Ground wire state monitoring and early warning system and method based on random forest algorithm

Abstract

The invention provides a grounding wire state monitoring and early warning system and method based on a random forest algorithm, relates to the technical field of electric power safety monitoring, and solves the limitation problems of low efficiency, poor data continuity, inaccurate early warning judgment and the like in the existing scheme. In the system, a multi-source sensor network acquires multi-dimensional operation data corresponding to a grounding wire and transmits the data to a data acquisition terminal; the data acquisition terminal performs feature extraction processing to generate a feature vector set and uploads the feature vector set to the cloud data processing platform; the cloud data processing platform uses a random forest classification model, based on the feature vector set, realizes dynamic feature weight adjustment through feature importance analysis, constructs a hierarchical decision tree structure, identifies and outputs a grounding wire fault mode, and issues the grounding wire fault mode to the local early warning terminal; and the local early warning terminal provides early warning information for the ground wire site by adopting a graded and differentiated early warning mode. According to the invention, high-precision early warning and predictive maintenance of the state of the grounding wire are effectively realized.

Description

Technical Field

[0001] This invention relates to the field of power safety monitoring technology, specifically to a grounding wire status monitoring and early warning system and method based on the random forest algorithm. Background Technology

[0002] Currently, monitoring the grounding wire status of power plants or substations mainly relies on regular manual inspections. Workers must inspect the grounding wire connections at close range under high-voltage conditions, which presents problems such as high safety risks, low efficiency, and poor data continuity. Especially under severe weather conditions, the difficulty and danger of inspections increase significantly, making it difficult to obtain real-time information on the grounding wire's operational status. Due to the long inspection cycle, sudden faults in the grounding wires, such as erosion or mechanical damage after lightning strikes, cannot be detected in a timely manner. Manual recording also introduces subjectivity, making data difficult to quantify and analyze, and failing to provide an effective basis for preventative maintenance.

[0003] Some substations use fixed sensors for online monitoring, but existing monitoring methods mostly focus on detecting single physical quantities, such as grounding resistance or surface temperature. Due to the complex electromagnetic environment of substations, single parameters are easily affected by interference and cannot comprehensively reflect the multi-dimensional changes in the grounding wire's condition. For example, a loose grounding wire may be accompanied by abnormal vibration characteristics, and corrosion processes may affect electrochemical properties, but existing systems lack the ability to comprehensively analyze the correlation between multiple parameters. Furthermore, some solutions require disconnecting the grounding wire for testing, making continuous online monitoring impossible. Additionally, non-contact sensors lack stability under strong electromagnetic interference, leading to a decrease in the reliability of monitoring data.

[0004] In fault diagnosis, existing technologies typically employ preset threshold rules for judgment, triggering alarms when parameters exceed fixed ranges. However, substation operating conditions are dynamically changing, and environmental factors significantly impact monitoring parameters. Fixed thresholds are ill-suited to complex conditions, leading to frequent false alarms or missed alarms. Long-term historical monitoring data recorded during substation operation has not been systematically used for pattern recognition, resulting in a lack of in-depth analysis of fault evolution patterns. For example, while grounding resistance naturally increases with load current, the system cannot distinguish between normal phenomena and potential faults, leading to a lack of scientific basis for maintenance decisions.

[0005] Furthermore, existing monitoring systems face challenges such as power supply difficulties and insufficient communication reliability during long-term operation. Some equipment relies on battery power and requires frequent replacement, increasing maintenance costs. Communication modules are prone to data loss in environments with strong electromagnetic interference, and the system cannot autonomously determine sensor faults, leading to a decrease in the reliability of monitoring data. Therefore, there is an urgent need for a grounding wire status monitoring method that can adapt to complex electromagnetic environments, integrate multi-dimensional information, and possess adaptive capabilities to improve the real-time performance, reliability, and data value of monitoring. Summary of the Invention

[0006] The purpose of this invention is to address the limitations of existing grounding wire monitoring schemes, such as low efficiency, poor data continuity, and inaccurate early warning judgments. Therefore, a grounding wire status monitoring and early warning system and method based on the random forest algorithm is proposed. This invention continuously collects multi-dimensional operational data through a multi-source sensor network deployed at the grounding wires of electrical equipment in power plants or substations. After feature extraction processing to generate a feature vector set, a trained random forest classification model dynamically adjusts feature weights through feature importance analysis and constructs a hierarchical decision tree structure to identify grounding wire fault modes. The system outputs a grounding wire fault probability assessment value and a fault cause analysis report, thereby achieving high-precision early warning and predictive maintenance of the grounding wire status.

[0007] The present invention employs the following technical solutions to achieve its objective: A grounding wire status monitoring and early warning system based on the random forest algorithm, comprising the following components: Deployed at the grounding wires of various electrical equipment in power plants or substations, it is used to collect multi-dimensional operational data corresponding to the grounding wires and transmit it to the data acquisition terminal; The data acquisition terminal is used to extract features from multi-dimensional operational data from a multi-source sensor network, generate corresponding feature vector sets, and then upload them to the cloud data processing platform. The cloud-based data processing platform is used to utilize a cloud-deployed random forest classification model. Based on the uploaded feature vector set, it dynamically adjusts feature weights through feature importance analysis and constructs a hierarchical decision tree structure to identify and output grounding wire fault modes, which are then sent to local early warning terminals. The local early warning terminal is deployed at the corresponding grounding wire and is used to provide early warning information to the grounding wire site using a graded and differentiated early warning method based on the grounding wire fault mode.

[0008] Preferably, the multi-dimensional operational data collected by the multi-source sensor network includes conductor temperature, conductor surface thermal radiation, vibration spectrum, mechanical tensile force, corrosion rate, current distribution, voltage waveform, and ambient temperature and humidity; wherein: conductor temperature is collected by a thermocouple fixed to the metal surface of the grounding wire; conductor surface thermal radiation is collected by an infrared thermal imager deployed above the grounding wire; vibration spectrum is collected by an accelerometer embedded inside the grounding terminal of the grounding wire; mechanical tensile force is collected by an S-shaped tensile sensor connected in series with the lower section of the grounding wire; corrosion rate is collected by an electrochemical potential sensor buried in the soil around the grounding terminal of the grounding wire; current distribution and voltage waveform are both collected by a Rogowski coil surrounding the conductor of the grounding wire; and ambient temperature and humidity are collected by a temperature and humidity sensor deployed at the grounding wire.

[0009] Preferably, the data acquisition terminal has a shielded shell with an optical fiber isolation interface, which establishes a communication connection with the multi-source sensor network through an optical fiber link. The data acquisition terminal has a built-in radio frequency transceiver unit based on narrowband wireless communication technology, which dynamically avoids the strong electromagnetic interference frequency band of the substation through an adaptive frequency hopping mechanism, realizing wireless communication between the data acquisition terminal and the cloud data processing platform. The data acquisition terminal is also used to preprocess the data after receiving multi-dimensional operating data by combining signal filtering and data compression algorithms, and then extract and generate the corresponding feature vector set.

[0010] Preferably, the cloud data processing platform is also used to pre-build and train the random forest classification model. First, after initializing a preset number of decision trees, the model is fitted using pre-acquired training set samples to complete basic training. After basic training, the importance scores of different features are calculated, so that the feature importance scores are converted into weight coefficients during the model retraining stage. At the same time, the decision tree splitting process is embedded in this stage, that is, when splitting nodes, the feature with the largest weighted information gain is selected first. Then, based on the feedback of the pre-acquired validation set samples after the model is applied, the model weights are dynamically updated. The above training process is iterated until the model performance converges, thus obtaining the random forest classification model.

[0011] Preferably, the cloud data processing platform is also used to organize the output labels of the random forest classification model into a hierarchical decision tree structure. During the training of the random forest classification model, the first-level classifier and multiple second-level classifiers under the first-level classifier are predetermined. The first-level classifier prioritizes the use of globally high-weight features, while the second-level classifier uses locally sensitive features.

[0012] Preferably, the training set samples and validation set samples pre-acquired by the cloud data processing platform both include real-scene samples and virtual-scene samples in terms of sample type. The real-scene samples are collected by temporary monitoring systems pre-deployed in multiple typical substations, collecting multi-dimensional operational data corresponding to the grounding wires, and then labeling the samples before uploading them to the cloud data processing platform. The virtual-scene samples are directly generated from simulation data, and the sensor inputs are randomly disturbed in the simulation environment. The core features of the corresponding multi-dimensional operational data are selected, labeled, and then uploaded to the cloud data processing platform.

[0013] Preferably, the local early warning terminal is also used to receive the grounding wire fault mode and, based on a preset multi-dimensional hierarchical strategy, automatically trigger the corresponding early warning level, including: Level 1 warning: Triggered when the grounding wire fault mode reaches the preset high-risk threshold; Level 2 warning: Triggered when the grounding wire fault mode is between the preset medium-risk threshold and the preset high-risk threshold; Level 3 warning: Triggered when the grounding wire fault mode is at the preset low-risk threshold and the preset medium-risk threshold; The local early warning terminal is also used to adaptively adjust the threshold parameters of preset high-risk threshold, preset medium-risk threshold and preset low-risk threshold based on historical early warning feedback.

[0014] This invention also provides a status monitoring and early warning method based on the aforementioned grounding wire status monitoring and early warning system, comprising the following steps: S1. Obtain training set samples for training the random forest classification model, including real scene samples and virtual scene samples; both real scene samples and virtual scene samples include multi-dimensional running data corresponding to the grounding wire; S2. Train the random forest classification model in the pre-constructed initial state, use the training set samples to fit the model, and the fitting objective is to minimize the classification error to complete the basic training of the model. S3. After basic training, calculate the importance score of each feature corresponding to the model on multi-dimensional running data, convert the feature importance score into weight coefficient, realize the dynamic weight adjustment of the model, and complete the retraining of the model. S4. Obtain multi-dimensional operational data corresponding to the grounding wire during actual operation of the substation, extract features, and input them into the trained random forest classification model. The model outputs the grounding wire fault mode. S5. Based on the output grounding wire fault mode, adopt a graded differentiated early warning method to provide early warning information to the corresponding grounding wire site.

[0015] In summary, due to the adoption of this technical solution, the beneficial effects of this invention are as follows: This invention improves the accuracy and reliability of grounding wire operation status monitoring by integrating multi-source sensor networks and intelligent analysis technology. The system integrates multiple physical field parameters such as conductor temperature, vibration spectrum, and current distribution, effectively overcoming the limitations of traditional single-parameter monitoring and achieving three-dimensional perception of the grounding wire's status throughout its entire lifecycle.

[0016] The cloud-based data processing platform introduces a hierarchical decision tree structure and a dynamic feature weight adjustment mechanism, enabling the random forest classification model to adapt to the feature distribution characteristics of different fault modes. This design not only improves the accuracy of identifying composite faults but also significantly enhances the model's robustness in the strong electromagnetic interference environment of substations, avoiding the problems of high false alarm rate and poor adaptability of traditional threshold alarm methods.

[0017] The local early warning terminal adopts a graded and differentiated early warning mechanism, which intelligently matches visual, auditory and tactile multimodal output strategies according to the severity of the fault, ensuring that on-site personnel can obtain key information in a timely manner under complex working conditions; its advantages also benefit from the accurate identification and output of grounding wire fault modes by the cloud data processing platform.

[0018] The overall architecture of this invention takes into account both data security and transmission reliability, and realizes intelligent operation of the entire process from data acquisition to early warning decision-making, thus promoting the development of power equipment condition monitoring technology towards refinement and precision. Attached Figure Description

[0019] The present invention further illustrates its embodiments and technical solutions with reference to the following figures, specifically including three figures as follows: Figure 1 This is a schematic diagram illustrating the overall composition and functional principle of the grounding wire status monitoring and early warning system of the present invention; Figure 2 This is a schematic diagram of the device composition and data acquisition of the multi-source sensor network in this invention; Figure 3 This is a schematic diagram of the overall process of the grounding wire status monitoring and early warning method of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0022] Example 1 A grounding wire status monitoring and early warning system based on random forest algorithm. Figure 1 The overall composition of the system is shown and can be viewed simultaneously; the system includes a multi-source sensor network, data acquisition terminals, a cloud data processing platform, and local early warning terminals, among which: Deployed at the grounding wires of various electrical equipment in power plants or substations, it is used to collect multi-dimensional operational data corresponding to the grounding wires and transmit it to the data acquisition terminal; The data acquisition terminal is used to extract features from multi-dimensional operational data from a multi-source sensor network, generate a corresponding set of feature vectors, and then upload them to the cloud data processing platform. The cloud-based data processing platform uses a cloud-deployed random forest classification model. Based on the uploaded feature vector set, it dynamically adjusts feature weights through feature importance analysis and constructs a hierarchical decision tree structure to identify and output grounding wire fault modes, which are then sent to local early warning terminals. Local early warning terminals are deployed at the corresponding grounding wires to provide early warning information to the grounding wire site using a graded and differentiated early warning method based on the grounding wire fault mode.

[0023] In this embodiment, as Figure 2 As shown, the multi-dimensional operational data collected by the multi-source sensor network includes conductor temperature, conductor surface thermal radiation, vibration spectrum, mechanical tension, corrosion rate, current distribution, voltage waveform, and ambient temperature and humidity. Among them: conductor temperature is collected by thermocouples fixed to the metal surface of the grounding wire; conductor surface thermal radiation is collected by infrared thermal imagers deployed above the grounding wire; vibration spectrum is collected by accelerometers embedded inside the grounding terminal of the grounding wire; mechanical tension is collected by S-type tension sensors connected in series with the lower section of the grounding wire; corrosion rate is collected by electrochemical potential sensors buried in the soil around the grounding terminal of the grounding wire; current distribution and voltage waveform are both collected by Rogowski coils surrounding the conductor of the grounding wire; and ambient temperature and humidity are collected by temperature and humidity sensors deployed at the grounding wire.

[0024] In this embodiment, the data acquisition terminal has a shielded shell with an optical fiber isolation interface. The optical fiber isolation interface establishes a communication connection with the multi-source sensor network through an optical fiber link. The data acquisition terminal has a built-in radio frequency transceiver unit based on narrowband wireless communication technology. It dynamically avoids the strong electromagnetic interference frequency band of the substation through an adaptive frequency hopping mechanism, realizing wireless communication between the data acquisition terminal and the cloud data processing platform. The data acquisition terminal is also used to preprocess the data after receiving multi-dimensional operating data by combining signal filtering and data compression algorithms, and then extract and generate the corresponding feature vector set.

[0025] In this embodiment, after receiving multi-dimensional operational data from a multi-source sensor network, the data acquisition terminal first performs preprocessing on the raw signals in the data to filter out electromagnetic interference and measurement noise commonly found in substation environments. This process can employ mature filtering and noise reduction techniques in the field, which will not be elaborated further. Subsequently, considering the characteristics of different physical quantities such as conductor temperature and vibration spectrum, in addition to retaining effective signal components using the previously employed adaptive filtering algorithm, compression techniques are also used to reduce data redundancy, thereby reducing the transmission load while ensuring information integrity.

[0026] The data acquisition terminal further performs deep feature mining on the preprocessed data, converting the time-domain waveform into frequency-domain feature parameters and extracting the statistical characteristics and transient change patterns of each signal. In this embodiment, the dominant frequency component and its energy distribution are separated from the vibration signal, and harmonic distortion characteristics and abnormal abrupt changes are identified in the current and voltage waveforms. These features, combined with other data features labeled according to actual needs, can characterize the specific fault mode of the grounding wire. These extracted feature parameters are systematically organized into a structured set of feature vectors. Each vector can accurately characterize the comprehensive operating state of the grounding wire at a specific moment, preserving the key fault information in the original data while eliminating the interference of irrelevant variables. The final set of feature vectors not only reduces the data dimensionality but also highlights the discriminative features closely related to the health status of the grounding wire, thus providing a high-quality analytical foundation for subsequent fault mode identification on the cloud platform and realizing the intelligent transformation from raw data to effective diagnostic information.

[0027] In this embodiment, the cloud data processing platform is also used to pre-build and train the random forest classification model. First, after initializing a preset number of decision trees, the model is fitted using pre-acquired training set samples to complete basic training. After basic training, the importance scores of different features are calculated, so that the feature importance scores are converted into weight coefficients during the model retraining stage. At the same time, the decision tree splitting process is embedded in this stage, that is, when splitting nodes, the feature with the largest weighted information gain is selected first. Then, based on the feedback of the pre-acquired validation set samples after the model is applied, the model weights are dynamically updated. The above training process is iterated until the model performance converges, thus obtaining the random forest classification model.

[0028] In the initialization of the random forest classification model, the number of decision trees is determined through cross-validation; in this embodiment, it is set to 500. Each tree uses Bootstrap sampling and a random feature subset, meaning that only about four features are considered at each split. During basic training, the training set samples are used to fit the model, with the fitting objective being to minimize the classification error. Gini impurity is used as the splitting criterion. After only basic training, it is found that the model achieves high accuracy on the validation set, but its recognition rate for complex faults is low, which does not meet the needs of practical applications. Therefore, a dynamic optimization mechanism needs to be introduced for retraining.

[0029] After basic training, the importance score of each feature is calculated by averaging down the Gini impurity. In most substation environments, the main frequency of the vibration spectrum and the harmonics of the current distribution contribute the most to fault identification, while the scores of environmental humidity and other factors are relatively low.

[0030] During the model retraining phase, this embodiment converts feature importance scores into weight coefficients. Specifically, the weights for high-importance features are set to 1.2, and those for low-importance features to 0.8. Simultaneously, a decision tree splitting process is embedded; when a node splits, the feature with the highest weighted information gain is selected first. Based on feedback from the validation set samples, if the recall rate of a certain fault category is below a threshold, the weights of relevant features are automatically increased. In this embodiment, the threshold can be set to 80%. This process iterates for 3 to 5 rounds to achieve basic model performance convergence. After the weights are adjusted during retraining, the overall robustness of the model to noise is significantly enhanced, and the composite fault recognition rate is improved, reaching the level required for practical deployment.

[0031] In this embodiment, the cloud data processing platform is also used to organize the output labels of the random forest classification model into a hierarchical decision tree structure. During the training of the random forest classification model, the first-level classifier and multiple second-level classifiers under the first-level classifier are predetermined. The first-level classifier preferentially uses global high-weight features, while the second-level classifier uses local sensitive features.

[0032] Hierarchical decision trees are mainly used to address the multi-level characteristics of grounding wire faults, where the first-level fault category encompasses the second-level specific patterns. In this embodiment, the five main fault categories distinguished by the first-level classifier are preset as mechanical, thermal, electrical, corrosive, and environmental. The specific types under the mechanical category can be further subdivided into loosening, breakage, etc., and other categories can be classified according to the actual fault conditions.

[0033] For training hierarchical decision tree models, a hierarchical random forest strategy can be used, building upon the overall model training process. This involves first training the first-level classifier to obtain five output classes, and then specifically training the corresponding second-level classifier for each class. For example, in mechanical engineering, the second-level classifier focuses on vibration and tension features, emphasizing these types of data from multi-dimensional operational data. During training, the first-level classifier prioritizes globally high-weight features, typically represented by voltage waveform data; the second-level classifier, on the other hand, utilizes locally sensitive features, such as weighting data from S-shaped tension sensors in mechanical engineering.

[0034] During training, the tree depth is automatically adjusted based on the sample distribution. More complex, compound faults are represented by deeper trees, while simple faults can be handled with shallower trees, thus avoiding wasted computational resources. This hierarchical structure improves the inference speed of the random forest classification model and supports the need for tiered, differentiated early warning systems. The final trained model can be compressed into lightweight formats such as ONNX, and feature weight parameters are fixed for the substation areas under deployment and management, ensuring that the cloud-based data processing platform can process the feature vector sets uploaded by the data acquisition terminals in real time.

[0035] Training set samples are the foundation of model training. However, even if a substation operates for a long time, the number of effective samples that can be obtained is relatively small. Therefore, as a preferred embodiment, the training set samples and validation set samples pre-acquired by the cloud data processing platform include both real-scene samples and virtual-scene samples in terms of sample type. Real-scene samples are collected by temporary monitoring systems pre-deployed in multiple typical substations, collecting multi-dimensional operating data corresponding to the grounding wires and labeling the samples before uploading them to the cloud data processing platform. Virtual-scene samples are directly generated from simulation data, and the sensor inputs are randomly disturbed in the simulation environment. The core features of the corresponding multi-dimensional operating data are selected and labeled in the same way before being uploaded to the cloud data processing platform.

[0036] In this embodiment, the data source for the real-world scenario samples is a temporary monitoring system deployed in multiple typical substations (covering specific climate zones, voltage levels, and grounding wire types). After the system runs continuously for a specific period, it synchronously records the corresponding multi-dimensional operational data. When labeling the real-world scenario samples, the following labeling scheme is preferred: Normal state samples: Collect steady-state data when the grounding wire is fault-free, and record a set of feature vectors after feature extraction every 5 minutes to accumulate a large number of samples.

[0037] Fault status samples are triggered and recorded in the following two ways: (1) Natural fault events: During the operation and maintenance of substations, real faults are monitored and captured in real time, including line breaks caused by lightning strikes, potential abnormalities caused by soil corrosion, etc. The fault mode is confirmed on-site by professional engineers and marked as a specific category. (2) Controlled small-scale fault injection: Under the premise of ensuring safety, minor faults are artificially simulated in accordance with the safety regulations of the power industry. The load can be adjusted to simulate overheating, and mechanical stress can be applied to simulate abnormal tension. For the simulation of corrosion faults, electrolytes can be added to the soil around the grounding terminal to accelerate electrochemical corrosion, and data from potential sensors and thermal imagers can be recorded simultaneously. For the simulation of vibration faults, a vibrator can be used to apply vibration at a specific frequency, and accelerometer spectrum data can be collected.

[0038] Ultimately, the dataset of real-world scenario samples should contain at least approximately 8,000 fault samples, covering 5 major fault categories, each further subdivided into 3 to 5 sub-patterns. The sample labels are double-verified by professional engineers. When this part of the training samples is collected by the temporary monitoring system, the raw data also undergoes preliminary filtering and compression, feature extraction, and then is uploaded to the training space of the cloud data processing platform. Before feature extraction, missing values ​​in the raw data can be filled using interpolation, and outliers are... The data is eliminated based on principles and then normalized to remove differences in sensor dimensions; the same method can be used to process raw data during actual monitoring and acquisition.

[0039] In this embodiment, virtual scene samples are acquired using simulation tools and related models. Based on the physical characteristics of the grounding wire, a high-fidelity digital twin model is constructed. COMSOL Multiphysics is used to establish a multiphysics coupling simulation, integrating modules of thermodynamics (conductor temperature distribution), structural dynamics (vibration spectrum), electrochemistry (corrosion rate), and electromagnetics (current distribution).

[0040] During simulation, fault scenarios are parameterized, and extreme or rare faults are simulated by adjusting key parameters, covering edge cases that are difficult to obtain in actual data, such as combined faults involving simultaneous vibration and overheating. Further reference can be made to the IEEE Std 80 standard to define a fault mode library and preset multiple fault scenarios, such as conductor breakage-instantaneous and insulation aging-gradual.

[0041] In the simulation environment, random disturbances are simulated from sensor inputs, such as Rogowski coil noise under strong electromagnetic interference, generating a large number of feature vectors. Each feature vector contains eight multi-dimensional operational data points, but core features are selected based on pre-analysis, prioritizing highly sensitive parameters such as vibration spectrum and current distribution, while ignoring redundant data such as fluctuations in ambient temperature and humidity under stable operating conditions. The simulation system automatically generates fault labels based on this, achieving non-manual sample annotation. Noise can also be added to simulate actual interference, characterizing sensor errors to some extent and increasing the generalization ability of the model training.

[0042] By mixing real-world and virtual-world samples at a ratio of 1:5, the deficiency of real-world data is largely compensated for, and a sample set that meets the requirements for model training is finally constructed, with 80% used for training, 10% for validation, and 10% for testing.

[0043] In this embodiment, the local early warning terminal is also used to receive the grounding wire fault mode and, based on a preset multi-dimensional hierarchical strategy, automatically trigger the corresponding early warning level, including: Level 1 warning: Triggered when the grounding wire fault mode reaches the preset high-risk threshold; Level 2 warning: Triggered when the grounding wire fault mode is between the preset medium-risk threshold and the preset high-risk threshold; Level 3 warning: Triggered when the grounding wire fault mode is at the preset low-risk threshold and the preset medium-risk threshold; The local early warning terminal is also used to adaptively adjust the threshold parameters of preset high-risk threshold, preset medium-risk threshold and preset low-risk threshold based on historical early warning feedback.

[0044] The local early warning terminal has a built-in lightweight rule engine, which can be implemented using mature microcontroller technology. It receives fault modes from the cloud, such as corrosion fault - level 3 and vibration loosening - instantaneous output by the random forest classification model, and automatically triggers the corresponding early warning level according to the preset multi-dimensional classification strategy.

[0045] Level 1 warnings indicate emergency faults, such as conductor breakage, and the terminal immediately activates the highest priority response; Level 2 warnings indicate serious faults, such as thermal overload, and the terminal initiates a medium-intensity warning; Level 3 warnings indicate potential faults, such as initial corrosion, and the terminal only provides mild alerts.

[0046] The local early warning terminal automatically fine-tunes the classification thresholds based on historical early warning feedback. For example, if the ambient temperature and humidity at a substation are consistently high, the terminal will lower the early warning threshold for temperature-related faults to avoid frequent false alarms. This mechanism ensures that the classification logic adaptively matches the actual on-site operating conditions.

[0047] As a key device that directly communicates and provides feedback to the grounding wire site when a fault occurs, the local early warning terminal adopts the following preferred method for its differentiated early warning output in this embodiment: For its hardware integration, the terminal adopts a modular output interface, activating differentiated physical signals for different warning levels to ensure that on-site personnel can still reliably perceive warnings in substation environments with strong noise and strong light, including: During a Level 1 warning, the visual solution is a high-brightness RGB LED array that flashes red light at a preset first flashing frequency and simultaneously drives the reflective coating on the terminal casing to enhance visibility; the auditory solution integrates an explosion-proof piezoelectric speaker that outputs a pulse alarm sound above a preset first decibel level, using a preset first audio frequency; the tactile solution is a built-in vibration motor that transmits vibration signals through a grounded metal structure for nearby workers to perceive.

[0048] During a Level 2 warning, the visual solution involves LEDs flashing yellow light at a preset second flashing frequency, along with an LCD screen displaying the fault type; the auditory solution involves a continuous warning tone at a preset second decibel level, using a preset second audio frequency; the volume is adaptively adjusted according to ambient noise; there is no tactile solution.

[0049] During a Level 3 alert, only the visual system is activated by a constantly lit green LED and a brief notification displayed on an LCD screen; there are no auditory or tactile systems.

[0050] Finally, as a preferred embodiment, the multi-source sensor network, data acquisition terminal, and local early warning terminal—these three types of devices deployed at the grounding wire site—all have independent power supply support and are equipped with redundant power supply methods such as lithium batteries. Regarding electromagnetic compatibility, the field equipment enclosure adopts a double-layer shielding structure, and local communication is achieved through fiber optic isolation interfaces, thereby blocking EMI coupling paths. Physical protection employs an IP67-level sealed design, and key components are coated with conformal coating to ensure long-term stable operation in outdoor substations. Under normal circumstances, the local early warning terminal can be in sleep mode to reduce power consumption, but it activates early warning output upon receiving instructions from the cloud, avoiding missing fault windows due to processing delays.

[0051] Example 2 Based on Example 1, this example provides a corresponding grounding wire status monitoring and early warning method based on its grounding wire status monitoring and early warning system. The hardware basis of this method is the system in Example 1. See also... Figure 3 A brief overview of the overall process of this method, in which the application of the random forest classification model is a key feature, including: S1. Obtain training set samples for training the random forest classification model, including real scene samples and virtual scene samples; both real scene samples and virtual scene samples include multi-dimensional running data corresponding to the grounding wire; multi-dimensional running data includes conductor temperature, conductor surface thermal radiation, vibration spectrum, mechanical tensile force, corrosion rate, current distribution, voltage waveform, and ambient temperature and humidity; S2. Train the random forest classification model in the pre-constructed initial state, use the training set samples to fit the model, and the fitting objective is to minimize the classification error to complete the basic training of the model. S3. After basic training, calculate the importance score of each feature corresponding to the model on multi-dimensional running data, convert the feature importance score into weight coefficient, realize the dynamic weight adjustment of the model, and complete the retraining of the model. S4. Obtain multi-dimensional operational data corresponding to the grounding wire during actual operation of the substation, extract features, and input them into the trained random forest classification model. The model outputs the grounding wire fault mode. S5. Based on the output grounding wire fault mode, adopt a graded differentiated early warning method to provide early warning information to the corresponding grounding wire site.

[0052] In step S2 of this embodiment, the initial random forest classification model is a two-level cascaded random forest, including a first-level classifier and multiple second-level classifiers under the first-level classifier; the first-level classifier is used to distinguish 5 types of main faults, including mechanical, thermal, electrical, corrosion and environmental faults; the second-level classifier is used to distinguish the specific sub-mode faults under each main fault.

[0053] In step S2 of this embodiment, during the basic training of the model, the class weights of the first-level classifier are calculated as follows:

[0054] In the formula, Represents the main fault category Weighting coefficients; The total number of samples in the training set; Represents the main fault category in the training set samples The total number of samples; when performing decision tree splitting on the first-level classifier, the node splitting criterion is to reduce the weighted Gini coefficient. The maximization method is as follows:

[0055]

[0056] In the formula, This represents the impurity of the parent node Gini. Represents the sample set of the current node; Represents the child nodes after splitting The sample set; Representative child node Middle main fault category The sample proportion; selection to make Biggest feature Perform the splitting process to complete the decision tree splitting process of the first-level classifier; The second-level classifier only uses the classifier that corresponds to the main fault category. sample subset During training, when splitting the decision tree for the second-level classifier, the node splitting criterion is the standard Gini reduction method, as shown in the following equation:

[0057] In the formula, The impurity of the parent node Gini at the sub-pattern level; Representative child node The Gini impurity is determined; finally, when the random forest classification model in the base training improves by less than 0.5% on the weighted accuracy of the validation set samples for three consecutive rounds, the base training is stopped and completed.

[0058] In step S3 of this embodiment, the formula for calculating the feature importance score of the first-level classifier is as follows:

[0059] In the formula, The total number of decision trees representing the first-level classifiers; Represents the first-level classifier All nodes of the tree; Representative node The number of samples; Representative node Use features The amount of Gini reduction during division; The formula for calculating the feature importance score of the second-level classifier is as follows:

[0060] In the formula, Represents the main fault category in the second-level classifier The corresponding number The nodes of a tree; Represents nodes in the second-level classifier Use features The amount of Gini reduction during division; Then, normalization is performed, as shown in the following formula:

[0061] In the formula, The numbers represent hierarchical levels, with 1 indicating the first level and 2 indicating the second level. The total number of representative features; This represents the normalized importance score, with a value ranging from 0 to 1 in a closed interval. Based on the normalized feature importance scores, the dynamic weight coefficients of the corresponding features are calculated as follows:

[0062] In the formula, This represents the adjustment strength coefficient used to prevent overfitting; in this embodiment, it is preferably 0.8. This represents the average importance of the current level. ; The standard deviation representing the importance of the current level. ; When the first-level classifier is retrained, the weight update formula corresponding to the reduction in weighted Gini coefficient is as follows:

[0063] When the second-level classifier is retrained, the weight update formula corresponding to the reduction in weighted Gini coefficient is as follows:

[0064] In the dynamic iterative mechanism, the main fault category is calculated. The recall rate is as follows:

[0065] In the formula, Represents the correctly identified primary fault category The number of samples; Represents the main fault category that was missed in identification. The number of samples; if If the value is less than 0.8, then the corresponding primary fault category is added. The weights of the relevant features; repeat the retraining process until... The accuracy reached 0.8; after all decision trees corresponding to the first-level and second-level classifiers in the random forest classification model were retrained as described above, the model training was completed, and it was deployed after passing the verification test.

[0066] In step S5 of this embodiment, when a fault occurs, it is necessary to directly communicate and provide feedback to the grounding wire site. The preferred method for the graded and differentiated early warning in this embodiment is as follows: Different physical signals are activated for different warning levels to ensure reliable detection by on-site personnel even in substation environments with strong noise and strong light, including: During a Level 1 warning, the visual solution is a high-brightness RGB LED array that flashes red light at a frequency of 2Hz, and simultaneously drives the reflective coating on the terminal casing to enhance visibility; the auditory solution integrates an explosion-proof piezoelectric speaker that outputs a pulse alarm sound of over 90dB at a frequency of 1kHz to avoid confusion with the background noise of the substation; the tactile solution is a built-in vibration motor that transmits vibration signals through a grounded metal structure for nearby workers to perceive.

[0067] During a Level 2 warning, the visual solution involves an LED flashing yellow light at a frequency of 1Hz, accompanied by a scrolling display of the fault type on an LCD screen; the auditory solution is an 80dB continuous warning tone at a frequency of 500Hz, with the volume adaptively adjusted according to ambient noise; there is no tactile solution.

[0068] During a Level 3 alert, only the visual system is activated by a constantly lit green LED and a brief notification displayed on the LCD screen; no other auditory or tactile systems are used to minimize false alarms.

Claims

1. A grounding wire status monitoring and early warning system based on the random forest algorithm, characterized in that, The system comprises the following components: Multi-source sensor networks are deployed at the grounding wires of various electrical equipment in power plants or substations to collect multi-dimensional operational data corresponding to the grounding wires and transmit them to the data acquisition terminal. The data acquisition terminal is used to extract features from multi-dimensional operational data from a multi-source sensor network, generate corresponding feature vector sets, and then upload them to the cloud data processing platform. The cloud-based data processing platform is used to utilize a cloud-deployed random forest classification model. Based on the uploaded feature vector set, it dynamically adjusts feature weights through feature importance analysis and constructs a hierarchical decision tree structure to identify and output grounding wire fault modes, which are then sent to local early warning terminals. The local early warning terminal is deployed at the grounding wire of the corresponding electrical equipment. It is used to provide early warning information to the grounding wire site according to the grounding wire fault mode and adopt a graded and differentiated early warning method.

2. The grounding wire status monitoring and early warning system according to claim 1, characterized in that: The multi-dimensional operational data collected by the multi-source sensor network includes conductor temperature, conductor surface thermal radiation, vibration spectrum, mechanical tension, corrosion rate, current distribution, voltage waveform, and ambient temperature and humidity. Specifically: conductor temperature is collected by thermocouples fixed to the metal surface of the grounding wire; conductor surface thermal radiation is collected by an infrared thermal imager deployed above the grounding wire; vibration spectrum is collected by an accelerometer embedded inside the grounding terminal of the grounding wire; mechanical tension is collected by an S-shaped tension sensor connected in series with the lower section of the grounding wire; corrosion rate is collected by an electrochemical potential sensor buried in the soil around the grounding terminal of the grounding wire; current distribution and voltage waveform are both collected by Rogowski coils surrounding the conductor of the grounding wire; and ambient temperature and humidity are collected by temperature and humidity sensors deployed at the grounding wire.

3. The grounding wire status monitoring and early warning system according to claim 1, characterized in that: The data acquisition terminal has a shielded shell with an optical fiber isolation interface. The optical fiber isolation interface establishes a communication connection with the multi-source sensor network through an optical fiber link. The data acquisition terminal has a built-in radio frequency transceiver unit based on narrowband wireless communication technology. It dynamically avoids the strong electromagnetic interference frequency band of the substation through an adaptive frequency hopping mechanism, realizing wireless communication between the data acquisition terminal and the cloud data processing platform. The data acquisition terminal is also used to preprocess the data after receiving multi-dimensional operating data by combining signal filtering and data compression algorithms, and then extract and generate the corresponding feature vector set.

4. The grounding wire status monitoring and early warning system according to claim 1, characterized in that: The cloud-based data processing platform is also used to pre-build and train the random forest classification model; after initializing a preset number of decision trees, the model is fitted using pre-acquired training set samples to complete basic training. After basic training, importance scores for different features are calculated, and these scores are converted into weight coefficients during the model retraining phase. This phase also incorporates the decision tree splitting process, where the feature with the largest weighted information gain is selected first when splitting nodes. Subsequently, the model weights are dynamically updated based on the feedback from the pre-acquired validation set samples after model application. The above training process is iterated until the model performance converges, thus obtaining the random forest classification model.

5. The grounding wire status monitoring and early warning system according to claim 4, characterized in that: The cloud-based data processing platform is also used to organize the output labels of the random forest classification model into a hierarchical decision tree structure. During the training of the random forest classification model, the first-level classifier and multiple second-level classifiers under the first-level classifier are predetermined. The first-level classifier prioritizes the use of globally high-weight features, while the second-level classifier uses locally sensitive features.

6. The grounding wire status monitoring and early warning system according to claim 4, characterized in that: The training set samples and validation set samples pre-acquired by the cloud data processing platform both include real scene samples and virtual scene samples in terms of sample type; Real-world scenario samples are collected from temporary monitoring systems pre-deployed in multiple typical substations. Multi-dimensional operational data corresponding to the grounding wires are collected, labeled, and then uploaded to the cloud data processing platform. Virtual scenario samples are directly generated from simulation data. Sensor inputs are randomly disturbed in the simulation environment. The core features of the corresponding multi-dimensional operational data are selected, labeled, and then uploaded to the cloud data processing platform.

7. The grounding wire status monitoring and early warning system according to claim 1, characterized in that: The local early warning terminal is also used to receive the grounding fault mode notification and, based on a preset multi-dimensional hierarchical strategy, automatically trigger the corresponding early warning level, including: Level 1 warning: Triggered when the grounding wire fault mode reaches the preset high-risk threshold; Level 2 warning: Triggered when the grounding wire fault mode is between the preset medium-risk threshold and the preset high-risk threshold; Level 3 warning: Triggered when the grounding wire fault mode is at the preset low-risk threshold and the preset medium-risk threshold; The local early warning terminal is also used to adaptively adjust the threshold parameters of preset high-risk threshold, preset medium-risk threshold and preset low-risk threshold based on historical early warning feedback.

8. A method for monitoring and warning the status of a grounding wire status monitoring and warning system according to any one of claims 1 to 7, characterized in that, Includes the following steps: S1. Obtain training set samples for training the random forest classification model, including real scene samples and virtual scene samples; both real scene samples and virtual scene samples include multi-dimensional running data corresponding to the grounding wire; S2. Train the random forest classification model in the pre-constructed initial state, use the training set samples to fit the model, and the fitting objective is to minimize the classification error to complete the basic training of the model. S3. After basic training, calculate the importance score of each feature corresponding to the model on multi-dimensional running data, convert the feature importance score into weight coefficient, realize the dynamic weight adjustment of the model, and complete the retraining of the model. S4. Obtain multi-dimensional operational data corresponding to the grounding wire during actual operation of the substation, extract features, and input them into the trained random forest classification model. The model outputs the grounding wire fault mode. S5. Based on the output grounding wire fault mode, adopt a graded differentiated early warning method to provide early warning information to the corresponding grounding wire site.

9. The status monitoring and early warning method according to claim 8, characterized in that: In step S1, the multi-dimensional operating data includes conductor temperature, conductor surface thermal radiation, vibration spectrum, mechanical tensile force, corrosion rate, current distribution, voltage waveform, and ambient temperature and humidity.

10. The status monitoring and early warning method according to claim 8, characterized in that: In step S2, the initial random forest classification model is a two-level cascaded random forest, including a first-level classifier and multiple second-level classifiers under the first-level classifier; the first-level classifier is used to distinguish 5 types of main faults, including mechanical, thermal, electrical, corrosion and environmental faults; the second-level classifier is used to distinguish the specific sub-mode faults under each main fault. In step S2, during the basic training of the model, the class weights of the first-level classifier are calculated as follows: In the formula, Represents the main fault category Weighting coefficients; The total number of samples in the training set; Represents the main fault category in the training set samples The total number of samples; when performing decision tree splitting on the first-level classifier, the node splitting criterion is to reduce the weighted Gini coefficient. The maximization method is as follows: In the formula, This represents the impurity of the parent node Gini. Represents the sample set of the current node; Represents the child nodes after splitting The sample set; Representative child node Middle main fault category The sample proportion; selection to make Biggest feature Perform the splitting process to complete the decision tree splitting process of the first-level classifier; The second-level classifier only uses the classifier that corresponds to the main fault category. sample subset During training, when splitting the decision tree for the second-level classifier, the node splitting criterion is the standard Gini reduction method, as shown in the following equation: In the formula, The impurity of the parent node Gini at the sub-pattern level; Representative child node The Gini impurity is determined; finally, when the improvement in the weighted accuracy of the random forest classification model on the validation set samples for a consecutive preset number of rounds is less than a threshold, the basic training is stopped and completed.

11. The status monitoring and early warning method according to claim 10, characterized in that: In step S3, the feature importance score of the first-level classifier is calculated as follows: In the formula, The total number of decision trees representing the first-level classifiers; Represents the first-level classifier All nodes of the tree; Representative node The number of samples; Representative node Use features The amount of Gini reduction during division; The formula for calculating the feature importance score of the second-level classifier is as follows: In the formula, Represents the main fault category in the second-level classifier The corresponding number The nodes of a tree; Represents nodes in the second-level classifier Use features The amount of Gini reduction during division; Then, normalization is performed, as shown in the following formula: In the formula, The numbers represent hierarchical levels, with 1 indicating the first level and 2 indicating the second level. The total number of representative features; This represents the normalized importance score, with a value ranging from 0 to 1 in a closed interval. Based on the normalized feature importance scores, the dynamic weight coefficients of the corresponding features are calculated as follows: In the formula, This represents the adjustment strength coefficient used to prevent overfitting; This represents the average importance of the current level; The standard deviation represents the importance of the current level; When the first-level classifier is retrained ; When the second-level classifier is retrained ; In the dynamic iterative mechanism, the main fault category is calculated. The recall rate is as follows: In the formula, Represents the correctly identified primary fault category The number of samples; Represents the main fault category that was missed in identification. The number of samples; if If the recall rate is less than the preset recall threshold, the corresponding primary fault category will be added. The weights of the relevant features; Repeat the retraining process until Once the preset recall threshold is reached; after all decision trees corresponding to the first-level and second-level classifiers in the random forest classification model have completed the above retraining, the model training is complete, and it is deployed after passing the verification test.

12. The status monitoring and early warning method according to claim 8, characterized in that, In step S5, the tiered differentiated early warning method specifically involves: using a modular output interface to activate differentiated physical signals for different early warning levels, including: During a Level 1 warning, the visual solution is a high-brightness RGB LED array that flashes red light at a preset first flashing frequency and simultaneously drives the reflective coating on the terminal casing to enhance visibility; the auditory solution integrates an explosion-proof piezoelectric speaker that outputs a pulse alarm sound above a preset first decibel level, using a preset first audio frequency; the tactile solution is a built-in vibration motor that transmits vibration signals through a grounded metal structure for nearby workers to perceive. During a Level 2 warning, the visual solution involves LEDs flashing yellow light at a preset second flashing frequency, combined with an LCD screen displaying the fault type; the auditory solution involves a continuous warning tone at a preset second decibel level, using a preset second audio frequency; the volume adaptively adjusts according to ambient noise; there is no tactile solution. During a Level 3 alert, only the visual system is activated by a constantly lit green LED and a brief notification displayed on an LCD screen; there are no auditory or tactile systems.

Images