An online monitoring system and method for power transmission conductor spacers

Abstract

The application discloses a kind of transmission conductor spacer rod online monitoring system and method, the present application relates to transmission line online monitoring technical field, solve the isolated collection of multi-source data, lack the technical problem of depth correlation and causal inference, the present application is spliced into joint feature vector by time alignment after spacer rod health state, conductor galloping parameter, sag variation and environmental meteorological data, context embedding vector is extracted by Transform encoder, and system-level anomaly detection is carried out in the mode of health mode prototype matching, compared with the anomaly judgment of single parameter in the prior art, the present application can realize multi-source data depth correlation and system-level anomaly detection, combine group concentration index and divide anomaly distribution into three categories of high concentration type, moderate concentration type and high dispersion type, respectively perform differentiated root cause positioning processing, realize the accurate positioning and differentiated disposal of abnormal reason, reduce false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of online monitoring technology for power transmission lines, specifically to an online monitoring system and method for power transmission conductor spacers. Background Technology

[0002] Transmission lines operate in complex outdoor environments for extended periods, enduring multiple stresses including wind loads, icing, temperature cycling, and electromagnetic stress. Conductor galloping can cause short circuits, tower collapses, line breaks, and even grid paralysis. Spacer bars, as key hardware for split conductors, function to maintain conductor spacing, suppress wind vibrations, and prevent whiplash and galloping. When spacer bars become loose, break, shift, or fail to dampen, it directly leads to increased abnormal conductor vibration, accelerates hardware fatigue damage, and in severe cases, causes phase-to-phase short circuits or line breaks.

[0003] Although existing monitoring systems have attempted to integrate various sensing components such as accelerometers, positioning modules, and meteorological sensors, the fusion depth between multi-source data is insufficient. Among them, sag monitoring, droop monitoring, and temperature monitoring are independent, and there is a lack of effective time alignment and collaborative analysis mechanisms between the data. Anomaly detection mainly relies on threshold judgment of a single parameter, which cannot capture the combined anomaly patterns formed by the collaborative changes of multiple parameters. In terms of anomaly cause localization, existing technologies can only determine that an anomaly has occurred, but it is difficult to quantify the contribution ratio of each factor and accurately trace the root cause. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an online monitoring system and method for transmission line spacers, which solves the problems of isolated acquisition of multi-source data and lack of in-depth correlation and causal inference.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an online monitoring method for transmission line spacers, comprising the following steps: Step 1: Obtain strain signal and vibration acceleration signal, extract strain time domain features and vibration frequency domain features respectively, combine the two into feature tensor and normalize them, then input them into lightweight neural network model, output stress anomaly probability, spectrum deviation probability and health index, and identify stress anomaly or vibration anomaly based on the output results. Step 2: Obtain the real-time displacement of the traverse in the geographic coordinate system, calculate the total amplitude of galloping, the horizontal composite amplitude, the vertical amplitude, and the dominant frequency of galloping based on the displacement time history, perform principal component analysis on the trajectory points in the horizontal plane, and calculate the ellipse tilt angle. Step 3: Obtain the real-time elevation of the conductor and calculate the elevation change in combination with the initial elevation. Obtain the conductor temperature. Calculate the current sag and sag rate based on the initial sag and elevation change. Obtain meteorological data through a micro-meteorological sensor. Step 4: Concatenate the spacer health status, conductor galloping parameters, sag changes and meteorological data into a joint feature vector, extract the context embedding vector, cluster the health pattern prototypes under normal line conditions, calculate the similarity between the current embedding vector and each health pattern prototype, and determine the system-level anomaly when the similarity is below the threshold and continues to exceed the set period. Step 5: For system-level anomalies, using the health model prototype as a baseline reference, calculate the marginal contribution value and contribution percentage, select the two data types with the highest contribution, decompose the contribution of each feature within each data type, and obtain the feature contribution ranking within the group. Step 6: Calculate the ratio of the maximum feature contribution within a group to the total contribution of the group to obtain the concentration index within the group, and identify highly concentrated, moderately concentrated and highly dispersed types based on this index. Perform differentiated root cause localization processing for different distribution types.

[0006] As a further aspect of the present invention, the strain time-domain characteristics include the average strain value. Peak strain and standard deviation The vibration frequency domain characteristics include the dominant frequency f0, the dominant frequency amplitude A0, and the center frequency offset Δf; If the probability of stress anomaly is ≥0.7, the peak strain ≥55% of the material's yield strain, or the average strain over three consecutive windows If the stress shows a monotonically increasing trend with a cumulative increase of ≥15%, or if the probability of stress anomaly is ≥0.85, then stress anomaly is determined. If the probability of spectrum deviation is ≥0.75, then the vibration is determined to be abnormal.

[0007] As a further aspect of the present invention, the total amplitude of the dance S E (t), S N (t) and S U (t) represents the horizontal, vertical, and lateral displacements of the traverse in the geographic coordinate system, respectively; Horizontal composite amplitude Vertical amplitude The main frequency of the dance is controlled by the S frequency within a 60-second window. U (t) After the sequence is decentered, a fast Fourier transform is performed to find the frequency point with the largest power spectrum peak in the 0.1 to 3 Hz frequency band; The method for calculating the tilt angle of an ellipse is as follows: Take the set of points P on the trajectory of the horizontal plane. i =[S E (i), S N (i)], calculate the covariance matrix , Let v represent the mean point of the horizontal displacement, and perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvector v corresponding to the largest eigenvalue. max Ellipse tilt angle v max-E and v max-N These represent the feature vectors in the horizontal direction and the vertical direction, respectively.

[0008] As a further aspect of the present invention, the elevation change H0 is the initial elevation, H z For actual elevation, sag rate ,in Indicates the time-varying window, current sag ; Meteorological data includes wind speed, wind direction, ambient temperature, humidity, and air pressure.

[0009] As a further aspect of the present invention, the context embedding vector extraction includes: aligning the obtained spacer health status H, conductor galloping parameters G, sag change S, and environmental meteorological data M in time, and concatenating them into a joint feature vector V(t) = [H(t), G(t), S(t), M(t)], and using a Transformer-based encoder to extract temporal features from the joint feature vector sequence of several past time steps; The method for determining system-level anomalies is as follows: collect the context embedding vectors under normal line operation, and cluster them to obtain several health mode prototypes. Calculate the cosine similarity between the current context embedding vector E and the context embedding vectors of each health mode prototype, and take the maximum value as the current health similarity value. If the health similarity value is <0.75 and continues for more than 3 analysis cycles, it is determined to be a system-level anomaly.

[0010] As a further aspect of the present invention, the marginal contribution value and contribution percentage are calculated as follows: Marginal contribution value: Using the data values ​​of various types in the health pattern prototype corresponding to the highest health similarity as the baseline reference value, and starting from the baseline reference value, each group of data is replaced from the baseline value to the current value one by one. The difference between the replaced health similarity and the original health similarity is calculated to obtain the marginal contribution value C. A ; Contribution percentage: Marginal contribution value C A The contribution percentage is obtained by normalization. .

[0011] As a further aspect of the present invention, the intra-group concentration index , where a represents the group feature type in data type A, and b represents the total number of group features; like If the value is greater than 0.6, it indicates that the group's characteristics are highly concentrated; if 0.4 ≤ If ≤0.6, it indicates moderate concentration. If the value is less than 0.4, it indicates a highly dispersed type.

[0012] As a further aspect of the present invention, for highly concentrated distributions: the feature with the largest contribution within the group is taken as the main control feature, the abnormal start time is determined, and the abnormal start time is compared with the external time axis to determine the root cause of the abnormality. For moderately concentrated distributions: the top three features in terms of contribution are selected as key anomaly features. The matching degree of each key feature with the anomaly records of the external timeline is checked one by one. When the matching degree of at least one key feature exceeds the preset threshold, the matching feature is determined as the root cause of the anomaly. When multiple key features match different external timelines, the feature with the highest matching degree is selected first. When there are no key features that meet the matching threshold, all key features are marked as items to be investigated and pushed to the operation and maintenance personnel.

[0013] As a further aspect of the present invention, for highly dispersed distributions: obtain all intra-group features to form a full-dimensional feature vector, and calculate the migration direction angle relative to the historical baseline. ,in Represents the historical baseline mean vector. This represents the main direction vector of seasonal migration. If the angle is less than 30°, it is considered normal seasonal migration; if the angle is less than or equal to 30°, it is considered normal seasonal migration. If the temperature is ≤60°, it is considered an abnormal climate drift. If the angle is greater than 60°, it is determined to be a systematic drift of the sensor group.

[0014] An online monitoring system for transmission line spacers includes: The data sensing and acquisition module includes strain sensors and vibration sensors for acquiring strain signals and vibration acceleration signals. It also includes temperature sensors, current sensors, and acceleration and attitude composite sensors, as well as micro-meteorological sensors and Beidou high-precision positioning units. All sensors acquire data synchronously with a unified clock source. The acquired data is preprocessed and stored in a local cache. The data intelligent processing module is used to extract the strain time-domain features and vibration frequency-domain features corresponding to the strain signal and vibration acceleration signal, combine the two into a feature tensor and normalize it before inputting it into a lightweight neural network model, and output the stress anomaly probability, spectrum deviation probability and health index. When the stress anomaly probability meets the first judgment condition, it is judged as stress anomaly, and when the spectrum deviation probability meets the second judgment condition, it is judged as vibration spectrum deviating from the normal range. It is also used to obtain the real-time displacement of the conductor in the geographic coordinate system and the real-time elevation of the conductor based on the Beidou high-precision positioning unit, calculate the total amplitude of the galloping, the horizontal composite amplitude, the vertical amplitude and the dominant frequency of the galloping based on the displacement time history, and perform principal component analysis on the trajectory points in the horizontal plane to calculate the ellipse tilt angle. The elevation change is calculated based on the real-time elevation and the initial elevation. The current sag and sag rate are calculated based on the initial sag and elevation change measured during installation. The multi-source data fusion and analysis module is used to align the health status of spacers, conductor galloping parameters, sag changes, and meteorological data by time and then stitch them together into a joint feature vector. The joint feature vector sequence of continuous time steps is processed by a Transformer encoder to extract context embedding vectors. The embedding vectors collected under normal line operating conditions are clustered to form health pattern prototypes. The similarity between the current embedding vector and each health pattern prototype is calculated. When the similarity is lower than the threshold and continues to exceed the set period, it is judged as a system-level anomaly. For system-level anomalies, using the health model prototype as a baseline reference, each type of data is replaced with the current value, and the similarity difference before and after the replacement is calculated as the marginal contribution value of each type of data. After normalization, the contribution percentage of each data type is obtained and sorted. The two types of data with the highest contribution are selected, and the contribution of each feature within them is decomposed to obtain the feature contribution ranking within the group. The intelligent monitoring and cause determination module is used to calculate the concentration index within a group based on the contribution of characteristics within the group. When the index is greater than the first threshold, it is determined to be highly concentrated; when the index is between the first and second thresholds, it is determined to be moderately concentrated; and when the index is less than the second threshold, it is determined to be highly dispersed. For highly concentrated distributions, the feature with the largest contribution within a group is used as the main control feature for single-cause tracing. For moderately concentrated distributions, multiple key features are matched and identified with an external time axis. For highly dispersed distributions, the migration direction angle of the full-dimensional feature vector relative to the historical baseline is calculated, and the anomaly type is determined based on the migration direction angle.

[0015] This invention provides an online monitoring system and method for transmission line spacers. Compared with the prior art, it has the following advantages: This invention aligns the health status of spacers, conductor galloping parameters, sag changes, and environmental meteorological data over time and concatenates them into a joint feature vector. The context embedding vector is then extracted using a Transformer encoder, and system-level anomaly detection is performed using a health pattern prototype matching method. Compared to the anomaly judgment of a single parameter in the prior art, this invention can achieve deep correlation of multi-source data and system-level anomaly detection.

[0016] This invention quantifies the percentage contribution of various data to anomalies by decomposing marginal contributions and, combined with the intragroup concentration index, classifies anomaly distribution into three categories: highly concentrated, moderately concentrated, and highly dispersed. Differentiated root cause localization processing is then performed on each category. Compared with existing technologies that uniformly process all anomalies, this invention achieves precise localization and differentiated handling of anomaly causes, thereby reducing the false alarm rate. Attached Figure Description

[0017] Figure 1 This is a flowchart of the online monitoring method for transmission line spacers according to the present invention; Figure 2 This is a system block diagram of the online monitoring method for transmission line spacers according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] First Embodiment Please see Figure 1 and Figure 2 This application provides an online monitoring method for transmission line spacers, which specifically includes the following steps: The main control module, intelligent processing module, sensing module, communication module and power supply module are arranged in separate cavities inside the spacer bar structure, and each module is integrated with the spacer bar body. Specifically, micro strain sensors and vibration sensors are embedded in the connection parts between the line clamp and the spacer bar body. The micro strain sensors are preferably resistive strain gauges, and the vibration sensors are preferably triaxial MEMS accelerometers. Temperature sensors, current sensors and acceleration and attitude composite sensors are embedded in each line clamp. A micro meteorological sensor and a Beidou high-precision positioning module are integrated at the top of the central body of the spacer bar. Based on the installed multi-source sensors, a unified clock source is used to synchronously acquire raw data from the multi-source sensors at a sampling frequency of not less than 30Hz. The acquired raw data is preprocessed to remove outlier data and perform data calibration. Based on the preprocessed raw data, the following implementation steps are performed: Step 1: Obtain strain signal and vibration acceleration signal, extract strain time domain features and vibration frequency domain features respectively, combine the two into feature tensor and normalize them, then input them into lightweight neural network model, output stress anomaly probability, spectrum deviation probability and health index, and identify stress anomaly or vibration anomaly based on the output results. Step 2: Obtain the real-time displacement of the traverse in the geographic coordinate system, calculate the total amplitude of galloping, the horizontal composite amplitude, the vertical amplitude, and the dominant frequency of galloping based on the displacement time history, perform principal component analysis on the trajectory points in the horizontal plane, and calculate the ellipse tilt angle. Step 3: Obtain the real-time elevation of the conductor and calculate the elevation change in combination with the initial elevation. Obtain the conductor temperature. Calculate the current sag and sag rate based on the initial sag and elevation change. Obtain meteorological data through a micro-meteorological sensor. Step 4: Concatenate the spacer health status, conductor galloping parameters, sag changes and meteorological data into a joint feature vector, extract the context embedding vector, cluster the health pattern prototypes under normal line conditions, calculate the similarity between the current embedding vector and each health pattern prototype, and determine the system-level anomaly when the similarity is below the threshold and continues to exceed the set period. Step 5: For system-level anomalies, using the health model prototype as a baseline reference, calculate the marginal contribution value and contribution percentage, select the two data types with the highest contribution, decompose the contribution of each feature within each data type, and obtain the feature contribution ranking within the group. Step 6: Calculate the ratio of the maximum feature contribution within a group to the total contribution of the group to obtain the concentration index within the group, and identify highly concentrated, moderately concentrated and highly dispersed types based on this index. Perform differentiated root cause localization processing for different distribution types.

[0020] Second Embodiment As a second embodiment of the present invention, it is implemented based on the first embodiment, and the difference from the first embodiment is as follows: For step one in the first embodiment, the specific processing procedure includes: the sensing module acquires real-time signals from the micro strain sensor and the vibration sensor within a fixed time window to obtain the original strain signal. The vibration acceleration signal a(t) is filtered by a low-pass filter with a cutoff frequency of 10Hz for the strain signal and a low-pass filter with a cutoff frequency of 50Hz for the vibration acceleration signal. For the filtered strain signal Calculate the average strain value within a fixed time window. Peak strain and standard deviation For the filtered vibration acceleration signal a(t), a fast Fourier transform is performed on the acceleration data within a fixed time window to extract the power spectral density in the 0 to 30 Hz frequency band, and the main frequency f0, the main frequency amplitude A0 and the center frequency offset Δf are calculated. The strain time-domain eigenvector and the vibration frequency-domain eigenvector are combined to obtain the characteristic tensor, which is then normalized using a sliding window method. The specific sliding normalization formula is x. norm =(x-μ窗口 ) / (σ 窗口 +q), where q is a constant to prevent division by zero. Then, the feature tensor is input into the lightweight neural network model, which outputs the stress anomaly probability and the spectrum deviation probability. The lightweight neural network model adopts an architecture that combines a temporal convolutional network (TCN) with a multi-branch output head. The input layer is a normalized feature tensor. The feature extraction layer contains three depthwise separable convolutional blocks with kernel sizes of [8, 4, 2] and a stride of 2. The attention fusion layer introduces a compression-excitation module to automatically assign weights to the six sensing channels. The output layer uses global average pooling and includes stress anomaly branch, spectrum deviation branch and health index branch. When the probability of stress anomaly is ≥0.7 and the peak strain is simultaneously satisfied ≥55% of the material's yield strain or the average strain value over three consecutive time windows When the stress shows a monotonically increasing trend and the cumulative increase is ≥15%, it is determined to be an abnormal stress; or when the probability of an abnormal stress is ≥0.85, it is determined to be an abnormal stress. When the probability of spectrum deviation is ≥0.75, the vibration spectrum is determined to be outside the normal range.

[0021] Unlike the first embodiment, in this embodiment, the specific processing steps of step two include: The real-time displacement of the conductor's centroid in the geographic coordinate system is obtained based on the BeiDou high-precision positioning unit, including horizontal and lateral displacements. E (t), horizontal and vertical S N (t) and vertical direction S U (t), and according to the formula The total amplitude A of the conductor's galloping was calculated. total ; According to the formula The horizontal composite amplitude A was calculated. 水平 (t), vertical amplitude ; For the S in the current preset window U (t) The sequence is decentered, and the preset window length is 60 seconds. A fast Fourier transform is performed to calculate the power spectral density. Within the physically constrained frequency band, specifically in the 0.1 to 3 Hz band, the frequency point with the largest power spectral peak is found and denoted as the dancing main frequency f. 主 ; Take the set P of trajectory points on the horizontal plane within the current time window. i =[S E (i), S N (i)], and calculate the covariance matrix C of the set of trajectory points, specifically the covariance matrix ,in Let T denote the mean point of the horizontal displacement, and let T denote the transpose. Then, perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvector v corresponding to the largest eigenvalue. max According to the formula The ellipse tilt angle was calculated. v max-E and v max-N These represent the horizontal and vertical feature vectors, respectively, and here they are the angles mapped to the interval [0°, 90°].

[0022] Third Embodiment As a third embodiment of the present invention, this embodiment is implemented based on the first embodiment. The difference from the first embodiment is that this embodiment provides a detailed description of the specific implementation process of step three: Obtain the initial elevation H0 of the traverse, and then obtain the actual elevation H based on the BeiDou high-precision positioning module. z And according to the formula The change in traverse elevation was calculated. The current temperature Ts of the conductor is obtained through an installed temperature sensor. Taking any point on the conductor as the measurement basis, the initial sag S0 of this measurement is obtained, and the initial sag is expressed as the measurement obtained during installation, according to the formula... Calculate the current sag Simultaneous calculation of sag rate ,in A window representing changes over time; Meteorological data is acquired through micro-meteorological sensors, including wind speed, wind direction, ambient temperature, humidity, and air pressure.

[0023] The obtained spacer health status H, conductor galloping parameters G, sag change S, and environmental meteorological data M are time-aligned, and the four types of data after alignment are concatenated into a joint feature vector V(t) = [H(t), G(t), S(t), M(t)] in the same time step. For the joint feature vector sequence of several past time steps, where the time step value is 600 time steps, a Transformer-based encoder is used to extract temporal features and output a fixed-length context embedding vector E. Meanwhile, after confirming that there are no abnormal operating conditions of the line, the context embedding vector under normal operating conditions is collected, and several health pattern prototypes are obtained by clustering using the DBSCAN algorithm. Then, the cosine similarity between the current context embedding vector E and the context embedding vector of each health pattern prototype is calculated, and the maximum value is taken as the current health similarity value. If the health similarity value is <0.75 and continues for more than 3 analysis cycles, it is judged as a system-level anomaly. For the identified system-level anomalies, the data values ​​of each type in the health pattern prototype corresponding to the highest health similarity are used as baseline reference values. Starting from the baseline reference values, each group of data is replaced from the baseline value to the current value one by one, while the remaining data remain unchanged. The difference between the replaced health similarity and the original health similarity is calculated. The above operation is performed on the four types of data in sequence to obtain the marginal contribution value C of the four types of data. A Where A represents the data type, and the marginal contribution value C is also included. A Normalization yields the contribution percentage. And sort them from largest to smallest according to their percentage of contribution; Based on the sorted analysis results, the two groups of data with the highest contribution percentage are selected and recorded as pre-selected outliers. At the same time, Shapley values ​​are calculated for each feature within the pre-selected outliers to obtain the contribution of the features within the groups of the pre-selected outliers, and then sorted in descending order of the contribution of the features within the groups.

[0024] Taking meteorological data as an example, it includes five features: wind speed, wind direction, temperature, humidity and air pressure. After in-group Shapley decomposition, the contribution ratio within the group is output as shown in Table 1 below. Table 1. Group Contribution Percentage

[0025] Fourth embodiment As a fourth embodiment of the present invention, this embodiment is implemented based on the third embodiment, specifically targeting the intra-group feature contribution ranking obtained in the third embodiment, according to the formula... The intragroup concentration index was calculated. Where 'a' represents the group feature type in data type A, and 'b' represents the total number of group features. Taking the meteorological data above as an example, 'a' here includes wind speed, wind direction, temperature, humidity, and air pressure, and 'b' = 5, meaning there are a total of five feature types. Simultaneously, based on the obtained group concentration index... Identify the distribution of contributions within the group, if If the value is greater than 0.6, it indicates that the group's characteristics are highly concentrated; if 0.4 ≤ If ≤0.6, it indicates moderate concentration. If the value is less than 0.4, it indicates a highly dispersed type, and differentiated processing will be performed based on different distribution conditions; For highly concentrated distribution cases, the feature with the largest contribution within the group is recorded as the main control feature. At the same time, the time of its anomaly is determined and compared with the external time axis to determine the root cause of the anomaly. The sources of the external time axis include the timestamp of the warning information issued by the meteorological department, the line load / current change records of the dispatching system, the lightning strike time records, and the timestamp of the anomaly of the adjacent spacer. For example, the main control feature obtained after screening is wind speed. Detecting the start time of the anomaly, it was found that the wind speed suddenly increased from 9.2 m / s to 18.5 m / s at 14:23. At the same time, after comparing it with the timestamp of the warning information issued by the meteorological department, it was found that the meteorological station issued a yellow warning for strong winds at 14:20. Therefore, the two match, and the corresponding root cause of the anomaly is a meteorological event.

[0026] For moderately concentrated distribution scenarios, the top three features in terms of contribution are selected as key anomaly features. The matching degree of each feature with the anomaly records of the external timeline is checked one by one. If at least one key feature has a matching degree with the timestamp of the external timeline that exceeds the preset matching degree threshold, and the preset matching degree threshold is set by the operator, then the matching feature is determined to be the root cause of the anomaly. If there is no key feature that meets the matching requirements, then all key features are marked as items to be investigated and pushed to the operation and maintenance personnel to assist in on-site verification. When multiple key features match different external timelines, the feature with the highest matching degree is selected as the root cause. For highly dispersed cases, all within-group features are obtained and a full-dimensional feature vector V is constructed. 全 For example, for meteorological data, the full-dimensional feature vector V 全 =[wind speed, wind direction, temperature, humidity, air pressure], and calculate its migration direction angle relative to the historical baseline. ,in This represents the historical baseline mean vector, specifically obtained by acquiring the mean values ​​of various features under normal operating conditions during the 72 hours prior to device commissioning. This represents the main direction vector of seasonal migration. If the angle is less than 30°, it is considered normal seasonal migration; if the angle is less than or equal to 30°, it is considered normal seasonal migration. If the temperature is ≤60°, it is considered an abnormal climate drift. If the angle is greater than 60°, it is determined to be a systematic drift of the sensor group.

[0027] Fifth embodiment As a fifth embodiment of the present invention, the focus is on combining the implementation processes of all the above embodiments.

[0028] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0029] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for online monitoring of spacers in transmission lines, characterized in that, Includes the following steps: Step 1: Obtain strain signal and vibration acceleration signal, extract strain time domain features and vibration frequency domain features respectively, combine the two into feature tensor and normalize them, then input them into lightweight neural network model, output stress anomaly probability, spectrum deviation probability and health index, and identify stress anomaly or vibration anomaly based on the output results. Step 2: Obtain the real-time displacement of the traverse in the geographic coordinate system, calculate the total amplitude of galloping, the horizontal composite amplitude, the vertical amplitude, and the dominant frequency of galloping based on the displacement time history, perform principal component analysis on the trajectory points in the horizontal plane, and calculate the ellipse tilt angle. Step 3: Obtain the real-time elevation of the conductor and calculate the elevation change in combination with the initial elevation. Obtain the conductor temperature. Calculate the current sag and sag rate based on the initial sag and elevation change. Obtain meteorological data through a micro-meteorological sensor. Step 4: Concatenate the spacer health status, conductor galloping parameters, sag changes and meteorological data into a joint feature vector, extract the context embedding vector, cluster the health pattern prototypes under normal line conditions, calculate the similarity between the current embedding vector and each health pattern prototype, and determine the system-level anomaly when the similarity is below the threshold and continues to exceed the set period. Step 5: For system-level anomalies, using the health model prototype as a baseline reference, calculate the marginal contribution value and contribution percentage, select the two data types with the highest contribution, decompose the contribution of each feature within each data type, and obtain the feature contribution ranking within the group. Step 6: Calculate the ratio of the maximum feature contribution within a group to the total contribution of the group to obtain the concentration index within the group, and identify highly concentrated, moderately concentrated and highly dispersed types based on this index. Perform differentiated root cause localization processing for different distribution types.

2. The online monitoring method for transmission line spacer bars according to claim 1, characterized in that, The strain time-domain characteristics include the average strain value. Peak strain and standard deviation The vibration frequency domain characteristics include the dominant frequency f0, the dominant frequency amplitude A0, and the center frequency offset Δf; If the probability of stress anomaly is ≥0.7, the peak strain ≥55% of the material's yield strain, or the average strain over three consecutive windows If the stress shows a monotonically increasing trend with a cumulative increase of ≥15%, or if the probability of stress anomaly is ≥0.85, then stress anomaly is determined. If the probability of spectrum deviation is ≥0.75, then the vibration is determined to be abnormal.

3. The online monitoring method for transmission line spacer bars according to claim 1, characterized in that, The total amplitude of the dance S E (t), S N (t) and S U (t) represents the horizontal, vertical, and lateral displacements of the traverse in the geographic coordinate system, respectively; Horizontal composite amplitude ; Vertical amplitude The main frequency of the dance is controlled by the S frequency within a 60-second window. U (t) After the sequence is decentered, a fast Fourier transform is performed to find the frequency point with the largest power spectrum peak in the 0.1 to 3 Hz frequency band; The method for calculating the tilt angle of an ellipse is as follows: Take the set of points P on the trajectory of the horizontal plane. i =[S E (i), S N (i)], calculate the covariance matrix , Let v represent the mean point of the horizontal displacement, and perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvector v corresponding to the largest eigenvalue. max Ellipse tilt angle v max-E and v max-N These represent the feature vectors in the horizontal direction and the vertical direction, respectively.

4. The method for online monitoring of transmission line spacers according to claim 1, characterized in that, The elevation change H0 is the initial elevation, H z For actual elevation, sag rate ,in Indicates the time-varying window, current sag ; Meteorological data includes wind speed, wind direction, ambient temperature, humidity, and air pressure.

5. The online monitoring method for transmission line spacer bars according to claim 1, characterized in that, The context embedding vector extraction includes: aligning the obtained spacer health status H, conductor galloping parameters G, sag change S, and environmental meteorological data M in time, and concatenating them into a joint feature vector V(t) = [H(t), G(t), S(t), M(t)]. For the joint feature vector sequence of several past time steps, a Transformer-based encoder is used to extract temporal features. The method for determining system-level anomalies is as follows: collect the context embedding vectors under normal line operation, and cluster them to obtain several health mode prototypes. Calculate the cosine similarity between the current context embedding vector E and the context embedding vectors of each health mode prototype, and take the maximum value as the current health similarity value. If the health similarity value is <0.75 and continues for more than 3 analysis cycles, it is determined to be a system-level anomaly.

6. The online monitoring method for transmission line spacer bars according to claim 1, characterized in that, The marginal contribution value and contribution percentage are calculated as follows: Marginal contribution value: Using the data values ​​of various types in the health pattern prototype corresponding to the highest health similarity as the baseline reference value, and starting from the baseline reference value, each group of data is replaced from the baseline value to the current value one by one. The difference between the replaced health similarity and the original health similarity is calculated to obtain the marginal contribution value C. A ; Contribution percentage: Marginal contribution value C A The contribution percentage is obtained by normalization. .

7. The online monitoring method for transmission line spacer bars according to claim 1, characterized in that, The group concentration index , where a represents the group feature type in data type A, and b represents the total number of group features; like If the value is greater than 0.6, it indicates that the group's characteristics are highly concentrated; if 0.4 ≤ If ≤0.6, it indicates moderate concentration. If the value is less than 0.4, it indicates a highly dispersed type.

8. The online monitoring method for transmission line spacer bars according to claim 7, characterized in that, For highly concentrated distributions: take the feature with the largest contribution within the group as the main control feature, determine the time of onset of the anomaly, compare the time of onset of the anomaly with the external time axis, and determine the root cause of the anomaly; For moderately concentrated distributions: the top three features in terms of contribution are selected as key anomaly features. The matching degree of each key feature with the anomaly records of the external timeline is checked one by one. When the matching degree of at least one key feature exceeds the preset threshold, the matching feature is determined as the root cause of the anomaly. When multiple key features match different external timelines, the feature with the highest matching degree is selected first. When there are no key features that meet the matching threshold, all key features are marked as items to be investigated and pushed to the operation and maintenance personnel.

9. The online monitoring method for transmission line spacer bars according to claim 7, characterized in that, For highly dispersed distributions: obtain all within-group features to construct a full-dimensional feature vector, and calculate the migration direction angle relative to the historical baseline. ,in Represents the historical baseline mean vector. This represents the main direction vector of seasonal migration. If the angle is less than 30°, it is considered normal seasonal migration; if the angle is less than or equal to 30°, it is considered normal seasonal migration. If the temperature is ≤60°, it is considered an abnormal climate drift. If the angle is greater than 60°, it is determined to be a systematic drift of the sensor group.

10. An online monitoring system for transmission line spacers, used to execute the online monitoring method for transmission line spacers according to any one of claims 1-9, characterized in that, include: The data sensing and acquisition module includes strain sensors and vibration sensors for acquiring strain signals and vibration acceleration signals. It also includes temperature sensors, current sensors, and acceleration and attitude composite sensors, as well as micro-meteorological sensors and Beidou high-precision positioning units. All sensors acquire data synchronously with a unified clock source. The acquired data is preprocessed and stored in a local cache. The data intelligent processing module is used to extract the strain time-domain features and vibration frequency-domain features corresponding to the strain signal and vibration acceleration signal, combine the two into a feature tensor and normalize it before inputting it into a lightweight neural network model, and output the stress anomaly probability, spectrum deviation probability and health index. When the stress anomaly probability meets the first judgment condition, it is judged as stress anomaly, and when the spectrum deviation probability meets the second judgment condition, it is judged as vibration spectrum deviating from the normal range. It is also used to obtain the real-time displacement of the conductor in the geographic coordinate system and the real-time elevation of the conductor based on the Beidou high-precision positioning unit, calculate the total amplitude of the galloping, the horizontal composite amplitude, the vertical amplitude and the dominant frequency of the galloping based on the displacement time history, and perform principal component analysis on the trajectory points in the horizontal plane to calculate the ellipse tilt angle. The elevation change is calculated based on the real-time elevation and the initial elevation. The current sag and sag rate are calculated based on the initial sag and elevation change measured during installation. The multi-source data fusion and analysis module is used to align the health status of spacers, conductor galloping parameters, sag changes, and meteorological data by time and then stitch them together into a joint feature vector. The joint feature vector sequence of continuous time steps is processed by a Transformer encoder to extract context embedding vectors. The embedding vectors collected under normal line operating conditions are clustered to form health pattern prototypes. The similarity between the current embedding vector and each health pattern prototype is calculated. When the similarity is lower than the threshold and continues to exceed the set period, it is judged as a system-level anomaly. For system-level anomalies, using the health mode prototype as a baseline reference, each type of data is replaced with the current value, and the similarity difference before and after the replacement is calculated as the marginal contribution value of each type of data. After normalization, the contribution percentage of each data type is obtained and sorted. Select the two data categories with the highest contribution, decompose the contribution of each feature within each category, and obtain the ranking of feature contribution within the group. The intelligent monitoring and cause determination module is used to calculate the concentration index within a group based on the contribution of characteristics within the group. When the index is greater than the first threshold, it is determined to be highly concentrated; when the index is between the first and second thresholds, it is determined to be moderately concentrated; and when the index is less than the second threshold, it is determined to be highly dispersed. For highly concentrated distributions, the feature with the largest contribution within a group is used as the main control feature for single-cause tracing. For moderately concentrated distributions, multiple key features are matched and identified with an external time axis. For highly dispersed distributions, the migration direction angle of the full-dimensional feature vector relative to the historical baseline is calculated, and the anomaly type is determined based on the migration direction angle.

Images