A high-low voltage switch cabinet early warning method and system

Abstract

This invention discloses an early warning method and system for high and low voltage switchgear, belonging to the field of power system automation. The method includes: constructing a multi-dimensional sensing network to collect multi-source heterogeneous data such as temperature rise, partial discharge, mechanical characteristics, and environmental conditions in real time; preprocessing the data and aligning it with spatiotemporal features, and dynamically correcting the sensitivity threshold of the partial discharge criterion based on humidity or pressure change rate; extracting multi-dimensional fault feature vectors and dynamically assigning weights to each component according to operating conditions and environmental conditions; inputting the weighted feature vectors into a long short-term memory network integrating an attention mechanism to generate a multi-timescale predicted trajectory of the operating state; comparing the predicted trajectory with preset multi-level state thresholds in real time, and executing a three-level early warning judgment (attention, abnormal, and severe) and corresponding linkage protection strategies. This invention can reduce the risk of false alarms and missed alarms, providing reliable support for the intelligent operation and maintenance of high and low voltage switchgear.

Description

Technical Field

[0001] This application belongs to the field of power system automation, specifically relating to a method and system for early warning of high and low voltage switchgear. Background Technology

[0002] With the rapid evolution of the power industry and smart grid technology, high- and low-voltage switchgear, as core infrastructure equipment in the distribution network that performs circuit switching, protection control, and power distribution, directly affects the reliability of the entire power system. Traditional manual periodic inspections and offline testing methods are no longer sufficient to meet the demands of modern power grids for real-time monitoring of equipment status. Deploying various sensors to achieve automated acquisition and monitoring of switchgear operating parameters has become an important technological path to improve the automation and intelligence level of power distribution systems.

[0003] Among them, the early warning method for high and low voltage switchgear is a key link in ensuring the full life cycle management of equipment. It aims to identify potential operational risks and trigger alarm signals in advance through in-depth analysis of collected multi-dimensional operational data. This technology mainly covers real-time tracking of key indicators such as abnormal temperature rise, partial discharge, changes in mechanical characteristics, and insulation performance degradation. By constructing an early warning algorithm model, it assesses the health status of the equipment, thereby providing decision support for condition-based maintenance and preventative maintenance.

[0004] However, existing early warning strategies largely rely on fixed threshold comparisons of single physical quantities, making it difficult to handle complex fault feature identification under the coupling of multiple parameters, resulting in insufficient accuracy of early warning results. Simultaneously, the operating environment of switchgear is subject to strong electromagnetic interference and complex, variable load conditions. Traditional data processing methods lack the ability to deeply integrate and analyze multi-source heterogeneous monitoring data, easily leading to false alarms or missed alarms in interference environments. Furthermore, existing early warning models often focus on static correlation analysis, failing to effectively capture nonlinear dynamic trend changes during equipment aging or abnormal deterioration, resulting in a significant lag in triggering early warning signals, making it difficult to meet the stringent forward-looking requirements of preventative maintenance for power equipment. Therefore, a new early warning scheme for high and low voltage switchgear is desired. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for early warning of high and low voltage switchgear, which can effectively solve the problems in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for early warning of high and low voltage switchgear includes the following specific steps: A multi-dimensional sensing network is constructed to collect multi-source heterogeneous data in real time during the operation of high and low voltage switchgear; the multi-source heterogeneous data includes temperature rise data, partial discharge data, mechanical characteristic data and environmental condition data; The collected multi-source heterogeneous monitoring data is preprocessed and aligned with spatiotemporal features. The preprocessing and spatiotemporal feature alignment include: using synchronization instructions issued based on a precise time protocol to make the synchronization error of each monitoring unit less than 1 microsecond, and using a reference time axis as the step size, using third-order Hermite interpolation to align the time dimension of monitoring data with sampling frequency lower than the reference time axis. A humidity correction coefficient is calculated based on the real-time relative humidity value in the environmental state data, and a pressure change correction coefficient is calculated based on the pressure change rate in the environmental state data. The calculated humidity correction coefficient or the pressure change correction coefficient is used to dynamically correct the sensitivity threshold of the partial discharge criterion. A multidimensional fault feature vector is extracted from the multi-source heterogeneous data after spatiotemporal feature alignment and dynamic correction, and the weights of each component in the multidimensional fault feature vector are dynamically allocated according to the current operating conditions and environmental status of the switchgear. The multi-dimensional fault feature vector after dynamic weight allocation is input into a long short-term memory network with integrated attention mechanism to generate a multi-timescale prediction trajectory of the switchgear's operating status. The predicted trajectory of the operating state is compared with the preset multi-level state thresholds in real time, and multi-level early warning judgment and corresponding linkage protection strategy are executed.

[0007] Furthermore, the step of calculating a humidity correction coefficient based on the real-time relative humidity value in the environmental state data and using the calculated humidity correction coefficient to dynamically correct the sensitivity threshold of the partial discharge criterion specifically includes: when the real-time collected environmental relative humidity value exceeds the preset humidity threshold, humidity correction is enabled; The humidity correction factor is calculated as follows: the humidity correction factor is equal to 1 plus the product of the difference between the real-time relative humidity value and the preset humidity threshold divided by 100 and then multiplied by 0.2. Multiply the original sensitivity threshold of the partial discharge criterion by the humidity correction coefficient to obtain the corrected sensitivity threshold of the partial discharge criterion.

[0008] Furthermore, the step of calculating a pressure change correction coefficient based on the pressure change rate in the environmental state data, and then dynamically correcting the sensitivity threshold of the partial discharge criterion using the calculated pressure change correction coefficient, specifically includes: When the real-time calculated pressure change rate exceeds the preset pressure change rate abnormal threshold, pressure correction is activated; the calculation formula for the pressure change correction coefficient is configured as follows: the pressure change correction coefficient is equal to 1 plus the product of the difference between the absolute value of the pressure change rate and the pressure change rate abnormal threshold, divided by the pressure change rate abnormal threshold and multiplied by 0.1. The original sensitivity threshold of the partial discharge criterion is multiplied by the pressure change correction coefficient to obtain the corrected sensitivity threshold of the partial discharge criterion.

[0009] Furthermore, the preprocessing and spatiotemporal feature alignment steps for the collected multi-source heterogeneous monitoring data also include noise reduction processing of the partial discharge signal: Daubechies wavelet was selected as the basis function to perform 5 to 8-level wavelet packet decomposition on the original partial discharge signal. At each decomposition scale, nonnegative matrix decomposition theory is used to extract background thermal noise features and environmental narrowband interference features to distinguish effective discharge pulses from background interference. For each decomposition scale, a threshold proportional to the noise level of that scale is set, and wavelet coefficients with absolute values ​​below the threshold are set to zero, while discharge pulse coefficients with abrupt change characteristics are retained. Wavelet reconstruction is performed on the retained wavelet coefficients to obtain the denoised partial discharge signal.

[0010] Furthermore, the steps for extracting multidimensional fault feature vectors from multi-source heterogeneous data after spatiotemporal feature alignment and dynamic correction specifically include: Pearson correlation analysis and mutual information theory are used to identify the correlation between various operating parameters and equipment health status, and to screen the core operating parameters that are sensitive to equipment health status. Based on the screening results, a multidimensional fault feature vector is extracted; the multidimensional fault feature vector specifically includes the following 5 components: temperature rise rate, average partial discharge quantity, discharge phase distribution characteristics, mechanical opening and closing speed fluctuation rate, and intrinsic mode energy ratio of vibration signal; The temperature rise rate is obtained by calculating the first and second order differences of the temperature rise sequence; the intrinsic mode energy ratio of the vibration signal is obtained by using variational mode decomposition technology to decompose the original vibration signal into several intrinsic mode functions with independent center frequencies, and calculating the energy distribution within a specific frequency band.

[0011] Furthermore, the step of dynamically allocating weights to each component of the multidimensional fault feature vector based on the current operating conditions of the switchgear and the environmental conditions specifically includes: By incorporating expert experience through the analytic hierarchy process, the importance of different parameters is subjectively assessed, forming a judgment matrix. At the same time, using the information entropy theory, the objective weights of each feature are calculated based on the degree of dispersion of the real-time collected data. The system weights are determined by combining the subjective assessments and the objective weights. The system weights can be dynamically adjusted according to the current operating conditions. Specifically, when the switchgear is under high load conditions of more than 80% of the rated load, the weight coefficient of the temperature rise characteristic is increased from the initial value of 0.25 to 0.55; when the ambient humidity monitoring value exceeds 85%, the combined weight coefficient of the partial discharge characteristic and the environmental characteristic is increased by more than 30% simultaneously.

[0012] Furthermore, the specific steps of inputting the dynamically weighted multidimensional fault feature vector into the long short-term memory network with integrated attention mechanism include: The weighted multidimensional feature vector sequence is input into the network in chronological order. The network contains an input layer, at least three hidden layers, and an output layer. Each hidden layer is configured with 128 to 512 neuron units. A temporal attention mechanism is introduced on the basis of the standard long short-term memory network. By calculating the correlation score between the current hidden state and each moment in the historical state sequence, higher weights are assigned to key mutation moments. The transfer learning mechanism is used to fine-tune the parameters of newly commissioned equipment: First, the model parameters are pre-trained using typical fault life cycle degradation data obtained from similar equipment in a simulated laboratory environment. After the model is deployed to the current inspected switchgear, the real-time operating data of the first 3 months after the equipment is put into operation is used to fine-tune the parameters of the last few layers of the pre-trained model.

[0013] Furthermore, the specific steps for implementing multi-level early warning judgment and corresponding linkage protection strategies include: The multi-level early warning judgment is divided into three levels: attention, abnormal, and severe. When a warning level alert is triggered, the alert information will be pushed to the mobile terminal of the maintenance personnel, and the alert record will be simultaneously saved in the device's electronic health record. When an abnormal level warning is triggered, the forced air cooling system built into the switchgear is activated, and the integrated automation backend is used to suggest to the dispatching system to reduce the load distribution of that branch. When a severe warning is triggered, a tripping command is sent directly to the logic controller of the relay protection system within a millisecond response time. The command is transmitted via hardwire or GOOSE network with a delay of no more than 10 milliseconds.

[0014] Furthermore, the triggering conditions for a level warning are: the predicted value of a certain operating parameter exceeds 110% but is less than 120% of its rated operating parameter, or the partial discharge pulse frequency shows an increasing trend of less than 5%. The triggering conditions for the abnormal level warning are: the temperature rise rate of the predicted parameters increases exponentially, that is, the first-order difference sequence of temperature rise shows an increasing trend and the second-order difference is positive for more than 3 sampling periods, or the partial discharge intensity continues to exceed 5 times the background noise. The triggering condition for the severe level warning is: the prediction parameters indicate that the insulation breakdown voltage or thermal runaway critical temperature will be triggered within the next 1 hour time threshold.

[0015] A high- and low-voltage switchgear early warning system includes: A multi-dimensional sensing network is deployed inside high and low voltage switchgear to collect multi-source heterogeneous data during the operation of the switchgear. The multi-source heterogeneous data includes temperature rise data, partial discharge data, mechanical characteristic data, and environmental condition data. An edge computing gateway is used to realize time synchronization and preprocessing of multi-sensor data; the edge computing gateway acts as a time reference source and sends synchronization commands to each monitoring unit through a precise time protocol, so that the synchronization error of each monitoring unit is less than 1 microsecond. The core processing platform performs the following operations: preprocessing and aligning the collected multi-source heterogeneous monitoring data with spatiotemporal features; calculating a humidity correction coefficient based on the real-time relative humidity value in the environmental state data, calculating a pressure change correction coefficient based on the pressure change rate in the environmental state data, and dynamically correcting the sensitivity threshold of the partial discharge criterion using the calculated humidity correction coefficient or the pressure change correction coefficient to obtain the corrected sensitivity threshold of the partial discharge criterion; then, extracting the multi-dimensional fault feature vector from the multi-source heterogeneous data after spatiotemporal feature alignment and dynamic correction, and dynamically assigning weights to each component of the multi-dimensional fault feature vector according to the current operating condition of the switchgear and the environmental state; inputting the dynamically weighted multi-dimensional fault feature vector into a long short-term memory network with an integrated attention mechanism to generate a multi-timescale predicted trajectory of the switchgear's operating state; and comparing the predicted trajectory of the operating state with preset multi-level state thresholds in real time to perform multi-level early warning judgment. The relay protection linkage system is used to receive early warning commands from the core processing platform and execute corresponding protection control strategies.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. By constructing a multi-dimensional sensing network encompassing temperature rise, partial discharge, mechanical characteristics, and environmental factors, this invention overcomes the limitations of traditional technologies that rely on single threshold judgments, enabling comprehensive identification of complex fault characteristics. Utilizing Pearson correlation analysis and dynamic weight allocation technology, this invention can effectively distinguish between normal load fluctuations and early latent faults, significantly improving early warning accuracy compared to traditional methods and substantially reducing false alarm and false negative rates under strong electromagnetic interference environments.

[0017] 2. Employing a long short-term memory network model with an integrated attention mechanism, this invention can capture the nonlinear dynamic trends in equipment operating parameters, achieving a leap from static correlation analysis to dynamic trend prediction. By outputting predicted trajectories within a preset time scale, it provides sufficient decision-making buffer time for condition-based maintenance and preventive maintenance of the power system, meeting the forward-looking and stringent requirements of modern power grids for safe operation.

[0018] 3. By utilizing an improved wavelet packet transform and adaptive filtering algorithm, this invention can accurately extract weak fault feature signals from strong background noise. Through high-frequency synchronous acquisition and spatiotemporal feature alignment technology, it ensures the logical consistency of monitoring data from different sources and physical quantities, resolving data conflicts and phase drift issues during multi-source heterogeneous data fusion, and guaranteeing stable system operation under complex and variable conditions.

[0019] 4. Through multi-level early warning judgment and automatic linkage protection strategies, this invention constructs a complete technical closed loop from state perception to risk identification and then to automatic intervention. Combining digital twin visualization technology and edge computing architecture, it not only improves the intuitiveness and accuracy of fault location, but also significantly shortens emergency response time, providing comprehensive intelligent support for the reliable operation of high and low voltage switchgear. Attached Figure Description

[0020] Figure 1 A schematic diagram of the overall scheme for early warning methods for high and low voltage switchgear; Figure 2 This is a schematic diagram illustrating the core principle of a dynamic trend prediction model based on an integrated attention mechanism long short-term memory network. Figure 3 A flowchart illustrating the logic of multidimensional fault feature vector extraction and dynamic weight allocation; Figure 4 This is a flowchart illustrating the logic of a multi-level early warning judgment and linkage protection control strategy. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1-4 The present invention will be further described in detail with reference to specific embodiments.

[0022] In the field of power system automation and smart grid technology, high- and low-voltage switchgear, as core basic equipment in distribution networks for performing circuit switching, protection control, and power distribution, directly affects the reliability of the entire power system due to its operational stability and security. Traditional manual periodic inspections and offline detection methods are no longer adequate for the real-time sensing of equipment status required by modern power grids. Existing early warning strategies mostly rely on fixed threshold comparisons of single physical quantities, which are insufficient to handle the identification of complex fault characteristics under the coupling of multiple parameters, resulting in inaccurate early warning results. Therefore, this application provides a high- and low-voltage switchgear early warning method and system to solve the above-mentioned technical problems.

[0023] Firstly, the high and low voltage switchgear early warning method provided in this application embodiment can be applied to application environments that include multi-dimensional sensing networks, edge computing gateways, core processing platforms, and relay protection linkage systems.

[0024] Among them, the multi-dimensional sensing network is deployed inside the high and low voltage switchgear to collect multi-source heterogeneous data during the operation of the switchgear; the edge computing gateway is used to realize the time synchronization and preprocessing of multi-sensor data; the core processing platform is used to execute feature extraction, model inference and early warning judgment logic; and the relay protection linkage system is used to receive early warning instructions and execute corresponding protection control strategies.

[0025] The core processing platform can be an independent industrial control server, a server cluster or distributed system consisting of multiple physical servers, or a dedicated cloud server for power that provides cloud computing services; the edge computing gateway can communicate with the core processing platform, and the multi-dimensional sensing network can be connected to the edge computing gateway through industrial Ethernet or a dedicated wireless communication link.

[0026] In an exemplary embodiment, a method for early warning of high and low voltage switchgear is provided. Taking the application of this method to the aforementioned core processing platform as an example, the method includes the following steps: S100. Construct a multi-dimensional sensing network and collect operational data of high and low voltage switchgear in real time. The operational data is multi-source heterogeneous data, including temperature rise data, partial discharge data, mechanical characteristic data, and environmental condition data; the multi-dimensional sensing network consists of a temperature rise monitoring unit, a partial discharge monitoring unit, a mechanical characteristic monitoring unit, and an environmental condition monitoring unit.

[0027] S200: Preprocess and align the acquired multi-source heterogeneous monitoring data with spatiotemporal features. Preprocessing and spatiotemporal feature alignment include noise reduction for partial discharge signals, filtering and correction for temperature rise data, and time dimension alignment for multi-sensor asynchronous sampling.

[0028] S300. Extract multi-dimensional fault feature vectors and dynamically assign weights. The multi-dimensional fault feature vectors contain core feature components that are sensitive to the evolution of switchgear faults, and the weights can be dynamically and adaptively adjusted according to the current operating conditions of the switchgear and the environmental conditions.

[0029] S400. Construct a deep learning-based dynamic trend prediction model to generate multi-timescale prediction trajectories of the switchgear's operating status. The dynamic trend prediction model is a long short-term memory network integrating an attention mechanism, which can output the predicted trajectory of the switchgear's operating status within a preset time scale.

[0030] S500 compares the generated predicted operating trajectory with preset multi-level status thresholds in real time, and executes multi-level early warning judgments and corresponding linkage protection strategies. The multi-level early warning judgments are divided into three levels: attention, abnormal, and severe, with each level corresponding to differentiated early warning actions and linkage protection strategies.

[0031] In an exemplary embodiment, the detailed process of constructing a multi-dimensional sensing network and collecting operational data of high and low voltage switchgear in S100 is explained, specifically including: S101. Deploy a multi-dimensional sensing network. Systematically deploy a multi-dimensional sensing network at the junction of each key topology node and the physical structure inside the switchgear. This sensing network consists of a temperature rise monitoring unit, a partial discharge monitoring unit, a mechanical characteristic monitoring unit, and an environmental condition monitoring unit.

[0032] Each monitoring unit is connected to the core processing platform via an industrial Ethernet or dedicated wireless communication link to achieve synchronous high-frequency acquisition of multi-source heterogeneous monitoring data.

[0033] S102. Implement temperature rise monitoring to obtain temperature rise-current coupling data sets of the switchgear. The temperature rise monitoring unit includes passive wireless temperature sensors deployed at circuit breaker contacts, busbar connections, and cable outlets. These sensors employ surface acoustic wave (SAW) technology, operating by receiving radio frequency signals emitted by a reader and feeding back reflected waves carrying temperature information.

[0034] The sensor operates at a preset center frequency, which is in the 433MHz or 2.4GHz industrial, scientific, and medical frequency band. Signal transmission is achieved through frequency-modulated continuous wave or phase modulation. The temperature sampling frequency is set to once every 500 milliseconds to 2 seconds, and the temperature measurement range covers -40 degrees Celsius to 150 degrees Celsius, with the measurement error controlled within ±1 degree Celsius.

[0035] Under high current operation, the abnormal temperature rise caused by the deterioration of contact resistance is tracked in real time. The original temperature value and the corresponding effective current value are associated and stored through a synchronous acquisition card to form a temperature rise-current coupled dataset.

[0036] S103. Implement partial discharge monitoring to acquire high-frequency partial discharge signals from the switchgear. The partial discharge monitoring unit consists of a built-in ultra-high frequency sensor and an external ultrasonic sensor. The ultra-high frequency sensor is installed on the inner wall of the switchgear or near the observation window. Its detection bandwidth is in the range of 300MHz to 1500MHz, and its sensitivity reaches above -70dBm. It is used to capture electromagnetic wave signals generated by early discharges inside the insulation due to air gaps, surface discharges, or floating potentials.

[0037] The ultrasonic sensor's detection center frequency is set to 40kHz±2kHz. It converts mechanical vibrations into electrical signals through the resonant effect of piezoelectric ceramic crystals, used to precisely locate the sound source position generated by surface discharge or the movement of metal particles. The ultra-high frequency sensor and the external ultrasonic sensor achieve sub-microsecond synchronous sampling through high-frequency synchronization pulses, ensuring a traceable causal relationship between the electrical and acoustic signals in the time domain.

[0038] S104. Implement mechanical characteristic monitoring to obtain mechanical motion characteristic data of the switchgear operating mechanism. The mechanical characteristic monitoring unit is implemented by installing a high-resolution rotary encoder on the transmission shaft of the operating mechanism and connecting a Hall current sensor in series in the opening and closing coil circuit; the resolution of the rotary encoder is not less than 1024 pulses per revolution, which is used to provide real-time feedback of the stroke-time curve of the moving contact; the current sensor captures the excitation current waveform of the opening and closing coil in real time during the transient operation.

[0039] The sampling rate for data acquisition is set to 50kHz to 100kHz to accurately analyze the characteristic points of opening and closing time, average contact speed, overtravel, rebound amplitude, and coil current waveform. Through these analyses, the mechanical performance degradation caused by mechanism jamming, fatigue aging of operating springs, or excessive residual magnetism of electromagnet core can be identified.

[0040] S105. Implement environmental condition monitoring to acquire environmental parameters of the switchgear and correct the sensitivity threshold of the partial discharge criterion based on the environmental parameters. The environmental condition monitoring unit includes an integrated temperature and humidity sensor and a pressure micromechanical sensor; the temperature and humidity sensor is deployed at the air inlet of the switchgear and the instrument room to collect the temperature and humidity gradient inside and outside the cabinet in real time; the pressure sensor is used to monitor the pressure fluctuation of the gas chamber in the SF6 gas-insulated switchgear.

[0041] When the humidity exceeds the preset humidity threshold of 75%, the system automatically corrects the sensitivity threshold of the partial discharge criterion according to the following method: the system has a built-in humidity correction function, which multiplies the original alarm threshold of partial discharge detection by a humidity correction coefficient.

[0042] The formula for calculating the humidity correction factor is: in, This is a humidity correction factor, dimensionless. The ambient relative humidity value is collected in real time, in percentage, and ranges from 75 to 100; the corrected partial discharge alarm threshold is the original alarm threshold multiplied by... When the humidity is 75%, the correction factor is 1, and the threshold remains unchanged. When the humidity rises to 100%, the correction factor is 1.05, and the threshold is increased by 5%, thus reducing sensitivity to avoid false alarms caused by excessive humidity. When the pressure change rate exceeds the abnormal threshold of 3% per minute, the system automatically corrects the sensitivity threshold of the partial discharge criterion according to the following method: the system multiplies the original alarm threshold for partial discharge detection by a pressure change correction factor.

[0043] The formula for calculating the pressure change correction factor is: in, This is a dimensionless correction factor for pressure changes. The pressure change rate is calculated in real time, and the unit is percentage per minute; Take its absolute value; this formula is only activated when the pressure change rate exceeds 3% per minute; when the pressure change rate is 3% per minute, the correction factor is 1; when the pressure change rate is 6% per minute, the correction factor is 1.1, the threshold is increased by 10%, that is, the sensitivity is reduced to avoid false alarms caused by pressure fluctuations.

[0044] If both humidity and pressure change rates exceed the threshold, the larger of the two correction coefficients is taken as the final sensitivity threshold adjustment factor.

[0045] In this embodiment, the above steps are used to achieve comprehensive acquisition of key operating parameters of high and low voltage switchgear. The entire acquisition scheme covers four types of physical quantities: temperature rise, partial discharge, mechanical characteristics and environmental conditions. The specific selection, installation location, sampling parameters and signal synchronization methods of various sensors are clearly defined.

[0046] Among them, temperature rise and current data are associated and stored through a synchronous acquisition card, partial discharge signals are ensured by sub-microsecond synchronous sampling of ultra-high frequency and ultrasonic sensors to ensure time-domain causal relationship, and mechanical characteristic monitoring is achieved by acquiring fine waveforms of the action process through high-resolution encoders and Hall current sensors.

[0047] The environmental monitoring unit not only collects temperature, humidity, and pressure data, but also provides a partial discharge sensitivity threshold correction mechanism based on the rate of change of humidity and pressure. The multi-source heterogeneous data obtained through this step are synchronized in time and cover the main fault-sensitive areas inside the switch cabinet in space, thus providing a complete raw data foundation for subsequent data processing.

[0048] In an exemplary embodiment, the detailed process of preprocessing and aligning the collected multi-source heterogeneous monitoring data with spatiotemporal features in S200 is explained, specifically including: S201. Perform wavelet packet transform denoising on the partial discharge signal to improve the signal-to-noise ratio of the partial discharge signal. The original high-frequency signal collected by the partial discharge monitoring unit has a sampling rate of not less than 100MHz. This signal first enters the preprocessing module of the core processing platform and is denoised using an improved wavelet packet transform algorithm.

[0049] The algorithm uses Daubechies wavelet or Symlets wavelet as basis functions to perform wavelet packet decomposition of the original signal at 5 to 8 levels. At each decomposition scale, nonnegative matrix factorization (NMF) theory is used to extract background thermal noise features and environmental narrowband interference features. NMF decomposition decomposes the wavelet packet coefficient matrix into the product of two nonnegative matrices, one of which represents the noise basis vector and the other represents the corresponding activation coefficient. By identifying and separating the noise components, the system can distinguish between effective discharge pulses and background interference.

[0050] Then, adaptive soft threshold filtering is implemented: for each decomposition scale, a threshold proportional to the noise level of that scale is set, wavelet coefficients with absolute values ​​below the threshold are set to zero, while discharge pulse coefficients with abrupt change characteristics are retained.

[0051] After the above processing, wavelet reconstruction is performed on the retained wavelet coefficients to obtain the denoised partial discharge signal. This method can improve the signal-to-noise ratio by more than 15dB.

[0052] S202. Adaptive moving average filtering is applied to the temperature rise data to highlight the true temperature rise trend. For the temperature data collected by the temperature rise monitoring unit, an adaptive filtering algorithm based on moving average is used to remove random fluctuation interference; the width of the moving average time window is set to 32 to 64 sampling points, and the window width can be automatically adjusted according to the degree of data fluctuation: when the standard deviation of the temperature data exceeds 1.5 times the historical average, the window width is increased towards 64 sampling points to enhance the smoothing effect.

[0053] Conversely, the sampling points are reduced towards 32 points to preserve detailed changes. A temperature compensation factor is introduced into the algorithm. This factor linearly corrects the measured values ​​based on the background temperature fluctuations obtained in real-time from the environmental condition monitoring unit. The correction formula combines the thermal constant and thermal resistance coefficient of the switchgear, specifically: Corrected temperature = Original temperature - Ambient temperature × Thermal constant / Thermal resistance coefficient × Sampling time interval.

[0054] Among them, the thermal capacity constant is the thermal capacity parameter of the main material of the switch cabinet, and the unit is joules per Kelvin; the thermal resistance coefficient is the equivalent thermal resistance of the internal heat conduction path of the switch cabinet, and the unit is Kelvin per watt; these two parameters can be pre-calibrated through factory testing or finite element simulation.

[0055] After the above filtering and correction, the standard deviation of the temperature data is reduced to less than 10% of the original data, thus highlighting the true temperature rise trend caused by resistive losses.

[0056] S203. Solve the asynchronous problem caused by inconsistent sampling frequencies of multiple sensors and achieve spatiotemporal feature alignment of multi-source data. To solve the data asynchronous problem caused by inconsistent sampling frequencies of different monitoring units, this step adopts linear interpolation processing technology based on a unified UTC timestamp; the edge computing gateway acts as a time reference source and sends synchronization commands to each monitoring unit through the PTP protocol to ensure that the synchronization error of the data collected by each unit is less than 1 microsecond.

[0057] During the data fusion phase, the system first determines a reference time axis, which takes the sampling interval corresponding to the highest sampling frequency among all sampling sequences as the step size, such as the sampling interval of a partial discharge signal; or it directly adopts a fixed step size, such as 10 milliseconds.

[0058] Then, for temperature data, environmental parameters and other sequences with low sampling frequency, the third-order Hermite interpolation method is used. At each sampling point on the reference time axis, the interpolated value is calculated based on the time position and value of the original sampling point. The third-order Hermite interpolation can ensure that the first derivative of the interpolation curve is continuous at the original sampling point, thus better preserving the monotonic change trend of the data.

[0059] In this embodiment, the quality of multi-source heterogeneous monitoring data is improved and the time domain is unified. For partial discharge signals, an improved wavelet packet transform combined with non-negative matrix decomposition and adaptive soft threshold filtering is used to effectively remove background thermal noise and environmental narrowband interference, thereby improving the signal-to-noise ratio by more than 15dB.

[0060] For the temperature rise data, adaptive moving average filtering and temperature compensation correction are used to eliminate random fluctuations and the influence of ambient temperature, compressing the data standard deviation to less than 10% of the original value, thereby highlighting the true temperature rise trend caused by resistive loss.

[0061] To address the issue of inconsistent sampling frequencies among multiple sensors, sub-microsecond synchronization is achieved through the PTP protocol. Furthermore, third-order Hermite interpolation is performed on low-frequency parameters based on the highest sampling rate or a fixed step size to ensure that all physical quantities are strictly aligned in the time dimension.

[0062] In an exemplary embodiment, the detailed process of extracting multidimensional fault feature vectors and dynamically assigning weights in S300 is explained, specifically including: S301. Identify key operating parameters and screen core features sensitive to equipment health status. The processed data first enters the feature engineering module. This step uses Pearson correlation analysis and mutual information theory to identify the correlation between each operating parameter and equipment health status.

[0063] Quantitative value of equipment deterioration Defined as a comprehensive indicator reflecting the current health status of the switchgear, the value ranges from 0 to 1, with a higher value indicating a higher degree of deterioration. The acquisition methods are as follows: For equipment with existing historical fault sample data, the data is obtained by normalizing the operating parameters at each fault moment and integrating expert scores; for newly commissioned equipment lacking historical fault data, the data can be obtained by expert system evaluation based on the equipment's operating years, periodic maintenance records, and statistical failure models of similar equipment.

[0064] The Pearson correlation coefficient is used to measure the degree of linear correlation, and its calculation formula is as follows: in, Representative monitoring parameter sequence Quantitative value of equipment deterioration The Pearson correlation coefficient between them; To monitor parameter sequences The first in Each sample value; To monitor parameter sequences The arithmetic mean; Quantification of equipment degradation degree sequence The first in A number; Quantification of equipment degradation degree sequence The arithmetic mean; This represents the total number of sampling points in the sequence.

[0065] S302. Construct a multi-dimensional fault feature vector to form the feature basis for model input. Based on the screening results of the previous step, the system extracts specific fault features to form a multi-dimensional vector; this vector specifically includes the following 5 components: temperature rise rate, average partial discharge quantity, discharge phase distribution characteristics, mechanical opening and closing speed fluctuation rate, and intrinsic mode energy ratio of vibration signal.

[0066] Among them, the temperature rise rate is obtained by calculating the first and second differences of the temperature rise sequence to further obtain the temperature rise acceleration information, so as to reflect the sudden change in the temperature rise trend; for mechanical vibration signals, the system uses variational mode decomposition technology to decompose the non-stationary original vibration signal into several intrinsic mode functions with independent center frequencies, and then calculates the energy distribution in a specific frequency band. The result is the "intrinsic mode energy ratio", which can effectively reflect the fastening state of the mechanism's connecting parts.

[0067] S303. Dynamically allocate feature weights to generate weighted feature vectors with adaptive operating conditions. After feature extraction, the system enters the weight allocation stage. This step uses the analytic hierarchy process (AHP) combined with information entropy theory. The AHP incorporates expert experience to subjectively assess the importance of different parameters and form a judgment matrix.

[0068] Meanwhile, information entropy theory calculates the objective weights of various characteristics based on the dispersion of real-time collected data. The final system weights are determined by a combination of the subjective and objective weights mentioned above, and can be dynamically adjusted according to the current operating conditions; for example, when the switchgear is under high load conditions of more than 80% of its rated load, ohmic loss dominates, and the weight coefficient of the temperature rise characteristic will increase from the initial 0.25 to 0.55.

[0069] For example, when the ambient humidity monitoring value exceeds 85%, the risk of insulation degradation increases significantly. The system will simultaneously increase the combined weighting coefficient of partial discharge characteristics and environmental characteristics by more than 30%. This differentiated focusing mechanism ensures that the early warning logic can adapt to changes in the physical environment and load stress.

[0070] In this embodiment, the transformation from multi-source data to a weighted feature vector with adaptive working conditions is completed, providing high-quality input for subsequent deep learning prediction models.

[0071] In an exemplary embodiment, the detailed process of S400 constructing a deep learning-based dynamic trend prediction model to generate a multi-timescale prediction trajectory of the switchgear's operating status is explained, specifically including: S401. Construct a Long Short-Term Memory network with an integrated attention mechanism to capture the nonlinear dynamic trend of device state. The weighted multidimensional feature vector sequence generated in step S300 is input sequentially into a Long Short-Term Memory network with an integrated attention mechanism. This network structure includes an input layer, at least three hidden layers, and an output layer. Each hidden layer is configured with 128 to 512 neurons. The forget gate, input gate, and output gate structures integrated within the hidden layers can effectively solve the gradient vanishing problem commonly encountered in training long sequence data.

[0072] To enhance the model's ability to capture abrupt changes in device state, this network introduces a temporal attention mechanism on the basis of the standard long short-term memory network. This mechanism calculates the correlation score between the current hidden state and each moment in the historical state sequence, and assigns higher weights to key abrupt changes, such as sudden increases in current or sudden jumps in discharge.

[0073] The attention score is calculated as follows: in, For the first Always be aware of history Attention score at any given moment; The attention weight vector to be trained transpose; This is the training weight matrix corresponding to the hidden state; For the Long Short-Term Memory Network in the first The hidden layer state vector at time step 1; This is the training weight matrix corresponding to the decoding state; For the network at the previous moment, i.e. the first moment... The decoded state vector at time step; The bias term to be trained.

[0074] Through this mechanism, the model can automatically identify key inducing factors that lead to equipment failure from lengthy operation logs.

[0075] S402. Design a model training scheme and complete the iterative optimization of the prediction model. The model training adopts the adaptive moment estimation optimization algorithm, with the initial learning rate set to 0.001, the batch size set to 64, and the number of training iterations not less than 500.

[0076] The training data is constructed as follows: continuous weighted feature vector sequences are extracted from historical running data as samples; the input of each sample is past data. At any given moment, for example The feature vector, and the corresponding output label is the future value of this sample. At any given moment, for example The equipment operating status parameters; all samples are divided into training set, validation set and test set in chronological order.

[0077] The loss function is designed by combining mean squared error and mean absolute error, and its calculation formula is as follows: ;in, This is the final total loss value; The mean squared error is used to ensure the overall smoothness of the prediction curve. The mean absolute error is used to improve the model's sensitivity to outliers; and The weighting coefficients used to balance the contributions of the two error terms are typically set to... , The goal of training is to minimize the loss function. This allows the model's predictions to approximate the actual future state values ​​as closely as possible.

[0078] S403. Implement transfer learning fine-tuning to improve the prediction accuracy of newly commissioned equipment. To avoid low prediction accuracy due to insufficient initial samples of newly commissioned equipment, this model introduces a transfer learning mechanism. First, the model parameters are pre-trained using typical fault lifecycle degradation data obtained from similar equipment in a simulated laboratory environment, enabling it to grasp the general physical laws of equipment degradation.

[0079] Subsequently, after the model is deployed to the current switch cabinet under inspection, the real-time operating data from the first three months after the equipment was put into operation is used to fine-tune the parameters of the last few layers of the pre-trained model.

[0080] This method enables the equipment to achieve a prediction accuracy of over 92% even in the early stages of operation, despite the lack of large-scale local fault samples. Ultimately, the trained model can output predicted trajectories of the operating status for the next 24, 48, and 72 hours.

[0081] In this embodiment, a predictive model capable of capturing the nonlinear dynamic degradation trend of equipment is constructed, and its output multi-step prediction trajectory provides a quantitative basis for subsequent early warning judgment.

[0082] In an exemplary embodiment, the detailed process of S500, which compares the generated predicted operating status trajectory with preset multi-level status thresholds in real time and executes multi-level early warning judgment and corresponding linkage protection strategy, is explained. Specifically, it includes: comparing the predicted operating status trajectory for the next 24 hours, 48 ​​hours, and 72 hours generated in step S400 with preset multi-level status thresholds stored in the cloud in real time; the setting of multi-level status thresholds comprehensively refers to the State Grid operation and maintenance standards and historical operating big data provided by equipment manufacturers; the early warning judgment logic is divided into three levels: attention, abnormal, and severe, with each level corresponding to different early warning actions and linkage protection strategies.

[0083] S501. Execute the attention level warning judgment and trigger the corresponding warning operation. The system triggers the attention level warning when any of the following conditions are met: Condition 1: The predicted value of a certain operating parameter exceeds 110% but is less than 120% of its rated operating parameter.

[0084] Rated operating parameters refer to the design standard values ​​of the equipment under normal operating conditions. For example, if the rated temperature of the circuit breaker contact is 75 degrees Celsius, then the predicted temperature is between 82.5 degrees Celsius and 90 degrees Celsius, triggering this level; Condition 2: The partial discharge pulse frequency shows an increasing trend of less than 5%.

[0085] The growth trend is calculated by comparing the average pulse frequency within the current time window with the average frequency within the previous window of equal duration, and the growth rate does not exceed 5%. After triggering the attention level warning, the system automatically performs the following operations: pushes the warning information to the mobile terminal of the maintenance personnel, which includes the fault location, parameter deviation value, and predicted degradation curve.

[0086] At the same time, the warning record will be saved in the device's electronic health record for reference in subsequent status inspections.

[0087] S502. Execute an anomaly level warning judgment and trigger the corresponding linkage protection strategy. The system triggers an anomaly level warning when any of the following conditions are met: Condition 1: The prediction parameters show a clear nonlinear growth trend; for example, the temperature rise rate shows an exponential increase, that is, the first difference sequence of temperature rise shows an increasing trend and the second difference is positive for more than 3 sampling periods. Condition 2: The partial discharge intensity continuously exceeds 5 times the background noise; the background noise is obtained by statistically analyzing the average signal amplitude during the non-discharge period, and the condition is met when the ratio of discharge intensity to background noise is greater than 5 for 5 consecutive sampling points.

[0088] After triggering an anomaly level warning, the system determines that the equipment has a mid-term failure risk and automatically activates the following linkage protection strategies: activating the switch cabinet's built-in forced air cooling system to force heat dissipation, specifically by sending a start command to the air cooling system controller and setting the fan speed to 80% of the rated speed; at the same time, through the integrated automation background, suggesting to the dispatch system to reduce the load distribution of this branch, suggesting a load reduction of 30% of the current load, and executing the load reduction operation strategy after confirmation by the dispatch personnel.

[0089] S503. Execute a severe level warning judgment and trigger emergency protection control. When the predicted parameters indicate that the insulation breakdown voltage or thermal runaway critical temperature will be triggered within the next 1 hour, the system triggers a severe level warning. The insulation breakdown voltage threshold is set according to the factory withstand voltage test value of the switchgear insulation material. For example, the breakdown voltage threshold of a 12kV switchgear is 42kV. The thermal runaway critical temperature is set according to the softening point of the contact material. For example, the critical temperature of copper contacts is 150 degrees Celsius.

[0090] If the predicted value at any moment in the predicted trajectory reaches the above threshold and the remaining time is less than 1 hour, the conditions for this level are met. After triggering a severe level warning, the system directly sends a tripping command to the logic controller of the relay protection system within a millisecond response time; this command is transmitted via hardwiring or a GOOSE network with a delay of no more than 10 milliseconds.

[0091] Upon receiving the instruction, the relay protection system disconnects the fault-side switch to prevent the accident from escalating into a busbar power outage or equipment explosion.

[0092] In this embodiment, a complete technical closed loop is constructed from state perception to risk identification and then to automatic intervention through the above-mentioned three-level early warning judgment and linkage protection control strategy. The system compares the predicted trajectory with the preset threshold in real time, and performs differentiated early warning information push, automatic equipment control and power grid dispatch suggestions according to the risk level, until the relay protection trips directly under the highest risk level.

[0093] This strategy ensures that maintenance personnel are notified in the early stages of equipment degradation, proactive load reduction and heat dissipation interventions are implemented under medium-term risks, and faults are cleared in milliseconds in emergency situations, thereby effectively ensuring the safe operation of switchgear and power distribution systems.

[0094] On the other hand, the high and low voltage switchgear early warning system disclosed in this application includes: A multi-dimensional sensing network is deployed inside high and low voltage switchgear to collect multi-source heterogeneous data during the operation of the switchgear. The multi-source heterogeneous data includes temperature rise data, partial discharge data, mechanical characteristic data, and environmental condition data. Edge computing gateways are used to achieve time synchronization and preprocessing of multi-sensor data. As a time reference source, the edge computing gateway sends synchronization commands to each monitoring unit through a precise time protocol, so that the synchronization error of each monitoring unit is less than 1 microsecond. The core processing platform performs the following operations: preprocessing and aligning the collected multi-source heterogeneous monitoring data with spatiotemporal features; calculating a humidity correction coefficient based on the real-time relative humidity value in the environmental state data, and a pressure change correction coefficient based on the pressure change rate in the environmental state data; dynamically correcting the sensitivity threshold of the partial discharge criterion using the calculated humidity correction coefficient or pressure change correction coefficient to obtain the corrected sensitivity threshold of the partial discharge criterion; then, extracting the multi-dimensional fault feature vector from the spatiotemporally aligned and dynamically corrected multi-source heterogeneous data, and dynamically assigning weights to each component of the multi-dimensional fault feature vector according to the current operating condition of the switchgear and the environmental state; inputting the dynamically weighted multi-dimensional fault feature vector into a long short-term memory network with an integrated attention mechanism to generate a multi-timescale predicted trajectory of the switchgear's operating state; and comparing the predicted trajectory of the operating state with preset multi-level state thresholds in real time to perform multi-level early warning judgments. The relay protection linkage system is used to receive early warning commands from the core processing platform and execute corresponding protection and control strategies.

[0095] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0096] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for early warning of high and low voltage switchgear, characterized in that, Includes the following steps: A multi-dimensional sensing network is constructed to collect multi-source heterogeneous data in real time during the operation of high and low voltage switchgear; the multi-source heterogeneous data includes temperature rise data, partial discharge data, mechanical characteristic data and environmental condition data; The collected multi-source heterogeneous monitoring data are preprocessed and aligned with spatiotemporal features; The preprocessing and spatiotemporal feature alignment includes: using synchronization instructions issued based on a precise time protocol to make the synchronization error of each monitoring unit less than 1 microsecond, and using a reference time axis as the step size to perform time dimension alignment processing on monitoring data with sampling frequency lower than the reference time axis using third-order Hermite interpolation. A humidity correction coefficient is calculated based on the real-time relative humidity value in the environmental state data, and a pressure change correction coefficient is calculated based on the pressure change rate in the environmental state data. The calculated humidity correction coefficient or the pressure change correction coefficient is used to dynamically correct the sensitivity threshold of the partial discharge criterion. A multidimensional fault feature vector is extracted from the multi-source heterogeneous data after spatiotemporal feature alignment and dynamic correction, and the weights of each component in the multidimensional fault feature vector are dynamically allocated according to the current operating conditions and environmental status of the switchgear. The multi-dimensional fault feature vector after dynamic weight allocation is input into a long short-term memory network with integrated attention mechanism to generate a multi-timescale prediction trajectory of the switchgear's operating status. The predicted trajectory of the operating state is compared with the preset multi-level state thresholds in real time, and multi-level early warning judgment and corresponding linkage protection strategy are executed.

2. The high and low voltage switchgear early warning method according to claim 1, characterized in that, The step of calculating a humidity correction coefficient based on the real-time relative humidity value in the environmental status data and using the calculated humidity correction coefficient to dynamically correct the sensitivity threshold of the partial discharge criterion specifically includes: when the real-time collected environmental relative humidity value exceeds the preset humidity threshold, humidity correction is enabled. The humidity correction factor is calculated as follows: the humidity correction factor is equal to 1 plus the product of the difference between the real-time relative humidity value and the preset humidity threshold divided by 100 and then multiplied by 0.

2. Multiply the original sensitivity threshold of the partial discharge criterion by the humidity correction coefficient to obtain the corrected sensitivity threshold of the partial discharge criterion.

3. The early warning method for high and low voltage switchgear according to claim 1, characterized in that, The steps of calculating a pressure change correction coefficient based on the pressure change rate in environmental state data, and then dynamically correcting the sensitivity threshold of the partial discharge criterion using the calculated pressure change correction coefficient, specifically include: When the real-time calculated pressure change rate exceeds the preset pressure change rate abnormal threshold, pressure correction is activated; the calculation formula for the pressure change correction coefficient is configured as follows: the pressure change correction coefficient is equal to 1 plus the product of the difference between the absolute value of the pressure change rate and the pressure change rate abnormal threshold, divided by the pressure change rate abnormal threshold and multiplied by 0.

1. The original sensitivity threshold of the partial discharge criterion is multiplied by the pressure change correction coefficient to obtain the corrected sensitivity threshold of the partial discharge criterion.

4. The high and low voltage switchgear early warning method according to claim 1, characterized in that, The steps of preprocessing and aligning the collected multi-source heterogeneous monitoring data with spatiotemporal features also include noise reduction processing of the partial discharge signal: Daubechies wavelet was selected as the basis function to perform 5 to 8-level wavelet packet decomposition on the original partial discharge signal. At each decomposition scale, nonnegative matrix decomposition theory is used to extract background thermal noise features and environmental narrowband interference features to distinguish effective discharge pulses from background interference. For each decomposition scale, a threshold proportional to the noise level of that scale is set, and wavelet coefficients with absolute values ​​below the threshold are set to zero, while discharge pulse coefficients with abrupt change characteristics are retained. Wavelet reconstruction is performed on the retained wavelet coefficients to obtain the denoised partial discharge signal.

5. The high and low voltage switchgear early warning method according to claim 1, characterized in that, The specific steps for extracting multidimensional fault feature vectors from multi-source heterogeneous data after spatiotemporal feature alignment and dynamic correction include: Pearson correlation analysis and mutual information theory are used to identify the correlation between various operating parameters and equipment health status, and to screen the core operating parameters that are sensitive to equipment health status. Based on the screening results, a multidimensional fault feature vector is extracted; the multidimensional fault feature vector specifically includes the following 5 components: temperature rise rate, average partial discharge quantity, discharge phase distribution characteristics, mechanical opening and closing speed fluctuation rate, and intrinsic mode energy ratio of vibration signal; The temperature rise rate is obtained by calculating the first and second order differences of the temperature rise sequence; the intrinsic mode energy ratio of the vibration signal is obtained by using variational mode decomposition technology to decompose the original vibration signal into several intrinsic mode functions with independent center frequencies, and calculating the energy distribution within a specific frequency band.

6. The high and low voltage switchgear early warning method according to claim 1 or 5, characterized in that, The specific steps for dynamically assigning weights to each component of the multidimensional fault feature vector based on the current operating conditions of the switchgear and the environmental conditions include: By incorporating expert experience through the analytic hierarchy process, the importance of different parameters is subjectively assessed, forming a judgment matrix. At the same time, using the information entropy theory, the objective weights of each feature are calculated based on the degree of dispersion of the real-time collected data. The system weights are determined by combining the subjective assessments and the objective weights. The system weights can be dynamically adjusted according to the current operating conditions. Specifically, when the switchgear is under high load conditions of more than 80% of the rated load, the weight coefficient of the temperature rise characteristic is increased from the initial value of 0.25 to 0.55; when the ambient humidity monitoring value exceeds 85%, the combined weight coefficient of the partial discharge characteristic and the environmental characteristic is increased by more than 30% simultaneously.

7. The early warning method for high and low voltage switchgear according to claim 1, characterized in that, The specific steps of inputting the dynamically weighted multidimensional fault feature vector into a long short-term memory network with an integrated attention mechanism include: The weighted multidimensional feature vector sequence is input into the network in chronological order. The network contains an input layer, at least three hidden layers, and an output layer. Each hidden layer is configured with 128 to 512 neuron units. A temporal attention mechanism is introduced on the basis of the standard long short-term memory network. By calculating the correlation score between the current hidden state and each moment in the historical state sequence, higher weights are assigned to key mutation moments. The transfer learning mechanism is used to fine-tune the parameters of newly commissioned equipment: First, the model parameters are pre-trained using typical fault life cycle degradation data obtained from similar equipment in a simulated laboratory environment. After the model is deployed to the current inspected switchgear, the real-time operating data of the first 3 months after the equipment is put into operation is used to fine-tune the parameters of the last few layers of the pre-trained model.

8. The early warning method for high and low voltage switchgear according to claim 1, characterized in that, The specific steps for implementing multi-level early warning judgment and corresponding linkage protection strategies include: The multi-level early warning judgment is divided into three levels: attention, abnormal, and severe. When a warning level alert is triggered, the alert information will be pushed to the mobile terminal of the maintenance personnel, and the alert record will be simultaneously saved in the device's electronic health record. When an abnormal level warning is triggered, the forced air cooling system built into the switchgear is activated, and the integrated automation backend is used to suggest to the dispatching system to reduce the load distribution of that branch. When a severe warning is triggered, a tripping command is sent directly to the logic controller of the relay protection system within a millisecond response time. The command is transmitted via hardwire or GOOSE network with a delay of no more than 10 milliseconds.

9. The high and low voltage switchgear early warning method according to claim 8, characterized in that, The triggering conditions for a level warning are: the predicted value of a certain operating parameter exceeds 110% but is less than 120% of its rated operating parameter, or the frequency of partial discharge pulses shows an increasing trend of less than 5%. The triggering conditions for the abnormal level warning are: the temperature rise rate of the predicted parameters increases exponentially, that is, the first-order difference sequence of temperature rise shows an increasing trend and the second-order difference is positive for more than 3 sampling periods, or the partial discharge intensity continues to exceed 5 times the background noise. The triggering condition for the severe level warning is: the prediction parameters indicate that the insulation breakdown voltage or thermal runaway critical temperature will be triggered within the next 1 hour time threshold.

10. A high- and low-voltage switchgear early warning system, applied to the method described in any one of claims 1-9, characterized in that, include: A multi-dimensional sensing network is deployed inside high and low voltage switchgear to collect multi-source heterogeneous data during the operation of the switchgear. The multi-source heterogeneous data includes temperature rise data, partial discharge data, mechanical characteristic data, and environmental condition data. An edge computing gateway is used to realize time synchronization and preprocessing of multi-sensor data; the edge computing gateway acts as a time reference source and sends synchronization commands to each monitoring unit through a precise time protocol, so that the synchronization error of each monitoring unit is less than 1 microsecond. The core processing platform is used to perform the following operations: preprocessing and aligning the collected multi-source heterogeneous monitoring data with spatiotemporal features; A humidity correction coefficient is calculated based on the real-time relative humidity value in the environmental state data, and a pressure change correction coefficient is calculated based on the pressure change rate in the environmental state data. The calculated humidity correction coefficient or the pressure change correction coefficient is used to dynamically correct the sensitivity threshold of the partial discharge criterion, resulting in a corrected sensitivity threshold for the partial discharge criterion. Then, a multi-dimensional fault feature vector is extracted from the multi-source heterogeneous data after spatiotemporal feature alignment and the dynamic correction. The weights of each component in the multi-dimensional fault feature vector are dynamically allocated according to the current operating condition of the switchgear and the environmental state. The dynamically weighted multi-dimensional fault feature vector is input into a long short-term memory network with an integrated attention mechanism to generate a multi-timescale predicted trajectory of the switchgear's operating state. The predicted trajectory of the operating state is compared with preset multi-level state thresholds in real time to perform multi-level early warning judgment. The relay protection linkage system is used to receive early warning commands from the core processing platform and execute corresponding protection control strategies.

Images