A cloud platform-based power grid remote monitoring and intelligent early warning system

Abstract

The application discloses a power grid remote monitoring and intelligent early warning system based on a cloud platform, comprising: a common-mode current data acquisition module for acquiring common-mode current data and frequency converter operation state data; a common-mode current feature extraction module for constructing a common-mode current deep feature sequence; a common-mode current feature baseline analysis module for constructing a common-mode current feature baseline model corresponding to each load interval; a grounding abnormality identification module for determining a feature offset event set of a potential grounding state abnormality; a grounding deterioration state identification module for determining a grounding state corresponding to the feature offset event of the potential grounding state abnormality; a model evaluation and optimization module for evaluating grounding deterioration state identification accuracy and performing parameter optimization. The application reduces dependence on artificial inspection and fixed thresholds, improves real-time performance, accuracy and intelligent level of power grid grounding state monitoring, and helps to reduce operation and maintenance costs and improve safety and reliability of power grid operation.

Description

Technical Field

[0001] This invention relates to the field of next-generation information technology, and in particular to a cloud-based remote monitoring and intelligent early warning system for power grids. Background Technology

[0002] With the continuous advancement of new power system construction, the power grid structure is becoming increasingly complex, and the penetration rate of power electronic equipment in the power grid is continuously increasing. Frequency converters, rectifiers, and power electronic interface devices have been widely integrated into industrial power grids, distribution networks, and new energy grid-connected systems. During power grid operation, the grounding system, as a key infrastructure ensuring personal safety and reliable equipment operation, directly affects the safety and stability of the power grid. However, power grid grounding systems are typically widely distributed and structurally complex. Grounding conductors, grounding electrodes, and connection nodes are exposed to complex electromagnetic environments and variable climatic conditions for extended periods, making them prone to corrosion, loosening, and aging. The deterioration process of grounding conditions is often highly concealed and slow-developing, making it difficult to detect in its early stages using traditional methods. Once deterioration intensifies, it can easily trigger equipment abnormalities, protection malfunctions, and even power grid failures. Currently, power grid grounding condition monitoring mainly relies on manual inspections, periodic grounding resistance tests, or local online monitoring devices. These methods generally suffer from long monitoring cycles, limited coverage, and insufficient real-time performance, making it difficult to meet the power grid's needs for continuous monitoring and early warning. Meanwhile, traditional monitoring methods often rely on single parameters or fixed thresholds as criteria, failing to fully consider the impact of changes in grid operating conditions on monitoring results. This is particularly problematic under conditions of frequent load fluctuations and concentrated operation of power electronic equipment, easily leading to misjudgments or omissions. Furthermore, in actual grid operation, load changes, operating mode switching, and high-frequency switching actions of power electronic devices cause significant differences in common-mode current characteristics across different times and power supply circuits. Existing technologies often lack systematic analysis of the multidimensional characteristics of common-mode current and fail to effectively distinguish between normal fluctuations caused by load changes, electromagnetic interference, or cable reflections and abnormal changes caused by grounding system degradation, thus limiting the accuracy and reliability of grid grounding status monitoring. With the continuous expansion of the grid and rising operation and maintenance costs, traditional methods relying on manual experience and distributed monitoring equipment are no longer sufficient to meet the real-time, accuracy, and intelligent requirements for safe grid operation. Therefore, there is an urgent need for a remote grid monitoring and intelligent early warning technology that can centrally process multi-source data on a cloud platform, combining multidimensional feature analysis and adaptive modeling, to achieve continuous monitoring, accurate identification, and intelligent maintenance of grounding status, thereby improving the safety and reliability of grid operation. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a cloud platform-based remote monitoring and intelligent early warning system for power grids, mainly comprising: The common-mode current data acquisition module is used to acquire common-mode current data and inverter operating status data through the cloud platform, form a continuous time series according to the preset sampling period, and perform data preprocessing. The common-mode current feature extraction module is used to extract time-domain features, frequency-domain features, and phase-domain features from the time-series data of common-mode current to construct a deep feature sequence of common-mode current. The common-mode current characteristic baseline analysis module is used to divide different load operating state intervals based on the historical operating state data of the frequency converter using the DBSCAN clustering algorithm, and to construct the common-mode current characteristic baseline model corresponding to each load interval. The grounding anomaly identification module is used to identify anomaly feature windows and determine the set of feature offset events for potential grounding anomalies based on the common-mode current depth feature sequence and load operating status. The grounding degradation status identification module is used to determine the grounding status corresponding to each potential grounding status anomaly based on the grounding status characteristic parameters corresponding to the characteristic offset events of potential grounding status anomalies, and to execute the corresponding grounding degradation maintenance strategy. The model evaluation and optimization module is used to evaluate the accuracy of grounding degradation status identification by comparing the grounding status identification results with maintenance record data, and to optimize the parameters and feature weights of the grounding degradation status classification identification model.

[0004] Furthermore, the common-mode current data acquisition module is used to acquire common-mode current data and inverter operating status data through a cloud platform, form a continuous time series according to a preset sampling period, and perform data preprocessing, including: The system accesses the variable frequency power supply system's operation data interface via the cloud platform, acquiring common-mode current data from current sensors deployed at the common-mode circuit location of the cable connecting the inverter output side and the load. It also acquires inverter operating status data uploaded synchronously with the common-mode current data, including carrier frequency, modulation ratio, output frequency, DC bus voltage, estimated load torque, and operating mode identifier. A unified time base is used to timestamp and calibrate data uploaded from different acquisition terminals, forming a continuous time series according to a preset sampling period. Data preprocessing is performed on the uploaded data, including removing lost sampling points, saturation values, and abnormal abrupt changes, and compensating for missing time periods. Amplitude normalization and time alignment are used to map data uploaded from different acquisition terminals to a unified time scale.

[0005] Furthermore, the common-mode current feature extraction module is used to extract time-domain features, frequency-domain features, and phase-domain features from the time-series data of the common-mode current, and construct a deep feature sequence of the common-mode current, including: Based on the time-series data of common-mode current, a short-time Fourier transform algorithm is used to segment the common-mode current signal into time-frequency components to obtain the spectral distribution data within each time window. A continuous wavelet transform algorithm is used to decompose the common-mode current data into multiple scales, extracting transient change features at different time scales to form a multi-scale feature representation. A Hilbert transform is used to calculate the instantaneous parameters of the common-mode current data within the corresponding time window, including instantaneous amplitude, instantaneous frequency, and instantaneous phase information. By constructing a fixed-length sliding time window, the high-frequency energy ratio, spectral centroid offset, phase difference distribution density, waveform symmetry index, and transient rate of change of the common-mode current data within each time window are statistically analyzed. Using a feature vector concatenation method, the time-domain features, frequency-domain features, and phase-domain features are combined into a high-dimensional feature matrix to form a deep feature sequence of the common-mode current, which is stored on a cloud platform. The time-domain features include the common-mode current amplitude change rate, waveform symmetry index, and high-frequency energy ratio; the frequency-domain features include spectral energy distribution and spectral centroid offset; and the phase-domain features include instantaneous phase parameters and phase difference distribution density.

[0006] Furthermore, the common-mode current characteristic baseline analysis module is used to divide different load operating state intervals based on the inverter's historical operating state data using the DBSCAN clustering algorithm, and to construct a common-mode current characteristic baseline model corresponding to each load interval, including: By acquiring historical operating status data of the frequency converter through the cloud platform and labeling normal grounding state samples, the DBSCAN clustering algorithm is used to divide different load operating status intervals based on the common-mode current deep feature sequence, and each sample is labeled with the corresponding load interval number. Within each load interval, an autoencoder neural network is used to reconstruct and train the normal grounding state samples. The model parameters are optimized by minimizing the reconstruction error between the input features and the reconstructed features. The autoencoder model is trained separately for different load intervals to construct the common-mode current feature baseline model corresponding to each load interval.

[0007] Furthermore, the grounding anomaly identification module is used to identify anomaly feature windows based on the common-mode current depth feature sequence and load operating status, and to determine a set of feature offset events for potential grounding anomalies, including: Based on the common-mode current depth feature sequence and its load operating status, the common-mode current depth feature under normal grounding conditions is reconstructed using the common-mode current feature baseline model corresponding to the current load range. The feature reconstruction residual vector is obtained by calculating the deviation between the real-time common-mode current depth feature and the reconstructed feature. A sliding window statistical method is used to calculate residual statistical indices, including the mean, variance, kurtosis, and skewness of the residuals. An isolated forest algorithm is used to detect anomalies in the residual statistical indices and identify abnormal feature windows that deviate from the historical normal distribution. The echo disturbance index is calculated based on real-time common-mode current data to identify echo disturbances in the abnormal feature windows. False anomaly signals identified as echo disturbances are removed to determine the set of feature offset events for potential grounding anomalies and to mark the power supply system locations with potential grounding anomalies.

[0008] It also includes calculating the echo disturbance index based on real-time common-mode current data, identifying echo disturbances in the abnormal feature window, and eliminating false abnormal signals identified as echo disturbances, specifically including: Based on real-time common-mode current data and a preset time window size, the echo disturbance index formula is used. The echo disturbance index within each sliding window is calculated. ,in, This represents real-time common-mode current data. t represents the start time of the anomaly feature window, and t represents the current time. The time delay variable represents the time difference between the current moment and its historical counterpart. The length of the echo delay window. The common-mode current historical mirror data refers to the common-mode current data reflected or fed back within a certain period after the start time of the abnormal feature window. If the calculated echo disturbance index is less than the preset index threshold, it is determined that the residual belongs to echo disturbance and is not an actual grounding anomaly. The pseudo-abnormal signal identified as echo disturbance is removed from the abnormal feature window identification results.

[0009] Furthermore, the grounding degradation state identification module is used to determine the grounding state corresponding to each potential grounding state anomaly's characteristic offset event based on the grounding state characteristic parameters corresponding to the characteristic offset events of potential grounding state anomalies, and to execute the corresponding grounding degradation maintenance strategy, including: Historical grounding degradation data samples are acquired through a cloud platform, and grounding status characteristic parameters are extracted. These samples include the characteristic offset amplitude, duration, change slope, and correlation information between multidimensional residuals under different grounding states. The grounding status characteristic parameters include characteristic offset amplitude, characteristic offset duration, characteristic change slope, and multidimensional residual correlation. Based on the historical data of grounding status characteristic parameters and corresponding grounding status annotation information, a random forest algorithm is used to train a model, constructing a grounding degradation status classification identification model. Grounding states include mild degradation, moderate degradation, and severe degradation. Based on the grounding status characteristic parameters corresponding to characteristic offset events of potential grounding status anomalies, the grounding status corresponding to each potential grounding status anomaly is determined using the grounding degradation status classification identification model. A graded early warning mechanism is activated based on the identified grounding status, pushing warning information to maintenance personnel. Corresponding grounding degradation maintenance strategies are then implemented based on different grounding states. These strategies include grounding connection repair, replacement of grounding conductors or grounding electrodes, local insulation repair, replacement of related components, increasing grounding monitoring points, or increasing the grounding status monitoring frequency.

[0010] Furthermore, the model evaluation and optimization module is used to evaluate the accuracy of grounding degradation state identification by comparing the grounding state identification results with maintenance record data, and to optimize the parameters and feature weights of the grounding degradation state hierarchical identification model, including: By comparing the grounding status identification results obtained based on the grounding degradation status classification identification model with the grounding status anomaly events and grounding status confirmation data recorded during the maintenance process, the accuracy of the grounding degradation status classification identification model and grounding status anomaly judgment parameters in identifying grounding degradation status is evaluated. If the identification accuracy of the model and judgment parameters is lower than the preset accuracy threshold, the grounding degradation status classification identification model and grounding status anomaly judgment parameters are optimized until the identification accuracy of the model and judgment parameters reaches the preset requirements. The optimized grounding degradation status classification identification model and grounding status anomaly judgment parameters are then implemented. The optimization process includes adjusting the feature input weights and reconfiguring the model parameters of the grounding degradation status classification identification model.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention provides a cloud-based remote monitoring and intelligent early warning system for power grids. The system utilizes a cloud platform to uniformly collect and time-series process common-mode current data and inverter operating status data. Combined with multi-dimensional feature extraction, adaptive load state modeling, and anomaly feature identification, it achieves continuous, accurate, and intelligent monitoring of power grid grounding status. The invention constructs a deep feature sequence of common-mode current using time-domain, frequency-domain, and phase-domain features, comprehensively depicting the variation patterns of common-mode current under different operating conditions. Simultaneously, by clustering analysis of historical inverter operating states, it automatically distinguishes different load operating ranges and establishes corresponding common-mode current feature baseline models, effectively eliminating the interference of load changes on grounding anomaly identification results. By analyzing the feature offset relationship between the deep features of common-mode current and load operating status, the invention accurately identifies feature offset events of potential grounding anomalies. Furthermore, it introduces an echo disturbance index to discriminate and eliminate echo disturbances in the anomaly feature window, effectively distinguishing between genuine grounding anomalies and false anomaly signals caused by electromagnetic interference or system echoes, significantly reducing the false alarm rate and improving the reliability of early warning results. By analyzing the grounding status characteristic parameters corresponding to characteristic offset events, this invention enables refined identification of grounding degradation states and automatically matches and executes corresponding maintenance strategies, thereby improving the intelligence level of power grid operation and maintenance. The cloud-based remote monitoring and intelligent early warning system for power grids provided by this invention achieves continuous perception of power grid grounding status, early identification of degradation trends, and intelligent linkage of maintenance strategies. This effectively reduces the power grid operation risks caused by grounding system degradation and significantly improves the safety, intelligence level, and operation and maintenance management efficiency of the power grid. Attached Figure Description

[0012] Figure 1 This is a flowchart of a cloud platform-based remote monitoring and intelligent early warning system for power grids according to the present invention; Figure 2 This is a schematic diagram of a cloud-based remote monitoring and intelligent early warning system for power grids according to the present invention. Figure 3 This is another schematic diagram of a cloud-based remote monitoring and intelligent early warning system for power grids according to the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] like Figure 1-3 This embodiment of a cloud platform-based remote power grid monitoring and intelligent early warning system may specifically include: Step S101: The common-mode current data acquisition module is used to acquire common-mode current data and inverter operating status data through the cloud platform, form a continuous time series according to the preset sampling period, and perform data preprocessing.

[0015] The system accesses the variable frequency power supply system's operation data interface via the cloud platform, acquiring common-mode current data from current sensors deployed at the common-mode circuit location of the cable connecting the inverter output side and the load. It also acquires inverter operating status data uploaded synchronously with the common-mode current data, including carrier frequency, modulation ratio, output frequency, DC bus voltage, estimated load torque, and operating mode identifier. A unified time base is used to timestamp and calibrate data uploaded from different acquisition terminals, forming a continuous time series according to a preset sampling period. Data preprocessing is performed on the uploaded data, including removing lost sampling points, saturation values, and abnormal abrupt changes, and compensating for missing time periods. Amplitude normalization and time alignment are employed to map data from different acquisition terminals to a unified time scale.

[0016] For example, during the actual operation of a variable frequency power supply system, common-mode current data is acquired from a current sensor located at the common-mode circuit of the cable connecting the inverter output side and the motor load via an operation data interface accessed through a cloud platform. The sampling frequency of the current sensor is 100kHz, and the instantaneous amplitude range of the common-mode current acquired at a certain moment is 0.8A to 1.6A. Simultaneously, the operating status data of the inverter is synchronously acquired through the same cloud platform interface, wherein the carrier frequency is 8kHz, the modulation ratio is 0.65, the output frequency is 45Hz, the DC bus voltage is 680V, ​​the estimated load torque is 120N·m, and the operating mode is identified as constant torque operating mode. Since the common-mode current data is uploaded by the field edge acquisition terminal, while the operating status data is uploaded by the inverter control unit, there is a millisecond-level time deviation between the two at the initial acquisition time. For example, the timestamp of the common-mode current data is 10:00:05.120, while the timestamp of the operating status data is 10:00:05.135. Therefore, the timestamp of the above data is calibrated by unifying the time base through the cloud platform, aligning them to 10:00:05.130, and forming a continuous time series according to the preset 1ms sampling period. During data preprocessing, it was found that there were two consecutive missing sampling points in the common-mode current data between 10:00:05.148 and 10:00:05.150. At the same time, an abnormal abrupt change occurred at 10:00:05.162, which instantly jumped to 3.5A. This abrupt change clearly exceeded the normal common-mode current range under this operating condition. Therefore, the abnormal abrupt change was removed, and the two missing sampling points were compensated by linear interpolation of adjacent time points to keep the common-mode current time series continuous. Subsequently, the processed common-mode current data is normalized, for example, by mapping 0-5A to a standardized range of 0-1, so that the normalized common-mode current amplitude range is 0.16-0.32. At the same time, the operating status data is time-aligned so that parameters such as carrier frequency, modulation ratio, and output frequency correspond one-to-one with the common-mode current data on the same time scale. This results in a set of multi-source standardized operating data sequences with a unified time base, unified sampling period, and unified numerical scale, which serve as input data for subsequent analysis and processing.

[0017] Step S102, common-mode current feature extraction module, is used to extract time-domain features, frequency-domain features and phase-domain features based on the time series data of common-mode current, and construct a common-mode current deep feature sequence.

[0018] Based on the time-series data of common-mode current, a short-time Fourier transform algorithm is used to segment the common-mode current signal into time-frequency components, obtaining the spectral distribution data within each time window. A continuous wavelet transform algorithm is then used to decompose the common-mode current data into multiple scales, extracting transient change features at different time scales to form a multi-scale feature representation. A Hilbert transform is employed to calculate the instantaneous parameters of the common-mode current data within the corresponding time window, including instantaneous amplitude, instantaneous frequency, and instantaneous phase information. By constructing a fixed-length sliding time window, the high-frequency energy ratio, spectral centroid offset, phase difference distribution density, waveform symmetry index, and transient rate of change of the common-mode current data within each time window are statistically analyzed. Using a feature vector concatenation method, the time-domain features, frequency-domain features, and phase-domain features are combined into a high-dimensional feature matrix, forming a deep feature sequence of the common-mode current, which is stored on a cloud platform. The time-domain features include the common-mode current amplitude change rate, waveform symmetry index, and high-frequency energy ratio; the frequency-domain features include spectral energy distribution and spectral centroid offset; and the phase-domain features include instantaneous phase parameters and phase difference distribution density.

[0019] For example, during the stable operation of a variable frequency power supply system, the cloud platform receives a 1-second common-mode current time series data. This data is collected by the output-side common-mode current sensor at a sampling frequency of 20kHz, corresponding to 20,000 continuous sampling points. The instantaneous current amplitude is mainly distributed between 1.0A and 1.8A. Based on this common-mode current time series data, a short-time Fourier transform algorithm is used to segment the signal for time-frequency expansion. Each time window is set to a length of 10ms, with adjacent windows overlapping by 5ms. Within each time window, a set of spectral distribution data is obtained. For instance, within a certain time window, the energy proportion of the common-mode current in the 5kHz–15kHz frequency band is significantly higher than that in the low-frequency band, with the spectral energy mainly concentrated around approximately 9kHz. Furthermore, a continuous wavelet transform algorithm is used to perform multi-scale decomposition on the same segment of common-mode current data. At a small scale, transient pulse features with a duration of less than 2ms are extracted, while at a large scale, current trend features that fluctuate slowly with load changes are extracted, thus forming a multi-scale feature representation that can simultaneously reflect both rapid and slow-changing behaviors. Meanwhile, the Hilbert transform was used to analyze the common-mode current signal within each time window. The instantaneous amplitude of the common-mode current within each time window was calculated to be approximately 1.45 A, with the instantaneous frequency fluctuating between 8.5 kHz and 9.3 kHz. The corresponding instantaneous phase changed continuously within the range of 0 to 2π, and the rate of phase change remained stable within this window. Subsequently, by constructing a fixed-length sliding time window, statistical analysis was performed on the common-mode current data within each time window. The results showed that the high-frequency energy proportion within this window was 0.62, the spectral centroid shifted approximately +0.8 kHz relative to the reference frequency, the phase difference distribution was mainly concentrated in the ±0.15 rad range, the amplitude difference between the positive and negative half-cycles of the waveform was less than 5%, the corresponding waveform symmetry index was 0.92, and the rate of change of the common-mode current amplitude between adjacent windows was approximately 0.03 A / ms. Finally, the common-mode current amplitude change rate, waveform symmetry index, and high-frequency energy ratio obtained from the time-varying statistics are used as time-domain features, the spectral energy distribution and spectral centroid offset obtained from the short-time Fourier transform are used as frequency-domain features, and the instantaneous phase parameters and phase difference distribution density obtained from the Hilbert transform are used as phase-domain features. The three types of features are combined in a unified manner using a feature vector concatenation method to form a set of high-dimensional feature vectors corresponding to the time window. As the time window slides continuously, a complete common-mode current depth feature sequence is finally formed, and the common-mode current depth feature sequence is stored in the cloud platform.

[0020] Step S103, the feature baseline analysis module, is used to divide different load operating state intervals based on the historical operating state data of the frequency converter using the DBSCAN clustering algorithm, and to construct the common mode current feature baseline model corresponding to each load interval.

[0021] Historical operating status data of the frequency converter is acquired through a cloud platform, and samples of normal grounding states are labeled. Combined with the common-mode current deep feature sequence, the DBSCAN clustering algorithm is used to divide different load operating status intervals, and each sample is labeled with a corresponding load interval number. Within each load interval, an autoencoder neural network is used to reconstruct and train the features of the normal grounding state samples. The model parameters are optimized by minimizing the reconstruction error between the input features and the reconstructed features. Autoencoder models are trained separately for different load intervals to construct a common-mode current feature baseline model corresponding to each load interval.

[0022] For example, historical operating status data of a variable frequency power supply system for three consecutive months is retrieved through a cloud platform. This includes operating status parameters of the frequency converter during stable operation and the common-mode current depth feature sequence at corresponding times. The operating status parameters include load torque, output frequency, and modulation ratio. Taking one segment of historical data as an example, approximately 200,000 time window samples were collected during the normal grounding status confirmation period. Each sample corresponds to a common-mode current depth feature vector and a set of operating status parameters. The load torque is mainly distributed between 40 N·m and 180 N·m, the output frequency is distributed between 20 Hz and 50 Hz, and the modulation ratio is distributed between 0.4 and 0.85. First, the above operating status parameters and the corresponding common-mode current depth features are jointly represented so that each sample simultaneously reflects the current load conditions and common-mode current behavior characteristics. Based on this, the DBSCAN clustering algorithm is used to perform cluster analysis on the joint feature samples. The distance between samples is determined by the comprehensive differences in load torque, output frequency, modulation ratio, and common-mode current characteristics. When the distance between samples in this feature space is less than a preset similarity threshold, they are determined to belong to the same operating status category. Through cluster analysis, historical samples were naturally divided into several load operating state intervals, forming three main load operating state intervals. The first interval corresponds to a light load operating state with load torque concentrated between 40 N·m and 80 N·m and output frequency concentrated between 20 Hz and 30 Hz. The second interval corresponds to a medium load operating state with load torque concentrated between 80 N·m and 130 N·m and output frequency concentrated between 30 Hz and 40 Hz. The third interval corresponds to a heavy load operating state with load torque concentrated between 130 N·m and 180 N·m and output frequency concentrated between 40 Hz and 50 Hz. Each sample was automatically labeled with its corresponding load interval number. Subsequently, within each load operating state interval, only samples labeled as being in a normal grounding state were selected as training data. Taking the second interval as an example, approximately 60,000 normal grounding state samples were selected from this interval, and their corresponding common-mode current deep feature vectors were input into an autoencoder neural network for feature reconstruction training. The common-mode current deep feature vector has a 32-dimensional dimension, consisting of time-domain features, frequency-domain features, and phase-domain features concatenated together. The input layer of the autoencoder neural network has 32 nodes, consistent with the dimension of the feature vector. The encoder part is a three-layer fully connected structure with 16, 8, and 4 hidden layer nodes, used to compress the feature representation layer by layer to form a low-dimensional latent feature vector. The decoder part adopts a symmetric structure with 8, 16, and 32 nodes, used to restore the low-dimensional latent representation to the reconstructed feature vector. Each hidden layer uses the ReLU activation function, and the output layer uses a linear activation function to maintain feature continuity. During training, mean squared error is used as the reconstruction loss function, and the overall reconstruction error is calculated by comparing the difference between the original input features and the reconstructed features.The optimization algorithm used was an adaptive gradient optimization algorithm for parameter updates, with an initial learning rate of 0.001 and a batch size of 256. All training samples were iterated once per training round, with a total of 80 training rounds. 10% of the samples were used as a validation set to monitor model convergence. Training was terminated early to prevent overfitting when the reconstruction error on the validation set stopped decreasing after five consecutive rounds. After training, the average reconstruction error of the autoencoder model corresponding to the second interval on the validation samples stabilized below 0.012, indicating that the model could accurately reconstruct the deep features of common-mode current under normal grounding conditions under medium load operation. Using the same network structure and training method, the normal grounding state samples in the first and third intervals were trained independently, resulting in multiple common-mode current feature baseline models corresponding one-to-one with different load operating state intervals. Each baseline model represents the typical distribution of common-mode current features under normal grounding conditions within its corresponding load operating state interval.

[0023] Step S104, the grounding anomaly identification module is used to identify anomaly feature windows and determine the feature offset event set of potential grounding anomalies based on the common-mode current depth feature sequence and load operating status.

[0024] Based on the common-mode current depth feature sequence and its load operating status, the common-mode current depth feature under normal grounding conditions is reconstructed using the common-mode current feature baseline model corresponding to the current load range. The feature reconstruction residual vector is obtained by calculating the deviation between the real-time common-mode current depth feature and the reconstructed feature. A sliding window statistical method is then used to calculate residual statistical indices, including the mean, variance, kurtosis, and skewness of the residuals. An isolated forest algorithm is employed to detect anomalies in the residual statistical indices, identifying anomalous feature windows that deviate from the historical normal distribution. The echo disturbance index is calculated based on real-time common-mode current data to identify echo disturbances within the anomalous feature windows. False anomaly signals identified as echo disturbances are removed, determining the set of feature offset events for potential grounding anomalies and marking the power supply system locations with potential grounding anomalies.

[0025] For example, during the operation of a variable frequency power supply system, the cloud platform receives a common-mode current depth characteristic sequence during the real-time monitoring phase, which corresponds to the medium-load operating state range, where the load torque is about 110 N·m, the output frequency is about 36 Hz, and the modulation ratio is about 0.72. Based on the load's operating status, a common-mode current characteristic baseline model corresponding to the load range is automatically selected. Within a certain time window, the obtained common-mode current depth feature vector is formed by concatenating time-domain feature vectors, frequency-domain feature vectors, and phase-domain feature vectors. The time-domain feature vector includes a common-mode current amplitude change rate of 0.031 A / ms, a waveform symmetry index of 0.94, and a high-frequency energy proportion of 0.58. The frequency-domain feature vector includes the spectral energy distribution ratio in multiple frequency bands, with a high-frequency energy proportion of 0.62, a mid-frequency energy proportion of 0.28, and a low-frequency energy proportion of 0.10. At the same time, the spectral centroid is offset by approximately +0.6 kHz relative to the reference frequency. The phase-domain feature vector includes the instantaneous phase parameter continuously changing within the range of 0 to 2π, and the phase difference distribution density mainly concentrated in the ±0.12 rad interval. After inputting the complete common-mode current depth feature vector into the corresponding common-mode current characteristic baseline model, the model outputs a set of reconstructed common-mode current depth feature vectors under normal grounding conditions. The common-mode current amplitude change rate is 0.028 A / ms, the waveform symmetry index is 0.96, and the high-frequency energy proportion is 0.55. The reconstruction results of the corresponding frequency domain feature vectors are: high-frequency energy proportion 0.58, mid-frequency energy proportion 0.30, and low-frequency energy proportion 0.12, with a spectral centroid offset of approximately +0.4 kHz. The reconstruction results of the corresponding phase domain feature vectors show that the instantaneous phase change remains continuous, and the phase difference distribution density is concentrated within ±0.08 rad. By calculating the deviation between the real-time common-mode current depth feature vector and the reconstructed common-mode current depth feature vector in each corresponding dimension, a complete feature reconstruction residual vector is obtained. This residual vector simultaneously reflects the comprehensive deviation of the time-domain features, frequency-domain features, and phase-domain features from the normal grounding state. Subsequently, a sliding window statistical method was used to analyze the feature reconstruction residual vectors within multiple consecutive time windows. Within a sliding window lasting 200ms, the calculated residual mean was 0.035, residual variance was 0.004, kurtosis was 4.8, and skewness was 1.6, indicating that the residual distribution had significantly changed compared to historical normal grounding conditions. After inputting the above residual statistical indicators into an isolated forest anomaly detection model trained based on residual data from historical normal grounding conditions, the model determined that the residual statistical features corresponding to this time window belonged to an anomaly region, thus identifying this time window as an anomalous feature window.Meanwhile, based on the real-time common-mode current raw data, the echo disturbance index is further calculated to characterize the short-term high-frequency oscillation characteristics caused by cable reflection, power device switching transients, or communication interference. The echo disturbance index calculated within this abnormal feature window is 0.85, while the preset index threshold is 0.3, indicating that this abnormal window is not caused by echo disturbance. If the echo disturbance index calculated within other abnormal feature windows is less than 0.3, the abnormal feature window is identified as a pseudo-abnormal signal and removed. In subsequent continuous time windows, when an abnormal feature window is detected again where the residual mean continuously increases to above 0.06, the residual variance steadily increases, and the corresponding echo disturbance index remains below 0.2, this type of abnormal window is determined as a real feature offset event, and these are aggregated to form a feature offset event set for potential grounding anomalies. Finally, based on the distribution of feature offset events in the system topology, the potential grounding anomalies are marked as occurring in the power supply system section between the inverter output side and the motor load.

[0026] Among them, the echo disturbance index is calculated based on real-time common-mode current data, the echo disturbance in the abnormal feature window is identified, and false abnormal signals identified as echo disturbances are eliminated.

[0027] Based on real-time common-mode current data and a preset time window size, the echo disturbance index formula is used. The echo disturbance index within each sliding window is calculated. ,in, This represents real-time common-mode current data. t represents the start time of the anomaly feature window, and t represents the current time. The time delay variable represents the time difference between the current moment and its historical counterpart. The length of the echo delay window. This refers to the common-mode current historical mirror data, which is the common-mode current data reflected or fed back within a certain period after the start time of the anomaly feature window. If the calculated echo disturbance index is less than the preset index threshold, the residual is determined to be an echo disturbance, not an actual grounding anomaly, and the pseudo-anomaly signal identified as an echo disturbance is removed from the anomaly feature window identification results.

[0028] For example, during the operation of a certain frequency converter power supply system, the cloud platform identified an anomaly feature window lasting 200ms during the anomaly detection phase. The starting time of this window is recorded as t0=10:00:05.200. Real-time common-mode current data was collected within this anomaly feature window. The instantaneous amplitude is mainly distributed between 1.2A and 1.6A. To determine whether this anomaly is caused by echo disturbance due to cable reflection or feedback, the echo disturbance index formula is introduced. The preset sliding time window length Set to 5ms, echo delay window The length is set to 10ms to characterize the typical delay range of the common-mode current signal after reflection in the propagation path. Real-time common-mode current data is selected at the current time t=10:00:05.230. Its corresponding historical mirror data For comparison, historical mirror data refers to reflected common-mode current signals acquired within 10ms after the start of the abnormal feature window, which may be caused by cable length mismatch or impedance abrupt changes. For example, in From 10:00:05.230 to 10:00:05.235, the average amplitude of the real-time common-mode current data was approximately 1.48A, while the average amplitude of the common-mode current data at the corresponding historical mirror time was approximately 1.45A. The two are highly similar in amplitude and waveform. Substituting these values ​​into the echo disturbance index formula, the echo disturbance index within this sliding window is finally obtained. The echo disturbance index is approximately 0.12. Comparing this index with the preset index threshold of 0.3, it is found that the echo disturbance index within the current sliding window is significantly smaller than the threshold. This indicates that the difference between the real-time common-mode current signal and its historical mirror signal is small. This abnormal feature is more consistent with the echo disturbance characteristics caused by signal reflection or feedback, rather than a real anomaly caused by ground path degradation. Therefore, the residual change corresponding to this abnormal feature window is determined to be a pseudo-abnormal signal caused by echo disturbance, and this abnormal window is removed from the potential grounding state anomaly identification results.

[0029] Step S105, the grounding degradation state identification module is used to determine the grounding state corresponding to each potential grounding state abnormality characteristic offset event based on the grounding state characteristic parameters corresponding to the characteristic offset events of potential grounding state abnormalities, and execute the corresponding grounding degradation maintenance strategy.

[0030] Historical grounding degradation data samples are acquired through a cloud platform, and grounding status characteristic parameters are extracted. These samples include the characteristic offset amplitude, duration, slope of change, and correlation information between multidimensional residuals under different grounding states. The grounding status characteristic parameters include characteristic offset amplitude, characteristic offset duration, characteristic slope of change, and correlation of multidimensional residuals. Based on the historical data of grounding status characteristic parameters and corresponding grounding status annotation information, a random forest algorithm is used to train a model to construct a grounding degradation status classification identification model, categorizing grounding states as mild, moderate, and severe degradation. Based on the grounding status characteristic parameters corresponding to characteristic offset events of potential grounding status anomalies, the grounding status corresponding to each potential grounding status anomaly is determined using the grounding degradation status classification identification model. A graded early warning mechanism is activated based on the identified grounding status, pushing warning information to maintenance personnel. Corresponding grounding degradation maintenance strategies are then implemented based on different grounding states, including grounding connection repair, replacement of grounding conductors or grounding electrodes, local insulation repair, replacement of related components, increasing grounding monitoring points, or increasing the frequency of grounding status monitoring.

[0031] For example, by summarizing and analyzing the operation and maintenance records of an industrial frequency converter power supply system over the past three years through a cloud platform, a total of 2100 sets of historical grounding degradation data samples with confirmed grounding states were obtained, including 800 sets of slightly degraded samples, 750 sets of moderately degraded samples, and 550 sets of severely degraded samples. For each set of historical samples, based on the aforementioned feature offset identification results, the corresponding grounding state feature parameters were extracted. The feature offset amplitude of a slightly degraded sample was shown to be approximately 0.038 in the mean value of the common-mode current feature reconstruction residual, which is slightly higher than the level of no more than 0.03 under normal grounding conditions. The feature offset lasted for about 3 hours and intermittently dropped during the monitoring process. The feature change slope was approximately 0.0004 / min, and the change trend was relatively gentle. At the same time, multidimensional residual correlation analysis showed that the correlation coefficient between the time domain residual and the frequency domain residual was approximately 0.32, indicating that no obvious coordinated offset of various features had occurred. Taking another moderately deteriorated sample as an example, its feature offset amplitude increased to approximately 0.062, lasting for more than 12 hours, with a feature change slope of approximately 0.0013 / min. Furthermore, the correlation coefficients between the time-domain, frequency-domain, and phase-domain residuals generally reached above 0.6, indicating that multiple features began to synchronously deteriorate. In a severely deteriorated sample, the feature offset amplitude remained above 0.085 for a long period, lasting for more than 48 hours, with a feature change slope of approximately 0.0024 / min. The multidimensional residual correlation coefficients were generally higher than 0.75, reflecting a systematic deterioration of the grounding condition. Using the aforementioned feature offset amplitude, duration, change slope, and multidimensional residual correlation as input features, and combining manual maintenance records and insulation resistance test results to label the grounding deterioration level of each sample, all 2100 samples were divided into training and validation sets, with 80% used for model training and 20% for model validation. In the model construction phase, a random forest algorithm was used to establish a grounding deterioration state classification and identification model. The random forest uses 150 decision trees, each with a maximum depth of 8 layers to prevent overfitting. A subset of features is randomly selected at each node split to enhance generalization. The minimum number of leaf node samples is set to 5 to ensure classification stability. The model uses the Gini index as the node splitting criterion, classifying input features through voting across multiple decision trees. The final output is a probability distribution of mild, moderate, or severe degradation, with the category corresponding to the highest probability being the final classification result. During training, cross-validation is used to fine-tune the number of trees and the maximum depth, achieving a classification accuracy of over 92% on the validation set. Validation results show a 91% accuracy rate for mild degradation, a 93% accuracy rate for moderate degradation, and a 94% accuracy rate for severe degradation, with an overall classification accuracy of 92.6%.When the cloud platform detects a feature offset event indicating a potential grounding anomaly, for example, if the feature offset amplitude is 0.041, the duration is approximately 2.5 hours, the feature change slope is 0.0005 / min, and the multidimensional residual correlation coefficient is approximately 0.35, the grounding status feature parameters are input into a trained random forest model. The model outputs a probability of 0.78 for mild degradation, 0.18 for moderate degradation, and 0.04 for severe degradation. The platform ultimately determines that the grounding status corresponding to this anomaly is mildly degraded. Based on this identification result, an early warning mechanism for mild degradation is automatically activated. A warning message is pushed to maintenance personnel via the cloud platform, and corresponding preventative maintenance strategies are implemented. These include increasing the grounding status monitoring frequency of the power supply circuit from once every 10 minutes to once every 2 minutes, temporarily adding monitoring points at key grounding connection points, and arranging for maintenance personnel to conduct focused inspections and status verifications of the grounding connection points during the next routine inspection. If subsequent monitoring reveals that the characteristic offset of the circuit further increases to above 0.065 and lasts for more than 10 hours, and the model re-determines the grounding status as moderately deteriorated, the system will upgrade the warning level, prompting maintenance personnel to implement planned maintenance strategies within the planned maintenance window. This includes repairing loose or aging grounding connections, patching areas of localized insulation damage, or replacing some grounding conductors. If the characteristic offset further develops to exceed 0.085 and lasts for more than 48 hours, and is determined by the model to be in a severely deteriorated state, the system will trigger the highest level warning, requiring immediate implementation of enhanced maintenance strategies, including replacing grounding conductors or grounding electrodes and replacing related aging components. This intervention will be completed before severe grounding deterioration occurs, ensuring the safe and stable operation of the frequency converter power supply system.

[0032] Step S106, Model Evaluation and Optimization Module, is used to evaluate the accuracy of grounding degradation status identification by comparing the grounding status identification results with maintenance record data, and to optimize the parameters and feature weights of the grounding degradation status classification identification model.

[0033] By comparing the grounding status identification results obtained based on the grounding degradation status classification identification model with the grounding status anomaly events and grounding status confirmation data recorded during maintenance, the accuracy of the grounding degradation status classification identification model and grounding status anomaly judgment parameters in identifying grounding degradation status is evaluated. If the identification accuracy of the model and judgment parameters is lower than the preset accuracy threshold, the grounding degradation status classification identification model and grounding status anomaly judgment parameters are optimized until the identification accuracy of the model and judgment parameters reaches the preset requirements. The optimized grounding degradation status classification identification model and grounding status anomaly judgment parameters are then implemented. The optimization process includes adjusting the feature input weights and reconfiguring the model parameters.

[0034] For example, after a variable frequency power supply system was put into operation, the cloud platform tracked and evaluated the operational effectiveness of the grounding degradation status classification identification model. During subsequent use, a total of 120 grounding status anomaly events were recorded, triggering maintenance procedures. For each anomaly event, manual inspection, grounding resistance testing, and component replacement results were recorded during subsequent maintenance, forming corresponding grounding status confirmation data. By comparing the grounding status results identified by the model with the actual grounding status confirmed during the maintenance process, the model's identification accuracy under three states: mild degradation, moderate degradation, and severe degradation was statistically analyzed. For instance, out of the 120 anomaly events, the model correctly identified the grounding status 102 times. The accuracy rate for mild degradation was approximately 88%, for moderate degradation approximately 82%, and for severe degradation approximately 75%, with an overall accuracy rate of approximately 85%. Comparing the overall recognition accuracy with the preset accuracy threshold of 90%, it was found that the recognition accuracy of the current grounding degradation state classification recognition model and grounding state anomaly judgment parameters did not meet the preset requirements, especially in the recognition of severe degradation states. Based on this evaluation result, the system automatically entered the model optimization process to jointly optimize the grounding degradation state classification recognition model and grounding state anomaly judgment parameters. During the optimization process, the system first analyzed the feature distribution of misjudged samples and found that some severe degradation samples had insufficient discriminative power in the feature offset duration and multidimensional residual correlation dimensions. Therefore, the feature input weights corresponding to feature offset duration and multidimensional residual correlation were appropriately increased in the model. At the same time, the number of trees and the maximum depth of a single tree in the random forest model were reconfigured to enhance the model's ability to learn long-term cumulative degradation features. After adjusting the model parameters, the system re-verified the optimized model using newly added grounding status anomaly events and their corresponding maintenance confirmation data from the past two months. This verification process involved 40 anomaly events, of which the model correctly identified 38, improving the overall recognition accuracy to 95%. The accuracy for identifying severely degraded grounding states increased to 92%, reaching and exceeding the preset accuracy threshold. Based on this, the optimized grounding degradation status classification identification model and grounding status anomaly judgment parameters were determined to meet application requirements. The optimized model and judgment parameters were then officially deployed to the cloud platform operating environment to replace the original model for continuous grounding status identification and early warning.

[0035] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the concept of this application. The exemplary features described above are examples of technical solutions formed by mutually substituting with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A cloud-based remote monitoring and intelligent early warning system for power grids, characterized in that, The system includes: The common-mode current data acquisition module is used to acquire common-mode current data and inverter operating status data through the cloud platform, form a continuous time series according to the preset sampling period, and perform data preprocessing. The common-mode current feature extraction module is used to extract time-domain features, frequency-domain features, and phase-domain features from the time-series data of common-mode current to construct a deep feature sequence of common-mode current. The common-mode current characteristic baseline analysis module is used to divide different load operating state intervals based on the historical operating state data of the frequency converter using the DBSCAN clustering algorithm, and to construct the common-mode current characteristic baseline model corresponding to each load interval. The grounding anomaly identification module is used to identify anomaly feature windows and determine the set of feature offset events for potential grounding anomalies based on the common-mode current depth feature sequence and load operating status. The grounding degradation status identification module is used to determine the grounding status corresponding to each potential grounding status anomaly based on the grounding status characteristic parameters corresponding to the characteristic offset events of potential grounding status anomalies, and to execute the corresponding grounding degradation maintenance strategy. The model evaluation and optimization module is used to evaluate the accuracy of grounding degradation status identification by comparing the grounding status identification results with maintenance record data, and to optimize the parameters and feature weights of the grounding degradation status classification identification model.

2. The system according to claim 1, wherein, The common-mode current data acquisition module is used to acquire common-mode current data and inverter operating status data through a cloud platform, form a continuous time series according to a preset sampling period, and perform data preprocessing, including: The system accesses the variable frequency power supply system's operation data interface via the cloud platform, acquiring common-mode current data from current sensors deployed at the common-mode circuit location of the cable connecting the inverter output side and the load. It also acquires inverter operating status data uploaded synchronously with the common-mode current data, including carrier frequency, modulation ratio, output frequency, DC bus voltage, estimated load torque, and operating mode identifier. A unified time base is used to timestamp and calibrate data uploaded from different acquisition terminals, forming a continuous time series according to a preset sampling period. Data preprocessing is performed on the uploaded data, including removing lost sampling points, saturation values, and abnormal abrupt changes, and compensating for missing time periods. Amplitude normalization and time alignment are used to map data uploaded from different acquisition terminals to a unified time scale.

3. The system according to claim 1, wherein, The common-mode current feature extraction module is used to extract time-domain features, frequency-domain features, and phase-domain features from the time-series data of the common-mode current, and construct a deep feature sequence of the common-mode current, including: Based on the time-series data of common-mode current, a short-time Fourier transform algorithm is used to segment the common-mode current signal into time-frequency components to obtain the spectral distribution data within each time window. A continuous wavelet transform algorithm is used to decompose the common-mode current data into multiple scales, extracting transient change features at different time scales to form a multi-scale feature representation. A Hilbert transform is used to calculate the instantaneous parameters of the common-mode current data within the corresponding time window, including instantaneous amplitude, instantaneous frequency, and instantaneous phase information. By constructing a fixed-length sliding time window, the high-frequency energy ratio, spectral centroid offset, phase difference distribution density, waveform symmetry index, and transient rate of change of the common-mode current data within each time window are statistically analyzed. Using a feature vector concatenation method, the time-domain features, frequency-domain features, and phase-domain features are combined into a high-dimensional feature matrix to form a deep feature sequence of the common-mode current, which is stored on a cloud platform. The time-domain features include the common-mode current amplitude change rate, waveform symmetry index, and high-frequency energy ratio; the frequency-domain features include spectral energy distribution and spectral centroid offset; and the phase-domain features include instantaneous phase parameters and phase difference distribution density.

4. The system according to claim 1, wherein, The common-mode current characteristic baseline analysis module is used to divide different load operating state intervals based on the inverter's historical operating state data using the DBSCAN clustering algorithm, and to construct a common-mode current characteristic baseline model corresponding to each load interval, including: By acquiring historical operating status data of the frequency converter through the cloud platform and labeling normal grounding state samples, the DBSCAN clustering algorithm is used to divide different load operating status intervals based on the common-mode current deep feature sequence, and each sample is labeled with the corresponding load interval number. Within each load interval, an autoencoder neural network is used to reconstruct and train the normal grounding state samples. The model parameters are optimized by minimizing the reconstruction error between the input features and the reconstructed features. The autoencoder model is trained separately for different load intervals to construct the common-mode current feature baseline model corresponding to each load interval.

5. The system according to claim 1, wherein, The grounding anomaly identification module is used to identify anomaly feature windows and determine a set of feature offset events for potential grounding anomalies based on the common-mode current depth feature sequence and load operating status, including: Based on the common-mode current depth feature sequence and its load operating status, the common-mode current depth feature under normal grounding conditions is reconstructed using the common-mode current feature baseline model corresponding to the current load range. The feature reconstruction residual vector is obtained by calculating the deviation between the real-time common-mode current depth feature and the reconstructed feature. A sliding window statistical method is used to calculate residual statistical indices, including the mean, variance, kurtosis, and skewness of the residuals. An isolated forest algorithm is used to detect anomalies in the residual statistical indices and identify abnormal feature windows that deviate from the historical normal distribution. The echo disturbance index is calculated based on real-time common-mode current data to identify echo disturbances in the abnormal feature windows. False anomaly signals identified as echo disturbances are removed to determine the set of feature offset events for potential grounding anomalies and to mark the power supply system locations with potential grounding anomalies.

6. The system according to claim 5, wherein, The step of calculating the echo disturbance index based on real-time common-mode current data, identifying echo disturbances in the abnormal feature window, and eliminating false abnormal signals identified as echo disturbances includes: Based on real-time common-mode current data and a preset time window size, the echo disturbance index formula is used. The echo disturbance index within each sliding window is calculated. ,in, This represents real-time common-mode current data. t represents the start time of the anomaly feature window, and t represents the current time. The time delay variable represents the time difference between the current moment and its historical counterpart. The length of the echo delay window. The common-mode current historical mirror data refers to the common-mode current data reflected or fed back within a certain period after the start time of the abnormal feature window. If the calculated echo disturbance index is less than the preset index threshold, it is determined that the residual belongs to echo disturbance and is not an actual grounding anomaly. The pseudo-abnormal signal identified as echo disturbance is removed from the abnormal feature window identification results.

7. The system according to claim 1, wherein, The grounding degradation state identification module is used to determine the grounding state corresponding to each potential grounding state anomaly's characteristic offset event based on the grounding state characteristic parameters corresponding to the characteristic offset events of potential grounding state anomalies, and to execute the corresponding grounding degradation maintenance strategy, including: Historical grounding degradation data samples are acquired through a cloud platform, and grounding status characteristic parameters are extracted. These samples include the characteristic offset amplitude, duration, change slope, and correlation information between multidimensional residuals under different grounding states. The grounding status characteristic parameters include characteristic offset amplitude, characteristic offset duration, characteristic change slope, and multidimensional residual correlation. Based on the historical data of grounding status characteristic parameters and corresponding grounding status annotation information, a random forest algorithm is used to train a model, constructing a grounding degradation status classification identification model. Grounding states include mild degradation, moderate degradation, and severe degradation. Based on the grounding status characteristic parameters corresponding to characteristic offset events of potential grounding status anomalies, the grounding status corresponding to each potential grounding status anomaly is determined using the grounding degradation status classification identification model. A graded early warning mechanism is activated based on the identified grounding status, pushing warning information to maintenance personnel. Corresponding grounding degradation maintenance strategies are then implemented based on different grounding states. These strategies include grounding connection repair, replacement of grounding conductors or grounding electrodes, local insulation repair, replacement of related components, increasing grounding monitoring points, or increasing the grounding status monitoring frequency.

8. The system according to claim 1, wherein, The model evaluation and optimization module is used to evaluate the accuracy of grounding degradation state identification by comparing the grounding state identification results with maintenance record data, and to optimize the parameters and feature weights of the grounding degradation state classification identification model, including: By comparing the grounding status identification results obtained based on the grounding degradation status classification identification model with the grounding status anomaly events and grounding status confirmation data recorded during the maintenance process, the accuracy of the grounding degradation status classification identification model and grounding status anomaly judgment parameters in identifying grounding degradation status is evaluated. If the identification accuracy of the model and judgment parameters is lower than the preset accuracy threshold, the grounding degradation status classification identification model and grounding status anomaly judgment parameters are optimized until the identification accuracy of the model and judgment parameters reaches the preset requirements. The optimized grounding degradation status classification identification model and grounding status anomaly judgment parameters are then implemented. The optimization process includes adjusting the feature input weights and reconfiguring the model parameters of the grounding degradation status classification identification model.

Images