Cable defect monitoring and positioning method based on big data sample statistical analysis

Abstract

The invention relates to the technical field of cable monitoring, and discloses a cable defect monitoring and positioning method based on big data sample statistical analysis, and the method comprises the following modules: a multi-source sensor array module which is used for setting a regular triangle array according to a specific interval along a cable path, an ultrahigh-frequency partial discharge sensor, a fiber grating temperature sensor and an acceleration vibration sensor are integrated, and partial discharge, temperature and vibration multi-dimensional physical signals are synchronously collected under the sampling frequency meeting the Shannon sampling theorem; and the space-time synchronization module is used for the GPS clock module and the I EEE1588 protocol to perform nanosecond synchronization of the sensor array, and eliminating inherent delay and global clock offset of the sensor through timestamp calibration. According to the cable defect monitoring system, comprehensive and accurate data acquisition, high-quality data preprocessing and multi-source information effective fusion are realized through multi-module collaborative operation, and the accuracy and reliability of the whole cable defect monitoring work are improved. The problem that cable defect monitoring is not comprehensive and accurate enough is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cable monitoring, in particular to a defect monitoring and positioning method for cables based on big data sample statistical analysis. BACKGROUND

[0002] In the field of power systems, cables are the key carriers for power transmission, and their safe and stable operation is crucial. However, current cable defect monitoring faces many challenges. Traditional monitoring methods often rely on data collected by a single type of sensor for analysis, such as using only partial discharge signals obtained by partial discharge sensors to determine the state of the cable. This single signal monitoring method has great limitations and can only reflect the situation of a certain aspect of the cable, and cannot fully and accurately grasp the real defect condition and overall operation state of the cable. Due to the complex and variable operating environment of the cable, various types of defects may occur inside the cable, and different defects will be reflected in aspects such as partial discharge, temperature change, and vibration. Relying on one-sided information makes it difficult to achieve accurate monitoring and timely warning, which affects subsequent operation and maintenance decisions and easily leads to cable faults that cannot be discovered and handled in time, seriously threatening the reliability of power transmission. SUMMARY

[0003] To overcome the shortcomings of the prior art, the present application provides a defect monitoring and positioning method for cables based on big data sample statistical analysis, which solves the problem of incomplete and inaccurate cable defect monitoring.

[0004] To achieve the above purpose, the present application realizes the following technical scheme: a defect monitoring system for cables based on big data sample statistical analysis, comprising the following modules:

[0005] A multi-source sensor array module is used to set up a regular triangle array along the cable path at a certain interval, integrate ultra-high frequency partial discharge sensors, fiber optic grating temperature sensors and acceleration vibration sensors, and synchronously collect multi-dimensional physical signals of partial discharge, temperature and vibration at a sampling frequency that meets the Shannon sampling theorem;

[0006] A time-space synchronization module is used for GPS clock module and IEEE1588 protocol to synchronize the sensor array at nanosecond level and eliminate the inherent delay of the sensor and the global clock offset through timestamp calibration;

[0007] A data preprocessing module is used to perform wavelet denoising, missing value repair and dimension reduction processing on the collected cable signals, remove noise through a soft threshold function, and repair missing data by combining LSTM and KNN algorithms;

[0008] A multi-source data fusion module is configured to define a defect identification framework, generate a basic probability assignment function by using a support vector machine, and fuse partial discharge, temperature, and vibration multi-source feature evidences by using a Dempster combination rule.

[0009] A deep learning defect identification module is configured to a CNN-LSTM network model, automatically extract and learn the fused multi-source data features, and classify and identify different types of cable defects.

[0010] A Bayesian early warning module is configured to a state space model and Kalman filter iterative calculation, evaluate the defect severity based on cable defect related data, divide early warning levels, and trigger corresponding early warnings.

[0011] By using the above technical solutions, through multi-module cooperation, accurate signal acquisition, high-precision time synchronization, data quality improvement, effective multi-source data fusion, accurate defect classification, and dynamic accurate early warning are achieved, which comprehensively improves the cable monitoring and operation level and ensures the stable and reliable power supply of the power system.

[0012] Preferably, the multi-source sensor array module includes a regular triangle array arranged at every L distance along the cable path, wherein L satisfies the formula υ p is the phase velocity of electromagnetic waves in the cable, f s is the sampling frequency and satisfies the Shannon sampling theorem f s = 2B, B is the signal bandwidth, and N is the number of sampling points per cycle. The array includes an ultra-high frequency partial discharge sensor, an optical fiber grating temperature sensor, and an acceleration vibration sensor.

[0013] Preferably, the space-time synchronization module includes a GPS clock module and an IEEE1588 protocol to realize sensor array synchronization, with a synchronization accuracy of ≤10 ns, and a timestamp calibration formula t' = t - Δt i + Δt0 is used to calibrate the data timestamp, wherein t is the original timestamp, Δt i is the inherent delay of the sensor, and Δt0 is the global clock offset.

[0014] Preferably, the data preprocessing module includes using a db4 wavelet to decompose a partial discharge signal for 5 layers, and using a soft threshold function to denoise, wherein the threshold value σ j is the standard deviation of the jth layer coefficient, and n is the signal length; missing data is repaired by combining time and space, the time dimension is predicted by an LSTM neural network, the spatial dimension is interpolated based on KNN algorithm using adjacent sensor data, and the fusion weight satisfies w t +w s = 1.

[0015] Preferably, the multi-source data fusion module comprises a definition identification framework Θ={D1, D2, D3, D4, θ}, D1-D4 are different types of defects, and θ is a normal state; a basic probability assignment function is generated by a support vector machine Wherein d i is the distance of the sample to the i-th hyperplane, and γ is a kernel function parameter; a Dempster combination rule is used to perform evidence fusion, K is a conflict coefficient and

[0016] Preferably, the deep learning defect identification module comprises a CNN-LSTM network model, which comprises three convolutional blocks, each convolutional block outputs x l =ReLU(W l *x l-1 +b l ), W l is a convolution kernel, b l is a bias, and * represents convolution operation; two bidirectional LSTM layers, the LSTM unit state update satisfies Wherein i t , f t , o t are input gate, forget gate and output gate respectively, c t is the cell state, h t is the hidden state, σ is the Sigmoid function, and is the element-level multiplication; two fully connected layers, the output layer is realized by a Softmax function to realize defect classification, and m is the number of defect types.

[0017] Preferably, the Bayesian early warning module comprises a state space model Wherein is an autoregressive coefficient, β is an observation sensitivity coefficient, ω t ~N(0,Q) is state noise, and υ t ~N(0,R) is observation noise; a state estimation value is calculated by Kalman filtering iteration, including a prediction step an update step Wherein is a state estimation value, P is an error covariance, and K is a Kalman gain; according to a defect severity index S, an early warning level is divided, and when S≥0.3, an early warning is triggered.

[0018] A positioning method for defect monitoring of a cable based on big data sample statistical analysis, for a kind of based on big data sample statistical analysis cable defect monitoring system, comprising the following steps:

[0019] S1, through the equilateral triangle multi-source sensor array arranged at intervals along the cable path, time information of signals generated by defects reaching different sensors is acquired, and a signal time difference is determined;

[0020] S2, high-precision sensor array synchronization realized by means of a space-time synchronization module is used to construct hyperbolic equation groups for describing the relationship between the defect position and the sensor position according to the signal time difference and the electromagnetic wave propagation speed;

[0021] S3, the Levenberg-Marquardt algorithm is used to solve the equation groups, and preliminary position coordinates of the defect are obtained;

[0022] S4, the preliminary positioning result is corrected by considering the attenuation characteristics of the signal in the propagation process with reference to the signal processing result of the data preprocessing module;

[0023] S5, the corrected defect position is verified by using the multi-source data fusion module to fuse the multi-source data feature information;

[0024] S6, whether the positioning result matches the defect feature is evaluated by the recognition result of the defect type of the deep learning defect recognition module;

[0025] S7, according to the analysis of the defect severity and development trend of the Bayesian warning module, if it is found that the positioning result may affect the warning accuracy or there is a potential error, the positioning process is checked and optimized again.

[0026] The present application provides a kind of based on big data sample statistical analysis cable defect monitoring and positioning method.There are the following beneficial effects:

[0027] 1, the present application realizes comprehensive accurate data acquisition, high-quality data preprocessing, multi-source information effective fusion by the cable defect monitoring system through the collaborative operation of multiple modules, improves the accuracy and reliability of the entire cable defect monitoring work.Solve the problem that cable defect monitoring is not comprehensive and accurate.

[0028] 2, the present application realizes synchronous accurate acquisition of the multi-dimensional physical signals such as cable partial discharge, temperature and vibration by the multi-source sensor array module through reasonable layout and collaborative work of multiple sensors, constructs a system for comprehensively sensing cable state, lays a solid foundation for subsequent data processing and analysis, improves the accuracy and reliability of cable defect monitoring.Solve the problem of the limitations of previous single signal monitoring and the lack of data correlation.

[0029] 3、The application improves signal quality through wavelet denoising, makes signal characteristics clearer, uses space-time joint repair of missing values to improve data integrity, improves data processing efficiency with dimension reduction processing, and provides high-quality, complete and easy-to-process data basis for subsequent cable monitoring and analysis. The problems of signal interference by noise, data missing value affecting analysis and high-dimensional data processing difficulty are solved.

[0030] 4、The application uses the CNN-LSTM network model of the deep learning defect identification module to realize deep feature extraction, high-precision classification and adapt to complex data changes, can comprehensively and accurately mine multi-source data features, accurately distinguish different defect types, and can adaptively adjust with data changes, providing reliable basis for cable defect identification and subsequent operation and maintenance. The problems of limitation of manual feature extraction, insufficient defect classification accuracy and poor adaptability to complex data changes are solved. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 A module architecture diagram of a defect monitoring system of a cable based on big data sample statistical analysis is provided for the application.

[0032] Figure 2 A positioning method flow chart of a defect monitoring system of a cable based on big data sample statistical analysis is provided for the application. DETAILED DESCRIPTION

[0033] The technical solutions of the application will be described clearly and completely below with reference to the drawings of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0034] Please refer to the drawings of the application Figure 1 The application provides a defect monitoring system of a cable based on big data sample statistical analysis, which comprises the following modules:

[0035] The multi-source sensor array module is used to set a regular triangle array at a specific interval along the cable path, integrate ultrahigh frequency partial discharge sensors, fiber Bragg grating temperature sensors and acceleration vibration sensors, and synchronously collect multi-dimensional physical signals of partial discharge, temperature and vibration under the sampling frequency satisfying the Shannon sampling theorem.

[0036] The multi-source sensor array module comprises a regular triangle array set at every L distance along the cable path, wherein L satisfies the formula υ p The phase velocity of electromagnetic wave in the cable is f s The sampling frequency satisfies the Shannon sampling theorem f s= 2B, B is signal bandwidth, N is single cycle sampling points, the array contains ultrahigh frequency partial discharge sensor, fiber grating temperature sensor and acceleration vibration sensor.

[0037] Specifically, the array layout principle: along the cable path, a regular triangle array is arranged at a certain distance, the interval distance L is determined according to the formula Where, υ p represents the phase velocity of electromagnetic wave in the cable (about 2x10 8 m / s, which is determined by the electromagnetic characteristics of the cable medium, and the value will be different for different materials and structures of the cable. In practical application, it needs to be measured in advance by professional measuring equipment), f s is the sampling frequency and strictly satisfies the Shannon sampling theorem (i.e. f s = 2B, B is signal bandwidth, which ensures that the collected signal can completely retain the original information and avoid problems such as spectrum aliasing. The value of bandwidth B is determined according to the signal frequency range corresponding to the monitored physical quantity. For example, for ultrahigh frequency partial discharge signal, the frequency range is 300MHz-3GHz, and the corresponding bandwidth is determined according to this range. N is the single cycle sampling point (usually a suitable value is selected, such as N = 256, which can ensure that there are enough sampling points in each signal period to accurately restore the signal characteristics). Through such a layout, the sensor array can be reasonably distributed in space and cover the cable area in all directions, creating good conditions for comprehensive signal acquisition.

[0038] Sensor selection and cooperative working principle: The array integrates three different types of sensors: ultra-high frequency partial discharge sensor, fiber Bragg temperature sensor and acceleration vibration sensor. The ultra-high frequency partial discharge sensor works on the principle of electromagnetic induction. Its built-in antenna structure can capture the high-frequency electromagnetic waves generated by the partial discharge inside the cable due to insulation defects, etc. and convert the electromagnetic wave signals into electrical signals. Since its frequency response range is set at 300MHz-3GHz, it can accurately detect the high-frequency signal changes generated by partial discharge in an instant and sensitively perceive even tiny discharge phenomena. The fiber Bragg temperature sensor utilizes the photo-thermal effect of optical fiber. When the temperature of the cable changes, it will cause the refractive index of the fiber Bragg to change, and in turn cause the wavelength of the reflected light to change. By detecting the offset of the reflected light wavelength, the temperature at the location of the cable can be accurately measured, with a temperature measurement accuracy of ±0.1℃, which can reflect the temperature state of the cable in real time and high precision. The acceleration vibration sensor is based on the piezoelectric effect. When the cable vibrates due to internal defects, external interference and other factors, the piezoelectric crystal inside the sensor is stressed to produce a change in electric charge, converting the mechanical energy of vibration into an electrical signal. Its frequency response range is 0.1Hz-10kHz, which can effectively capture the vibration of the cable at different frequencies. The three sensors work synchronously, each collecting corresponding physical signals, and in time, through high-precision synchronization mechanism (such as high-precision clock module and synchronization protocol, etc., to ensure that the sampling time of each sensor is accurately synchronized), ensuring that the multi-dimensional physical signals collected are strictly corresponding in time dimension, providing accurate and complete data basis for subsequent fusion analysis.

[0039] Through the reasonable layout of the multi-source sensor array module and the cooperative work of various sensors, the synchronous and accurate collection of multi-dimensional physical signals such as cable partial discharge, temperature and vibration is realized, building a comprehensive sensing system for cable state, laying a solid foundation for subsequent data processing and analysis, and improving the accuracy and reliability of cable defect monitoring. It solves the limitations of previous single signal monitoring and the lack of data correlation.

[0040] The space-time synchronization module is used for GPS clock module and IEEE1588 protocol, to realize nanosecond-level synchronization of sensor array, and eliminate sensor inherent delay and global clock offset through timestamp calibration.

[0041] The space-time synchronization module includes GPS clock module and IEEE1588 protocol to realize synchronization of sensor array, with synchronization accuracy ≤10ns, and through timestamp calibration formula t' = t-Δt i +Δt0to calibrate data timestamp, where t is the original timestamp, Δt i is the inherent delay of the sensor, and Δt0is the global clock offset.

[0042] Specifically, the working mechanism of the GPS clock module: the GPS clock module can receive high-precision time signals from satellites, and its core lies in the atomic clocks carried on the satellites, which have extremely high timing accuracy. GPS satellites continuously send signals containing accurate time information to the ground, and the GPS clock module in the sensor array on the ground receives these signals and parses the accurate time reference from them to calibrate its own clock. In this way, each sensor can obtain a unified and high-precision external time reference, laying the foundation for the synchronization of the entire sensor array.

[0043] Application principle of IEEE 1588 protocol: IEEE 1588 protocol, also known as Precision Time Protocol (PTP), plays a key role in local area network environment. In the network where the sensor array is located, each sensor device acts as a node and follows the message interaction mechanism specified by IEEE 1588 protocol to perform time synchronization. Specifically, the master clock node (usually acted by a device with a high-precision clock source, here the clock calibrated by the GPS clock module can be used as the master clock) periodically sends synchronization messages containing time stamps to the slave clock nodes (i.e. each sensor), and the slave clock nodes calculate the time deviation from the master clock according to the algorithm specified by the protocol after receiving these messages, and adjust their own clocks accordingly, thereby realizing the clock synchronization among the sensors. Through this protocol, the clocks of each sensor in the local network composed of sensors can be accurately aligned, ensuring the consistency of their time, and the protocol can achieve nanosecond-level synchronization accuracy, i.e. the synchronization accuracy can be controlled within ≤10 ns.

[0044] Timestamp calibration principle: t' = t - Δt i + Δt0. Where t represents the original timestamp recorded by the sensor when collecting data, which reflects the time mark of the data generation time in the local clock of the sensor. Δt iThe inherent delay of the sensor is a fixed time difference between the actual occurrence of the signal and the recorded timestamp, caused by factors such as the hardware characteristics and signal processing circuits of each sensor. Different sensors have different inherent delays due to differences in their internal structure and performance, requiring specialized calibration experiments to accurately determine this parameter value for each sensor. Δt0 represents the global clock offset, which reflects the overall deviation of the clock of the entire sensor array from an ideal, absolutely accurate reference clock. This offset is measured and calculated using specific time synchronization calibration algorithms such as bidirectional time transfer. In practical applications, this formula is used to calibrate the original timestamps accompanying the data collected by each sensor, thereby eliminating time errors caused by the inherent delay of the sensors and the global clock offset. This ensures that the data collected by all sensors are accurately aligned in the time dimension, guaranteeing data time consistency.

[0045] By employing a time-space synchronization module with the GPS clock module and the IEEE 1588 protocol to achieve nanosecond-level synchronization, and eliminating related time errors through timestamp calibration, a high-precision and accurate time reference is provided for the data collected by the sensor array. This ensures the accuracy and reliability of subsequent multi-source data fusion, defect location, and other analyses. It solves the problem of time inconsistency among multiple sensors and the impact of accumulated time errors on analytical accuracy.

[0046] The data preprocessing module is used to perform wavelet denoising, missing value repair, and dimensionality reduction on the acquired cable signals. It removes noise through a soft threshold function and repairs missing data by combining LSTM and KNN algorithms.

[0047] The data preprocessing module includes performing a 5-level decomposition of the partial discharge signal using the db4 wavelet and applying a soft thresholding function. Denoising, thresholding σ j Let w be the standard deviation of the coefficients in the j-th layer, and n be the signal length. For missing data, a spatiotemporal joint repair method is used. The temporal dimension is predicted using an LSTM neural network, and the spatial dimension is interpolated using adjacent sensor data based on the KNN algorithm. The fusion weights satisfy w. t +w s =1.

[0048] Specifically, the wavelet transform basis: for the collected partial discharge signal, db4 wavelet is used for 5-layer decomposition. Wavelet transform is a time-frequency analysis method, which can decompose the signal at different scales and time positions, and represent the original signal as a series of wavelet coefficients with different frequency characteristics and time positioning. Db4 wavelet has good time-frequency localization characteristics, which is suitable for analyzing non-stationary signals such as partial discharge. Through 5-layer decomposition, the signal is gradually decomposed into detail components of different frequency bands and an approximate component. Each layer of decomposition will further refine the signal of the previous layer according to the wavelet function and scale function, reflecting the characteristic information of the signal in different frequency ranges.

[0049] Soft threshold function denoising mechanism: after obtaining the wavelet coefficients of each layer, the soft threshold function is used to remove noise. Among them, σ j is the standard deviation of the jth layer coefficient, which reflects the dispersion degree of the wavelet coefficient of this layer. It is determined by statistical analysis of the wavelet coefficient of each layer. The longer the signal, the longer the calculated threshold value will be adjusted. When the absolute value of the wavelet coefficient is greater than or equal to the threshold value λ j , the soft threshold value is processed, that is, the sign is retained after subtracting the threshold value. In this way, the details of the signal can be preserved while removing noise; while when the absolute value of the wavelet coefficient is less than the threshold value, it is directly set to 0, considering that these coefficients are mainly contributed by noise, so as to achieve the purpose of removing noise and purifying the signal.

[0050] Space-time joint repair idea: for the missing value in the data, the space-time joint repair method is adopted. In the time dimension, LSTM (Long Short Term Memory) neural network is used for prediction. LSTM has a unique gating structure, including input gate, forget gate and output gate, which can effectively handle the long-term dependence relationship in time series data. It can learn the effective data features of historical time and predict the value of the missing time according to the time sequence change rule. For example, for temperature signals that have a certain trend over time, LSTM can estimate the temperature value at the missing time according to the previously collected temperature values and their trends.

[0051] Spatial dimension repair method: in the spatial dimension, KNN (K-Nearest Neighbor Algorithm) is used to interpolate the data of adjacent sensors. The core idea of KNN algorithm is to find the K nearest neighbors (here K is usually selected according to the actual data situation and experience, such as K = 5) of the sensor where the missing value is located according to the spatial distribution characteristics of the data, and then the effective data collected by the K nearest neighbors at the corresponding time is weighted and averaged (the weight distribution can be determined according to the distance and other factors), so as to estimate the missing value. And set the fusion weight to satisfy w tw s = 1, by reasonably allocating the weight of the time dimension and the space dimension, the missing values are repaired more accurately by comprehensively considering the time and space information, so that the repair result is more in line with the actual data situation.

[0052] Dimension reduction processing principle: through a suitable dimension reduction algorithm (such as principal component analysis), the correlation between the characteristics of the data is analyzed, the high-dimensional data is mapped to a low-dimensional space, and under the premise of preserving the main information of the original data, the dimension of the data is reduced, the computational complexity of subsequent data processing and analysis is reduced, and the operation efficiency is improved.

[0053] Through wavelet denoising to improve signal quality, the signal characteristics are clearer, the data integrity is improved by using space-time joint repair of missing values, and the data processing efficiency is improved by means of dimension reduction processing, which provides a high-quality, complete and easy-to-process data basis for subsequent cable monitoring and analysis. The problems of signal interference by noise, data missing value affecting analysis and high-dimensional data processing difficulty are solved.

[0054] The multi-source data fusion module is used to define a defect recognition framework, generate a basic probability assignment function by using a support vector machine, and fuse local discharge, temperature, and vibration multi-source feature evidence by using a Dempster combination rule.

[0055] The multi-source data fusion module includes defining a recognition framework Θ = {D1, D2, D3, D4, θ}, D1-D4 are different types of defects, and θ is a normal state; a basic probability assignment function is generated by using a support vector machine where d i is the distance from the sample to the i-th hyperplane, and γ is a kernel function parameter; a Dempster combination rule is used to fuse the evidence, K is a conflict coefficient and

[0056] Specifically, the recognition framework is defined as follows: first, the recognition framework Θ = {D1, D2, D3, D4, θ} is defined, where D1, D2, D3, and D4 represent different types of cable defects, such as D1 being an internal discharge defect, D2 being a surface discharge defect, D3 being a shielding layer defect, and D4 being a joint aging defect, and θ representing the cable being in a normal state. This recognition framework provides a clear category definition for subsequent classification and judgment of the cable state reflected by the multi-source data, so that all data fusion and analysis work is carried out around these possible state categories, like building a classification "framework structure", so that each data evidence can be corresponded to the corresponding category.

[0057] The basic probability assignment function is generated based on a support vector machine: a support vector machine (SVM) is used to generate a basic probability assignment function Support vector machine is a classification algorithm based on statistical learning theory, whose core idea is to find an optimal hyperplane that can separate data samples of different classes in feature space as much as possible. For a given cable monitoring data sample (containing features extracted from multi-source data such as partial discharge, temperature, vibration, etc.), the distance d i of each sample in feature space to the hyperplane of different classes (i.e. D1, D2, etc. each defect type and normal state θ in the identification framework) is different. Here γ is the kernel function parameter, which determines the range of action of the kernel function and the mapping effect on the data. By adjusting the value of γ, the distribution of samples in feature space and the shape of classification boundary can be changed (for example, a larger γ value will make the range of action of the kernel function smaller, and the classification boundary more consistent with the training samples, which may lead to overfitting; while a smaller γ value will have a larger range of action, and the boundary will be smoother). By calculating the distance of each sample to each hyperplane, and according to the above formula, the basic probability assignment value of the sample for different classes (i.e. A i ) can be obtained, which is an initial probability measure of the likelihood of the sample belonging to each class, providing a basic probability basis for subsequent evidence fusion.

[0058] Dempster combination rule fusion evidence principle: Dempster combination rule is used to fuse multi-source evidence, where the conflict coefficient In the multi-source data environment, after feature extraction and basic probability assignment based on support vector machine, multiple evidences about the cable state belonging to different classes (such as different defect types or normal state) are obtained (i.e. different basic probability assignment functions m i (A i )). Dempster combination rule is a mechanism for synthesizing and fusing evidence from different sources. When there is an intersection between the class sets pointed by multiple evidences (i.e. ∩A iWhen the conflict coefficient K is greater than 1, it means that the evidences are in conflict, and the fusion result needs to be treated with caution or some special conflict processing strategy (such as weighting different evidences according to their reliability) is needed to reasonably fuse the evidences to obtain a comprehensive judgment probability of the cable state that is more in line with the actual situation.

[0059] By defining the identification framework, the multi-source feature evidences are fused based on the support vector machine and the Dempster combination rule, the effects of integrating multi-source information, improving judgment accuracy and quantifying uncertainty are realized, a more comprehensive, scientific and reliable basis for cable state judgment is provided, and scientific and reasonable operation and maintenance decisions are assisted. The problems of inconsistency of multi-source data, insufficient utilization and lack of comprehensive basis for cable state judgment are solved.

[0060] The deep learning defect identification module is used for the CNN-LSTM network model, and the fused multi-source data features are automatically extracted and learned, and the cables of different types of defects are classified and identified.

[0061] The deep learning defect identification module includes a CNN-LSTM network model, which includes three convolutional blocks, each convolutional block outputs x l = ReLU(W l *x l-1 +b l ), W l is a convolution kernel, b l is a bias, and * represents a convolution operation; two bidirectional LSTM layers, the LSTM unit state update satisfies where i t , f t , o t are input gate, forget gate and output gate respectively, c t is the cell state, h t is the hidden state, σ is the Sigmoid function, and is the element-level multiplication; two fully connected layers, the output layer realizes defect classification through the Softmax function m is the number of defect types.

[0062] Specifically, the working principle of the convolutional layer is: in the CNN-LSTM network model, it first includes three convolutional blocks. For each convolutional block, its output follows the formula x l=ReLU(W l *x l-1 +b l The convolution kernel W here l It is a matrix with a specific size (e.g., common 3×3, 5×5, etc., the specific size is set according to the actual application and the characteristics of the input data) and weight values, which is applied to the input data x. l-1 (The initial input data is multi-source feature data fused from multiple sources. Its dimensionality depends on the representation of the fused data, such as a two-dimensional image feature matrix or a multi-dimensional feature vector.) Sliding the kernel upwards, a convolution operation is performed (* indicates convolution). This process is equivalent to using a convolution kernel to extract features from different local regions of the input data. For example, for fused feature data such as partial discharge, temperature, and vibration, the convolution kernel can capture the spatial correlation patterns of different physical quantity features, such as the spatial distribution characteristics between the intensity of partial discharge signals and the corresponding temperature changes. Bias b l This is a learnable constant term used to adjust the offset of the convolution result, enabling the model to better fit the data. After the convolution operation, a nonlinear transformation is performed using the ReLU (Rectified Linear Unit) activation function. The ReLU function is expressed as f(x) = max(0,x). It increases the model's nonlinear expressive power because if only linear convolution operations are used, the model's expressive power is limited and it cannot learn complex data feature relationships. ReLU can set values ​​less than 0 to 0 and keep values ​​greater than 0 unchanged, giving the output data nonlinear characteristics. Through these three layers of convolutional blocks, layer-by-layer feature extraction and nonlinear transformation, more abstract and higher-level spatial feature representations are gradually extracted from the original multi-source fused data.

[0063] Working principle of bidirectional LSTM layer: Next are two bidirectional LSTM layers, whose cell state updates satisfy the following set of formulas: Where i t f t o t These are the input gate, forget gate, and output gate, all of which utilize the Sigmoid function. The calculated values ​​between 0 and 1 are used by these gating mechanisms to filter and control information. Input gate i t This determines how much of the new information input at the current moment can enter the cell state c. t Middle; Forget Gate f t c controls the cell state at the previous moment t-1 How much information needs to be forgotten, that is, how much historical information needs to be preserved and passed on; output gate o t This determines the current cell state c.t How much information in h can be output to the hidden state h t which in turn affects the subsequent calculations. The cell state c t is the core memory unit of the LSTM unit, which dynamically updates its value through the forgetting operation of the previous time cell state by the forgetting gate and the update operation of the current input information combined with the input gate, as in the above formula c t = f t ⊙ c t-1 + i t ⊙ tanh(W cx x t + W ch h t-1 + b c ), where ⊙ represents element-wise multiplication, and the tanh function is also a nonlinear activation function that is involved in the update calculation of the cell state after nonlinear transformation of the input information. Bidirectional LSTM is to perform the above LSTM unit calculation in both forward and reverse directions, which can capture the forward and reverse time sequence dependency of the input sequence data (here, it is the feature sequence processed by the convolutional layer, which contains the multi-source data feature relationship changing over time, such as the change of each physical quantity feature of the cable at different times). Compared with unidirectional LSTM, it can better learn the long-term dependency information in the data and mine more comprehensive feature change rules.

[0064] The working principle of the fully connected layer: after the learning of the time sequence features by the bidirectional LSTM layer, the data enters the two fully connected layers. In the fully connected layer, each neuron is connected to all neurons of the previous layer, and after linear transformation such as weighted summation, the data is processed by a nonlinear activation function (such as ReLU, etc.), which is selected according to the actual model design), further integrates and refines the features. Finally, the output layer realizes the classification of defects through the Softmax function , where m is the number of defect types, that is, the number of different cable defect categories that the model ultimately wants to predict. The Softmax function converts the feature vector output by the fully connected layer into a probability distribution, which calculates a corresponding probability value for each possible defect category, representing the possibility of the input data belonging to the defect category, for example, outputting a probability vector [0.1, 0.3, 0.6] can be understood as the probabilities of the current cable data corresponding to the state belonging to the three defect types are 0.1, 0.3, and 0.6, respectively, thereby realizing the classification and identification of different types of defects, and determining the specific defect condition of the cable according to the category with the highest probability.

[0065] The deep learning defect recognition module realizes deep feature extraction, high-precision classification and adaptation to complex data changes by means of a CNN-LSTM network model, can comprehensively and accurately mine multi-source data features, accurately distinguish different defect types, and adaptively adjust with data changes, thereby providing a reliable basis for cable defect recognition and subsequent operation and maintenance. The problems of limitations of manual feature extraction, insufficient defect classification accuracy and poor adaptability to complex data changes are solved.

[0066] The Bayesian early warning module is used for state space model and Kalman filter iterative calculation, evaluates the defect severity according to the cable defect related data, divides the early warning level and triggers the corresponding early warning.

[0067] The Bayesian early warning module includes establishing a state space model wherein is an autoregressive coefficient, β is an observation sensitivity coefficient, ω t is state noise, and υ t is observation noise; the state estimation value is calculated by Kalman filter iteration, including a prediction step an update step wherein is a state estimation value, P is an error covariance, and K is a Kalman gain; the early warning level is divided according to the defect severity index S, and the early warning is triggered when S≥0.3.

[0068] Specifically, the state space model is established as follows: first, a state space model is established, and its expression is wherein S t represents the system state vector at time t, which covers various internal state information related to cable defects, such as the aging degree of cable internal insulation, the development stage of potential defects, etc. (these state information is usually not directly observable, and needs to be inferred through model and related data). is an autoregressive coefficient, which describes the evolution law of the system state from time t-1 to time t, and its value depends on the physical properties of the cable itself and the running environment, etc. For example, for the state of cable insulation aging, if its aging process is relatively stable, the value of will be close to 1, indicating a slow and smooth change of the state; if affected by external complex factors, the aging speed is unstable, the value of will fluctuate according to the actual change. ω t is state noise, which represents the interference introduced in the system state evolution process due to various unpredictable and random factors (such as the uncertain influence of small random fluctuations of environmental temperature on cable aging, random changes of material internal microstructure, etc.), which is usually assumed to follow a normal distribution with mean value 0, i.e. ω t~N(0, Q), where Q is the covariance matrix of state noise, which describes the variance of state noise and the correlation of noise between different state variables, and its specific value is determined by statistical analysis of a large amount of historical data and related experiments.

[0069] And D t represents the observation vector at time t, that is, the data related to the cable that can be actually measured by the sensor, such as the amplitude of partial discharge, temperature value, vibration amplitude, etc. β is the observation sensitivity coefficient, which reflects the influence of the observation value D t on the real state S t of the system. The observation sensitivity coefficients corresponding to different observation variables are different, for example, for the observation variable of the amplitude of partial discharge, which is closely related to the internal insulation defect state of the cable, the corresponding β value may be larger, meaning that a small change in the insulation defect state will have a more obvious manifestation in the amplitude of partial discharge. υ t is the observation noise, which is the error existing in the observation data due to factors such as the precision limitation of the measuring instrument and environmental interference, and is also assumed to follow a normal distribution with a mean of 0 υ t ~N(0, R), where R is the covariance matrix of observation noise, and its value is determined by factors such as the precision of the sensor and the stability of the measurement environment, and is obtained by calibrating the sensor and measuring experiments in different environments.

[0070] Prediction step: the iterative calculation of Kalman filtering includes a prediction step and an update step. In the prediction step, the state estimation value is first predicted, and the formula is Here represents the predicted value of the state at time t based on the state estimation value at time t-1, which is a linear prediction based on the evolution law of the state in the state space model (reflected by the autoregressive coefficient ). At the same time, the error covariance matrix needs to be updated, and the formula is where P t|t-1 is the state estimation error covariance matrix at time t, which reflects the error uncertainty between the predicted state value and the real state value, P t-1|t-1 is the state estimation error covariance matrix at time t-1, which is updated in each iteration, and Q is the state noise covariance matrix mentioned earlier. Through this calculation, the change of the prediction error with the evolution of the state can be estimated.

[0071] Update step: in the update step, the Kalman gain K is first calculated. The Kalman gain plays a role in weighing the contribution of the predicted value and the observation value to the final state estimation value, and it considers the prediction error covariance P t|t-1, the observation sensitivity coefficient β and the observation noise covariance R, dynamically adjust the weight of the fusion of the predicted value and the observation value. Then, the predicted state estimation value is updated using the observation value, and the formula is That is, the predicted value is added to the deviation of the observation value adjusted according to the Kalman gain from the predicted value, to obtain a more accurate state estimation value at time t Finally, the error covariance matrix is updated again, and the formula is P t|t =(1-K t β)P t|t-1 to reflect the change of the error uncertainty between the updated state estimation value and the true state value, complete one iteration of Kalman filtering, and so on. With the continuous acquisition of new observation data, the state estimation value and the error covariance are continuously updated, so that the estimation of the cable state becomes more and more accurate.

[0072] The principle of warning level division: According to the defect severity index S, the warning level is divided, and the defect severity index S here is an index that can quantitatively reflect the severity of the cable defect based on the state estimation value and other related information obtained through the above Kalman filtering iteration calculation, and is obtained through a specific calculation or mapping relationship (for example, a calculation function can be set according to the corresponding key physical quantity threshold, state change trend, etc. of the cable in different defect states to obtain the value of S). When S≥0.3, the warning is triggered. This threshold 0.3 is obtained through analysis of a large amount of historical cable failure data, simulation experiments and actual operation and maintenance experience, which means that when the defect severity index reaches this value, the risk of cable failure or serious performance decline significantly increases, and timely warning notification is needed to inform relevant operation and maintenance personnel to take appropriate inspection, maintenance or repair measures.

[0073] Through the state space model and Kalman filtering iteration calculation of the Bayesian warning module, accurate state estimation, dynamic warning adjustment and comprehensive consideration of uncertainty are realized, which can more accurately grasp the cable defect state, adjust the warning level in real time, and help the operation and maintenance decision-making to be more scientific and reasonable. The problems of inaccurate cable state estimation, untimely and lack of dynamic warning, and lack of uncertainty consideration in operation and maintenance decision-making are solved.

[0074] Please refer to the attached Figure 2 A positioning method for defect monitoring of a cable based on big data sample statistical analysis, for a defect monitoring system of a cable based on big data sample statistical analysis, comprising the following steps:

[0075] S1, through the equilateral triangle multi-source sensor array arranged at a certain distance along the cable path, the time information of the signal generated by the defect reaching different sensors is obtained, and the signal time difference is determined;

[0076] S2, high-precision sensor array synchronization realized by means of the space-time synchronization module, according to the signal time difference and the electromagnetic wave propagation speed, a hyperbolic equation set is constructed for describing the relationship between the defect position and the sensor position;

[0077] S3, the Levenberg-Marquardt algorithm is used to solve the equation set to obtain the preliminary position coordinates of the defect;

[0078] S4, the signal processing result is referenced by the data preprocessing module, and the preliminary positioning result is corrected considering the attenuation characteristics of the signal in the propagation process;

[0079] S5, the corrected defect position is verified by using the multi-source data fusion module fused multi-source data feature information;

[0080] S6, the positioning result is evaluated whether it matches the defect feature through the recognition result of the defect type of the deep learning defect recognition module;

[0081] S7, according to the analysis of the defect severity and development trend of the Bayesian warning module, if it is found that the positioning result may affect the warning accuracy or there is potential error, the positioning process is checked and optimized again.

[0082] The multi-source sensor array module is arranged in a regular triangle array at a specific interval along the cable path, integrates multiple sensors, and synchronously collects multi-dimensional physical signals according to the Shannon sampling theorem; the space-time synchronization module realizes nanosecond-level synchronization of the sensor array by means of the GPS clock module and the IEEE1588 protocol, and uses the timestamp calibration formula to eliminate the inherent delay of the sensor and the global clock offset; the data preprocessing module sequentially processes the collected signals through wavelet denoising, missing value repair based on the LSTM and KNN algorithms, and dimension reduction processing, to improve the signal quality, perfect the data integrity and improve the processing efficiency; the multi-source data fusion module first defines an identification framework containing different defects and normal states, then generates a basic probability assignment function using a support vector machine, and finally fuses multi-source feature evidence through the Dempster synthesis rule; the deep learning defect recognition module adopts a CNN-LSTM network model, which sequentially extracts and learns features through convolution layers, bidirectional LSTM layers and fully connected layers, and classifies and recognizes cable defects through a Softmax function; the Bayesian warning module first establishes a state space model to describe the relationship between cable state evolution and observation, then iteratively calculates and updates the state estimation value using Kalman filtering, and finally divides and triggers the warning according to the defect severity index. The modules cooperate with each other, aiming to realize accurate monitoring and warning of cable defects.

[0083] Through the cable defect monitoring system, through the cooperative operation of multiple modules, comprehensive and accurate data acquisition, high-precision time synchronization, high-quality data preprocessing, effective fusion of multi-source information, accurate defect classification and identification, and dynamic and accurate early warning are realized, which provides strong support for comprehensively mastering the cable operation state and scientific operation and decision-making, and improves the accuracy and reliability of the entire cable defect monitoring work. The problem that the cable defect monitoring is not comprehensive and accurate enough is solved.

[0084] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications, changes, variations, substitutions, and equivalents will occur to those of ordinary skill in the art without departing from the spirit and scope of the application as defined by the following claims and their equivalents.

Claims

1. A system for monitoring defects in a cable based on statistical analysis of a large data sample, characterized by, Comprise the following modules: A multi-source sensor array module for setting a regular triangle array at certain intervals along the cable path, integrating ultra-high frequency partial discharge sensors, fiber optic temperature sensors, and acceleration vibration sensors, and synchronously collecting multi-dimensional physical signals of partial discharge, temperature, and vibration at a sampling frequency that meets the Shannon sampling theorem; A space-time synchronization module for synchronizing the sensor array at the nanosecond level through the GPS clock module and IEEE1588 protocol, and eliminating the inherent delay of the sensor and the global clock offset through timestamp calibration; A data preprocessing module for wavelet denoising, missing value repair, and dimensionality reduction processing of the collected cable signals, eliminating noise through a soft threshold function, and repairing missing data using the LSTM and KNN algorithms; A multi-source data fusion module for defining a defect recognition framework, generating a basic probability assignment function through a support vector machine, and fusing multi-source feature evidence of partial discharge, temperature, and vibration using the Dempster combination rule; A deep learning defect recognition module for a CNN-LSTM network model to automatically extract and learn the fused multi-source data features and classify and recognize different types of cable defects; A Bayesian warning module for state space model and Kalman filter iterative calculation to evaluate the severity of the defect based on cable defect-related data, divide the warning level, and trigger the corresponding warning.

2. The system for monitoring defects in the cable based on statistical analysis of large data samples as claimed in claim 1, wherein: The multi-source sensor array module comprises a regular triangle array arranged at every L distance along the cable path, wherein L satisfies the formula υ p is the phase velocity of electromagnetic wave in the cable, f s is the sampling frequency and satisfies the Shannon sampling theorem f s =2B, B is the signal bandwidth, N is the number of sampling points per cycle, and the array comprises an ultrahigh frequency partial discharge sensor, an optical fiber grating temperature sensor and an acceleration vibration sensor.

3. The system for monitoring defects in the cable based on statistical analysis of large data samples as claimed in claim 1, wherein: The space-time synchronization module comprises: a GPS clock module and an IEEE1588 protocol are used to realize synchronization of the sensor array, and the synchronization precision is less than or equal to 10 ns; and a timestamp calibration formula t'=t-Δt+Δt0 is used to calibrate the data timestamp, wherein t is an original timestamp, Δt is inherent delay of the sensor, and Δt0 is a global clock offset. i i ​​ 4. The system for monitoring defects in the cable based on statistical analysis of large data samples as claimed in claim 1, wherein: The data preprocessing module includes adopting db4 wavelet to perform 5-layer decomposition on the partial discharge signal, and performing soft threshold function de-noising, threshold sigma j is the standard deviation of the jth layer coefficient, and n is the length of the signal; the missing data is repaired by combining time and space, the time dimension is predicted by an LSTM neural network, the space dimension is interpolated based on a KNN algorithm using adjacent sensor data, and the fusion weight satisfies w t +w s =1.

5. The system for monitoring defects in the cable based on statistical analysis of large data samples as claimed in claim 1, wherein: The multi-source data fusion module comprises a definition identification framework Θ={D1, D2, D3, D4, θ}, D1-D4 are different types of defects, and θ is a normal state; a basic probability assignment function is generated by a support vector machine Wherein d i is the distance of the sample to the i-th hyperplane, γ is a kernel function parameter; evidence fusion is performed by using a Dempster combination rule , K is a conflict coefficient and 6. The system for monitoring defects in the cable based on statistical analysis of large data samples as claimed in claim 1, wherein: The deep learning defect recognition module comprises a CNN-LSTM network model, the model comprising three convolutional blocks, each convolutional block outputting x l = ReLU(W l *x l-1 +b l ), W l is a convolution kernel, b l is a bias, and * represents convolution operation; Two layers of bidirectional LSTM layers, the LSTM unit state update satisfies where i t , f t , o t are input gate, forget gate and output gate respectively, c t is cell state, h t is hidden state, σ is Sigmoid function, and is element-wise multiplication; two layers of fully connected layers, the output layer is realized by Softmax function to realize defect classification, and m is the number of defect types.

7. The system for monitoring defects in the cable based on statistical analysis of large data samples as claimed in claim 1, wherein: The Bayesian early warning module comprises establishing a state space model wherein is an autoregressive coefficient, β is an observation sensitivity coefficient, ω t ~N(0, Q) is state noise, υ t ~N(0, R) is observation noise; a state estimation value is calculated by Kalman filtering iteration, comprising a prediction step an update step P t|t = (1 - K t β) P t|t-1 wherein is a state estimation value, P is an error covariance, and K is a Kalman gain According to the defect severity index S, the warning level is divided, and the warning is triggered when S≥0.

3.

8. A method for locating a defect monitoring of a cable based on a large data sample statistical analysis, characterized by, A defect monitoring system for a cable based on big data sample statistical analysis according to any one of claims 1-7, comprising the following steps: S1, through the regular triangle multi-source sensor array set at certain intervals along the cable path, the time information of the signal generated by the defect reaching different sensors is obtained, and the signal time difference is determined; S2, with the help of high-precision sensor array synchronization realized by the space-time synchronization module, a hyperbolic equation set is constructed based on the signal time difference and the electromagnetic wave propagation speed, which is used to describe the relationship between the defect position and the sensor position; S3, the Levenberg-Marquardt algorithm is used to solve the equation set and obtain the preliminary position coordinates of the defect; S4, referring to the results of signal processing by the data preprocessing module, considering the attenuation characteristics of the signal in the propagation process, the preliminary positioning result is corrected; S5, the multi-source data feature information fused by the multi-source data fusion module is used to verify the corrected defect position; S6, through the recognition result of the defect type by the deep learning defect recognition module, it is evaluated whether the positioning result matches the defect characteristics; S7, according to the analysis of the defect severity and development trend by the Bayesian warning module, if it is found that the positioning result may affect the accuracy of the warning or there is a potential error, the positioning process is checked and optimized again.

Images