A GIL support vibration intelligent monitoring and early warning system
By using a distributed fiber optic vibration sensing module and an improved Levenberg-Marquardt algorithm, combined with undirected weighted node distribution graphs and Jacobian matrix weighting, the problems of low positioning accuracy and weak anti-interference capability in the GIL support vibration monitoring system were solved, achieving accurate positioning and real-time monitoring of vibration sources and reducing operation and maintenance costs.
Patent Information
- Application Number
- CN202611619385.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing GIL support vibration monitoring systems suffer from low positioning accuracy, weak anti-interference capabilities, and poor algorithm convergence stability, making it impossible to accurately locate and monitor vibration sources in real time. This results in high maintenance costs, inaccurate positioning results, and an inability to meet the needs of engineering sites.
By employing a distributed fiber optic vibration sensing module and an improved Levenberg-Marquardt algorithm, combined with an undirected weighted node distribution map and Jacobian matrix weighting, node reliability coefficients are constructed to achieve precise positioning of the vibration source of the GIL support.
It improves the positioning accuracy and anti-interference capability of GIL support vibration monitoring, ensures the convergence stability of the algorithm, realizes real-time monitoring and early warning of vibration sources, and reduces operation and maintenance costs.
Smart Images

Figure CN123111451A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of GIL engineering, specifically to a GIL support vibration intelligent monitoring and early warning system. Background Technology
[0002] Gas-insulated metal-enclosed transmission lines (GILs), as core equipment for high-voltage, high-capacity, and long-distance power transmission, have become widely used in key power projects such as large hydropower stations, substations, and high-voltage power grids due to their significant advantages of large transmission capacity, low loss, small footprint, and long service life. They are a crucial infrastructure for ensuring the stable operation of power systems. With the continuous growth of electricity demand, the transmission current of GILs is constantly increasing, and the electromagnetic force between the three-phase conductors is significantly enhanced, leading to increasingly prominent vibration problems between the conductors and the outer casing. Simultaneously, factors such as foundation settlement and external environmental disturbances can also cause abnormal vibration of the GIL supports. Vibration of the GIL supports not only reduces the stable operating time of the equipment and generates radiated noise pollution, affecting the physical and mental health of surrounding personnel, but may also lead to structural fatigue, loose bolts, and damage to insulation components, potentially causing GIL equipment failure or even power outages, resulting in huge economic losses ranging from millions to tens of millions of yuan. Therefore, effective monitoring and early warning of the vibration status of GIL supports is a key requirement for ensuring the safety of power transmission and reducing operation and maintenance costs.
[0003] Currently, the main shortcomings of power industry monitoring technologies for GIL (Gas Infrared Lamp) are as follows: 1. Existing monitoring systems mostly focus on the entire GIL tunnel or pipeline body, lacking specific monitoring designs for support vibration, resulting in insufficient monitoring targeting and accuracy; 2. Traditional monitoring methods rely mainly on manual inspections, which are inefficient, costly, and unable to achieve real-time continuous acquisition of vibration signals, making it difficult to capture instantaneous abnormal vibrations; 3. Some online monitoring solutions use distributed sensors or single-sensor technology, which have problems such as complex installation, high cost, weak resistance to electromagnetic interference, and limited positioning accuracy; 4. Data processing is mostly limited to simple acquisition and linear display, lacking intelligent analysis and feature extraction capabilities, making it impossible to automatically identify vibration anomalies, trace fault sources, and predict trends, and making it difficult for maintenance personnel to quickly locate potential hazards.
[0004] Existing distributed monitoring solutions can only identify and roughly locate vibration events, and cannot accurately solve for the coordinates of vibration sources based on the vibration propagation characteristics of GIL steel structures. In situations where there is a large amount of environmental noise and abnormal monitoring data on site, the positioning algorithm is prone to iterative divergence and excessive deviation in positioning results, which cannot meet the actual operation and maintenance needs of engineering sites. Summary of the Invention
[0005] To address the problems of low positioning accuracy, weak anti-interference ability, poor algorithm convergence stability, and insufficient adaptability to engineering scenarios in the current traditional GIL vibration monitoring schemes in practical applications, this application provides a method for intelligent monitoring and early warning of GIL support vibration.
[0006] The first aspect of this application provides a GIL (Gas Insulator) support vibration intelligent monitoring and early warning system, comprising: The distributed optical fiber vibration sensing module includes multiple optical fiber vibration sensing units and optical fibers arranged on the GIL support to be monitored. Each optical fiber vibration sensing unit collects vibration signals at the corresponding sampling location point. The communication array, comprising multiple communicators, is used to transmit the vibration signals collected by each fiber optic vibration sensing unit to a cloud server. A cloud server is used to determine the vibration source based on the received vibration signal and an improved Levenberg-Marquardt algorithm. The Jacobian matrix of the improved Levenberg-Marquardt algorithm is obtained by weighting the node reliability coefficients. The node reliability coefficients are obtained by weighting the edge weights between nodes in an undirected weighted node distribution graph. The undirected weighted node distribution graph is generated based on the vibration location corresponding to the vibration signal.
[0007] Optionally, the optical fiber of the distributed optical fiber sensing vibration module is arranged along the support legs, crossbeam connection and insulator mounting part of the GIL bracket.
[0008] Optionally, the communicators of the communication array are configured in a one-to-one correspondence or a one-to-many correspondence with the fiber optic vibration sensing units.
[0009] A second aspect of this application provides a method for intelligent monitoring and early warning of GIL (Glass Insulation Layer) support vibration. The method is applied to a GIL support vibration intelligent monitoring and early warning system as described above, and includes: Based on the vibration signals and the vibration positions corresponding to each vibration signal, an undirected weighted node distribution map is generated, wherein the coordinates of each node in the undirected weighted node distribution map correspond to the vibration positions. Based on the edge weights between nodes in the undirected weighted node distribution graph, the node reliability coefficient of each node is obtained. Based on the node reliability coefficients of each node, the initial Jacobian matrix of the Levenberg-Marquardt algorithm is updated again. An improved Levenberg-Marquardt algorithm is formed based on the weighted Jacobian matrix; The vibration source is determined based on the undirected weighted node distribution map and the improved Levenberg-Marquardt algorithm.
[0010] Optionally, the improved Levenberg-Marquardt algorithm based on the weighted Jacobian matrix includes: Replace the initial Jacobian matrix of the Levenberg-Marquardt algorithm with the weighted Jacobian matrix; Based on the column sum of squares of the weighted Jacobian matrix, a reliability orientation regularization matrix is constructed, and the identity matrix regularization term in the Levenberg-Marquardt algorithm is replaced with the reliability orientation regularization matrix. Based on the weighted Jacobian matrix and the reliability orientation regularization matrix, an improved Levenberg-Marquardt algorithm iterative step size formula is constructed. Combined with an adaptively adjusted damping factor, multiple rounds of iterative solutions are performed until the convergence condition is met, and the optimal coordinates of the vibration source are output.
[0011] Optionally, generating an undirected weighted node distribution map based on the vibration signals and the corresponding vibration positions of each vibration signal includes: The received vibration signals are filtered by amplitude threshold to remove invalid signals from faulty nodes and retain valid signals whose vibration amplitude exceeds the threshold. A node set is constructed by combining the vibration locations corresponding to the effective signals, and a fully connected undirected weighted node distribution map is generated based on the node set.
[0012] Optionally, constructing a node set by combining the vibration locations corresponding to the valid signals includes: Calculate the vibration amplitude similarity between any two nodes in the node distribution graph, determine the edge weights between the corresponding nodes based on the vibration amplitude similarity, and form the weight set of the node distribution graph.
[0013] Optionally, the step of obtaining the node reliability coefficient of each node based on the edge weights between nodes in the undirected weighted node distribution graph includes: Calculate the average edge weight between a single node and all other nodes, and use this average value as the node reliability coefficient for that node.
[0014] Optionally, updating the initial Jacobian matrix of the Levenberg-Marquardt algorithm based on the node reliability coefficients of each node includes: Based on the physical model of cylindrical wave attenuation in GIL steel structure, the original Jacobian matrix is calculated. The weighted Jacobian matrix is obtained by weighting the corresponding rows of the original Jacobian matrix using the node reliability coefficients of each node.
[0015] Optionally, the processing module of the cloud server compares the determined vibration source location with the preset safety threshold, generates a warning message of the corresponding level based on the comparison result, and then pushes the warning message to the on-site operation and maintenance terminal and the remote management platform.
[0016] As can be seen from the above technical solution, this application combines an undirected weighted node distribution graph and the Levenberg-Marquardt (LM) algorithm, constructs an approximate Hessian matrix using the Jacobian matrix, and solves for the iteration step size. The Jacobian matrix is the first-order partial derivative matrix of the residual vector with respect to the optimization variables (vibration source coordinates, vibration source amplitude). Its number of rows equals the number of effective sensing nodes, and its number of columns equals the dimension of the optimization variables. Each row corresponds to a sensing node, representing the contribution of the node's residual to the partial derivative of the optimization variables. The core improvement of this method is the construction of a node... Point reliability weighted Jacobian matrix: The corresponding rows of the Jacobian matrix are weighted using the node reliability coefficient of each node. In this way, the reliability coefficient of high-reliability nodes is close to 1, and their corresponding Jacobian rows are almost unchanged. The residual partial derivatives are completely preserved and dominate the optimization direction during the algorithm iteration process, so that the iteration always revolves around the real and reliable measurement data. The Jacobian rows of medium and low-reliability nodes are compressed by the corresponding coefficients, and the contribution of the residual partial derivatives is reasonably reduced. This eliminates the interference of false gradients on the iteration direction and solves the positioning deviation problem caused by abnormal data. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a schematic diagram of the structure of the GIL stent provided in the embodiments of this application.
[0019] Figure 2 This is a flowchart illustrating the intelligent monitoring and early warning method for GIL support vibration provided in the embodiments of this application.
[0020] Figure 3 This is a schematic diagram of the undirected weighted node distribution diagram provided in the embodiments of this application.
[0021] Reference numerals: 1-GIL pipe body; 2- Insulator; 3- Insulator mounting part; 4- Crossbeam; 5- Crossbeam connection part; 11- Support leg; 12- Base; 13- Sensor fiber. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0023] This application provides a GIL support vibration intelligent monitoring and early warning system, specifically including a distributed fiber optic vibration sensing module, a communication array, and a cloud server. In this application embodiment, edge computing gateways, on-site maintenance terminals, remote management platforms, and other supporting equipment can be added according to the actual GIL engineering scenario. The number of hardware modules and the networking method can also be adjusted according to the monitoring range. This application is not limited to these.
[0024] In this embodiment of the application, the cloud server determines the vibration source based on the received vibration signal and the improved Levenberg-Marquardt algorithm. The Jacobian matrix of the improved Levenberg-Marquardt algorithm is obtained by weighting the node reliability coefficients. The node reliability coefficients are obtained by weighting the edge weights between nodes in the undirected weighted node distribution graph. The node distribution graph is generated based on the effective signals in the vibration signal and the vibration positions corresponding to each effective signal.
[0025] The following is a detailed description of this application.
[0026] like Figure 1 As shown, the GIL support provided in this embodiment is a steel structure support for supporting the GIL pipe body 1, including structural components such as support legs 11, crossbeam connection parts 5, insulator mounting parts 3 and base 12; the optical fibers of the distributed optical fiber vibration sensing module are arranged along the support legs 11, crossbeam connection parts 5 and insulator mounting parts 3 to realize vibration sensing of key structural parts of the GIL support.
[0027] In this embodiment, the distributed optical fiber vibration sensing module specifically includes an optical fiber vibration sensing unit and optical fibers arranged on the GIL support to be monitored. It should be noted that the optical fiber vibration sensing unit in this application acquires vibration signals based on the principle of phase-sensitive optical time-domain reflectometry. The optical fiber itself serves as the sensing medium, enabling continuous distributed vibration sensing along the optical fiber deployment path. Furthermore, those skilled in the art will understand that in this embodiment, the spacing of the optical fiber vibration sensing units can be adjusted according to the structure of the GIL support, and the sampling frequency of the sensing system can be adjusted according to the monitored vibration frequency band. This application is not limited to these limitations.
[0028] In this embodiment, the distributed fiber optic vibration sensing unit is the smallest front-end sensing and execution unit of the GIL support vibration intelligent monitoring and early warning system, and is deployed one-to-one with the sampling points of the sensing fiber 13 arranged on the GIL support to be monitored. This unit consists of a data acquisition unit, a signal amplification unit, and an analog-to-digital conversion unit, which completes the physical acquisition, linear conversion, and digital signal output of vibration signals. All signal analysis and data processing work, including phase demodulation, noise reduction filtering, feature extraction, over-threshold node screening, construction of undirected weighted node distribution map, calculation of node reliability coefficient, and iterative solution of the improved Levenberg-Marquardt algorithm, is completed on a cloud server. The unit itself does not have any signal analysis, algorithm calculation, fault judgment, or parameter adaptive adjustment functions.
[0029] In the embodiments of this application, the hardware parameter configuration and deployment quantity of each unit can be flexibly adjusted according to the line length, support structure form and monitoring point density requirements of the GIL project, but this application is not limited thereto.
[0030] In this embodiment, the data acquisition unit acquires the backscattered Rayleigh light signal carrying the vibration information of the GIL support transmitted by the sensing fiber 13 at the corresponding sampling location point. The signal amplification unit and the analog-to-digital conversion unit complete the amplification of the optical signal and the linear photoelectric conversion of the optical signal to the analog electrical signal, and output the original analog electrical signal that is strictly linearly corresponding to the power of the vibration optical signal.
[0031] In this embodiment, the data acquisition unit consists of an optical fiber coupling docking subunit and a photoelectric conversion subunit. The optical fiber coupling docking subunit uses an FC / APC (optical fiber connector / angled physical contact) type low-loss optical fiber adapter as the core docking component, with a built-in zirconia ceramic ferrule and a micron-level precision alignment structure, and is compatible with G.652D (the standard single-mode fiber type recommended by the International Telecommunication Union) standard single-mode sensing fiber 13.
[0032] In this embodiment, the photoelectric conversion subunit uses an InGaAs (Indium Gallium Arsenide) PIN (P-type-intrinsic-N-type) photodiode as the core photoelectric conversion device, adapted to the 1550nm operating band of the sensing fiber, with a response spectrum covering 1200nm~1600nm, matching the operating band of the light source in the distributed fiber optic sensing system; the photoelectric conversion responsivity is ≥0.9A / W, capable of linearly converting incident light power into photocurrent of corresponding amplitude, with a linear correlation coefficient between light power and output photocurrent ≥0.999, ensuring distortion-free conversion of vibration amplitude information; the response bandwidth is ≥100MHz. It can fully cover the effective vibration signal frequency band of GIL brackets from 10Hz to 100kHz, ensuring linear conversion of vibration signals of different frequencies without distortion. It is suitable for the full-band acquisition requirements of GIL bracket power frequency vibration, low-frequency vibration of mechanical loosening, impact vibration of structural cracking, and high-frequency ultrasonic vibration of partial discharge. The dark current is ≤1nA, which can effectively reduce the dark current drift caused by changes in ambient temperature and improve the conversion accuracy of weak vibration optical signals in the early stage of GIL bracket failure. The photosensitive surface diameter is 300μm, which is compatible with the optical fiber mode field diameter, ensuring that the incident light signal can completely cover the photosensitive surface and maximize the photoelectric conversion efficiency.
[0033] In this embodiment, the data acquisition unit is arranged in a one-to-one correspondence with the sampling points of the GIL bracket. The core arrangement points include the flange connection of the GIL bracket support leg 11, the connection between the crossbeam 4 and the crossbeam 5, the insulator mounting part 3, the supporting bracket for the line expansion joint / settlement joint, and the connection between the support leg 11 and the ground.
[0034] In this embodiment, the optical signal input end of the data acquisition unit is connected to the corresponding sampling position point of the sensing optical fiber 13 through optical fiber fusion splicing to realize the acquisition of optical signals carrying vibration information; the analog electrical signal output end of the data acquisition unit is directly connected to the signal input end of the signal amplification unit through a 50Ω impedance-matched double-shielded coaxial cable, and then enters the analog-to-digital conversion unit to convert the analog signal into a digital signal, and then sends the vibration signal in digital form to the cloud server.
[0035] The cloud server in this application embodiment will be described in detail below. First, after receiving the vibration signals, the cloud server processes the vibration signals. Specifically, this application further provides a method for intelligent monitoring and early warning of GIL bracket vibration, which includes: S1: Based on the vibration signal and the vibration position corresponding to each vibration signal, an undirected weighted node distribution map is generated, wherein the coordinates of each node in the undirected weighted node distribution map correspond to the vibration position; S2: Based on the edge weights between nodes in the undirected weighted node distribution graph, obtain the node reliability coefficient of each node; S3: Based on the node reliability coefficients of each node, update the initial Jacobian matrix of the Levenberg-Marquardt algorithm; S4: An improved Levenberg-Marquardt algorithm based on the weighted Jacobian matrix; S5: Determine the vibration source based on the undirected weighted node distribution map and the improved Levenberg-Marquardt algorithm.
[0036] like Figure 2 As shown in the embodiments of this application, the overall process of the intelligent monitoring and early warning method for GIL support vibration includes: sensor deployment, signal acquisition, data transmission, and vibration feature extraction. The sensor deployment stage simultaneously performs vibration reduction performance extraction, thereby completing vibration reduction performance evaluation. The vibration reduction performance evaluation results are used to determine whether an early warning triggers the signal acquisition stage, and also to enter the anomaly judgment and graded early warning stage. This stage combines frequency domain analysis and threshold comparison to output graded early warnings and complete data storage. The anomaly judgment and graded early warning stage also points to anomaly handling tracking, which is fed back to the vibration feature extraction stage, forming a closed-loop monitoring and early warning processing system.
[0037] This application embodiment combines an undirected weighted node distribution graph and the Levenberg-Marquardt (LM) algorithm. It constructs an approximate Hessian matrix using the Jacobian matrix to solve for the iteration step size. The Jacobian matrix is the first-order partial derivative matrix of the residual vector with respect to the optimization variables (vibration source coordinates, vibration source amplitude). Its number of rows equals the number of effective sensing nodes, and its number of columns equals the dimension of the optimization variables. Each row corresponds to a sensing node, representing the contribution of that node's residual to the partial derivative of the optimization variables. The core improvement of this method is the construction of a node reliability-weighted Jacobian matrix: the node reliability coefficient of each node is used to weight the corresponding rows of the Jacobian matrix. This results in a high-reliability node with a reliability coefficient close to 1, and its corresponding Jacobian row remains almost unchanged, preserving the residual partial derivative completely. This derivative dominates the optimization direction during algorithm iteration, ensuring that the iteration always revolves around reliable measurement data. The Jacobian rows of medium- and low-reliability nodes are compressed by their corresponding coefficients, reasonably reducing the contribution of the residual partial derivative. This eliminates the interference of false gradients on the iteration direction and solves the positioning deviation problem caused by abnormal data.
[0038] In this embodiment, the original digital signal completely preserves the time-domain amplitude information of the vibration signal. The cloud server can extract the peak value of the vibration amplitude at each sampling location point, complete the screening of nodes exceeding the threshold, and construct an effective node set.
[0039] After receiving the vibration signal, the cloud server performs a full-dimensional time-domain statistical analysis on the vibration signal of each sensing node, calculating time-domain characteristic parameters such as peak value, mean, root mean square value, kurtosis, margin, peak factor, impulse factor, and waveform factor. Among these, the peak vibration amplitude is a core parameter for constructing the node graph and calculating edge weights and node reliability coefficients. The processing unit normalizes the peak amplitude of each sensing node, linearly mapping it to the [0,1] interval. Simultaneously, based on the 3σ criterion and combined with background vibration data under normal operating conditions of the GIL support, the processing unit performs threshold filtering on the vibration signal, identifying valid sensing nodes with vibration amplitudes exceeding the threshold and eliminating invalid data from dead nodes, no-signal nodes, and nodes with strong noise.
[0040] The processing unit converts the time-domain vibration signal to the frequency domain using Fast Fourier Transform (FFT) and calculates frequency domain characteristic parameters such as spectral distribution, dominant frequency, harmonic components, sidebands, power spectral density, and spectral kurtosis. Based on the frequency domain analysis results, it distinguishes different types of vibration signals, including 50Hz / 100Hz power frequency vibration of the GIL support, low-frequency vibration caused by mechanical loosening, impact vibration caused by structural cracking, and high-frequency ultrasonic vibration caused by partial discharge. Simultaneously, it uses digital filtering technology to filter out fixed-frequency noise components such as power frequency harmonics and electromagnetic interference, further improving the signal-to-noise ratio of the vibration signal. For non-stationary impact vibration signals, the processing unit can also identify the transmission path and resonant frequency of the vibration signal through power spectral cepstrum analysis, providing a basis for structural fault diagnosis of the GIL support.
[0041] The processing unit employs time-frequency domain analysis methods such as wavelet transform, ensemble empirical mode decomposition (EEMD), and variational mode decomposition (VMD) to perform multi-scale decomposition and reconstruction of the non-stationary impact vibration signal caused by GIL support faults. This separates the effective fault component and environmental noise component from the vibration signal, extracting vibration features at different time scales. Specifically, the wavelet transform uses the db4 wavelet basis function to perform a 5-level decomposition of the vibration signal, extracting the wavelet coefficient energy features at each scale. VMD decomposition adaptively decomposes the vibration signal into multiple intrinsic mode functions (IMFs), calculating the energy entropy and correlation coefficient of each IMF component, and filtering out the effective components related to the fault. This accurately captures transient impact vibration signals caused by faults such as loose flange bolts, weld cracking, and insulator 2 detachment in the GIL support, providing more comprehensive feature dimensions for the pattern recognition stage and improving the accuracy of fault identification.
[0042] The processing unit standardizes, normalizes, and reduces the dimensionality of high-dimensional feature parameters obtained from time-domain, frequency-domain, and time-frequency-domain analyses to construct a standardized vibration feature dataset. First, the Z-score (standard score) standardization method is used to eliminate differences in the dimensions and amplitude ranges of different features, mapping all feature parameters to a standard interval with a mean of 0 and a variance of 1. Then, principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional features, setting the cumulative contribution rate to 95% while retaining core feature components. This significantly reduces the feature dimensionality without losing effective feature information, thereby reducing the computational load in the pattern recognition process and improving recognition efficiency and accuracy.
[0043] The processing unit, for the selected valid nodes exceeding the threshold, combines the spatial coordinate information of each node with the normalized peak amplitude to complete the standardized construction of the node set, which is completely consistent with the node set definition. Simultaneously, based on the amplitude data in the node set, it pre-calculates the vibration amplitude similarity between any two nodes, providing pre-processed data for subsequent node graph construction, edge weight calculation, and node reliability coefficient classification on the cloud server, significantly improving the operating efficiency of the subsequent vibration source localization algorithm. It should be noted that this application is not limited to this; those skilled in the art will understand that the edge weight calculation, node reliability coefficient calculation, and classification can be pre-complete in this processing stage, and the classified node data can be directly uploaded to the signal analysis module, further simplifying the calculation process of the subsequent vibration source localization algorithm.
[0044] In this embodiment, the communication array specifically includes multiple communicators for transmitting the vibration signals collected by each fiber optic vibration sensing unit to a cloud server. It should be noted that the communicators and fiber optic vibration sensing units can be configured in a one-to-one correspondence or a one-to-many correspondence. This application is not limited to this. Those skilled in the art will understand that in this embodiment, a one-to-one industrial Ethernet communicator can be used in a GIL monitoring room scenario with wired network coverage, while a one-to-many wireless communicator can be used in an outdoor long-distance GIL line scenario without wired network coverage. Different communication methods such as LoRa, 5G, and fiber optic ring network can also be selected according to the transmission distance.
[0045] Specifically, in some embodiments, the improved Levenberg-Marquardt algorithm based on the weighted Jacobian matrix includes: Replace the initial Jacobian matrix of the Levenberg-Marquardt algorithm with the weighted Jacobian matrix; Based on the column sum of squares of the weighted Jacobian matrix, a reliability orientation regularization matrix is constructed, and the identity matrix regularization term in the Levenberg-Marquardt algorithm is replaced with the reliability orientation regularization matrix. Based on the weighted Jacobian matrix and the reliability orientation regularization matrix, an improved Levenberg-Marquardt algorithm iterative step size formula is constructed. Combined with an adaptively adjusted damping factor, multiple rounds of iterative solutions are performed until the convergence condition is met, and the optimal coordinates of the vibration source are output.
[0046] Specifically, in some embodiments, generating an undirected weighted node distribution map based on the vibration signals and the vibration positions corresponding to each vibration signal includes: The received vibration signals are filtered by amplitude threshold to remove invalid signals from faulty nodes and retain valid signals whose vibration amplitude exceeds the threshold. A node set is constructed by combining the vibration locations corresponding to the effective signals, and a fully connected undirected weighted node distribution map is generated based on the node set.
[0047] Specifically, in some embodiments, constructing a node set by combining the vibration locations corresponding to the valid signals includes: Calculate the vibration amplitude similarity between any two nodes in the node distribution graph, determine the edge weights between the corresponding nodes based on the vibration amplitude similarity, and form the weight set of the node distribution graph.
[0048] Specifically, in some embodiments, obtaining the node reliability coefficient of each node based on the edge weights between nodes in the undirected weighted node distribution graph includes: Calculate the average edge weight between a single node and all other nodes, and use this average value as the node reliability coefficient for that node.
[0049] Specifically, in some embodiments, updating the initial Jacobian matrix of the Levenberg-Marquardt algorithm based on the node reliability coefficients of each node includes: Based on the physical model of cylindrical wave attenuation in GIL steel structure, the original Jacobian matrix is calculated. The weighted Jacobian matrix is obtained by weighting the corresponding rows of the original Jacobian matrix using the node reliability coefficients of each node.
[0050] Specifically, in some embodiments, the processing module of the cloud server compares the determined vibration source location with a preset safety threshold, generates a warning message of the corresponding level based on the comparison result, and then pushes the warning message to the on-site operation and maintenance terminal and the remote management platform.
[0051] In this embodiment of the application, the cloud server is specifically used to determine the vibration source based on the received vibration signal and the improved Levenberg-Marquardt algorithm. The Jacobian matrix of the improved Levenberg-Marquardt algorithm is obtained by weighting the node reliability coefficients. The node reliability coefficients are obtained by weighting the edge weights between nodes in the undirected weighted node distribution graph. The node distribution graph is generated based on the effective signals in the vibration signal and the vibration positions corresponding to each effective signal.
[0052] In this embodiment, the cloud server's processing module specifically performs amplitude threshold filtering on the received vibration signals, eliminating invalid signals from faulty nodes and retaining valid signals with vibration amplitudes exceeding the threshold. Then, it constructs a node set based on the vibration locations corresponding to the valid signals, and generates a fully connected undirected weighted node distribution map based on the node set. It should be noted that invalid signals specifically include dead node data with no vibration signal output, abnormal node data with amplitudes exceeding the normal range, and noise node data with a signal-to-noise ratio lower than a preset threshold. Such data cannot reflect the true vibration patterns of the GIL support; eliminating it avoids interference from false data in subsequent vibration source localization. Furthermore, each element in the node set corresponds to a valid fiber optic vibration sensing unit node, containing three core attributes: a unique node identifier, spatial coordinates, and measured vibration amplitude. Moreover, those skilled in the art will understand that in this embodiment, the amplitude threshold setting method can be adjusted according to the background vibration level of the GIL support.
[0053] It should be noted that dead node identification specifically identifies nodes that have no vibration signal output for multiple consecutive acquisition cycles and whose signal amplitude is always 0. These nodes are considered dead nodes, as their data cannot reflect the true vibration state of the support and are marked as invalid signals and discarded. Over-range abnormal node identification identifies nodes whose vibration amplitude exceeds the normal measurement range of the sensor and whose amplitude differs from that of surrounding nodes by more than a preset multiple. These nodes are considered abnormal nodes, as their data is likely caused by signal interference or sensor failure and are marked as invalid signals and discarded. Low signal-to-noise ratio (SNR) node identification specifically involves calculating the SNR of the vibration signal for each node and marking node data with an SNR lower than a preset threshold as invalid signals and discarding them. The SNR threshold can be set according to the background noise level at the site, typically set to 3dB~10dB, which can effectively remove invalid data that is drowned out by noise and ensure the validity of subsequent analysis data. After invalid signal removal, the preprocessing module filters valid signals. Based on the background vibration data of the GIL stent under normal vibration-free conditions, it calculates the mean μ and standard deviation σ of the background vibration amplitude. Based on the 3σ criterion, it determines the amplitude threshold, i.e., the threshold is set to μ+3σ. This threshold can statistically cover 99.73% of the background vibration fluctuations. When the peak value of the vibration amplitude of a node exceeds this threshold, it is determined to be a valid signal exceeding the threshold, indicating that the node has detected abnormal vibration of the GIL stent and is an effective analysis object for subsequent vibration source localization. Those skilled in the art will understand that the 3σ criterion is a classic outlier identification method based on the normal distribution, suitable for conventional GIL engineering scenarios with stable background vibration. This application is not limited to this. For complex scenarios with large background vibration fluctuations, such as hydropower stations and GIL lines near major transportation routes, other outlier identification methods such as box plots, DBSCAN (density-based noisy spatial clustering) clustering algorithm, and isolated forest algorithm can be used for effective signal screening. Among them, the box plot method is based on quartiles for outlier identification, is not affected by extreme values, and has stronger robustness; the DBSCAN (density-based noisy spatial clustering) clustering algorithm can identify outliers based on the spatial distribution characteristics of data and is suitable for joint anomaly identification of multi-node data; the isolated forest algorithm is based on the principle of ensemble learning and can effectively identify anomalous samples in high-dimensional data, suitable for weak anomalous signal identification in complex backgrounds. The appropriate effective signal screening method can be selected according to the actual field conditions.
[0054] In this embodiment, after filtering valid signals, the processing module organizes the fiber optic vibration sensing units corresponding to all valid signals into a node set. Each node element in the node set contains three core attributes: a unique node identifier, which corresponds one-to-one with the positioning identifier on the optical fiber, used to uniquely identify the sensing node; node spatial coordinates, which are the physical coordinates of the sensing node on the GIL support; in a one-dimensional linear GIL scenario, these are one-dimensional coordinates along the line direction, and in a two-dimensional planar GIL scenario, these are two-dimensional planar coordinates; and the measured peak vibration amplitude, which is the peak amplitude of the abnormal vibration detected by the node. After completing the node set construction, the node graph construction module generates a fully connected undirected weighted node distribution graph based on the node set, providing a data structure foundation for the subsequent calculation of node reliability coefficients. In this embodiment, the cloud server processing module specifically calculates the vibration amplitude similarity between any two nodes in the node distribution graph, determines the edge weights between corresponding nodes based on the vibration amplitude similarity, and forms a weight set for the node distribution graph. It should be noted that the undirected weighted node distribution graph consists of three parts: a node set, an edge set, and a weight set. The edge set is a fully connected structure, meaning that an undirected edge is set between any two different nodes to represent the correlation of vibration amplitude between the two nodes. Furthermore, the edge weight is calculated based on the similarity of the vibration amplitude of the two nodes. The closer the measured vibration amplitude of the two nodes is, the larger the corresponding edge weight, indicating a higher correlation between the vibration signals of the two nodes and a greater likelihood that they are under the same vibration attenuation law. Conversely, the greater the difference in the measured vibration amplitude of the two nodes, the smaller the corresponding edge weight, indicating a lower correlation and confirming that at least one node may have noise interference. In addition, those skilled in the art will understand that other similarity calculation methods can be used to determine the edge weight in the embodiments of this application, and this application is not limited to this.
[0055] Those skilled in the art will understand that in large-scale monitoring scenarios with a large number of effective nodes, in order to reduce the amount of computation, a K-nearest neighbor graph structure can also be adopted, which connects each node to the K nearest nodes. While preserving the core relationships, this significantly reduces the number of edges and improves computational efficiency. This application is not limited to this. In this embodiment, the edge weights between nodes are specifically calculated using a negative exponential function. The absolute difference in amplitude between two nodes is normalized using the maximum measured vibration amplitude in the node set to eliminate the influence of amplitude magnitude on similarity calculation. Then, the normalized amplitude difference is mapped to the interval between 0 and 1 using a negative exponential function, ensuring that the edge weight value always falls within the range of (0,1). When the measured vibration amplitudes of two nodes are exactly the same, the amplitude difference is 0, and the corresponding edge weight is 1, representing the highest correlation between the vibration amplitudes of the two nodes. The larger the amplitude difference between the two nodes, the closer the corresponding edge weight is to 0, indicating a lower correlation between the two nodes. Using a negative exponential function to calculate the edge weights can produce a sensitive response to small changes in amplitude difference, while ensuring that the weight value will not be negative or exceed a reasonable range.
[0056] Those skilled in the art will understand that, in addition to the negative exponential function method used in this application, various similarity calculation methods can be used to determine edge weights, and this application is not limited to these. In the embodiments of this application, after calculating the edge weights of all node pairs, a weight set of the node distribution graph is formed. The weight set is stored in the form of an adjacency matrix. The number of rows and columns of the adjacency matrix are equal to the number of valid nodes. The element in the i-th row and j-th column of the adjacency matrix is the edge weight between the i-th node and the j-th node. Since the node distribution graph is an undirected graph, the adjacency matrix is a symmetric matrix, that is, the element in the i-th row and j-th column is completely equal to the element in the j-th row and i-th column. For the diagonal elements corresponding to the node itself, they can be set to 1 or 0, which does not affect the calculation of the subsequent node reliability coefficient. In scenarios with a large number of valid nodes, a sparse matrix can be used to store the adjacency matrix to reduce memory usage and improve computational efficiency.
[0057] It is understood that, in this embodiment, for the actual vibration source on the GIL support, the generated elastic waves propagate in the uniform steel structure, following the physical law of cylindrical wave attenuation. That is, the vibration amplitude decreases monotonically with increasing propagation distance. Therefore, normal nodes under the same vibration propagation law will exhibit a distance-related regularity in their vibration amplitude, with high amplitude similarity between nodes and corresponding high edge weights. In contrast, invalid nodes affected by noise interference or local anomalies will deviate from the normal attenuation law in their vibration amplitude, exhibiting extremely low amplitude similarity with most surrounding nodes and corresponding extremely low edge weights. Therefore, by calculating the average edge weight of a single node and all other nodes, the correlation between the node and the overall node set can be comprehensively quantified. The higher the average value, the more the node conforms to the overall vibration attenuation law, and the higher the data reliability; conversely, the lower the average value, the lower the reliability.
[0058] Based on this, in this embodiment, the cloud server's processing module specifically calculates the average edge weight between a single node and all other nodes, using this average value as the node reliability coefficient. Then, based on the numerical range of the node reliability coefficient, all nodes are classified into reliability levels. It should be noted that the node reliability coefficient is used to quantify the effectiveness and reliability of the vibration signal collected by a single fiber optic vibration sensing unit node. A higher coefficient indicates a higher correlation between the node's vibration amplitude and that of most surrounding nodes, meaning the data closely matches the vibration propagation law of the GIL steel structure, resulting in stronger measurement reliability. Conversely, a lower coefficient indicates a lower correlation between the node and surrounding nodes, greater susceptibility to noise and local interference, and poorer measurement reliability. Furthermore, the node reliability coefficient only needs to be calculated once before the iterative solution begins and does not need to be updated during the iteration process, significantly reducing the computational load of the algorithm and improving the efficiency of vibration source localization.
[0059] Furthermore, those skilled in the art will understand that, in the embodiments of this application, the threshold for classifying reliability levels can be adjusted based on the statistical results of measured data from the engineering site, and the number of level classifications can also be adjusted based on the number of nodes.
[0060] Those skilled in the art will understand that, in addition to the arithmetic mean used in this application, other statistical measures can be used to calculate the node reliability coefficient, and this application is not limited to these. In the embodiments of this application, the node reliability coefficient ranges from (0,1). Based on the statistical results of measured data from a large number of GIL engineering sites in China, the embodiments of this application divide the node reliability level into three levels. The specific division rules are as follows: High-reliability nodes, corresponding to node reliability coefficients ≥0.8, these nodes are highly correlated with the vibration amplitude of the vast majority of nodes, and the data completely conforms to the vibration attenuation law of GIL. They are the core data source for solving the vibration source problem and dominate the optimization direction in the subsequent algorithm iteration process; Medium-reliability nodes, corresponding to node reliability coefficients between 0.4 and ... Some nodes exhibit high correlation in vibration amplitude, possessing a certain degree of data reliability, though they suffer from slight noise interference. They provide auxiliary reference in subsequent iterations, but their contribution is lower than that of highly reliable nodes. Low-reliability nodes, with a reliability coefficient <0.4, have extremely low correlation in vibration amplitude with the vast majority of nodes, exhibiting significant noise or local interference, and thus have poor reliability. Their contribution is significantly suppressed in subsequent iterations to avoid false gradients interfering with the iteration direction. It should be noted that the above-mentioned threshold classification is optimized based on some engineering measurement data, which will not be elaborated upon in this application.
[0061] In this embodiment, the cloud server's processing module is specifically based on the cylindrical wave attenuation physical model of the GIL steel structure. It calculates the original Jacobian matrix and then weights the corresponding rows of the original Jacobian matrix using the node reliability coefficients of each node to obtain the weighted Jacobian matrix. It should be noted that the GIL support can be approximated as a uniform and continuous metal steel structure. The vibrating elastic wave propagates in the steel structure in the form of a cylindrical wave, and the vibration amplitude monotonically decreases with the square root of the propagation distance. Based on this physical law, a theoretical vibration amplitude calculation formula can be constructed, thereby obtaining the residual vector and the original Jacobian matrix. Furthermore, the original Jacobian matrix is the first-order partial derivative matrix of the residual vector with respect to the optimization variables. Its number of rows equals the number of effective nodes, and its number of columns equals the dimension of the optimization variables. Each row corresponds to an effective sensing node, and each column corresponds to an optimization variable to be solved. The weighted Jacobian matrix is obtained by weighting the corresponding rows of the nodes using the node reliability coefficients. Weighting the corresponding rows of the original Jacobian matrix strengthens the contribution of residual partial derivatives from high-reliability nodes to the iterative process and suppresses the interference of spurious gradients from low-reliability nodes on the iterative direction. This allows the algorithm's iterative gradient to be dominated by high-reliability nodes, fundamentally improving the accuracy of vibration source solving. Furthermore, those skilled in the art will understand that the elements of the Jacobian matrix can be calculated analytically or numerically in this embodiment. This application is not limited to these methods; using analytical formulas further improves the accuracy of the Jacobian matrix calculation and avoids calculation errors caused by numerical differences. In this embodiment, the vibration source localization problem is transformed into a nonlinear least-squares optimization problem using the above technical solution. Based on the GIL cylindrical wave attenuation physical model, the optimal vibration source coordinates and the maximum amplitude of the vibration source are found, minimizing the sum of squared residuals between the theoretical and measured vibration amplitudes of all effective nodes.
[0062] In this embodiment, the elements of the original Jacobian matrix are the first-order partial derivatives of the residual vector with respect to the optimization variables. The number of rows equals the number of effective nodes, and the number of columns equals the dimension of the optimization variables. The element in the i-th row and j-th column is the first-order partial derivative of the residual of the i-th node with respect to the j-th optimization variable, calculated at the value of the optimization variable in the current iteration. It should be noted that this application, based on the theoretical amplitude model of cylindrical wave attenuation, can directly derive the analytical calculation formula for the elements of the Jacobian matrix without using numerical differencing, thus avoiding truncation and rounding errors caused by numerical differencing. For the one-dimensional linear GIL scenario, the optimization variables are the one-dimensional coordinates and maximum amplitude of the vibration source, and analytical formulas for the partial derivatives of the residuals with respect to the two optimization variables can be derived separately. For the two-dimensional planar GIL scenario, the optimization variables are the two-dimensional planar coordinates and maximum amplitude of the vibration source, and analytical formulas for the partial derivatives of the residuals with respect to the three optimization variables can be derived separately. Substituting the value of the optimization variable in the current iteration into the analytical formula yields all elements of the original Jacobian matrix.
[0063] In this embodiment, the reliability-weighted Jacobian matrix is constructed as follows: for each row of the original Jacobian matrix, multiply by the node reliability coefficient of the corresponding node to obtain the corresponding row of the weighted Jacobian matrix. That is, the i-th row of the weighted Jacobian matrix is equal to the i-th row of the original Jacobian matrix multiplied by the reliability coefficient of the i-th node. It should be noted that the core physical meaning and mathematical principle of this weighting method is that the reliability coefficient of a high-reliability node is close to 1, and its corresponding Jacobian row hardly changes, the residual partial derivative is completely preserved, and it dominates the optimization direction in subsequent iterations; the reliability coefficient of a medium-reliability node is between 0.4 and 0.8, and its corresponding Jacobian row is moderately compressed, and the contribution of the residual partial derivative is reasonably reduced; the reliability coefficient of a low-reliability node is less than 0.4, and its corresponding Jacobian row is significantly compressed, the contribution of the residual partial derivative is greatly suppressed, and the interference of false gradients on the iteration direction is effectively avoided. This weighted approach constructs a weighted nonlinear least squares problem, giving greater weight to the residuals of high-reliability nodes, making the optimization results more closely match the real data of high-reliability nodes. This is equivalent to assuming that the measurement noise variance of high-reliability nodes is smaller. Weighted least squares is a nonlinear extension of the optimal linear unbiased estimation in heteroscedastic noise scenarios, which can guarantee the unbiasedness and optimality of the optimization results.
[0064] In this embodiment, the cloud server's processing module specifically constructs a reliability-oriented regularization matrix based on the column sum of squares of the weighted Jacobian matrix, replacing the identity matrix regularization term in the Levenberg-Marquardt algorithm with the reliability-oriented regularization matrix. It should be noted that the core function of the regularization term is to improve the condition number of the iteration matrix, avoiding instability and iteration divergence problems caused by matrix singularity or ill-conditioned states. The original Levenberg-Marquardt algorithm uses the identity matrix as the regularization term, which imposes indiscriminate constraints on all optimization variables. In contrast, the reliability-oriented regularization matrix in this application is a diagonal matrix, with only diagonal elements being non-zero and all other elements being 0. The dimension of the diagonal matrix is consistent with the dimension of the optimization variables, and each diagonal element corresponds to the regularization constraint strength of an optimization variable. Furthermore, the diagonal elements of the regularization matrix are determined by... The sum of squares of a column in the weighted Jacobian matrix is calculated from the reciprocal of the sum of squares of the corresponding columns. The sum of squares of a column in the weighted Jacobian matrix reflects the effective information content of the corresponding optimization variable. The larger the sum of squares, the greater the contribution of the residual partial derivative of the optimization variable by the high-reliability node, the higher the effective information content, the smaller the corresponding diagonal element, and the weaker the regularization constraint on the variable, thus preserving the flexibility of iteration. Conversely, the smaller the sum of squares, the more the optimization variable is mainly affected by the low-reliability node, the lower the effective information content, the larger the corresponding diagonal element, and the stronger the regularization constraint on the variable, thus suppressing the interference caused by spurious gradients.
[0065] During the iterative process of the LM algorithm, when the approximate Hessian matrix (i.e., the product of the transpose of the Jacobian matrix and itself) obtained by the Gauss-Newton method becomes singular or ill-conditioned, it will lead to an ill-posed problem, causing instability in the inverse matrix solution and resulting in significant fluctuations or even divergence in the iterative results. The singularity or ill-conditioning of the matrix is essentially caused by an excessively large condition number. The condition number is the ratio of the maximum to the minimum singular value of the matrix. The larger the condition number, the more severe the ill-conditioning of the matrix, the more sensitive the inverse matrix solution is to small perturbations in the input data, and the larger the error in the calculation results. The original LM algorithm increases the diagonal elements of the matrix by adding a damping factor multiplied by the identity matrix to the approximate Hessian matrix, thereby reducing the condition number and improving the numerical stability of the matrix. This is a classic Tikhonov regularization method. However, the identity matrix regularization term used in the original algorithm is isotropic, imposing the same strength of constraint on all optimization variables. While suppressing ill-conditioned problems, it also restricts the iterative degrees of freedom of optimization variables with high effective information content, leading to slower convergence speed and even causing the optimization results to deviate from the true values. In contrast, the reliability-oriented regularization matrix designed in this application is an anisotropic Tikhonov regularization, which can differentiate the strength of regularization constraints according to the effective information content of each optimization variable. While ensuring the numerical stability of the matrix, it maximizes the preservation of the iterative degrees of freedom of effective optimization variables, achieving a balance between convergence stability and convergence speed.
[0066] In this embodiment, the reliability orientation regularization matrix is a diagonal matrix. The calculation logic for its diagonal elements is as follows: For the j-th column of the weighted Jacobian matrix, the sum of squares of all elements in that column is calculated. This sum of squares reflects the weighted accumulation of the residual partial derivatives of all nodes with respect to the j-th optimization variable. Since the rows of the Jacobian matrix are already weighted by the node reliability coefficients, this sum of squares essentially reflects the contribution of highly reliable nodes to the j-th optimization variable. The larger the sum of squares, the higher the effective information content of the j-th optimization variable, the stronger the dominance of its iteration direction by highly reliable nodes, and the smaller the influence of spurious gradients. Therefore, a smaller regularization constraint needs to be applied to retain its iteration flexibility. Conversely, the smaller the sum of squares, the lower the effective information content of the j-th optimization variable, and the greater the influence of spurious gradients from less reliable nodes. Therefore, a larger regularization constraint needs to be applied to suppress spurious interference and ensure the stability of the iteration. Infinitesimals are added to the calculation formula for the diagonal elements to avoid the problem of a denominator of 0 when the column sum of squares is 0, further improving the stability of numerical calculation. It should be noted that in the improved iterative step size formula, the product λk·ωm of the adaptive damping factor λk and the diagonal element ωm of the reliability orientation regularization matrix maintains the same dimensions as the elements of the Hessian approximation matrix, ensuring that the regularization term is additive with the Hessian approximation matrix, and the value of λk is adaptively adjusted during the iteration process. Those skilled in the art will understand that in a one-dimensional linear GIL scenario, a simplified orientation regularization matrix can be used, for example, by setting different regularization constraint weights for the vibration source coordinate variables and the maximum amplitude variable, or by adjusting the overall constraint strength of the regularization matrix based on the average reliability coefficient of all nodes; this application is not limited to these methods.
[0067] In this embodiment, the cloud server's processing module specifically constructs an improved Levenberg-Marquardt algorithm iterative step size formula based on the weighted Jacobian matrix and the reliability-oriented regularization matrix. This is combined with an adaptively adjusted damping factor for multiple iterations until the convergence condition is met, outputting the optimal coordinates of the vibration source. It should be noted that in the improved iterative step size formula, the original Jacobian matrix of the original algorithm is replaced by a reliability-weighted Jacobian matrix, and the identity matrix regularization term of the original algorithm is replaced by a reliability-oriented regularization matrix. Simultaneously, the adaptively adjusted damping factor from the original algorithm is retained. This damping factor enables adaptive switching between Gaussian-Newton mode and gradient descent mode. When the iteration is effective, the damping factor is decreased, switching to Gaussian-Newton mode to accelerate convergence; when the iteration is ineffective, the damping factor is increased, switching to gradient descent mode to ensure numerical stability of the iteration. Furthermore, before iterative solving, the initial values of the optimization variables, the initial damping factor, the convergence threshold, and the maximum number of iterations need to be set. The initial vibration source coordinates can be obtained by using multiple highly reliable nodes with the largest amplitudes. The weighted centroid of the point is used, with the weight being the reliability coefficient of the corresponding node. The initial maximum amplitude of the vibration source can be the maximum measured amplitude in the node set. In each iteration, the residual vector and objective function value are calculated sequentially, the original Jacobian matrix is calculated, the weighted Jacobian matrix is constructed, the directional regular matrix is constructed, the iteration step size is calculated, the optimization vector is updated, and the damping factor is adaptively adjusted until the change in the objective function value between two adjacent iterations is less than the convergence threshold, or the number of iterations reaches the maximum number of iterations. Then, the iteration stops and the optimal coordinates of the vibration source are output. In addition, those skilled in the art will understand that in the embodiments of this application, the convergence threshold and the maximum number of iterations can be adjusted according to the actual solution accuracy, and the adjustment step size of the damping factor can also be adjusted according to the optimization scenario. This application is not limited to this.
[0068] It should be noted that in the improved iterative formula, the approximate Hessian matrix is the product of the transpose of the weighted Jacobian matrix and itself, plus the product of the damping factor and the directional regularization matrix. The product of the transpose of the weighted Jacobian matrix and itself is a positive semi-definite matrix, and the directional regularization matrix is a positive definite diagonal matrix. Therefore, regardless of whether the weighted Jacobian matrix is full rank, the sum of the two is always a strictly positive definite matrix, and its inverse matrix always exists. This fundamentally solves the problem of instability in inverse matrix solving caused by matrix singularity and ill-conditionedness in the original algorithm, greatly reducing the probability of iterative divergence and ensuring that the algorithm can still converge stably in engineering scenarios with strong noise and multiple abnormal nodes.
[0069] In this embodiment, the complete iterative process of the improved LM algorithm is divided into three main stages: iterative initialization, single-round iterative execution, and convergence judgment. The design logic and implementation details of each stage are as follows: The first stage is iterative initialization, which requires setting hyperparameters such as the initial values of optimization variables, initial damping factors, convergence threshold, and maximum number of iterations. It should be noted that the selection of the initial values of optimization variables is crucial to the convergence speed of the algorithm and whether it can converge to the global optimum. Random initial values not only increase the number of iterations but may also cause the algorithm to converge to an incorrect local optimum. Therefore, in this embodiment, the initial vibration source coordinates are the weighted centroids of multiple highly reliable nodes with the largest amplitudes, with the weights being the reliability coefficients of the corresponding nodes. This initial value can be as close as possible to the actual vibration source location, significantly reducing the number of iterations, accelerating the convergence speed, and effectively avoiding the algorithm from getting trapped in a local optimum. The initial maximum amplitude of the vibration source is the maximum measured amplitude in the node set, which conforms to the physical law of cylindrical wave attenuation. The initial damping factor can be set to 0.01 to 0.1 and can be adjusted according to the actual scenario. The convergence threshold is set according to the required positioning accuracy and can be set to 1×10. -6 Up to 1×10 - ³, the smaller the convergence threshold, the higher the solution accuracy, but the more iterations are required; the maximum number of iterations is usually set to 50~200 to avoid the algorithm from iterating infinitely.
[0070] The second step involves a single-round iterative execution. Let the current iteration number be k, and the current optimization vector be the value of the k-th iteration. The following steps are executed sequentially: First, calculate the residual vector and objective function value of the current iteration. Substitute the optimization variables of the current iteration into the theoretical amplitude model to calculate the theoretical vibration amplitude of each effective node, thereby obtaining the residual of each node, forming a residual vector. Then calculate the current objective function value, which is half the weighted L2 norm square of the residual vector. Second, calculate the original Jacobian matrix. Based on the analytical partial derivative formula, substitute the optimization variable values of the current iteration to calculate all elements of the original Jacobian matrix. Third, construct the reliability-weighted Jacobian matrix. Multiply the reliability coefficient of each node by the corresponding row of the original Jacobian matrix to obtain the weighted Jacobian matrix. Fourth, construct the reliability-oriented regularization matrix. Calculate the sum of squares of each column of the weighted Jacobian matrix. Calculate the diagonal elements of the regularization matrix based on the reciprocal of the column sum of squares to generate an oriented regularized diagonal matrix. Fifth, calculate the iteration step size. Based on the improved iteration step size formula, substitute... The optimization step size for the current iteration is calculated using the weighted Jacobian matrix, the orientation regularization matrix, the current damping factor, and the residual vector. The sixth step updates the optimization vector by adding the calculated iteration step size to the current iteration's optimization vector, resulting in a candidate optimization vector for the next iteration. The seventh step adaptively adjusts the damping factor by calculating the objective function value corresponding to the candidate optimization vector and comparing it with the objective function value of the current iteration. If the candidate objective function value is less than the current objective function value, the iteration is considered valid, and the candidate optimization vector is accepted as the optimization vector for the next iteration. Simultaneously, the damping factor is reduced (in this embodiment, it is reduced to one-tenth of its original value), switching the algorithm to Gauss-Newton mode to accelerate convergence. If the candidate objective function value is greater than or equal to the current objective function value, the iteration is considered invalid, rejecting the candidate optimization vector and keeping the current optimization vector unchanged. Simultaneously, the damping factor is increased (in this embodiment, it is increased to ten times its original value), switching the algorithm to gradient descent mode, correcting the iteration direction, and recalculating the iteration step size until the iteration is valid. It should be noted that the 10x scaling factor of the damping factor is based on optimization in engineering practice. Those skilled in the art will understand that other scaling factors such as 2x or 5x can be used depending on the actual scenario, and this application is not limited to this.
[0071] The third step is convergence determination. After each iteration, it is determined whether the convergence condition is met. If the absolute difference between the objective function values of two adjacent iterations is less than the convergence threshold, or the maximum number of iterations is reached, the iteration stops, the current optimization vector is output, and the optimal coordinates of the vibration source are extracted from it. If the convergence condition is not met, the iteration count is incremented by one, and the system returns to the first step of the single iteration to continue the next iteration. It should be noted that when the maximum number of iterations is reached and the convergence threshold is still not met, the system will automatically output the current optimal result and issue an algorithm iteration anomaly alarm, reminding maintenance personnel to check the data validity and algorithm parameter settings to ensure the controllability of the algorithm.
[0072] In this embodiment, the improved Levenberg-Marquardt algorithm solves a nonlinear least squares problem with the specific optimization objective of finding an optimal set of vibration source coordinates and the maximum amplitude of the vibration source, such that the sum of squared errors between the theoretical and measured vibration amplitudes of all effective nodes is minimized. The theoretical vibration amplitude is constructed based on the GIL cylindrical wave attenuation physical model, and each element of the residual vector corresponds to the difference between the theoretical and measured vibration amplitudes of an effective node. The objective function is half of the square of the L2 norm of the residual vector. It should be noted that multiplying by half can eliminate the coefficient after subsequent differentiation without affecting the optimization result. Furthermore, the improved objective function incorporates the node reliability coefficient as a weight, forming a weighted nonlinear least squares problem, giving greater weight to the residuals of high-reliability nodes, making the optimization result more closely match the real data of high-reliability nodes.
[0073] The improved Levenberg-Marquardt algorithm of this application is described in detail below.
[0074] It should be noted that for approximately uniform and continuous GIL linear / planar steel structures, the energy of the vibrating elastic wave diffuses cylindrically with the propagation distance, causing the vibration amplitude to decrease monotonically with the square root of the distance. The theoretical amplitude formula is:
[0075] in, Virtual seismic source To the Nodes Euclidean distance in a two-dimensional scene In a one-dimensional scene That is, the closer to the epicenter, the greater the theoretical amplitude; the farther away, the smaller the theoretical amplitude.
[0076] It should be noted that ε is a very small positive regularization constant used to avoid the denominator of the theoretical amplitude formula being zero when the virtual seismic source and the sensing node position coincide. Its dimension is consistent with the distance dimension, and its value is much smaller than the characteristic distance scale in the GIL support monitoring scenario. This application does not limit the specific value of ε.
[0077] In this embodiment, the Levenberg-Marquardt (LM) algorithm is a nonlinear optimization algorithm between gradient descent and Gauss-Newton methods. For nonlinear least squares problems, the conventional Gauss-Newton method modifies the residual vector... Perform a first-order Taylor expansion to linearize the problem, and then solve the linear least squares problem to obtain the iteration step size.
[0078] Let the first The optimization variables for the next iteration are For the residual vector in Perform a first-order Taylor expansion:
[0079] in, For the first The iteration step size of the next iteration, i.e. ; The Jacobian matrix is the first-order partial derivative matrix of the residual vector with respect to the optimization variables, with dimension 1. ( To optimize variable dimensions, two-dimensional One-dimensional ), its first Line 1 The elements of the column are:
[0080] Substituting the first-order Taylor expansion into the objective function, we obtain a linearized approximate objective function:
[0081] For the approximate objective function with respect to Taking the derivative and setting it to zero, we obtain the iteration step size formula for the Gauss-Newton method:
[0082]
[0083] In this embodiment of the application, for the sake of simplicity, the following will use replace , replace .
[0084] The core problem of the Gauss-Newton method is: when (Approximation of the Hessian matrix) When the inverse matrix is singular or ill-conditioned, it does not exist or is numerically unstable, leading to iterative divergence.
[0085] The formula for the iteration step size of the gradient descent method is:
[0086] in, The learning rate is used to ensure convergence by directly descending along the negative gradient direction, but the convergence speed is extremely slow.
[0087] The core improvement of the LM algorithm is the addition of a damping factor to the Hessian approximation matrix of the Gauss-Newton method. With the identity matrix The product of these terms preserves the fast convergence of the Gauss-Newton method while ensuring numerical stability through a damping term:
[0088] in, For the first The damping factor in the next iteration is adaptively adjusted. If this iteration reduces the objective function ( This indicates that the iteration is effective. Reduce damping and switch to Gauss-Newton mode to accelerate convergence; If this iteration increases the objective function ( This indicates that the iteration is invalid. Increase the damping, switch to gradient descent mode, and pull back to the correct direction.
[0089] In this embodiment of the application, based on the selected For nodes exceeding the threshold, construct a fully connected undirected weighted node graph. This is the underlying data structure for all improvements.
[0090] like Figure 3 As shown, the nodes in the undirected weighted node distribution graph are labeled V1 to V10, corresponding one-to-one with the effective sensing nodes exceeding the threshold. The undirected edges between nodes are labeled with edge weights. Figure 3 The edge weights shown range from 0.15 to 0.92. The larger the edge weight, the higher the correlation between the vibration amplitudes of the two nodes.
[0091] The node set corresponds exactly to the over-threshold sensor nodes:
[0092] Each node The attributes include: known coordinates Measured amplitude Where the coordinates are known For the first i Physical coordinates of each node on the GIL support, measured amplitude For the first i The peak value of vibration amplitude detected by each node.
[0093] Construct a fully connected graph, that is, any two nodes and ( There is an undirected edge between each of them. :
[0094] Physical meaning: There is a correlation in vibration amplitude between any two nodes, and the existence of an edge represents this correlation.
[0095] side weight Defined as the similarity of vibration amplitudes between two nodes, the quantification formula is:
[0096] in, The maximum measured amplitude in the set of nodes exceeding the threshold is... , used to normalize the amplitude difference; This represents the absolute difference between the measured amplitudes at the two nodes. Physical meaning: The closer the measured amplitudes of the two nodes are, the better. The smaller, The closer it is to 1, the higher the correlation between the vibration amplitudes of the two nodes, and the more likely they are to be under the same decay law; The greater the difference in measured amplitude between the two nodes, The closer to 0, the lower the correlation, and at least one node may be subject to noise interference.
[0097] The measurement reliability of each node is quantified by the average edge weight of the node graph. The higher the reliability of the node, the greater its contribution to the iteration.
[0098] For the Nodes Its reliability coefficient Defined as the average of the edge weights between this node and all other nodes:
[0099] in, The total number of all nodes other than itself; For the node graph and The edge weights (defined as described above); Physical meaning: , The closer the value is to 1, the higher the correlation between the amplitude of the node and most of the surrounding nodes. The data perfectly matches the global decay law of GIL, and the measurement reliability is extremely high. The closer the value is to 0, the lower the correlation between the node and its surrounding nodes, indicating potential issues such as fiber optic bends, signal noise, and localized interference, resulting in extremely poor measurement reliability.
[0100] Based on the reliability coefficient, all nodes are divided into 3 levels. The thresholds are optimized based on statistical data from the GIL project site, as shown in Table 1. Table 1 Node Reliability Level Classification Table
[0101] Key Note: Reliability Coefficient It only needs to be calculated once before the iteration begins, and does not need to be updated with each iteration, because the measurement reliability of the nodes is determined by the edge weights of the node graph, which is independent of the virtual source location of the current iteration, thus greatly reducing the amount of computation.
[0102] By using the aforementioned reliability coefficients and the rows of the Jacobian matrix weighted by the reliability coefficients, the contribution of the residual partial derivatives of high-reliability nodes to the iteration is enhanced, while the spurious gradient interference of low-reliability nodes is suppressed.
[0103] In the traditional LM algorithm, the Jacobian matrix under the GIL cylindrical wave attenuation model The analytical partial derivative formula is taken as an example in a two-dimensional GIL scenario (optimization vector). ), residual The partial derivatives with respect to the three optimization variables are as follows: For virtual seismic sources x coordinate Partial derivative:
[0104] For virtual seismic sources y coordinate Partial derivative:
[0105] Maximum amplitude of virtual source Partial derivative:
[0106] For a one-dimensional GIL scenario (optimized vector) The partial derivative simplifies to: right Partial derivative:
[0107] in, It is a symbolic function.
[0108] right Partial derivative:
[0109] Original Jacobi Matrix The The line represents these three (or two) partial derivatives in Values at:
[0110] Specifically, for each row of the Jacobian matrix above, multiply by the reliability coefficient of the corresponding node. The reliability-weighted Jacobian matrix is obtained. :
[0111] in, The first weighted Jacobian matrix OK; For the first The reliability coefficient of each node (as defined above); Physical meaning: High reliability node ( ): The Jacobian line remains almost unchanged, the residual partial derivatives are preserved, and they dominate the direction of iteration; Medium reliable node ( ): The Jacobian line is moderately compressed, and the contribution of the residual partial derivative is reduced; Low reliability node ( ): The Jacobian line is significantly compressed, and the contribution of the residual partial derivative is greatly suppressed, avoiding false gradient interference with the iteration.
[0112] In the LM algorithm, the regularization term increases the diagonal element values of the Hessian approximation matrix, improves the condition number of the matrix, avoids matrix singularity or ill-conditionedness, and ensures the numerical stability of the inverse matrix.
[0113] The regularization term of the traditional identity matrix is ,in yes identity matrix ( To optimize variable dimensions, the constraints on all diagonal elements are the same.
[0114] This application upgrades the regularization term from an "identity matrix constraint" to a "reliability-oriented constraint," and designs a reliability-oriented regularization matrix. :
[0115] in, This is a diagonal matrix generating function, where only the diagonal elements are non-zero, and the rest are 0; To optimize variable dimensions (two-dimensional) One-dimensional ); Key design: Diagonal elements of a regular matrix ( Instead of directly relying on the reliability of a single node, it indirectly relates node reliability through the column squares of the weighted Jacobian matrix, ensuring that the regularization term matches the dimension of the optimization variable, resulting in a clearer physical meaning.
[0116] in, Reliability-weighted Jacobian matrix The Line 1 Column elements; For the weighted Jacobian matrix, the first... The sum of squares of a column reflects the "effective information content" of the optimization variable. The larger the sum of squares, the greater the contribution of the high-reliability node to the residual partial derivative of the variable, and the higher the effective information content. If the effective information content of an optimization variable is high (large sum of squares), then Smaller size means less constraint on the variable, preserving iteration flexibility; If the effective information content of a certain optimization variable is low (small sum of squares, indicating that it is mainly affected by low-reliability nodes), then A large value imposes a strong constraint on the variable, thus suppressing spurious gradient interference.
[0117] If the GIL scenario is linear and one-dimensional, and only one of the optimization variables is used... (Virtual source coordinates) are affected by node reliability. (Maximum amplitude) is a global variable; a simpler orienting regularization matrix can be used instead.
[0118] in The average reliability coefficient is the sum of all nodes' reliability coefficients. A higher average reliability coefficient indicates better reliability for all nodes. The smaller the constraint.
[0119] The improved reliability-weighted Jacobian matrix and reliability orientation regularization term Substituting into the traditional LM iterative formula, we obtain the improved complete iterative formula.
[0120] First, the improved objective function (incorporating reliability weights, corresponding to the weighted Jacobian) is:
[0121] in, This is the diagonal matrix representing node reliability.
[0122] Then, a first-order Taylor expansion is performed on the improved objective function, the derivative is taken and set to 0, and combined with the orientation regularization term, the improved LM iteration step size formula is finally obtained:
[0123] in, A reliability-weighted Jacobian matrix; For reliability orientation regularization matrix; The damping factor is adjusted according to the same rules as traditional LM (the objective function decreases). If it increases ); Traditional residual vector; weighted Jacobian matrix Allow the iterative gradient to be dominated by highly reliable nodes; directional regularization term Apply stronger constraints to optimization variables with low effective information content to ensure numerical stability; damping factor Adaptively switch convergence modes to balance speed and stability.
[0124] The following specific scenario examples will be provided to explain this application in detail.
[0125] Filtering nodes exceeding the threshold: Filtering out Approximately 30 fiber optic sensor nodes had vibration amplitudes exceeding the threshold (background value + 3 standard deviations), and obvious faulty nodes (such as those with amplitudes more than 10 times that of other nodes) were removed. Construct a fully connected weighted node graph: Calculate the edge weights for all pairs of nodes. Build a node graph ; Calculate and classify node reliability coefficients: Calculate the reliability coefficient for each node. It is divided into three levels: high, medium, and low reliability. Set initial values and hyperparameters: Initial virtual source coordinates Take the weighted centroids of the three nodes with the largest amplitudes (weights are...). ); Initial maximum amplitude Take the maximum measured amplitude from the node set. ; Initial optimization vector ; Initial damping factor ; Convergence threshold (Convergence occurs when the change in the objective function is less than this value); Maximum number of iterations (Avoid infinite iteration).
[0126] Iterative process (executed in each iteration) Let the current iteration number be... The current optimization vector is Perform the following steps: Step 1: Calculate the residual vector and the objective function value: Calculate the residual vector ; Calculate the improved objective function value .
[0127] Step 2: Calculate the traditional Jacobian matrix: Calculate the traditional Jacobian matrix .
[0128] Step 3: Construct the reliability-weighted Jacobian matrix: right Multiply each row ,get .
[0129] Step 4: Construct the reliability orientation regularization matrix: Calculate the column sum of squares of the weighted Jacobian matrix and construct the directional regularization matrix. .
[0130] Step 5: Calculate the iteration step size: Calculate the iteration step size .
[0131] Step 6: Update and optimize vectors: Obtain the optimization vector for the next iteration. .
[0132] Step 7: Calculate the new objective function value and adjust the damping factor: calculate The corresponding residual vector and objective function value ; like (Iteration valid): accept ; Reduce damping factor: ; like (Iteration invalid): reject ,Keep constant; Increase the damping factor: ; Return to step 5 and use the new one. Recalculate the iteration step size.
[0133] Step 8: Convergence check: like or : Stop iteration and output the optimal virtual source coordinates. (forward The first element is a coordinate, and the last element is... ); otherwise: Return to step 1 and continue iterating.
[0134] In this embodiment, the Hessian approximation matrix is: ,in: It is a positive semi-definite matrix; It is a positive definite diagonal matrix (all diagonal elements) ); therefore, It is a strictly positive definite matrix, its inverse matrix always exists, and it has high numerical stability, greatly reducing the probability of iterative divergence.
[0135] The following verification is based on field measurement data from a domestic 500kV GIL project, and the results are shown in Table 2: Table 2 Comparison of the effects of the improved scheme in this application and the traditional LM algorithm
[0136] Furthermore, in this embodiment, the cloud server's processing module specifically compares the determined vibration source location with a preset safety threshold, generates a corresponding level of warning information based on the comparison result, and then pushes the warning information to the on-site maintenance terminal and the remote management platform. It should be noted that the safety threshold can be preset based on the GIL device's operating procedures and historical fault data, and can be divided into multiple threshold levels, corresponding to different warning levels. Different warning levels correspond to different handling strategies. Furthermore, the warning information specifically includes the precise location of the vibration source, vibration characteristic parameters, warning level, fault risk analysis, and handling suggestions. It can be simultaneously pushed to the on-site maintenance terminal and the remote management platform through various methods to ensure that maintenance personnel can obtain the warning information in a timely manner.
[0137] In this embodiment, the warning level can be divided into three levels based on the safety threshold: Level 1, Level 2, and Level 3. Each level corresponds to a different fault risk level and handling procedure: Level 1 is a warning of caution, corresponding to a vibration amplitude slightly exceeding the safety threshold with no obvious structural fault risk. Only routine inspections are needed to focus on the vibration source location and record vibration trends. Level 2 is an alarm warning, corresponding to a vibration amplitude significantly exceeding the safety threshold, indicating potential structural loosening, component abnormalities, or other fault risks. On-site investigation is required within a specified time (e.g., 72 hours) to confirm the cause of the abnormality and take appropriate measures. Level 3 is an emergency warning, corresponding to a vibration amplitude significantly exceeding the safety threshold, indicating an emergency fault risk such as structural cracking, insulator detachment, or pipeline deformation. Immediate shutdown of the relevant equipment and emergency response are required to avoid serious safety accidents. Those skilled in the art will understand that the classification of warning levels and corresponding handling procedures can be adjusted according to the voltage level, importance, and on-site conditions of the GIL equipment; this application is not limited to this.
[0138] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
Claims
1. A GIL (Gas Inertia Reinforced Insulator) support vibration intelligent monitoring and early warning system, characterized in that, include: The distributed optical fiber vibration sensing module includes multiple optical fiber vibration sensing units and optical fibers arranged on the GIL support to be monitored. Each optical fiber vibration sensing unit collects vibration signals at the corresponding sampling location point. The communication array, comprising multiple communicators, is used to transmit the vibration signals collected by each fiber optic vibration sensing unit to a cloud server. A cloud server is used to determine the vibration source based on the received vibration signal and an improved Levenberg-Marquardt algorithm. The Jacobian matrix of the improved Levenberg-Marquardt algorithm is obtained by weighting the node reliability coefficients. The node reliability coefficients are obtained by weighting the edge weights between nodes in an undirected weighted node distribution graph. The undirected weighted node distribution graph is generated based on the vibration location corresponding to the vibration signal.
2. The system according to claim 1, characterized in that, The optical fibers of the distributed optical fiber sensing vibration module are arranged along the support legs, crossbeam connection parts, and insulator mounting parts of the GIL bracket.
3. The system according to claim 1, characterized in that, The communicator of the communication array is configured in a one-to-one correspondence or a one-to-many correspondence with the fiber optic vibration sensing unit.
4. A method for intelligent monitoring and early warning of vibration of a GIL (Gas Insulator) support, characterized in that, The method described in claim 1 is applied to a GIL support vibration intelligent monitoring and early warning system, and the method includes: Based on the vibration signals and the vibration positions corresponding to each vibration signal, an undirected weighted node distribution map is generated, wherein the coordinates of each node in the undirected weighted node distribution map correspond to the vibration positions. Based on the edge weights between nodes in the undirected weighted node distribution graph, the node reliability coefficient of each node is obtained. Based on the node reliability coefficients of each node, the initial Jacobian matrix of the Levenberg-Marquardt algorithm is updated again. An improved Levenberg-Marquardt algorithm is formed based on the weighted Jacobian matrix; The vibration source is determined based on the undirected weighted node distribution map and the improved Levenberg-Marquardt algorithm.
5. The method according to claim 4, characterized in that, The improved Levenberg-Marquardt algorithm based on the weighted Jacobian matrix includes: Replace the initial Jacobian matrix of the Levenberg-Marquardt algorithm with the weighted Jacobian matrix; Based on the column sum of squares of the weighted Jacobian matrix, a reliability orientation regularization matrix is constructed, and the identity matrix regularization term in the Levenberg-Marquardt algorithm is replaced with the reliability orientation regularization matrix. Based on the weighted Jacobian matrix and the reliability orientation regularization matrix, an improved Levenberg-Marquardt algorithm iterative step size formula is constructed. Combined with an adaptively adjusted damping factor, multiple rounds of iterative solutions are performed until the convergence condition is met, and the optimal coordinates of the vibration source are output.
6. The method according to claim 4, characterized in that, The step of generating an undirected weighted node distribution map based on the vibration signals and the corresponding vibration positions includes: The received vibration signals are filtered by amplitude threshold to remove invalid signals from faulty nodes and retain valid signals whose vibration amplitude exceeds the threshold. A node set is constructed by combining the vibration locations corresponding to the effective signals, and a fully connected undirected weighted node distribution map is generated based on the node set.
7. The method according to claim 6, characterized in that, The construction of the node set by combining the vibration positions corresponding to the effective signals includes: Calculate the vibration amplitude similarity between any two nodes in the node distribution graph, determine the edge weights between corresponding nodes based on the vibration amplitude similarity, and form the weight set of the node distribution graph.
8. The method according to claim 4, characterized in that, The node reliability coefficient of each node is obtained based on the edge weights between nodes in the undirected weighted node distribution graph, including: Calculate the average edge weight between a single node and all other nodes, and use this average value as the node reliability coefficient for that node.
9. The method according to claim 4, characterized in that, The process of updating the initial Jacobian matrix of the Levenberg-Marquardt algorithm based on the node reliability coefficients of each node includes: Based on the physical model of cylindrical wave attenuation in GIL steel structure, the original Jacobian matrix is calculated. The weighted Jacobian matrix is obtained by weighting the corresponding rows of the original Jacobian matrix using the node reliability coefficients of each node.
10. The method according to claim 4, characterized in that, The method further includes: The determined vibration source location is compared with the preset safety threshold. Based on the comparison result, a corresponding level of early warning information is generated and then pushed to the on-site operation and maintenance terminal and the remote management platform.
