuates between 0.95 and 1.05. In a loose state, the dominant frequency decreases by more than 10%, and the high-frequency energy increases significantly. Through this embodiment of the invention, bolt vibration features can be effectively extracted, providing high-quality data support for subsequent condition assessment and significantly improving the accuracy and reliability of transmission line operation risk prediction.

[0103] S103 analyzes the shift of the dominant frequency component in the frequency feature set, combines wavelet transform to decompose the original high-frequency vibration signal, extracts multi-scale vibration energy distribution features, and constructs a feature dataset reflecting the degree of bolt preload attenuation.

[0104] In this embodiment of the invention, when the dominant frequency component in the frequency feature set deviates from a preset range, it indicates that the bolt may have preload attenuation. Wavelet transform is used to perform multi-scale decomposition of the original vibration signal, extracting energy distribution characteristics of different frequency bands. Combined with statistical analysis and regression models, the relationship between frequency offset and energy characteristics is quantified, providing a reliable basis for bolt condition assessment.

[0105] S1031 uses spectral analysis to detect the deviation between the main frequency component in the frequency feature set and the preset reference frequency, extracts 4096 vibration signal segments, and applies wavelet transform to perform four-level decomposition to generate multi-scale decomposition coefficients.

[0106] In this embodiment of the invention, the preset reference frequency range is 200Hz to 800Hz, determined based on the inherent vibration characteristics of the bolt structure. For example, the reference frequency of a double-angle steel connecting bolt is approximately 350Hz. When the main frequency deviation exceeds 50Hz, wavelet transform analysis is triggered. A 4096-point signal segment is extracted, with a sampling frequency of 2000Hz, covering approximately a 2-second vibration cycle. A four-level decomposition is performed using the db4 wavelet basis function, generating four detail coefficients and one approximation coefficient, respectively reflecting high-frequency details and low-frequency trends.

[0107] S1032 calculates the energy values ​​of the low-frequency band (0Hz to 200Hz), the mid-frequency band (200Hz to 500Hz), and the high-frequency band (500Hz to 1000Hz) based on the decomposition coefficients, sets the energy threshold for each frequency band to 1.5 times the mean, and screens out significant vibration characteristics.

[0108] In this embodiment of the invention, the low-frequency band includes environmental vibration and the structural fundamental frequency, the mid-frequency band reflects the main vibration modes of the bolt, and the high-frequency band indicates local stress changes. The energy of each frequency band is calculated by the sum of squares of the decomposition coefficients, and the threshold is set to 1.5 times the mean value to identify abnormal vibrations. When the energy in the high-frequency band exceeds the threshold, it indicates that the bolt may have localized loosening, requiring further analysis.

[0109] For the over-threshold frequency band, S1033 extracts wavelet coefficients and calculates the mean, variance, and kurtosis factor to generate a frequency band energy feature vector that reflects the intensity, fluctuation, and impulse characteristics of the vibration signal.

[0110] In this embodiment of the invention, the mean of the wavelet coefficients characterizes the vibration intensity, the variance reflects the stability of the energy distribution, and the kurtosis factor captures the sharp pulse characteristics of the signal. When the bolts are loose, the kurtosis factor in the high-frequency band can increase to more than four times the normal level. The eigenvector integrates these statistics, providing multidimensional input for energy distribution analysis.

[0111] Based on the frequency band energy feature vector, S1034 calculates the energy centroid frequency, frequency band energy proportion, and energy dispersion. It then uses a support vector regression model to establish a mapping relationship between frequency offset and energy distribution characteristics, generating a feature dataset that includes preload attenuation indicators.

[0112] In this embodiment of the invention, the energy centroid frequency is calculated using a weighted average to reflect the concentration location of vibration energy; the frequency band energy proportion shows the relative contribution of each frequency band, for example, the mid-frequency band accounts for about 60% under normal conditions; the energy discretization measure quantifies the uniformity of distribution, increasing to over 0.5 when loosened. The support vector regression model uses a radial basis function kernel function with a kernel parameter set to 1.0 and a penalty factor of 50. After training, it maps the relationship between frequency offset and energy features, generating a feature dataset containing frequency offset, energy distribution, and preload attenuation indicators.

[0113] In this embodiment of the invention, bolt vibration characteristic analysis is a key step in the transmission line operation risk prediction method. The selection of the reference frequency is based on a large amount of experimental data to ensure coverage of typical vibration modes of bolts connecting two-piece and four-piece angle steel. A 4096-point signal segment provides sufficient temporal resolution, and the four-level decomposition of wavelet transform balances computational efficiency and feature extraction accuracy. The frequency band division considers the physical characteristics of the bolt vibration signal: the low-frequency band filters environmental noise, the mid-frequency band captures structural response, and the high-frequency band is sensitive to local stress concentration caused by loosening. The energy threshold is set based on statistical laws; 1.5 times the mean effectively distinguishes between normal and abnormal states. The introduction of statistical features enhances the expressive power of the feature vector, especially the sensitivity of the kurtosis factor to the loosening state, making early warning more accurate. The energy concentration index quantifies the vibration distribution change through the center of gravity frequency and discrete quantity; for example, when the center of gravity frequency drops from 400Hz to below 300Hz during loosening, it indicates a significant attenuation of the preload. The support vector regression model ensures the robustness of the mapping relationship by optimizing parameters, with training data covering normal and multi-level loosening states, achieving a classification accuracy of over 93%.

[0114] In practical applications, the mid-frequency energy ratio of normal bolts is stable, while the high-frequency energy ratio is relatively low. When bolts loosen, the low-frequency energy ratio increases to over 35%, and the preload may drop to 75% of the initial value. Through the embodiments of this invention, the extraction and analysis of vibration energy characteristics provide high-precision data support for the dynamic monitoring of bolt conditions, effectively improving the reliability and real-time performance of transmission line operation risk prediction.

[0115] S104 performs feature selection and dimensionality reduction on the initial feature set, combines the bolt finite element model to simulate the stress distribution under different preload attenuation levels, extracts key stress features, and generates a feature dataset containing stress distribution and deformation features.

[0116] In this embodiment of the invention, the computational efficiency of the preliminary feature set is optimized by feature dimensionality reduction, a high-precision bolt finite element model is constructed, the influence of preload variation on stress distribution is simulated, and deep learning methods are used to extract stress concentration features, providing reliable stress feature data for transmission line operation risk prediction methods.

[0117] like Figure 2S1041 calculates the correlation between features in the preliminary feature set and performs dimensionality reduction using principal component analysis to generate low-dimensional feature vectors, providing efficient input for subsequent finite element analysis.

[0118] In this embodiment of the invention, the initial feature set includes multidimensional features such as vibration frequency, strain, and temperature. By calculating the Pearson correlation coefficient and constructing a correlation matrix, it was found that the correlation between vibration frequency and strain is as high as 0.87, indicating the existence of redundant information. Principal component analysis was used to select feature combinations with a cumulative contribution rate of 90%, reducing the feature dimension from 12 dimensions to 4 dimensions, reducing computational complexity by approximately 70%, while retaining the main information.

[0119] S1042 constructs a hexahedral mesh model based on bolt geometric parameters, sets preload boundary conditions and material properties, and calculates the nodal displacement field and stress field distribution using the finite element method to generate stress distribution data.

[0120] In this embodiment of the invention, taking an M20 bolt as an example, the geometric parameters include a diameter of 20 mm, a pitch of 2.5 mm, and a length of 100 mm. A hexahedral mesh is used with a mesh size of 2 mm, and the thread engagement zone is refined to 1 mm, resulting in a total of approximately 16,000 meshes. The preload decreases by 10% progressively from the nominal value of 120 kN, simulating six working conditions. The material used is Q345 steel with an elastic modulus of 206 GPa, a Poisson's ratio of 0.3, and a yield strength of 345 MPa. The friction coefficient of the threaded contact surface is set to 0.15, and normal constraints and tangential friction forces are applied. Finite element calculations show that under the nominal preload, the bolt's axial deformation is approximately 0.18 mm, and the maximum equivalent stress at the thread root reaches 275 MPa, approaching 80% of the yield strength.

[0121] S1043 performs grid resampling on stress distribution data, extracts the coordinates and features of stress concentration areas, and uses a deep autoencoder for feature extraction to generate a dimensionality-reduced feature dataset containing stress distribution and deformation features.

[0122] In this embodiment of the invention, mesh resampling is used to filter mesh cells with stress values ​​exceeding 50% of the yield strength, identifying stress concentration regions, primarily located at the root of the thread and the transition zone between the bolt shank and nut. The depth autoencoder sets the number of nodes in the input layer to match the number of mesh cells, and includes two hidden layers with node counts of 1 / 2 and 1 / 4 of the input layer, respectively. The dimensionality-reduced feature vector integrates stress distribution patterns, deformation amounts, and stress concentration location information, reflecting the impact of preload attenuation on the stress field.

[0123] In this embodiment of the invention, bolt finite element analysis is an important component of the method for predicting the operational risks of transmission lines. Feature reduction, through principal component analysis, effectively reduces redundant information and improves the computational efficiency of the model, making it suitable for processing high-dimensional vibration and stress data. The construction of the finite element model fully considers the bolt geometry and material properties, and mesh refinement ensures high-precision simulation of the threaded region.

[0124] The preload reduction scenario simulates the bolt loosening process, revealing the trend of stress concentration extending from the thread root towards the shank. For example, when the preload decreases to 70% of the nominal value, the stress concentration region expands by 20%, and the stress gradient increases by approximately 15%. Hooke's law provides the theoretical basis for the transformation from displacement to stress fields, ensuring the accuracy of stress calculations. The application of a deep autoencoder further enhances feature extraction capabilities, capturing complex patterns of stress distribution through nonlinear mapping. The dimensionality-reduced feature dataset retains over 95% of the original information.

[0125] In practical applications, the stress at the root of the thread is uniform under normal preload. When loosening occurs, the stress concentration area expands and the non-uniformity intensifies, consistent with the failure mode of bolt fatigue fracture in engineering. Through the embodiments of this invention, the generated stress feature dataset provides high-precision input for bolt condition assessment, effectively supporting the dynamic monitoring and early warning of transmission line operation risks.

[0126] S105 integrates the preliminary feature set and the stress distribution feature set, and uses the support vector machine algorithm to construct a bolt state evaluation model, outputting the degree of bolt preload attenuation.

[0127] In this embodiment of the invention, a method for predicting operational risks of transmission lines is developed by integrating vibration and stress characteristic data. Through outlier detection, feature filtering, and model training, a support vector machine model is constructed to quantify the bolt preload attenuation level, providing technical support for the dynamic assessment of bolt conditions. The entire process includes data standardization, feature optimization, and model validation to ensure the quality of the feature data and the robustness of the model predictions.

[0128] S1051 performs outlier detection and standardization on the preliminary feature set and stress distribution feature set to generate a feature matrix with uniform dimensions, providing high-quality input for feature selection.

[0129] In this embodiment of the invention, the original feature set contains approximately 25 dimensions, such as vibration frequency, stress value, and temperature. Outliers are removed using a three-standard-deviation criterion; approximately 0.2% of data points are removed, for example, points whose vibration frequency exceeds the mean ± 3σ. Maximum and minimum value normalization maps the features to the 0-1 range, eliminating dimensional differences; for example, vibration frequencies from 0 to 1000 Hz, stresses from 0 to 500 MPa, and temperatures from -20 to 80°C are all standardized to a uniform scale.

[0130] S1052 optimizes the feature matrix through variance screening and correlation analysis, generating a feature subset with low redundancy and reducing the computational complexity of the model.

[0131] In this embodiment of the invention, variance analysis is performed on the standardized feature matrix to remove features with variances below 0.05, such as some low-contribution stress distribution features. Pearson correlation coefficient analysis removes redundant features with correlations higher than 0.9; for example, when the correlation between vibration frequency and strain reaches 0.88, only one feature is retained. The final feature subset is reduced from 25 dimensions to 10 dimensions, retaining more than 90% of the information content.

[0132] S1053 classifies and labels the feature subsets according to the degree of preload decay, and uses stratified sampling to divide the training set and validation set to provide a balanced dataset for model training.

[0133] In this embodiment of the invention, the preload decay is divided into five levels: 0-20% is slight loosening, 20-40% is moderate loosening, 40-60% is significant loosening, 60-80% is severe loosening, and 80-100% is complete loosening. Stratified sampling is used to divide the training set and validation set in an 8:2 ratio to ensure a balanced distribution of samples at each level. The training set contains approximately 4000 samples, and the validation set contains 1000 samples.

[0134] S1054 utilizes cross-validation and grid search to optimize support vector machine parameters, constructs a classification hyperplane, and generates a high-precision bolt condition assessment model.

[0135] In this embodiment of the invention, five-fold cross-validation is used to optimize the radial basis function parameters of the support vector machine, with a search range of gamma from 0.01 to 10 and a penalty coefficient C from 0.1 to 100. The optimal parameter combination is gamma = 0.6 and C = 12, achieving a classification accuracy of 92%. The model fits the classification hyperplane in the feature space using training data to distinguish different preload attenuation levels.

[0136] S1055 makes predictions on the validation set, calculates the mean square error, constructs confidence intervals, and generates an evaluation result that includes the preload attenuation level and the prediction reliability.

[0137] In this embodiment of the invention, the validation set prediction results are compared with the true labels, and the mean squared error is used to evaluate the model accuracy. A 95% confidence interval is constructed based on the error distribution; the prediction interval width for slight loosening is approximately ±4%, while for severe loosening it reaches ±12%, reflecting that the higher the degree of loosening, the stronger the nonlinear characteristics. The evaluation results include the preload attenuation level and the confidence interval.

[0138] In this embodiment of the invention, the construction of the bolt condition assessment model is the core of the transmission line operation risk prediction method. Outlier detection and standardization ensure data quality, eliminating outliers to reduce noise interference, and normalization to eliminate dimensional differences, laying the foundation for feature selection. Variance screening and correlation analysis optimize the feature subset, reducing the computational load by approximately 60%, while retaining key information, such as the contribution rates of vibration dominant frequency and stress concentration to preload assessment, which are 35% and 30%, respectively. Stratified sampling ensures the representativeness of the dataset and avoids model bias towards a particular level.

[0139] Support Vector Machines (SVMs) optimize parameters through grid search, while radial basis function (RBF) kernels enhance nonlinear mapping capabilities, and a classification hyperplane effectively separates different loosening states. In practical applications, the model demonstrates high prediction accuracy and a narrow error range for slight and moderate loosening; severe loosening exhibits a wider prediction range due to the nonlinearity of its mechanical behavior, consistent with the stress concentration and fatigue failure patterns caused by bolt loosening in engineering. Through this embodiment, the bolt condition assessment model can output the preload attenuation level in real time, providing accurate early warning for the safe operation of transmission lines and improving power grid reliability.

[0140] S106 integrates bolt preload attenuation with historical inspection records and uses Bayesian inference to optimize the parameters of the bolt condition assessment model, generating an assessment model that can dynamically reflect the bolt's operating status.

[0141] In this embodiment of the invention, the transmission line operation risk prediction method utilizes historical inspection data to supplement and update model parameters. Through data interpolation, statistical modeling, and probabilistic inference, it enhances the adaptability and prediction accuracy of the bolt condition assessment model. The processing includes data completion, attenuation characteristic analysis, and iterative parameter updates to ensure that the model can reflect the dynamic changes in bolt preload in real time.

[0142] S1061 extracts key information from historical bolt inspection records, uses cubic spline interpolation to fill in missing data, and generates a continuous time series dataset.

[0143] In this embodiment of the invention, the inspection record includes information such as inspection timestamp, preload measurement value, ambient temperature, and humidity, and is typically collected every 30 days. If the data gap is less than 90 days, cubic spline interpolation is used to complete the data, generating a smooth curve based on the trend of the preceding and following data points; missing segments exceeding 90 days are marked as invalid. The time series dataset generated after completion covers at least 12 months of monitoring data, providing a complete foundation for subsequent analysis.

[0144] S1062 segments the time series dataset according to the inspection cycle, calculates the statistical characteristics of the preload decay rate, constructs a normal distribution model, and generates prior probability information.

[0145] In this embodiment of the invention, the dataset is segmented according to a 30-day inspection cycle, with each segment containing approximately 10 to 15 data points. The mean and variance of the preload decay rate for each segment are calculated. For example, under normal operating conditions, the average annual decay rate of an M20 bolt is approximately 6%, with a standard deviation of 1%. Based on the statistical results of multiple data segments, a normal distribution model of the decay rate is constructed, generating a prior probability density function. The mean and covariance matrix serve as the initial inputs for Bayesian inference.

[0146] S1063 uses the Gibbs sampling method to iteratively update the prior probability distribution, and combines the likelihood function to calculate the posterior probability, generating an optimized state evaluation parameter matrix.

[0147] In this embodiment of the invention, Gibbs sampling extracts 1200 sample points per iteration, ensuring posterior distribution convergence with a variation range of less than 0.8%. The likelihood function is based on the deviation between the measured preload and the model's predicted value, assuming the deviation follows a zero-mean normal distribution. Historical data is weighted by time distance, with the weights decaying exponentially over time, with a decay coefficient of 0.12, reducing the weight of data from 180 days ago to 0.15 of the initial value. The parameter matrix is ​​updated based on the posterior probability, including the preload decay threshold, confidence interval, and warning level limits; for example, a 15% decay indicates a state of concern, and a 30% decay indicates an alarm state.

[0148] In this embodiment of the invention, Bayesian inference optimization is the core component of the transmission line operation risk prediction method. Cubic spline interpolation completes the data through nonlinear fitting, preserving the physical trend of preload changes, and is more suitable for capturing nonlinear fluctuations caused by environmental factors compared to linear interpolation. Piecewise statistical analysis reveals the temporal law of bolt preload decay; for example, the decay rate can increase to 8% under high temperature and high humidity conditions, providing the model with environmental adaptability.

[0149] The iterative process of Gibbs sampling approximates the true distribution through multiple rounds of sampling, offering higher computational efficiency compared to the Markov chain Monte Carlo method, making it suitable for real-time update scenarios. The exponential decay design of time weights emphasizes the priority of recent data, ensuring the model's sensitivity to the current bolt condition; for example, the weight of the latest 30 days' data accounts for over 60% of the total weight. The optimized parameter matrix dynamically adjusts the warning threshold. For instance, after one year of data optimization for a certain line, the attention threshold increased from 15% to 16.5%, reflecting the differences in bolt materials and connection structures in actual operation. Confidence intervals are calculated based on the quantiles of the posterior distribution; the 95% confidence interval width narrowed from the initial ±12% to ±9%, indicating that data accumulation improves prediction reliability. In practical applications, the model can be updated in real-time based on newly added inspection data, dynamically adjusting the decay threshold and warning level, providing technical support for accurate monitoring and risk warning of transmission line bolt conditions.

[0150] S107 uses the state score output by the optimized bolt condition assessment model, combined with vibration signals and stress analysis, to calculate the load sharing characteristics, vibration frequency drift, and stiffness changes at the bolt connection, generating a comprehensive feature dataset that includes stress distribution, frequency deviation, and load transfer characteristics.

[0151] In this embodiment of the invention, the transmission line operation risk prediction method utilizes bolt condition scoring, combined with mechanical analysis and neural network models, to quantify the uneven load distribution and dynamic response characteristics of the connection structure, providing data support for the accurate assessment of bolt loosening risk. The processing encompasses stress distribution calculation, frequency analysis, and load transfer modeling, ensuring the high quality and reliability of the feature dataset.

[0152] S1071 constructs a stress distribution model of the connection structure based on bolt condition scoring, calculates the axial and circumferential stress of each bolt, generates a stress distribution matrix, and quantifies the load distribution unevenness through stress concentration factor.

[0153] In this embodiment of the invention, taking a four-bolt angle steel connection as an example, the preload data from the condition scoring is used to calculate the axial and circumferential stresses of each bolt through the stress superposition principle, generating a stress distribution matrix. Under normal conditions, the axial stress is distributed between 270 and 310 MPa, and the ratio of the maximum stress to the average stress, i.e., the stress concentration factor, is approximately 1.1. When the bolts loosen, the stress concentration factor can rise to 1.9, the stress of the loosened bolt decreases by about 55%, and the stress of adjacent bolts increases by 35%, indicating that the unevenness of load distribution is significantly aggravated.

[0154] S1072 performs spectral analysis on the vibration signal at the connection, extracts the frequency response characteristics, calculates the stiffness attenuation vector, and reflects the influence of bolt preload changes on structural stiffness.

[0155] In this embodiment of the invention, the vibration signal is analyzed using Fourier transform to extract the amplitude and phase spectra within the range of 0 to 1000 Hz, generating a frequency response curve. Under normal conditions, the natural frequency is between 200 and 400 Hz. When the preload is reduced by 20%, the frequency drifts by approximately 6 Hz, and when reduced by 50%, it drifts to 15 Hz. The stiffness attenuation vector is calculated based on the amplitude reduction of the resonance peak and the curve broadening. Loosening the bolts causes an overall stiffness decrease of approximately 28%, providing dynamic characteristics for the model.

[0156] S1073 constructs a gridded stress field model, adopts a hexahedral element discretization connection structure, and combines cubic spline interpolation to extract the deformation characteristics of the contact surface, generating refined stress distribution and deformation data.

[0157] In this embodiment of the invention, based on material fatigue and preload deviation parameters, hexahedral elements are used to spatially discretize the M20 bolt connection. The initial mesh size is 2 mm, and the contact surface is refined to 1 mm, resulting in a total of approximately 22,000 meshes. A stress interpolation algorithm is used to calculate the normal and tangential stresses at the refined mesh points, generating refined stress distribution data. Cubic spline interpolation is used to fit the deformation surface of the contact surface, extracting the deformation of 64 feature points. Under normal conditions, the deformation at the hole edge is approximately 0.18 mm, while the local deformation increases to 0.3 mm when loosened.

[0158] S1074 utilizes neural networks and recurrent neural networks to fuse load sharing unevenness, stiffness attenuation vector, and deformation characteristics, calculates the load transfer coefficient between nodes, and generates a comprehensive feature dataset.

[0159] In this embodiment of the invention, a three-layer feedforward neural network is constructed. The input layer has 12 nodes, containing the stress, frequency, and stiffness features of four bolts. The hidden layer has 8 nodes using the hyperbolic tangent activation function, outputting the load transfer relationship. Simultaneously, a five-layer recurrent neural network is constructed. The input layer has 68 nodes (64 deformation feature points plus 4 frequency drift values), and the number of hidden layer nodes is 32, 16, 8, and 4 respectively. Stochastic gradient descent is used for optimization, with a learning rate of 0.01. After 6000 iterations, the mean square error is less than 0.04. The two networks are jointly outputting the stress distribution ratio, frequency deviation, stiffness change, and load transfer coefficient of each bolt, generating a comprehensive feature dataset.

[0160] In this embodiment of the invention, load sharing characteristic analysis is a key step in the method for predicting the operational risks of transmission lines. The stress distribution matrix accurately quantifies the load distribution among bolts using mechanical principles. Under loosening conditions, stress redistribution leads to overload of adjacent bolts, increasing the risk of fatigue fracture. Frequency response analysis captures the nonlinear frequency drift caused by preload decay, and combined with the stiffness decay vector, reveals changes in the structure's dynamic characteristics. For example, when a single bolt is completely loosened, the stiffness decreases by 30%, directly affecting vibration stability.

[0161] Mesh-based stress field modeling, through high-resolution discretization combined with cubic spline interpolation, accurately describes the non-uniformity of contact surface deformation. The significant increase in deformation near loose bolts provides a sensitive indicator for early warning. The fusion of neural networks and recurrent neural networks leverages the fast mapping capability of feedforward networks and the temporal modeling ability of recurrent networks to accurately characterize the dynamic process of load transfer. For example, the transfer coefficient of loose bolts decreases to 0.6, while that of adjacent bolts increases to 1.4, reflecting an imbalance in load distribution. Cross-validation shows that the model achieves a 96% accuracy rate in identifying severe loosening and 88% in identifying slight loosening, with a mean square error of less than 0.05, validating the reliability of feature extraction. In practical applications, this method can monitor the load distribution and vibration characteristics of bolted connections in real time, providing high-precision data support for the safe operation of transmission lines.

[0162] S108 constructs a Monte Carlo simulation model using a load-sharing feature set to simulate the failure probability of bolts under service conditions. It then combines a deep neural network to predict failure trends and generate a comprehensive bolt condition assessment dataset that includes condition levels and confidence intervals.

[0163] In this embodiment of the invention, the transmission line operation risk prediction method utilizes load sharing and vibration frequency drift data, and through probabilistic modeling and neural network analysis, quantifies bolt failure risk, providing an accurate assessment basis for the safe operation of transmission lines. The processing includes probability distribution modeling, Monte Carlo sampling, and failure probability prediction to ensure the reliability and operability of the assessment results.

[0164] S1081 constructs a sample space based on the load-sharing feature set, uses the maximum likelihood estimation method to generate a joint probability density function, and generates initial state probability distribution data to provide input for Monte Carlo simulation.

[0165] In this embodiment of the invention, the load-sharing feature set includes multi-dimensional features such as stress distribution ratio, frequency deviation, and stiffness variation. Maximum likelihood estimation is used to analyze the load distribution and frequency drift, determining the joint probability density function. The load distribution approximately follows a log-normal distribution with a mean of 0.92 times the design preload and a standard deviation of 0.12 times; the frequency drift follows a normal distribution with a standard deviation of approximately 5 Hz. The generated initial state probability distribution data provides a statistical basis for subsequent sampling.

[0166] S1082 randomly samples from the initial probability distribution using the Monte Carlo method, calculates the stress amplitude and fatigue cycle number of the sample points, and generates fatigue damage values ​​based on the cumulative damage criterion.

[0167] In this embodiment of the invention, the Monte Carlo sampling number is set to 12,000 points. Samples are drawn from the probability distribution, and the stress amplitude and cycle number of each sample point are calculated. For Q345 steel bolts, a stress amplitude of 200 MPa corresponds to a fatigue life of approximately 1.2 million cycles. For every 50 MPa increase, the life decreases by approximately 80%. When the preload drops to 70% of the initial value, the stress amplitude increases by approximately 35%, accelerating fatigue damage. Based on the linear cumulative damage criterion, the damage value of each sample point is calculated, providing data for failure probability modeling.

[0168] The S1083 constructs a deep convolutional neural network, trains sample point data, predicts the failure probability under different working conditions, generates a failure probability curve, and reflects the service reliability of the bolt.

[0169] In this embodiment of the invention, the deep neural network comprises three convolutional layers. The input layer dimension is consistent with the number of features. The first layer has 64 3×3 convolutional kernels, while the second and third layers have halved to 32 and 16 kernels, respectively. Each layer is followed by a batch normalization layer and a ReLU activation function to prevent gradient vanishing. Training employs stochastic gradient descent with an initial learning rate of 0.008, which decays exponentially and converges after 8000 iterations. The failure probability curve shows that the failure probability of new bolts is close to zero, increasing to 0.06 after one year of service, reaching 0.18 after two years, and rising to 0.40 after three years, consistent with the actual failure trend.

[0170] S1084 uses cross-validation to evaluate the model's prediction accuracy, classifies bolt condition levels based on failure probability curves, and generates a comprehensive evaluation dataset that includes timestamps and evaluation indicators.

[0171] In this embodiment of the invention, the relative error between the predicted and measured values ​​is calculated using five-fold cross-validation. The error for bolts within two years of service is approximately 12%, increasing to 22% after three years, reflecting the complex failure mechanism of long-term service. The condition levels are divided into intact, failure probability <0.1, warning 0.1-0.3, and danger >0.3. The 95% confidence interval is calculated based on the error distribution, with a width of approximately ±10% of the predicted value. The comprehensive evaluation dataset includes timestamps accurate to the hour, evaluation indicators, stress levels, frequency offsets, failure probabilities, prediction results, condition levels, and confidence intervals.

[0172] In this embodiment of the invention, the combination of Monte Carlo simulation and deep neural networks is the core advantage of the transmission line operation risk prediction method. Maximum likelihood estimation accurately describes the probabilistic characteristics of load and frequency through statistical modeling, and the generated joint probability density function provides a reliable basis for sampling in the high-dimensional feature space. Monte Carlo sampling simulates the mechanical behavior of bolts under various working conditions using a large sample size. The cumulative damage criterion quantifies the impact of stress amplitude and cycle number on fatigue life; for example, a 30% reduction in preload shortens fatigue life by approximately 50%, providing a physical basis for calculating failure probability.

[0173] Deep convolutional neural networks, through multi-layer convolution and batch normalization, effectively capture the nonlinear relationships between features, improving prediction accuracy by approximately 15% compared to traditional regression models. The generated failure probability curve intuitively reflects the aging trend of bolts over service time; for example, bolts that have been in service for 3 years have a failure probability of up to 0.45 under high load conditions, indicating the need for early maintenance. Cross-validation ensures the model's generalization ability, and error analysis reveals the complexity of long-term service bolts affected by environmental and load fluctuations. The state level classification and confidence interval design facilitate engineering applications, and timely identification of early warning states can reduce the risk of sudden failures by more than 70%. In practical applications, the comprehensive evaluation dataset supports real-time monitoring and maintenance decisions, providing efficient and accurate technical support for the safe management of transmission line bolts.

[0174] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the concept of this application. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for predicting the operational risks of transmission lines based on big data mining, characterized in that, Mainly includes: (1) Real-time acquisition of high-frequency vibration data and environmental parameters at the bolted connection of transmission line towers to generate a time series dataset containing vibration frequency, amplitude and temperature; (2) Extract the main frequency component of the vibration frequency from the time series dataset using the Fourier transform algorithm, and perform linear correction of the vibration frequency based on the thermal expansion coefficient and temperature of the bolt material to generate a frequency feature set; (3) If the main frequency component in the frequency feature set deviates from the preset frequency range, the original high-frequency vibration signal is decomposed by wavelet transform, the local vibration energy distribution features of different frequency bands are extracted, and the local vibration energy distribution features are processed to obtain a preliminary feature set reflecting the degree of bolt preload attenuation. (4) Perform feature selection and dimensionality reduction on the preliminary feature set, input it into the pre-established bolt finite element model, simulate the stress distribution of the bolt connection under different preload attenuation levels, extract stress distribution features, and generate a stress distribution feature set; (5) Construct a bolt condition assessment model based on the preliminary feature set and stress distribution feature set, assess the bolt condition, and output the degree of bolt preload attenuation. (6) Combine the bolt preload attenuation with the obtained bolt historical inspection record data, and use Bayesian inference method to update the parameters of the bolt condition assessment model to obtain the optimized bolt condition assessment model. (7) Based on the state scores of each bolt output by the optimized state assessment model, calculate the overall load sharing and vibration frequency drift of the bolt connection and generate a load sharing feature set. (8) Establish a Monte Carlo simulation model based on the load sharing feature set, simulate the failure probability of bolts under service conditions based on the current load sharing and vibration frequency drift, and generate a comprehensive evaluation result of bolt status based on the failure probability.

2. The method for predicting the operational risk of transmission lines based on big data mining according to claim 1, characterized in that, Step (1) specifically includes: The vibration signal output by the sensor array is collected, and the vibration signal is filtered by a digital bandpass filter to obtain a filtered vibration signal. Fourier transform is performed on the filtered vibration signal to obtain spectral features. The main vibration frequency points and their amplitudes are extracted from the spectral features, and the ambient temperature data is recorded to obtain the original monitoring dataset. The support vector regression algorithm is used to establish a mapping relationship between vibration feature vectors and temperature changes in the original monitoring dataset; the sliding window method is used to segment the original monitoring dataset, with the window length set to one thousand data points, and outliers are removed by median filtering to obtain the preprocessed monitoring dataset.

3. The method for predicting the operational risk of transmission lines based on big data mining according to claim 1, characterized in that, Step (2) is as follows: The Hanning window function is used to process vibration time series data. The windowed signal is subjected to spectrum analysis by fast Fourier transform to obtain the dominant frequency feature vector. The collected vibration time series data is windowed using the Hanning window function with the window length set to the length of the vibration data sequence. The spectrum analysis of the windowed signal is performed by fast Fourier transform, and the three frequency peaks with the largest amplitude are extracted from the spectrum as the dominant frequency feature vector. Based on the displacement of the frequency peak value in the main frequency feature vector as a function of temperature, and combined with the thermal expansion coefficient of the bolt material, the temperature-frequency mapping relationship is obtained, and the frequency drift vector is obtained. The frequency drift vector is curve-fitted by the least squares method to obtain a temperature compensation function containing a slope term and an intercept term. The temperature compensation function is then applied to the main frequency feature vector to eliminate the influence of temperature on the frequency and obtain the corrected frequency feature vector. Wavelet decomposition is performed on the corrected frequency feature vector. Wavelet basis functions are selected to perform three-level decomposition of the signal and extract the energy coefficients of each frequency component. The energy coefficients are then classified using the support vector machine method to obtain a frequency feature set containing the frequency value of the main frequency component, the energy value, and the temperature compensation coefficient.

4. The method for predicting the operational risk of transmission lines based on big data mining according to claim 1, characterized in that, Step (3) specifically includes: The original vibration signal is decomposed using wavelet transform to obtain the multi-scale decomposition coefficients of the vibration signal. The frequency band energy values ​​of the low-frequency band, mid-frequency band, and high-frequency band are calculated based on the multi-scale decomposition coefficients, and the frequency band energy values ​​are compared with preset energy thresholds. For frequency bands exceeding the preset energy threshold, the corresponding wavelet coefficients are extracted, and the frequency band energy feature vector is obtained by calculating the mean, variance, and kurtosis factor of the wavelet coefficients. Based on the frequency band energy feature vector, calculate the energy distribution characteristics including the energy centroid frequency value, frequency band energy ratio, and energy dispersion. Use a support vector regression model to establish the mapping relationship between the frequency offset and the energy distribution characteristics to obtain a preliminary feature set reflecting the degree of bolt preload attenuation.

5. The method for predicting the operational risk of transmission lines based on big data mining according to claim 1, characterized in that, Step (4) specifically includes: A hexahedral mesh model is constructed based on the bolt diameter and thread pitch, with the mesh size set to one-tenth of the bolt diameter. A preload boundary condition is applied to the bolt connection surface, with the preload value decreasing stepwise from the nominal preload. The force model is obtained by applying the preload boundary condition to the bolt connection surface of the mesh model. The elastic modulus, Poisson's ratio, and yield strength parameters of the bolt material are set. Threaded pair contact constraints are applied to the mesh model. The nodal displacement field is calculated using the finite element method. The nodal displacement field is then converted into stress field distribution data using Hooke's law. The data includes the normal stress and shear stress values ​​at the contact surface between the bolt shank and the nut. The stress field distribution data is resampled by grid, and the coordinates of the stress concentration region are obtained by extracting grid cells whose stress values ​​exceed the yield strength. A deep autoencoder is used to extract features from the coordinates of the stress concentration region. By setting the number of hidden layer nodes of the encoder to half of the input dimension, a stress distribution feature set is obtained. The stress distribution feature set mainly includes stress distribution, deformation, and stress concentration location.

6. The method for predicting the operational risk of transmission lines based on big data mining according to claim 1, characterized in that, Step (5) specifically includes: Outlier detection is performed on the preliminary feature set and stress distribution feature set to remove data points that exceed three times the standard deviation. The maximum and minimum value normalization method is used to map the feature values ​​to the zero-to-one interval to generate a standardized feature matrix. Variance screening is performed based on the standardized feature matrix to remove feature columns with variance values ​​less than a preset threshold. Pearson correlation coefficient analysis is then used to obtain a feature subset; that is, redundant features with correlation values ​​higher than 0.9 are removed using the Pearson correlation coefficient to obtain a feature subset. The feature subset is graded and labeled according to the preload attenuation amount, and the preload attenuation amount is divided into grade intervals. The training set and validation set are obtained by stratified sampling. Based on the training set, the parameters of the support vector machine are selected using cross-validation, and the parameters of the classification hyperplane are obtained through grid search to form a bolt state evaluation model. For the validation set, a prediction is made, and the mean square error is calculated based on the predicted value and the true label. The mean square error is used to construct a confidence interval, and an evaluation result containing the preload attenuation degree and the confidence interval is generated.

7. The method for predicting the operational risk of transmission lines based on big data mining according to claim 1, characterized in that, Step (6) specifically includes: Based on the bolt inspection records, the preload measurement value and environmental parameter records are obtained. The missing data in the inspection records are supplemented by cubic spline interpolation to obtain a time series dataset. The time series dataset is segmented according to the monitoring time, and the mean and variance of the preload decay rate in the dataset are obtained by statistical calculation method, so as to obtain the normal distribution model of the preload decay parameter. The prior probability density function of the preload attenuation parameter is calculated based on the normal distribution model. The mean and covariance matrix of the function are obtained by calculating the probability density function, which can be used as prior information for Bayesian inference. The prior probability density function is iteratively updated using the Gibbs sampling method. In each iteration, a preset number of sample points are extracted, and the likelihood function value corresponding to the sample points is calculated. The posterior probability distribution is obtained by calculating the likelihood function value and the prior probability. Weighting coefficients are set for historical data according to time distance, and the time weight decreases exponentially with the increase of distance. The bolt condition assessment model parameters are updated based on the posterior probability distribution, including the preload attenuation threshold, confidence interval range, and warning level limit, to obtain the optimized bolt condition assessment model.

8. The method for predicting the operational risk of transmission lines based on big data mining according to claim 1, characterized in that, Step (7) specifically includes: The stress of the connection structure is calculated based on the bolt condition assessment score. The axial and circumferential stresses of each bolt are calculated. The overall stress distribution at the connection is obtained by using the stress superposition principle, and the stress distribution matrix at the connection is obtained. The stress concentration factor is calculated using the maximum stress value and the average stress value in the stress distribution matrix, and the load distribution non-uniformity is obtained based on the stress concentration factor. Fourier transform is performed on the vibration signal at the connection point. The Fourier transform is then used to perform spectral analysis on the vibration signal at the connection point, extract the amplitude spectrum and phase spectrum of the vibration signal, and calculate the deviation value between the vibration frequency of each bolt and the frequency of the initial assembly state. The stiffness attenuation vector is calculated based on the frequency response curve of the vibration signal; the amplitude response at the natural frequency point is extracted, the change in structural stiffness is calculated, and the stiffness attenuation vector is generated. A neural network is constructed, which includes an input layer, a hidden layer, and an output layer. The input layer receives the load sharing unevenness and stiffness attenuation vector. The load sharing feature set is obtained through iterative calculation of the neural network. The load sharing feature set includes the stress distribution ratio, frequency deviation value, and stiffness change of each bolt.

9. The method for predicting the operational risk of transmission lines based on big data mining according to claim 8, characterized in that, Also includes: A dynamic stress distribution matrix is ​​established by obtaining material fatigue parameters and preload deviation parameters from the bolt condition assessment score. This matrix contains contact pressure gradient data for each node in the bolt group. Contact surface deformation characteristics are extracted from the dynamic stress distribution matrix. These characteristics, along with vibration frequency drift, are input into the load distribution calculation model, outputting the dynamic load transfer coefficient for each bolt node. Based on these coefficients, a load sharing feature set is generated, specifically including: A gridded nodal stress field is constructed based on the material fatigue parameters and preload deviation parameters in the bolt condition score. The initial stress distribution matrix is ​​obtained by spatial discretization of the bolt connection using hexahedral elements. Discretization involves spatial discretization of the bolt connection using hexahedral elements and calculation of the normal and tangential stresses at each grid point. The initial stress distribution matrix is ​​subjected to mesh refinement processing, and the stress values ​​of the refined mesh points are calculated using a stress interpolation algorithm to obtain refined stress distribution data; The refined stress distribution data is fitted with a cubic spline interpolation method to fit the deformation surface of the contact surface, and the deformation of key feature points is extracted from the deformation surface. A recurrent neural network is trained based on the deformation and frequency drift data. The network parameters are optimized using stochastic gradient descent to obtain the load transfer coefficient between nodes. A load sharing feature set is generated based on the load transfer coefficient and stress gradient data. The recurrent neural network is a five-layer recurrent neural network. The input layer contains deformation features and frequency drift data. The number of nodes in the hidden layer is halved with each layer. The network parameters are optimized using stochastic gradient descent. The load transfer coefficient between nodes is calculated based on the trained neural network. The coefficient value represents the load distribution ratio between adjacent nodes. A load distribution matrix is ​​constructed based on the load transfer coefficient, and the stress distribution weight of each node is calculated by combining the stress gradient data to generate a feature dataset containing node stress distribution, deformation, and transfer coefficient. Cross-validation is used to verify the accuracy of the feature dataset, and the mean square error between the predicted and measured values ​​is calculated to ensure the reliability of the dataset.

10. The method for predicting the operation risk of transmission lines based on big data mining according to claim 1, characterized in that, Step (8) specifically includes: The maximum likelihood estimation method is used to process the load distribution and frequency drift data to obtain the joint probability density function, and the initial state probability distribution data is generated based on the joint probability density function. Random sampling is performed based on the initial state probability distribution data to obtain a set of sample points. The stress amplitude and the number of cycles are calculated for the set of sample points. During random sampling, the number of Monte Carlo samplings is set, and sample points are randomly selected from the initial state probability distribution. The stress amplitude and the number of cycles corresponding to each sample point are calculated, and the fatigue damage value is calculated according to the linear cumulative damage criterion. A prediction model is constructed using a deep neural network, which contains three convolutional layers. The input dimension of each convolutional layer is the same as the number of features. The number of kernels in each convolutional layer is set to half that of the previous layer. Batch normalization is used to prevent gradient vanishing. The deep neural network is used to train the sample point set to obtain the trained neural network model. The failure probability under working conditions is calculated for the trained neural network model to generate a comprehensive evaluation result of the bolt condition. The failure probability under different working conditions is calculated based on the trained neural network to generate a failure probability curve. The accuracy of the prediction result is evaluated by using the five-fold cross-validation method, and the relative error between the predicted value and the measured value is calculated. Based on the failure probability curve, the bolt status is divided into three levels: intact, warning, and dangerous. The confidence level of each level is calculated. A comprehensive evaluation dataset containing status level, failure probability value, and confidence interval is generated. The dataset format includes three fields: timestamp, evaluation index, and prediction result.

Images