A lightning protection grounding monitoring system and method

Abstract

This invention relates to the technical field of lightning protection grounding monitoring in power systems, and in particular to a lightning protection grounding monitoring system and method, comprising: a sensing layer, a site monitoring layer, a site aggregation layer, and a segment-level management layer; it proposes a four-level hierarchical architecture and a distributed deployment mode of segment-level training and site inference, which is accurately adapted to the multi-level operation and maintenance system of power systems; even when the network is down, the site can still independently complete intelligent analysis and alarms, significantly improving the reliability of system operation; it adopts a multi-factor deep correction model with embedded seasonal feature codes to fit the nonlinear effects of soil temperature and humidity, air temperature, and season, significantly improving the correction accuracy of grounding resistance under complex meteorological scenarios; based on an unsupervised autoencoder, it realizes early hidden fault identification of the grounding system, without the need for a large number of fault labeling samples, and can detect weak hidden dangers such as local corrosion and poor contact of the grounding grid in advance, supporting preventive operation and maintenance.

Description

Technical Field

[0001] This invention relates to the technical field of lightning protection grounding monitoring in power systems, and in particular to a lightning protection grounding monitoring system and method. Background Technology

[0002] Lightning protection grounding systems (such as the lightning protection grounding system for a distribution substation disclosed in patent CN105071061B and the lightning protection grounding system disclosed in patent CN111585203B) are core infrastructures for ensuring the insulation safety of equipment and personnel safety in power facilities such as substations and power transmission and distribution networks. Their grounding performance directly determines the discharge efficiency of lightning current and the overvoltage suppression effect.

[0003] With the development of intelligent power systems, online monitoring technology has been gradually applied to the field of lightning protection and grounding. However, during its use, the following systemic defects have been found in the existing technology: 1. Poor architecture adaptability and unreasonable computing power deployment: Existing intelligent lightning protection monitoring systems mostly adopt the general IoT three-layer architecture of "sensing layer - transmission layer - application layer", which is not matched with the power industry's two-level operation and maintenance system of "site-based operation and maintenance - section-level centralized management"; and the intelligent analysis computing power is all concentrated on the upper-layer platform, while the site side only undertakes the data transmission function. Once the communication link is interrupted, the site side immediately loses its intelligent analysis capability, which cannot meet the high reliability operation requirements of industrial sites. 2. Insufficient accuracy of grounding resistance environmental correction: Existing online grounding resistance correction mostly uses single-factor linear empirical formulas, which only simply correlate soil temperature and humidity, without taking into account long-term coupling factors such as ambient temperature and seasonal periodicity. The correction error is significant in scenarios such as seasonal changes and extreme temperatures, which directly leads to the distortion of grounding status assessment and easily causes false alarms or missed alarms. 3. The subsystems are isolated and lack an intelligent linkage mechanism: The analysis logic of the three systems, namely grounding resistance monitoring, surge protection device (SPD) status monitoring and lightning warning, is independent. The lightning warning can only trigger basic audible and visual alarms or hard contact actions. There is no coordinated linkage mechanism of "lightning risk level - sensing sampling frequency - model inference density". Under normal circumstances, high-frequency sampling causes a waste of computing power and energy consumption, while during periods of high lightning risk, there are problems of monitoring response lag and insufficient accuracy. 4. Weak ability to identify early hidden faults: Existing grounding fault diagnosis relies mainly on fixed threshold over-limit alarms, which can only identify serious failure faults; for weak hidden faults such as local corrosion of the grounding grid, poor conductor contact, and early deterioration of SPD, there is a lack of intelligent identification methods based on multi-parameter features, the fault detection window is lagging and cannot support preventive operation and maintenance.

[0004] Therefore, there is an urgent need for a lightning protection grounding monitoring system and method to improve the above-mentioned problems. Summary of the Invention

[0005] To address the aforementioned technical challenges, this invention provides a distributed deployment model with a four-level hierarchical architecture and segment-level training and site inference, accurately adapting to multi-level power operation and maintenance systems. Even during network outages, sites can independently perform intelligent analysis and alarms, significantly improving system reliability. A multi-factor deep correction model embedding seasonal feature codes is employed to fit the nonlinear effects of soil temperature and humidity, air temperature, and season, significantly improving the correction accuracy of grounding resistance under complex meteorological scenarios. An unsupervised autoencoder enables early latent fault identification in grounding systems, eliminating the need for numerous fault-labeled samples and enabling early detection of minor issues such as localized corrosion and poor contact in the grounding network, supporting preventative maintenance. A dual dynamic linkage mechanism of sampling and inference is constructed, adaptively adjusting the sampling frequency and inference density according to the lightning risk level, balancing low power consumption and economy during normal operation with rapid monitoring response during high-risk periods. An incremental training and incremental update model iteration is established, continuously optimizing model accuracy as operational data accumulates, extending the system's effective lifespan without requiring on-site hardware upgrades. This is a lightning protection grounding monitoring system and method.

[0006] The present invention provides a lightning protection grounding monitoring system and method, comprising: Sensing layer: includes grounding parameter monitoring unit, return current status monitoring unit, surge protection monitoring unit, lightning early warning sensing unit, and soil temperature and humidity monitoring unit; It is used to collect grounding electrical parameters, return circuit status parameters, surge protection device operating parameters, lightning sensing parameters, and soil temperature and humidity environmental parameters. Site monitoring layer: includes grounding and return current status monitoring screen, integrated lightning protection screen and lightning early warning monitoring host; The grounding and return current status monitoring panel is connected to the grounding parameter monitoring unit and the return current status monitoring unit via a MODBUS bus. The integrated lightning protection panel is connected to the surge protection monitoring unit via a MODBUS bus. The lightning warning monitoring host is connected to the lightning warning sensing unit via a MODBUS bus; Site aggregation layer: includes site servers and communication management units; The site server has a built-in deep learning inference engine, and the site server and communication management unit are connected to the grounding and return status monitoring screen, the integrated lightning protection screen and the lightning warning monitoring host respectively through the MODBUS / TCPIP protocol to complete the aggregation of multi-source monitoring data and local intelligent inference within a single site; Segment-level management layer: includes switches, segment-level servers, segment-level workstations, and large-screen display terminals; The switch connects to the site server and the communication management unit via fiber optic links. The segment-level server has a built-in deep learning training engine. The segment-level server, segment-level workstations and large-screen display terminals are all connected to the switch. The deep learning training engine has built-in a multi-factor correction model for grounding resistance, a hidden grounding fault identification model, a SPD lifetime time series prediction model, and a lightning disaster early warning model. The above models are trained and iterated based on the full historical dataset, and the trained model weights are distributed to the servers of each site. The deep learning inference engine is used to load the corresponding model weights and perform inference calculations on real-time monitoring data; The system has a dual dynamic linkage mechanism of sampling and inference driven by lightning risk: the sampling frequency of the perception layer and the inference frequency of the site inference engine are adjusted synchronously according to the lightning warning level to achieve adaptive matching between risk level and monitoring density. A four-level hierarchical architecture and a distributed deployment mode of segment-level training and site inference are proposed to accurately adapt to the multi-level operation and maintenance system of the power industry. Even during network outages, sites can still independently complete intelligent analysis and alarms, significantly improving system reliability. A multi-factor deep correction model with embedded seasonal feature coding is adopted to fit the nonlinear effects of soil temperature and humidity, air temperature, and season, significantly improving the correction accuracy of grounding resistance under complex meteorological scenarios. An unsupervised autoencoder enables early latent fault identification in the grounding system, eliminating the need for a large number of fault-labeled samples and enabling early detection of subtle hidden dangers such as localized corrosion and poor contact in the grounding grid, supporting preventative operation and maintenance. A dual dynamic linkage mechanism of sampling and inference is constructed, adaptively adjusting the sampling frequency and inference density according to the lightning risk level, balancing low power consumption and economy during normal operation with monitoring response speed during high-risk periods. An incremental training and down-heating update model iteration is established, continuously optimizing model accuracy as operational data accumulates, extending the effective lifespan of the system without on-site hardware upgrades.

[0007] Preferably, the grounding parameter monitoring unit includes a grounding resistance monitoring unit, a step voltage monitoring unit, a contact voltage monitoring unit, and an electrical integrity monitoring unit; The grounding resistance monitoring unit is used to collect the grounding resistance value of the grounding grid in real time; The step voltage monitoring unit is used to collect step voltage parameters around the grounding grid; The contact voltage monitoring unit is used to collect contact voltage parameters around the grounding grid; The electrical integrity monitoring unit is used to detect the continuity of the grounding grid conductor; The return current status monitoring unit includes a transformer core grounding status monitoring unit and a return current cable status monitoring unit; The transformer core grounding status monitoring unit is used to collect the leakage current and continuity status of the transformer core grounding circuit; The return cable status monitoring unit is used to collect the current parameters and on / off status of the lightning protection return cable; The surge protection monitoring unit includes multi-level SPD devices, SCB backup protection devices, and an SPD intelligent management unit; Multi-level SPD devices and SCB backup protection devices are used to form a graded surge protection path; The SPD intelligent management unit is used to collect leakage current, number of operations, device temperature and SCB on / off status of each level of SPD; The lightning warning sensing unit includes an atmospheric electric field intensity monitoring device and a lightning rod protection device with an integrated lightning current acquisition module. The atmospheric electric field intensity monitoring device is used to collect the atmospheric electric field intensity and rate of change in the monitoring area; The lightning rod protection device is used to perform lightning current discharge and collect parameters such as peak value, polarity, and number of lightning strikes.

[0008] Preferably, the multi-factor correction model for grounding resistance is a fully connected deep neural network with embedded seasonal feature encoding; The input dimensions of the multi-factor correction model for grounding resistance include the measured grounding resistance output by the grounding parameter monitoring unit, the soil temperature and soil humidity output by the soil temperature and humidity monitoring unit, as well as the ambient temperature and month coding vector. The output is the grounding resistance correction value under standard operating conditions. The multi-factor correction model for grounding resistance eliminates the interference of soil environment, meteorological conditions, and seasonal periodicity on the measurement results of grounding resistance through multi-layer nonlinear fitting.

[0009] Preferably, the grounding latent fault identification model is an unsupervised anomaly detection model that combines stacked sparse autoencoders and deep belief networks; The input dimensions of the grounding hidden fault identification model include the grounding resistance correction value output by the grounding resistance multi-factor correction model, the step voltage, contact voltage and electrical integrity parameters output by the grounding parameter monitoring unit, and the core grounding leakage current and return cable current output by the return current status monitoring unit. The grounding latent fault identification model fits the feature distribution of the normal operating state through unsupervised learning, and outputs the reconstruction error and latent fault probability of the current state. It is used to identify early weak faults such as local corrosion of the grounding grid and poor conductor contact.

[0010] Preferably, the SPD lifetime time series prediction model is a long short-term memory time series neural network with multiple feature inputs; The input to the SPD lifetime timing prediction model includes the leakage current timing sequence of each level of SPD, the cumulative number of actions, device temperature data, ambient temperature sequence and lightning strike frequency data output by the surge protection monitoring unit, and the output is the remaining lifetime of the SPD and the degradation trend curve. When the lightning warning level is raised, the model automatically encrypts the inference frequency to achieve real-time tracking of the SPD status during the risk period.

[0011] Preferably, the lightning disaster early warning model is a Bi-long short-term memory bidirectional time-series prediction model that incorporates a channel attention mechanism; The input to the lightning disaster early warning model includes the atmospheric electric field intensity time series data, electric field change rate sequence and historical lightning strike labeled samples output by the lightning early warning sensing unit. The output is the probability of lightning occurrence within a set time period in the future and the three-level warning level. The lightning disaster early warning model uses an attention mechanism to focus on the characteristic segments of electric field abrupt changes, thereby improving the lead time and accuracy of disaster early warning.

[0012] Preferably, the specific rules of the sampling and inference dual dynamic linkage mechanism are as follows: When the warning level is the attention level, the basic sampling frequency of the perception layer is increased to twice the original frequency, and the inference frequency of the grounding hidden fault identification model is increased to twice the original frequency. When the warning level is warning level, the basic sampling frequency of the perception layer is increased to 5 times, and the SPD lifetime time series prediction model starts continuous real-time inference; When the warning level is dangerous, the basic sampling frequency of the perception layer is increased to 10 times, all deep learning models are switched to high-frequency continuous inference mode, and on-site alarm linkage is triggered simultaneously.

[0013] Preferably, the deep learning training engine of the segment-level server is configured to aggregate newly added monitoring data and fault labeling samples from each site at fixed intervals, and to perform incremental fine-tuning training on the basis of the original model weights using transfer learning. After training is completed, a new version of the model weight file is generated and distributed to each site server via fiber optic network. The site server performs a hot update to load the new model without interrupting on-site monitoring services.

[0014] Preferably, the deep learning inference engine of the site server supports offline independent operation: when the fiber optic link between the site and the segment is interrupted, the site server can still complete all intelligent inference calculations and alarm outputs based on the locally loaded model weights, ensuring the site-level intelligent monitoring capability under network outage conditions;

[0015] Preferably, the lightning protection grounding monitoring system based on any one of claims 1-9 is implemented by including the following steps: S1. Multi-source sensing data acquisition: Each monitoring unit in the sensing layer collects grounding parameters, return current status, surge protection status, lightning electric field and soil temperature and humidity parameters in real time according to the basic sampling frequency, and converts them into standard MODBUS data frames and uploads them to the corresponding monitoring terminal. S2, Site-level partition preprocessing: Grounding and return current status monitoring screen, integrated lightning protection screen, and lightning early warning monitoring host respectively receive the corresponding monitoring data, perform data denoising, outlier removal and hard threshold initial judgment, and trigger local audible and visual alarms and store event data when the limit is exceeded. S3, Site Fusion and Local Intelligent Inference: The site server timestamps and fuses the data from the three types of monitoring terminals, calls the local deep learning inference engine, and sequentially performs multi-factor correction of grounding resistance, identification of grounding latent faults, and SPD lifetime prediction inference, and outputs quantitative evaluation results; after the communication management unit completes the protocol encapsulation, it uploads the original data and inference results to the segment-level switch through the fiber optic link. S4. Segment-level centralized control and incremental model iteration: The segment-level server aggregates the monitoring data and inference results of all sites under its jurisdiction, and generates the results of global grounding health distribution, surge device life status, and lightning risk classification. At the same time, it aggregates all historical data and fault-labeled samples at fixed intervals, performs incremental fine-tuning training on each deep learning model, and distributes the updated model weights to each site server to complete hot updates. S5. Sampling and Inference Dual Dynamic Linkage and Closed-Loop Operation and Maintenance: Based on the warning level output by the lightning disaster early warning model, the sampling frequency of the perception layer and the inference frequency of the site inference engine are adjusted synchronously. When monitoring parameters exceed limits or the probability of hidden faults exceeds the set threshold, alarm information and maintenance suggestions are pushed according to the fault level, the entire handling process is recorded, and the verification samples are sent back to the segment-level training set to form a closed loop of maintenance.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A four-level hierarchical architecture and a distributed deployment mode of segment-level training and site inference are proposed to accurately adapt to the multi-level operation and maintenance system of the power industry; the site can still independently complete intelligent analysis and alarm when the network is down, which greatly improves the reliability of system operation. 2. A multi-factor deep correction model with embedded seasonal feature coding is adopted to fit the nonlinear effects of soil temperature and humidity, air temperature and season, which significantly improves the correction accuracy of grounding resistance under complex meteorological scenarios. 3. Based on unsupervised self-encoders, early hidden faults in grounding systems can be identified without the need for a large number of fault labeling samples. This allows for the early detection of subtle hidden dangers such as local corrosion and poor contact in the grounding grid, supporting preventive operation and maintenance. 4. Construct a dual dynamic linkage mechanism for sampling and inference, which adaptively adjusts the sampling frequency and inference density according to the lightning risk level, taking into account both the low power consumption and economy of normal operation and the monitoring response speed during high-risk periods. 5. Establish incremental training and heat-up update model iteration, continuously optimize model accuracy as running data accumulates, and extend the effective life cycle of the system without on-site hardware upgrades. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0018] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. The present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0019] Example 1: As Figure 1 As shown, a lightning protection grounding monitoring system and method includes: Sensing layer: includes grounding parameter monitoring unit, return current status monitoring unit, surge protection monitoring unit, lightning early warning sensing unit, and soil temperature and humidity monitoring unit; It is used to collect grounding electrical parameters, return circuit status parameters, surge protection device operating parameters, lightning sensing parameters, and soil temperature and humidity environmental parameters. Site monitoring layer: includes grounding and return current status monitoring screen, integrated lightning protection screen and lightning early warning monitoring host; The grounding and return current status monitoring panel is connected to the grounding parameter monitoring unit and the return current status monitoring unit via a MODBUS bus. The integrated lightning protection panel is connected to the surge protection monitoring unit via a MODBUS bus. The lightning warning monitoring host is connected to the lightning warning sensing unit via a MODBUS bus; Site aggregation layer: includes site servers and communication management units; The site server has a built-in deep learning inference engine, and the site server and communication management unit are connected to the grounding and return status monitoring screen, the integrated lightning protection screen and the lightning warning monitoring host respectively through the MODBUS / TCPIP protocol to complete the aggregation of multi-source monitoring data and local intelligent inference within a single site; Segment-level management layer: includes switches, segment-level servers, segment-level workstations, and large-screen display terminals; The switch connects to the site server and the communication management unit via fiber optic links. The segment-level server has a built-in deep learning training engine. The segment-level server, segment-level workstations and large-screen display terminals are all connected to the switch. The deep learning training engine has built-in a multi-factor correction model for grounding resistance, a hidden grounding fault identification model, a SPD lifetime time series prediction model, and a lightning disaster early warning model. The above models are trained and iterated based on the full historical dataset, and the trained model weights are distributed to the servers of each site. The deep learning inference engine is used to load the corresponding model weights and perform inference calculations on real-time monitoring data; The system has a dual dynamic linkage mechanism of sampling and inference driven by lightning risk: the sampling frequency of the perception layer and the inference frequency of the site inference engine are adjusted synchronously according to the lightning warning level to achieve adaptive matching between risk level and monitoring density. The grounding parameter monitoring unit includes a grounding resistance monitoring unit, a step voltage monitoring unit, a contact voltage monitoring unit, and an electrical integrity monitoring unit; The grounding resistance monitoring unit is used to collect the grounding resistance value of the grounding grid in real time; The step voltage monitoring unit is used to collect step voltage parameters around the grounding grid; The contact voltage monitoring unit is used to collect contact voltage parameters around the grounding grid; The electrical integrity monitoring unit is used to detect the continuity of the grounding grid conductor; The return current status monitoring unit includes a transformer core grounding status monitoring unit and a return current cable status monitoring unit; The transformer core grounding status monitoring unit is used to collect the leakage current and continuity status of the transformer core grounding circuit; The return cable status monitoring unit is used to collect the current parameters and on / off status of the lightning protection return cable; The surge protection monitoring unit includes multi-level SPD devices, SCB backup protection devices, and an SPD intelligent management unit; Multi-level SPD devices and SCB backup protection devices are used to form a graded surge protection path; The SPD intelligent management unit is used to collect leakage current, number of operations, device temperature and SCB on / off status of each level of SPD; The lightning warning sensing unit includes an atmospheric electric field intensity monitoring device and a lightning rod protection device with an integrated lightning current acquisition module. The atmospheric electric field intensity monitoring device is used to collect the atmospheric electric field intensity and rate of change in the monitoring area; The lightning rod protection device is used to perform lightning current discharge and collect parameters such as peak value, polarity and number of lightning strikes. The multi-factor correction model for grounding resistance is a fully connected deep neural network with embedded seasonal feature encoding; The input dimensions of the multi-factor correction model for grounding resistance include the measured grounding resistance output by the grounding parameter monitoring unit, the soil temperature and soil humidity output by the soil temperature and humidity monitoring unit, as well as the ambient temperature and month coding vector. The output is the grounding resistance correction value under standard operating conditions. The multi-factor correction model for grounding resistance eliminates the interference of soil environment, meteorological conditions and seasonal periodicity on the grounding resistance measurement results through multi-layer nonlinear fitting; The grounding latent fault identification model is an unsupervised anomaly detection model that combines stacked sparse autoencoders and deep belief networks. The input dimensions of the grounding hidden fault identification model include the grounding resistance correction value output by the grounding resistance multi-factor correction model, the step voltage, contact voltage and electrical integrity parameters output by the grounding parameter monitoring unit, and the core grounding leakage current and return cable current output by the return current status monitoring unit. The grounding latent fault identification model fits the feature distribution of the normal operating state through unsupervised learning, and outputs the reconstruction error and latent fault probability of the current state, which is used to identify early weak faults such as local corrosion of the grounding grid and poor conductor contact. The SPD lifetime time series prediction model is a long short-term memory time series neural network with multiple feature inputs; The input to the SPD lifetime timing prediction model includes the leakage current timing sequence of each level of SPD, the cumulative number of actions, device temperature data, ambient temperature sequence and lightning strike frequency data output by the surge protection monitoring unit, and the output is the remaining lifetime of the SPD and the degradation trend curve. When the lightning warning level is raised, the model automatically encrypts the inference frequency to achieve real-time tracking of the SPD status during the risk period; The lightning disaster early warning model is a Bi-long short-term memory bidirectional time-series prediction model that incorporates a channel attention mechanism; The input to the lightning disaster early warning model includes the atmospheric electric field intensity time series data, electric field change rate sequence and historical lightning strike labeled samples output by the lightning early warning sensing unit. The output is the probability of lightning occurrence within a set time period in the future and the three-level warning level. The lightning disaster early warning model focuses on the characteristic segments of electric field abrupt changes through an attention mechanism, thereby improving the lead time and accuracy of disaster early warning. The specific rules of the sampling and inference dual dynamic linkage mechanism are as follows: When the warning level is the attention level, the basic sampling frequency of the perception layer is increased to twice the original frequency, and the inference frequency of the grounding hidden fault identification model is increased to twice the original frequency. When the warning level is warning level, the basic sampling frequency of the perception layer is increased to 5 times, and the SPD lifetime time series prediction model starts continuous real-time inference; When the warning level is dangerous, the basic sampling frequency of the perception layer is increased to 10 times, all deep learning models are switched to high-frequency continuous inference mode, and on-site alarm linkage is triggered at the same time. The deep learning training engine of the segment-level server is configured to aggregate newly added monitoring data and fault labeling samples from each site at fixed intervals, and to incrementally fine-tune the training based on the original model weights using transfer learning. After training is completed, a new version of the model weight file is generated and distributed to each site server via fiber optic network. The site server performs hot update to load the new model without interrupting the on-site monitoring business. The site server's deep learning inference engine supports offline independent operation: when the fiber optic link between the site and the segment is interrupted, the site server can still complete all intelligent inference calculations and alarm outputs based on the locally loaded model weights, ensuring site-level intelligent monitoring capabilities under network outage conditions.

[0020] The lightning protection grounding monitoring system based on any one of claims 1-9 includes the following steps: S1. Multi-source sensing data acquisition: Each monitoring unit in the sensing layer collects grounding parameters, return current status, surge protection status, lightning electric field and soil temperature and humidity parameters in real time according to the basic sampling frequency, and converts them into standard MODBUS data frames and uploads them to the corresponding monitoring terminal. S2, Site-level partition preprocessing: Grounding and return current status monitoring screen, integrated lightning protection screen, and lightning early warning monitoring host respectively receive the corresponding monitoring data, perform data denoising, outlier removal and hard threshold initial judgment, and trigger local audible and visual alarms and store event data when the limit is exceeded. S3, Site Fusion and Local Intelligent Inference: The site server timestamps and fuses the data from the three types of monitoring terminals, calls the local deep learning inference engine, and sequentially performs multi-factor correction of grounding resistance, identification of grounding latent faults, and SPD lifetime prediction inference, and outputs quantitative evaluation results; after the communication management unit completes the protocol encapsulation, it uploads the original data and inference results to the segment-level switch through the fiber optic link. S4. Segment-level centralized control and incremental model iteration: The segment-level server aggregates the monitoring data and inference results of all sites under its jurisdiction, and generates the results of global grounding health distribution, surge device life status, and lightning risk classification. At the same time, it aggregates all historical data and fault-labeled samples at fixed intervals, performs incremental fine-tuning training on each deep learning model, and distributes the updated model weights to each site server to complete hot updates. S5. Sampling and Inference Dual Dynamic Linkage and Closed-Loop Operation and Maintenance: Based on the warning level output by the lightning disaster early warning model, the sampling frequency of the perception layer and the inference frequency of the site inference engine are adjusted synchronously. When monitoring parameters exceed limits or the probability of hidden faults exceeds the set threshold, alarm information and maintenance suggestions are pushed according to the fault level, the entire handling process is recorded, and the verification samples are sent back to the segment-level training set to form a closed loop of maintenance.

[0021] The main functions achieved by this invention are: 1. A four-level hierarchical architecture and a distributed deployment mode of segment-level training and site inference are proposed to accurately adapt to the multi-level operation and maintenance system of the power industry; the site can still independently complete intelligent analysis and alarm when the network is down, which greatly improves the reliability of system operation. 2. A multi-factor deep correction model with embedded seasonal feature coding is adopted to fit the nonlinear effects of soil temperature and humidity, air temperature and season, which significantly improves the correction accuracy of grounding resistance under complex meteorological scenarios. 3. Based on unsupervised self-encoders, early hidden faults in grounding systems can be identified without the need for a large number of fault labeling samples. This allows for the early detection of subtle hidden dangers such as local corrosion and poor contact in the grounding grid, supporting preventive operation and maintenance. 4. Construct a dual dynamic linkage mechanism for sampling and inference, which adaptively adjusts the sampling frequency and inference density according to the lightning risk level, taking into account both the low power consumption and economy of normal operation and the monitoring response speed during high-risk periods. 5. Establish incremental training and heat-up update model iteration, continuously optimize model accuracy as running data accumulates, and extend the effective life cycle of the system without on-site hardware upgrades.

[0022] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A lightning protection grounding monitoring system, characterized in that, include: Sensing layer: includes grounding parameter monitoring unit, return current status monitoring unit, surge protection monitoring unit, lightning early warning sensing unit, and soil temperature and humidity monitoring unit; It is used to collect grounding electrical parameters, return circuit status parameters, surge protection device operating parameters, lightning sensing parameters, and soil temperature and humidity environmental parameters. Site monitoring layer: includes grounding and return current status monitoring screen, integrated lightning protection screen and lightning early warning monitoring host; The grounding and return current status monitoring panel is connected to the grounding parameter monitoring unit and the return current status monitoring unit via a MODBUS bus. The integrated lightning protection panel is connected to the surge protection monitoring unit via a MODBUS bus. The lightning warning monitoring host is connected to the lightning warning sensing unit via a MODBUS bus; Site aggregation layer: includes site servers and communication management units; The site server has a built-in deep learning inference engine, and the site server and communication management unit are connected to the grounding and return status monitoring screen, the integrated lightning protection screen and the lightning warning monitoring host respectively through the MODBUS / TCPIP protocol to complete the aggregation of multi-source monitoring data and local intelligent inference within a single site; Segment-level management layer: includes switches, segment-level servers, segment-level workstations, and large-screen display terminals; The switch connects to the site server and the communication management unit via fiber optic links. The segment-level server has a built-in deep learning training engine. The segment-level server, segment-level workstations and large-screen display terminals are all connected to the switch. The deep learning training engine has built-in a multi-factor correction model for grounding resistance, a hidden grounding fault identification model, a SPD lifetime time series prediction model, and a lightning disaster early warning model. The above models are trained and iterated based on the full historical dataset, and the trained model weights are distributed to the servers of each site. The deep learning inference engine is used to load the corresponding model weights and perform inference calculations on real-time monitoring data; The system has a dual dynamic linkage mechanism of sampling and inference driven by lightning risk: the sampling frequency of the perception layer and the inference frequency of the site inference engine are adjusted synchronously according to the lightning warning level to achieve adaptive matching between risk level and monitoring density.

2. The lightning protection grounding monitoring system as described in claim 1, characterized in that, The grounding parameter monitoring unit includes a grounding resistance monitoring unit, a step voltage monitoring unit, a contact voltage monitoring unit, and an electrical integrity monitoring unit; The grounding resistance monitoring unit is used to collect the grounding resistance value of the grounding grid in real time; The step voltage monitoring unit is used to collect step voltage parameters around the grounding grid; The contact voltage monitoring unit is used to collect contact voltage parameters around the grounding grid; The electrical integrity monitoring unit is used to detect the continuity of the grounding grid conductor; The return current status monitoring unit includes a transformer core grounding status monitoring unit and a return current cable status monitoring unit; The transformer core grounding status monitoring unit is used to collect the leakage current and continuity status of the transformer core grounding circuit; The return cable status monitoring unit is used to collect the current parameters and on / off status of the lightning protection return cable; The surge protection monitoring unit includes multi-level SPD devices, SCB backup protection devices, and an SPD intelligent management unit; Multi-level SPD devices and SCB backup protection devices are used to form a graded surge protection path; The SPD intelligent management unit is used to collect leakage current, number of operations, device temperature and SCB on / off status of each level of SPD; The lightning warning sensing unit includes an atmospheric electric field intensity monitoring device and a lightning rod protection device with an integrated lightning current acquisition module. The atmospheric electric field intensity monitoring device is used to collect the atmospheric electric field intensity and rate of change in the monitoring area; The lightning rod protection device is used to perform lightning current discharge and collect parameters such as peak value, polarity, and number of lightning strikes.

3. The lightning protection grounding monitoring system as described in claim 2, characterized in that, The multi-factor correction model for grounding resistance is a fully connected deep neural network with embedded seasonal feature encoding; The input dimensions of the multi-factor correction model for grounding resistance include the measured grounding resistance output by the grounding parameter monitoring unit, the soil temperature and soil humidity output by the soil temperature and humidity monitoring unit, as well as the ambient temperature and month coding vector. The output is the grounding resistance correction value under standard operating conditions. The multi-factor correction model for grounding resistance eliminates the interference of soil environment, meteorological conditions and seasonal periodicity on the grounding resistance measurement results through multi-layer nonlinear fitting.

4. The lightning protection grounding monitoring system as described in claim 2, characterized in that, The grounding latent fault identification model is an unsupervised anomaly detection model that combines stacked sparse autoencoders and deep belief networks. The input dimensions of the grounding hidden fault identification model include the grounding resistance correction value output by the grounding resistance multi-factor correction model, the step voltage, contact voltage and electrical integrity parameters output by the grounding parameter monitoring unit, and the core grounding leakage current and return cable current output by the return current status monitoring unit. The grounding latent fault identification model fits the feature distribution of the normal operating state through unsupervised learning, and outputs the reconstruction error and latent fault probability of the current state. It is used to identify early weak faults such as local corrosion of the grounding grid and poor conductor contact.

5. The lightning protection grounding monitoring system as described in claim 1, characterized in that, The SPD lifetime time series prediction model is a long short-term memory time series neural network with multiple feature inputs; The input to the SPD lifetime timing prediction model includes the leakage current timing sequence of each level of SPD, the cumulative number of actions, device temperature data, ambient temperature sequence and lightning strike frequency data output by the surge protection monitoring unit, and the output is the remaining lifetime of the SPD and the degradation trend curve. When the lightning warning level is raised, the model automatically encrypts the inference frequency to achieve real-time tracking of the SPD status during the risk period.

6. The lightning protection grounding monitoring system as described in claim 1, characterized in that, The lightning disaster early warning model is a Bi-long short-term memory bidirectional time-series prediction model that incorporates a channel attention mechanism; The input to the lightning disaster early warning model includes the atmospheric electric field intensity time series data, electric field change rate sequence and historical lightning strike labeled samples output by the lightning early warning sensing unit. The output is the probability of lightning occurrence within a set time period in the future and the three-level warning level. The lightning disaster early warning model uses an attention mechanism to focus on the characteristic segments of electric field abrupt changes, thereby improving the lead time and accuracy of disaster early warning.

7. The lightning protection grounding monitoring system as described in claim 1, characterized in that, The specific rules of the sampling and inference dual dynamic linkage mechanism are as follows: When the warning level is the attention level, the basic sampling frequency of the perception layer is increased to twice the original frequency, and the inference frequency of the grounding hidden fault identification model is increased to twice the original frequency. When the warning level is warning level, the basic sampling frequency of the perception layer is increased to 5 times, and the SPD lifetime time series prediction model starts continuous real-time inference; When the warning level is dangerous, the basic sampling frequency of the perception layer is increased to 10 times, all deep learning models are switched to high-frequency continuous inference mode, and on-site alarm linkage is triggered at the same time.

8. The lightning protection grounding monitoring system as described in claim 1, characterized in that, The deep learning training engine of the segment-level server is configured to aggregate newly added monitoring data and fault labeling samples from each site at fixed intervals, and to incrementally fine-tune the training based on the original model weights using transfer learning. After training is completed, a new version of the model weight file is generated and distributed to each site server via fiber optic network. The site server performs hot update to load the new model without interrupting the on-site monitoring business.

9. A lightning protection grounding monitoring system as described in claim 1, characterized in that, The site server's deep learning inference engine supports offline independent operation: when the fiber optic link between the site and the segment is interrupted, the site server can still complete all intelligent inference calculations and alarm outputs based on the locally loaded model weights, ensuring site-level intelligent monitoring capabilities under network outage conditions.

10. A method for monitoring lightning protection grounding, characterized in that, The lightning protection grounding monitoring system based on any one of claims 1-9 includes the following steps: S1. Multi-source sensing data acquisition: Each monitoring unit in the sensing layer collects grounding parameters, return current status, surge protection status, lightning electric field and soil temperature and humidity parameters in real time according to the basic sampling frequency, and converts them into standard MODBUS data frames and uploads them to the corresponding monitoring terminal. S2, Site-level partition preprocessing: Grounding and return current status monitoring screen, integrated lightning protection screen, and lightning early warning monitoring host respectively receive the corresponding monitoring data, perform data denoising, outlier removal and hard threshold initial judgment, and trigger local audible and visual alarms and store event data when the limit is exceeded. S3, Site Fusion and Local Intelligent Inference: The site server timestamps and fuses the data from the three types of monitoring terminals, calls the local deep learning inference engine, and sequentially performs multi-factor correction of grounding resistance, identification of grounding latent faults, and SPD lifetime prediction inference, and outputs quantitative evaluation results; after the communication management unit completes the protocol encapsulation, it uploads the original data and inference results to the segment-level switch through the fiber optic link. S4. Segment-level centralized control and incremental model iteration: The segment-level server aggregates the monitoring data and inference results of all sites under its jurisdiction, and generates the results of global grounding health distribution, surge device life status, and lightning risk classification. At the same time, it aggregates all historical data and fault-labeled samples at fixed intervals, performs incremental fine-tuning training on each deep learning model, and distributes the updated model weights to each site server to complete hot updates. S5. Sampling and Inference Dual Dynamic Linkage and Closed-Loop Operation and Maintenance: Based on the warning level output by the lightning disaster early warning model, the sampling frequency of the perception layer and the inference frequency of the site inference engine are adjusted synchronously. When monitoring parameters exceed limits or the probability of hidden faults exceeds the set threshold, alarm information and maintenance suggestions are pushed according to the fault level, the entire handling process is recorded, and the verification samples are sent back to the segment-level training set to form a closed loop of maintenance.

Images