Intelligent operating system of low-voltage switch cabinet

Abstract

The invention belongs to the technical field of switch cabinets, and particularly relates to an intelligent operating system of a low-voltage switch cabinet, which comprises a data acquisition module used for acquiring multi-source operating data of the low-voltage switch cabinet in real time, and the multi-source operating data comprises operating parameters and state information; the storage module is used for storing configuration parameters and control programs and recording historical operation data, fault records and operation logs of the switch cabinet, and the data can be used for trend analysis and preventive maintenance; the prediction module is used for predicting early-stage characteristics of potential faults of the switch cabinet based on the multi-source operation data acquired by the data acquisition module and a pre-trained digital twin model of the prediction module; the early warning module is used for outputting an early warning signal based on the fault prediction result output by the prediction module; and the diagnosis module is used for analyzing the fault of the switch cabinet based on the alarm output by the early warning module and generating a diagnosis report and decision support.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of switch cabinets, and particularly relates to an intelligent operation system of a low-voltage switch cabinet. BACKGROUND

[0002] The low-voltage switch cabinet is a device for controlling, protecting and distributing electric power; it is usually installed in a low-voltage power system and used for managing power distribution and controlling circuits.

[0003] At present, the operation and maintenance management of the low-voltage switch cabinet mainly depends on the traditional planned maintenance and after-service maintenance mode.

[0004] Periodic inspection and preventive test: the operation and maintenance personnel periodically (such as every quarter or every year) perform on-site inspection on the switch cabinet, and check the appearance, temperature and abnormal sound through naked eye observation, infrared temperature detector sampling inspection and the like; meanwhile, the preventive test (such as measurement of loop resistance, insulation resistance and the like) needs to be performed with power off; this mode has obvious disadvantages: firstly, the inspection is discontinuous, and the equipment state cannot be grasped in real time, and the sudden failure or rapidly developing defect occurring between two inspections cannot be effectively warned; secondly, the power-off test affects normal power supply, and frequent operation may cause unnecessary damage to the equipment; finally, it is easy to miss detection and misjudge, and it is difficult to find latent hidden dangers such as slight oxidation of the contact, early deterioration of the insulation material and the like.

[0005] Some improved schemes attempt to install a single online monitoring device such as a temperature sensor or a current transformer on the switch cabinet to realize continuous measurement of a specific parameter; however, this kind of system has single function, and can usually only issue an alarm when the parameter exceeds a fixed threshold, which belongs to the nature of after-service alarm; its limitation lies in: data isolation, lack of correlation analysis: only individual parameters (such as temperature) are monitored, and it is difficult to accurately determine the fault source by correlating the temperature change with load current, contact resistance, partial discharge and other factors.

[0006] Therefore, the application provides an intelligent operation system of a low-voltage switch cabinet. SUMMARY

[0007] In order to make up for the deficiencies of the prior art and solve at least one technical problem proposed in the background art.

[0008] The technical scheme adopted by the application to solve the technical problems is: the intelligent operation system of the low-voltage switch cabinet comprises a data acquisition module: the data acquisition module is used for real-time acquisition of multi-source operation data of the low-voltage switch cabinet, and the multi-source operation data comprises operation parameters and state information.

[0009] Storage module: The storage module is used to store configuration parameters and control programs, and also records the historical operating data, fault records and operation logs of the switchgear. This data can be used for trend analysis and preventive maintenance.

[0010] Prediction Module: The prediction module predicts early characteristics of potential faults in the switchgear based on multi-source operational data acquired by the data acquisition module and its own pre-trained digital twin model.

[0011] Early warning module: The early warning module outputs an early warning signal based on the predicted fault results output by the prediction module;

[0012] Diagnostic module: The diagnostic module analyzes the faults of the switchgear itself based on the alarms output by the early warning module, and generates a diagnostic report and decision support, including disconnecting the switchgear.

[0013] Preferably, the prediction module predicts potential faults in the switchgear based on a digital twin model as follows:

[0014] Step S1: Construct a joint geometric-physical model of the switchgear;

[0015] Step S2: Construct a physical information neural network model based on the switch cabinet's geometric-physical joint model, and generate a digital twin model;

[0016] Step S3: Inject the fault mode into the digital twin model to simulate the entire process of the switchgear gradually deteriorating from a healthy state under various potential operating conditions;

[0017] Step S4: Continuously input the real-time sensor data, such as current and temperature, into the digital twin model and calculate the overall health index of the switchgear.

[0018] Preferably, in step S1, constructing the geometric-physical joint model of the switchgear specifically involves the following steps:

[0019] Step S1.1: First, perform digital modeling and data acquisition for the switchgear:

[0020] This process utilizes computer-aided design drawings of switchgear, combined with 3D laser scanning technology, to obtain high-precision geometric models of the switchgear and key components (such as contacts, busbars, or insulation components). Based on these high-precision geometric models, accurate material properties are specified for each component, such as contacts, busbars, or insulation components. These properties include parameters such as density, specific heat capacity, thermal conductivity, and electrical conductivity. Subsequently, according to the actual operating environment of the switchgear, boundary conditions are set on the corresponding surfaces or regions of the geometric model. For example, in thermal field analysis, the convective heat dissipation coefficient of the cabinet surface and the ambient temperature need to be set.

[0021] Step S1.2: Deploy the sensor array and establish a data channel:

[0022] Multiple sensor arrays, including temperature sensors, current transformers, voltage sensors, and partial discharge sensors, are deployed in key areas inside the switch cabinet.

[0023] In step S2, constructing the physical information neural network model specifically involves the following steps:

[0024] Step S2.1: The network architecture directly embeds the physical laws from step S1.1, such as energy conservation and heat transfer equations, as constraints into the loss function of the neural network, instead of relying solely on pure data-driven processes. The input layer of the network receives real-time sensor data, such as current and ambient temperature, while the hidden layer learns the complex mapping between physical laws and observed data. The output layer predicts key state parameters, such as the real-time temperature and equivalent resistance of the contacts, resulting in a digital twin model.

[0025] The neural network refers to a fully connected deep learning model that takes sensor data as input and state parameters as output. Embedding physical laws into the loss function means using automatic differentiation technology to calculate the derivative of the network output with respect to the input and constructing a composite loss function that includes physical equation residuals, data density and error, and boundary condition error. By minimizing this total loss, the model learns data while its output is strictly constrained within the range allowed by physical laws, thus becoming a standard and reliable digital twin.

[0026] Step S2.2: Next, model parameter calibration and verification are performed. Using the switchgear factory test data, historical normal operation data, and some known minor anomaly data, the constructed physical information neural network model is trained and calibrated. The internal parameters of the model are adjusted using the Simulink Design Optimization tool to minimize the data error of the digital twin simulation output, such as minimizing the error between the contact temperature rise curve and the actual measurement data of the sensor, so as to ensure that the virtual model can reflect the real-time status of a specific switchgear with high fidelity.

[0027] Preferably, in step S3, injecting the fault mode into the digital twin model specifically involves:

[0028] In the calibrated digital twin model in step S2.2 above, the system injects a preset fault mode, that is, introduces a specific fault mode into the digital twin model, such as directly adjusting the model parameters, such as gradually increasing the contact resistance of the contacts;

[0029] In the virtual environment of the digital twin model, by accelerating simulation and running a digital twin injected with fault modes, the entire process of the switch cabinet gradually deteriorating from a healthy state under various potential operating conditions can be simulated. This process can generate a large amount of fault state data that is difficult to obtain in the real world, especially potential fault data that develops slowly and has no obvious early signs.

[0030] Preferably, in step S4, the real-time collected sensor data, such as current and temperature, is continuously input into the digital twin model, and the comprehensive health index of the switchgear is calculated, including the following steps:

[0031] Step S4.1: First, extract key health feature parameters;

[0032] From the synchronously updated digital twin model, key feature parameters that can quantify equipment performance degradation are extracted, including instantaneous deviation features and trend evolution features; instantaneous deviation features: reflect the absolute difference between the current state and the ideal health benchmark;

[0033] For example, contact temperature deviation ΔT: Calculates the contact temperature T simulated by the digital twin under the current load current. S The reference temperature T of a brand new contact under the same operating conditions b The deviation, i.e., ΔT = T S -T b ;

[0034] Contact resistance deviation rate, calculate the simulated contact resistance R of the contact. s Relative to its initial health value R i The rate of change, i.e., ΔR% = (R s -R i ) / R i ×100%;

[0035] Trend evolution characteristics: Reflecting the long-term rate of change of equipment condition parameters, predicting potential aging processes, such as the contact resistance trend slope kR, over a continuous period, such as the past 30 days, for the contact resistance time series data (R1, R2, ..., R...) output by the digital twin. n Linear regression analysis (least squares method) is performed, and the slope kR of the fitted line is the slope of the trend. A positive value indicates that the resistance is increasing.

[0036] The acceleration of temperature change αT is used to perform quadratic curve fitting or calculate the second derivative of the time series data of the contact temperature deviation ΔT to determine whether there is an acceleration in the deterioration process.

[0037] Preferably, step S4.1, which involves extracting key health feature parameters, further includes:

[0038] Step S4.1.1: First, preprocess the key feature parameters and assign weights to the feature data;

[0039] Data standardization is necessary because different feature parameters have different dimensions and orders of magnitude. Z-score standardization is used to transform all feature values ​​to the same dimension, eliminating their incommensurability. The formula is x... std = (x-μ) / σ, where X is the original feature value, μ is the historical mean of the feature, and σ is the standard deviation;

[0040] Step S4.1.2, Feature weight configuration: Based on the degree of influence of each feature parameter on the overall health status of the switchgear, assign weights to them. The weight configuration can be based on the feature importance assessment and dynamically adjusted according to the equipment type or operating history to ensure that the sum of all feature weights is 100%.

[0041] Step S4.1.3: Based on the preprocessed key feature parameters and their feature weight configuration, calculate the comprehensive health index HI using a weighted average model;

[0042] The health index is based on a composite model that comprehensively considers the physical model of the equipment and its real-time status.

[0043] HI=∑(ω i ×F i (P i )); where P i Let be the i-th preprocessed feature parameter value;

[0044] ω i Let ω be the weight coefficient of the i-th feature, satisfying ∑ω i =1. The allocation of weights is crucial and can be based on expert experience, such as temperature, where local factors contribute more to the failure and therefore have a larger weight, or obtained through correlation analysis or learning from historical failure data via machine learning.

[0045] The calculated health index HI needs to be compared with a preset threshold to determine the current health status and decide whether to trigger an alarm. Collect the switch cabinet's full life cycle data in terms of health status, known fault status, and aging process, including historical sensor data and corresponding maintenance records.

[0046] The system is operating normally when the health threshold HI ≥ 80.

[0047] If the warning threshold is 50≤HI<80, the system is experiencing performance degradation; therefore, monitoring should be strengthened.

[0048] The alarm threshold, HI<50, indicates a severe deterioration in the system's health status, and a signal is transmitted to the diagnostic module.

[0049] Preferably, when the diagnostic module receives the comprehensive health index as the alarm threshold, the diagnostic module will analyze which characteristic parameters (such as a surge in C-phase current, an abnormal increase in temperature at a certain point in C-phase, accompanied by a strong partial discharge signal) caused the change in HI;

[0050] Step 1: The diagnostic module acquires all raw sensor data of HI within a time window before and after the change; this data includes, but is not limited to, three-phase current, voltage waveform data, temperature sequence of key points such as contacts and busbars, vibration acceleration signals, and partial discharge pulse signals; the time window before and after the change is 10 seconds.

[0051] Since different sensors have different sampling frequencies and data acquisition times, data synchronization and alignment are performed first. The engine uses interpolation to unify all data onto the same timestamp sequence, forming a strictly synchronized multi-dimensional time series data matrix.

[0052] Data cleaning is performed on the aligned data, including removing impulse noise obviously caused by interference and filling in a small amount of missing data caused by brief communication interruptions, so as to provide a high-quality data foundation for subsequent analysis.

[0053] Step II, the diagnostic module further extracts time-domain and frequency-domain features:

[0054] Then, all the extracted time-domain and frequency-domain features are combined with the raw values ​​directly measured by the sensors to construct a high-dimensional time-series feature matrix, which fully describes the overall operating status of the switchgear within the analysis time window;

[0055] Step III: Then, the time series feature matrix is ​​input into a pre-trained deep learning model based on the attention mechanism. This model can automatically learn and assign different feature parameters to the contribution or attention weight of HI change in a specific fault mode. For example, in the case of suspected contact overheating, the model will assign higher attention weights to parameters such as contact temperature and contact current, thereby focusing on the core feature most likely to cause HI to decrease among many parameters.

[0056] Step IV: The diagnostic module includes a fault mode knowledge base, which is derived from historical fault data, expert experience and digital twin simulation data. The knowledge base defines various typical faults, such as poor contact of contacts, moisture in insulation and jamming of mechanisms, and the corresponding characteristic parameter combinations and typical change modes.

[0057] The current high-weight feature combination and its changing trend obtained in steps II and III are compared with the patterns in the fault knowledge base for similarity calculation. Through retrieval and matching, the diagnostic module finally outputs the most likely fault diagnosis conclusion. For example, there is a 92% probability that the increased contact resistance of the C-phase cable joint caused the joint temperature to rise sharply from 75°C to 112°C within 15 minutes under a load current of 1250A, which led to a significant increase in partial discharge activity and was the root cause of this sudden drop in HI.

[0058] Step V: The diagnostic module generates diagnostic reports and decision support.

[0059] The report includes the primary cause, i.e. the first abnormal parameter that leads to the decrease in HI, associated parameters, other parameters that are strongly correlated with the cause, such as C-phase current or partial discharge, failure probability, and the reliability of the diagnostic results.

[0060] The report, which includes the evolution trend and historical change curves of key parameters, is directly pushed to the human-machine interface.

[0061] Preferably, in step II, the time-series feature matrix includes:

[0062] Step II-I: First, clean the raw data of the multi-sensor array to remove noise and obvious outliers; handle missing values, fill them with interpolation, and normalize the data.

[0063] Step II-II: Use the sliding window technique to divide the continuous time series data into multiple time segments. For example, set a window size (e.g., 60 minutes) and a step size (e.g., 5 minutes), and extract data according to this rule to form a series of continuous, possibly overlapping data windows.

[0064] Sliding window segmentation: The data within each window constitutes the basic unit for feature extraction;

[0065] Steps II and III calculate the mean, standard deviation, maximum value, minimum value, quantiles, etc. of the data within the window to describe the basic distribution characteristics of the data within that time period;

[0066] After steps II-IV, historical data is introduced as features, such as the values ​​of the previous moment and the two moments before, to capture short-term dependencies; then, finer-grained sliding is performed within the window, and calculations such as moving average and moving standard deviation are performed to smooth noise and highlight trends.

[0067] Steps II-V involve concatenating all types of features extracted from the same time window along the feature dimension to form a wide vector representing the overall state of that window; stacking the feature vectors corresponding to all time windows in chronological order to form a complete temporal feature matrix. Each row of this matrix corresponds to a time window, and each column corresponds to a feature variable, providing standardized input for the subsequent diagnostic engine.

[0068] Preferably, the deep learning model building in step III includes the following steps:

[0069] Step III-I: Data preparation and preprocessing, basic feature extractor, using one-dimensional convolutional neural network or long short-term memory network as the backbone network. 1D-CNN is good at extracting local feature patterns, and LSTM is good at capturing long-term dependencies in time series.

[0070] Step III-II: After the feature extractor, an attention layer is connected. This attention layer will calculate an attention weight for each feature parameter at each time step in the temporal feature matrix. These weights are normalized by functions such as Softmax to represent the importance of each feature to the final decision at a specific time.

[0071] Step III-III: Multiply the learned attention weights with the original features and perform a weighted sum to obtain a "context vector" that can represent the key information of the entire time window. Finally, this vector is fed into a fully connected layer to output the predicted value of the health index or the probability of fault classification.

[0072] Steps III-IV: After model training and optimization, if the predicted HI value is used, the mean squared error loss function is used to measure the difference between the model's predicted value and the true label.

[0073] Preferably, the specific method for establishing the fault mode knowledge base in step IV is as follows:

[0074] The switchgear system is considered as a whole, and then decomposed layer by layer into subsystems, assemblies and individual components. For example, it can be decomposed into subsystems such as "circuit breaker mechanism", "insulation system" and "contact system". The "contact system" can be further decomposed into "moving contact" and "stationary contact". This hierarchical decomposition helps to accurately identify the specific location of the fault.

[0075] Furthermore, based on importance and failure frequency, identify the key components and subsystems that need to be focused on in the initial stage of the knowledge base to avoid making the scope too broad and difficult to start.

[0076] By collecting and organizing historical work orders, maintenance records, and accident reports, and injecting fault modes into digital twin models, we can obtain fault data that is difficult to obtain in actual applications, especially slow-developing and early-stage latent fault data, which greatly enriches the breadth and depth of the knowledge base.

[0077] Establish the association between fault modes and combinations of characteristic parameters and change patterns. For example, a fault mode usually triggers changes in multiple characteristic parameters. "Poor contact" may simultaneously lead to "abnormal increase in contact temperature", "temperature rise exceeding normal value when passing the same current", and "increased activity of partial discharge signal". The knowledge base needs to record these combinations of characteristic parameters.

[0078] Then, an SQL database was used for storage, and multiple related data tables were designed. Offline confirmed fault cases and real feature data were fed back into the knowledge base and continuously evolved.

[0079] The beneficial effects of this invention are as follows:

[0080] 1. The intelligent operating system for low-voltage switchgear described in this invention constructs a high-fidelity digital twin model and injects fault modes for accelerated simulation, enabling the simulation of the entire process of switchgear from health to degradation. Combined with real-time data, the system can predict early characteristics of potential faults and calculate equipment lifespan. This allows maintenance personnel to take intervention measures before a fault occurs, or even when performance just begins to degrade (such as when HI enters the warning threshold of 50-80), avoiding unplanned downtime, preventing small defects from evolving into major accidents, and greatly improving the continuity of power supply and the safety of equipment.

[0081] 2. The intelligent operating system for low-voltage switchgear described in this invention extracts information from multiple dimensions (time domain, frequency domain, statistical features) to construct a comprehensive feature matrix that fully describes the equipment status. The deep learning model can automatically learn and assign different feature parameters "contribution" or "attention weight" under specific faults, thereby focusing on the core features causing the anomaly from massive amounts of data. The system matches the current high-weight feature combination with typical fault modes in the knowledge base to achieve accurate diagnosis. The system can not only alarm, but also clearly point out the "most likely cause of the fault" and its probability, and generate a detailed diagnostic report containing the main causes and related parameters. This significantly shortens the fault investigation time and provides strong data support for maintenance decisions. Attached Figure Description

[0082] The invention will now be further described with reference to the accompanying drawings.

[0083] Figure 1 This is a system diagram of the present invention;

[0084] Figure 2 This is a flowchart of the digital twin model establishment method in this invention. Detailed Implementation

[0085] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0086] like Figure 1 and Figure 2 As shown in the embodiment of the present invention, a low-voltage switchgear intelligent operating system includes a data acquisition module, a storage module, a prediction module, an early warning module, and a diagnostic module.

[0087] Data acquisition module: The data acquisition module is used to acquire multi-source operating data of low-voltage switchgear in real time. The multi-source operating data includes operating parameters and status information, including current transformers, voltage sensors, temperature sensors and partial discharge sensors;

[0088] Storage module: The storage module is used to store configuration parameters and control programs, and also records the historical operating data, fault records and operation logs of the switchgear. This data can be used for trend analysis and preventive maintenance.

[0089] Prediction Module: The prediction module predicts early characteristics of potential faults in the switchgear based on multi-source operational data acquired by the data acquisition module and its own pre-trained digital twin model.

[0090] Early warning module: The early warning module outputs an early warning signal based on the predicted fault results output by the prediction module;

[0091] Diagnostic module: The diagnostic module analyzes the faults of the switchgear itself based on the alarms output by the early warning module, and generates a diagnostic report and decision support, including disconnecting the switchgear.

[0092] The decision support includes maintenance suggestions displayed to operation and maintenance personnel, emergency stop operation suggestions, such as immediately disconnecting the C-phase circuit breaker, or sending standardized circuit breaker command signals to the circuit breaker controller of the switchgear via the communication interface.

[0093] The prediction module predicts potential faults in the switchgear based on a digital twin model, specifically as follows:

[0094] Step S1: Construct a joint geometric-physical model of the switchgear;

[0095] Step S2: Construct a physical information neural network model based on the switch cabinet's geometric-physical joint model, and generate a digital twin model;

[0096] Step S3: Inject the fault mode into the digital twin model to simulate the entire process of the switchgear gradually deteriorating from a healthy state under various potential operating conditions;

[0097] Step S4: Continuously input the real-time sensor data, such as current and temperature, into the digital twin model and calculate the overall health index of the switchgear.

[0098] In step S1, constructing the geometric-physical joint model of the switchgear specifically involves the following steps:

[0099] Step S1.1: First, perform digital modeling and data acquisition of the switchgear;

[0100] This process utilizes computer-aided design drawings of the switchgear, combined with 3D laser scanning technology, to obtain high-precision geometric models of the switchgear and key components (such as contacts, busbars, or insulation components). Based on these high-precision geometric models, precise material properties are specified for each component, including parameters such as density, specific heat capacity, thermal conductivity, and electrical conductivity. Then, according to the actual operating environment of the switchgear, boundary conditions are set on the corresponding surfaces or regions of the geometric model. For example, in thermal field analysis, the convective heat dissipation coefficient of the cabinet surface and ambient temperature need to be set; in electric field analysis, the boundary conditions of the busbars need to be set. The voltage excitation conditions of the contacts are embedded with physical laws, such as contact resistance models describing the electrical contact characteristics of contacts, empirical models characterizing the aging process of insulating materials, and Joule's law models for calculating the heat generation of conductors. Among them, physical law equations are embedded through multi-physics field coupling equations, such as embedding them with electromagnetic-thermal coupling models. Based on Joule's law and the differential equation of heat conduction, mathematical models can be established to show the heat generated by the current flowing in the conductor, as well as the conduction, convection, and radiation of heat inside and on the surface of the object. For example, in the temperature field simulation of the plum blossom contact, the coupled equations of heat conduction, heat convection, and heat radiation can be solved to simulate the temperature rise under different currents.

[0101] Step S1.2: Deploy the sensor array and establish a data channel;

[0102] A multi-sensor array, including temperature sensors, current transformers, voltage sensors, and partial discharge sensors, is deployed in key areas inside the switch cabinet. These sensors collect operational data in real time and continuously transmit the data to the core control unit through a communication module. This step ensures the foundation for data flow between the physical entity and the virtual model.

[0103] Based on CAD drawings and 3D laser scanning point cloud data, a high-precision 3D real-scene model of key components, from the external cabinet to the internal contacts and busbars, was realized. The dynamic model, which integrates physical laws, embeds multi-physics field models based on Joule's law and electrical contact theory, such as electric field, magnetic field and temperature field, into the geometric model, which can simulate the actual working state. Furthermore, by deploying a sensor array and establishing a stable data channel, the virtual model can be updated synchronously with the changes in the state of the physical equipment.

[0104] Thus, traditional switchgear maintenance relies heavily on regular inspections and manual experience, which can not only fail to detect potential hazards in a timely manner, but may also lead to over-maintenance or neglect of maintenance. By building such a digital twin model, we can shift from passively responding to faults to proactively predicting and warning. It can simulate the state of switchgear under various operating conditions, deeply integrate the massive amount of operational data collected by sensors with models based on physical laws, and use data analysis to assess the health status of equipment and generate valuable maintenance suggestions.

[0105] In step S2, constructing the physical information neural network model specifically involves the following steps:

[0106] Step S2.1: This network architecture directly embeds the physical laws from Step S1.1, such as energy conservation and heat transfer equations, as constraints into the loss function of the neural network, rather than relying solely on pure data-driven processes. The input layer of the network receives real-time sensor data, such as current and ambient temperature. The hidden layer learns the complex mapping between physical laws and observed data, while the output layer predicts key state parameters, such as the real-time temperature and equivalent resistance of the contacts. This results in a digital twin model, which is a virtual simulation model based on a joint geometry-physics architecture, used to map the switchgear status in real time.

[0107] The neural network refers to a fully connected deep learning model that takes sensor data as input and state parameters as output. It embeds physical laws into a loss function, specifically implemented as follows: the overall loss function L of the physical information neural network consists of the data inverse sum loss and the physical equation residual loss, expressed as L=λ1*L data +λ2*L phy L data L represents the mean square error between the network-predicted state parameters, such as temperature, and the sensor's measured values. phy To calculate the mean square error of the residual term obtained by substituting the network's predicted output into the physical control equation, λ1 and λ2 are weight coefficients. λ1 can be 0.7 and λ2 can be 0.3. The gradient of the loss function with respect to the network weights is calculated using automatic differentiation, and the network parameters are optimized using the backpropagation algorithm so that the network satisfies physical laws while fitting the data.

[0108] Step S2.2: Next, model parameter calibration and verification are performed. Using the switchgear factory test data, historical normal operation data, and some known minor anomaly data, the constructed physical information neural network model is trained and calibrated. The internal parameters of the model are adjusted using the Simulink Design Optimization tool to minimize the data error of the digital twin simulation output, such as minimizing the error between the contact temperature rise curve and the actual sensor measurement data, to ensure that the virtual model can reflect the real-time status of a specific switchgear with high fidelity. The historical normal data can be 100,000 sensor data points from the switchgear over the past year as the training set, with a sampling frequency of 1Hz.

[0109] In step S3, injecting the fault mode into the digital twin model specifically involves:

[0110] In the calibrated digital twin model in step S2.2 above, the system injects a preset fault mode, that is, introduces a specific fault mode into the digital twin model, such as directly adjusting the model parameters, such as gradually increasing the contact resistance of the contacts;

[0111] The definition and library construction of failure modes include the establishment of failure modes based on FMEA and historical failure data. For example, for switch cabinet contacts, typical failure modes include "increased contact resistance due to oxidation and wear", for mechanical components, it may be "sticking of operating mechanism", and for insulating materials, it is "decreased dielectric strength". These failure modes need to be represented by corresponding parameters or sub-models in the digital twin model.

[0112] For example, in order to simulate the trend of increased resistance due to contact aging, the resistance parameter value can be gradually increased in the sub-model representing the contact to simulate the increase in contact resistance caused by oxidation and wear. At the same time, it can also simulate other common faults such as the decrease in dielectric strength of insulating materials and wear of mechanical parts.

[0113] In the virtual environment of the digital twin model, by accelerating simulation, the digital twin injected with fault modes is run to simulate the entire process of the switch cabinet gradually deteriorating from a healthy state under various potential operating conditions. This process can generate a large amount of fault state data that is difficult to obtain in the real world, especially potential fault data that develops slowly and has no obvious early signs.

[0114] The most direct way to inject is to directly modify the input parameters of the digital twin model through the application programming interface or script. For example, in order to simulate contact aging, a script can be written to gradually increase the parameter value representing the contact resistance during the simulation process, or algorithms such as support vector regression can be used to dynamically calculate and compensate for the increase in contact resistance based on factors such as simulation running time, number of actions, and load current, so that the aging simulation is more in line with physical laws.

[0115] Further, time-scale acceleration can be performed. Time-scale acceleration is a common technique, which involves making virtual time faster than real time in the simulation, thereby obtaining aging data equivalent to several years of actual operation in a few hours. Stress intensification accelerates deterioration by applying higher loads. For example, in the simulation, the switchgear is made to operate continuously at a higher rated current to quickly expose thermal aging problems. This method is particularly suitable for obtaining potential fault data in the early stages of fault development, those that change very slowly under normal conditions and are difficult to detect.

[0116] In step S4, the real-time sensor data, such as current and temperature, is continuously input into the digital twin model, and the comprehensive health index of the switchgear is calculated, including the following steps:

[0117] Step S4.1: First, extract key health feature parameters;

[0118] From the synchronously updated digital twin model, key feature parameters that can quantify equipment performance degradation are extracted, including instantaneous deviation features and trend evolution features; instantaneous deviation features: reflect the absolute difference between the current state and the ideal health benchmark;

[0119] For example, contact temperature deviation ΔT: Calculates the contact temperature T simulated by the digital twin under the current load current. S The reference temperature T of a brand new contact under the same operating conditions b The deviation, i.e., ΔT = T S -T b ;

[0120] Contact resistance deviation rate, calculate the simulated contact resistance R of the contact. s Relative to its initial health value R i The rate of change, i.e., ΔR% = (R s -R i ) / R i ×100%;

[0121] Trend evolution characteristics: Reflecting the long-term rate of change of equipment condition parameters, predicting potential aging processes, such as the contact resistance trend slope kR, over a continuous period, such as the past 30 days, for the contact resistance time series data (R1, R2, ..., R...) output by the digital twin. n Linear regression analysis (least squares method) is performed, and the slope kR of the fitted line is the slope of the trend. A positive value indicates that the resistance is increasing.

[0122] The acceleration of temperature change αT is used to perform quadratic curve fitting or calculate the second derivative of the time series data of the contact temperature deviation ΔT to determine whether there is an acceleration in the deterioration process.

[0123] Step S4.1, which involves extracting key health feature parameters, further includes:

[0124] Step S4.1.1: First, preprocess the key feature parameters and assign weights to the feature data;

[0125] Data standardization is necessary because different feature parameters have different dimensions and orders of magnitude. Z-score standardization is used to transform all feature values ​​to the same dimension, eliminating their incommensurability. The formula is x... std = (x-μ) / σ, where X is the original feature value, μ is the historical mean of the feature, and σ is the standard deviation;

[0126] Step S4.1.2, Feature weight configuration: Based on the degree of influence of each feature parameter on the overall health status of the switchgear, assign weights to them. The weight configuration can be based on the feature importance assessment and dynamically adjusted according to the equipment type or operating history to ensure that the sum of all feature weights is 100%.

[0127] Step S4.1.3: Based on the preprocessed key feature parameters and their feature weight configuration, calculate the comprehensive health index HI using a weighted average model;

[0128] The health index is based on a composite model that comprehensively considers the physical model of the equipment and its real-time status;

[0129] HI=∑(ω i ×F i (P i )); where P i Let be the i-th preprocessed feature parameter value;

[0130] F i () is a characteristic function for the i-th parameter, used to map the actual value of the parameter to a sub-score or health score that reflects its health level. For example, for the temperature parameter, F T (T) may be an inverse S-shaped function, scoring highly when the temperature is below the safe threshold and dropping sharply when it approaches or exceeds the danger threshold.

[0131] ω i Let ω be the weight coefficient of the i-th feature, satisfying ∑ω i =1. The allocation of weights is crucial and can be based on expert experience, such as temperature, where local factors contribute more to the failure and therefore have a larger weight, or obtained through correlation analysis or learning from historical failure data via machine learning.

[0132] The calculated health index HI needs to be compared with a preset threshold to determine the current health status and decide whether to trigger an alarm. Collect the switch cabinet's full life cycle data in terms of health status, known fault status, and aging process, including historical sensor data and corresponding maintenance records.

[0133] The system is operating normally when the health threshold HI ≥ 80.

[0134] If the warning threshold is 50≤HI<80, the system is experiencing performance degradation; therefore, monitoring should be strengthened.

[0135] When the alarm threshold HI < 50, the system health status has seriously deteriorated, and the early warning module automatically sends an interrupt signal to the diagnostic module.

[0136] The system first collects raw data in real time from a multi-sensor array deployed in key parts of the switchgear, such as circuit breaker contacts, busbar connection points, and cable joints. These characteristic parameters mainly include:

[0137] Electrical parameters: three-phase current, voltage, zero-sequence current, and leakage current;

[0138] Thermal parameters: Temperatures of key points such as contacts, busbars, and cable joints, labeled as [T1, T2, ..., T...]. n ];

[0139] Mechanical and insulation parameters: vibration signal (used to detect looseness or abnormality in the mechanism), partial discharge PD signal (used to assess insulation degradation);

[0140] Operating history parameters: number of short-circuit current interruptions, cumulative number of operations, and historical load current curve;

[0141] Before calculation, the raw data needs to be preprocessed, including data cleaning (removing obvious error outliers), standardization (eliminating the influence of dimensions, such as converting temperature, current, etc. to the [0,1] interval) and feature extraction, such as extracting effective values, peak values, and frequency components from vibration signals.

[0142] Real-time data is input into a digital twin model built on physical laws such as thermodynamics and electrodynamics. This model can not only output the simulated value of the current state, but also predict the state evolution trend over a period of time. The health index can be calculated comprehensively based on the deviation between the current measured value and the simulated value, as well as the rate of deterioration of the predicted trend.

[0143] When the diagnostic module receives the comprehensive health index as the alarm threshold, it will analyze which characteristic parameters (such as a surge in C-phase current, an abnormal increase in temperature at a certain point in C-phase, accompanied by a strong partial discharge signal) caused the change in HI.

[0144] Step 1: The diagnostic module acquires all raw sensor data of HI within a time window before and after the change; this data includes, but is not limited to, three-phase current, voltage waveform data, temperature sequence of key points such as contacts and busbars, vibration acceleration signals, and partial discharge pulse signals; the time window before and after the change is 10 seconds.

[0145] Since different sensors have different sampling frequencies and data acquisition times, data synchronization and alignment are performed first. The engine uses interpolation to unify all data onto the same timestamp sequence, forming a strictly synchronized multi-dimensional time series data matrix.

[0146] Data cleaning is performed on the aligned data, including removing impulse noise obviously caused by interference and filling in a small amount of missing data caused by brief communication interruptions, so as to provide a high-quality data foundation for subsequent analysis.

[0147] Step II, the diagnostic module also extracts time-domain and frequency-domain features: the diagnostic module not only focuses on the instantaneous values ​​of parameters, but also extracts feature quantities that can characterize the status of feature devices in greater depth;

[0148] For example, by performing a fast Fourier transform on the vibration signal, the energy value of a specific frequency band can be extracted to identify the characteristic frequencies of mechanical loosening or bearing wear.

[0149] Analyze current and voltage signals to calculate harmonic content and total harmonic distortion rate in order to determine load characteristics or power quality issues.

[0150] For partial discharge signals, extract characteristic patterns such as discharge quantity, discharge frequency, and phase distribution;

[0151] For the temperature data, calculate its temperature rise frequency and temperature difference relative to the reference current;

[0152] Then, all the extracted time-domain and frequency-domain features are combined with the raw values ​​directly measured by the sensors to construct a high-dimensional time-series feature matrix, which fully describes the overall operating status of the switchgear within the analysis time window;

[0153] Step III: Then, the time series feature matrix is ​​input into a pre-trained deep learning model based on the attention mechanism. This model can automatically learn and assign different feature parameters to the contribution or attention weight of HI change in a specific fault mode. For example, in the case of suspected contact overheating, the model will assign higher attention weights to parameters such as contact temperature and contact current, thereby focusing on the core feature most likely to cause HI to decrease among many parameters.

[0154] Step IV: The diagnostic module includes a fault mode knowledge base, which is derived from historical fault data, expert experience and digital twin simulation data. The knowledge base defines various typical faults, such as poor contact of contacts, moisture in insulation and jamming of mechanisms, and the corresponding characteristic parameter combinations and typical change modes.

[0155] The current high-weight feature combination and its changing trend obtained in steps II and III are compared with the patterns in the fault knowledge base for similarity calculation. Through retrieval and matching, the diagnostic module finally outputs the most likely fault diagnosis conclusion. For example, there is a 92% probability that the increased contact resistance of the C-phase cable joint caused the joint temperature to rise sharply from 75°C to 112°C within 15 minutes under a load current of 1250A, which led to a significant increase in partial discharge activity and was the root cause of this sudden drop in HI.

[0156] Step V: The diagnostic module generates a diagnostic report and decision support.

[0157] The report includes the primary cause, namely the first abnormal parameter that leads to the decrease in HI, other parameters that are strongly correlated with the cause, such as C-phase current or partial discharge, failure probability, and the reliability of the diagnostic results.

[0158] The report, which includes the evolution trend and historical change curves of key parameters, is directly pushed to the human-machine interface.

[0159] If the input C-phase current is 1250A and the temperature rises from 75℃ to 112℃, the output diagnostic report indicates a 92% probability of poor contact.

[0160] In step II, the time-series feature matrix includes:

[0161] Step II-I: First, clean the raw data of the multi-sensor array to remove noise and obvious outliers; handle missing values, fill them with interpolation, and normalize the data.

[0162] Step II-II: Use the sliding window technique to divide the continuous time series data into multiple time segments. For example, set a window size (e.g., 60 minutes) and a step size (e.g., 5 minutes), and extract data according to this rule to form a series of continuous, possibly overlapping data windows.

[0163] Sliding window segmentation: The data within each window constitutes the basic unit for feature extraction. Assuming there are n monitoring points, such as different monitoring points in a switch cabinet, and the time step within the window is t. w If each time step has f original features, then a preliminary four-dimensional tensor representation of the time series data block can be obtained, whose dimension can be represented as [n, n]. windows , t w ,f];

[0164] Multi-dimensional feature extraction and enhancement: This is the core of building an information-rich feature matrix, which aims to extract multiple types of features from each data window, including statistical features, time-series features, domain-specific features, and automated features;

[0165] Raw sensor values: The sensor readings at each time point within the window are directly retained as the basic features;

[0166] Steps II and III calculate the mean, standard deviation, maximum, minimum, and quantiles of the data within the window to describe the basic distribution characteristics of the data over that time period;

[0167] After steps II-IV, historical data is introduced as features, such as the values ​​of the previous moment and the two moments before, to capture short-term dependencies; then, finer-grained sliding is performed within the window, and calculations such as moving average and moving standard deviation are performed to smooth noise and highlight trends.

[0168] Steps II-V involve concatenating all types of features extracted from the same time window along the feature dimension to form a wide vector representing the overall state of that window; stacking the feature vectors corresponding to all time windows in chronological order to form a complete temporal feature matrix. Each row of this matrix corresponds to a time window, and each column corresponds to a feature variable, providing standardized input for the subsequent diagnostic engine.

[0169] The deep learning model establishment in step III includes the following steps:

[0170] Step III-I: Data preparation and preprocessing, basic feature extractor, using one-dimensional convolutional neural network or long short-term memory network as the backbone network. 1D-CNN is good at extracting local feature patterns, and LSTM is good at capturing long-term dependencies in time series.

[0171] Step III-II: After the feature extractor, an attention layer is connected. This attention layer will calculate an attention weight for each feature parameter at each time step in the temporal feature matrix. These weights are normalized by functions such as Softmax to represent the importance of each feature to the final decision at a specific time.

[0172] Step III-III: Multiply the learned attention weights with the original features and perform a weighted sum to obtain a "context vector" that can represent the key information of the entire time window. Finally, this vector is fed into a fully connected layer to output the predicted value of the health index or the probability of fault classification.

[0173] Steps III-IV: Following this, model training and optimization are performed. If predicting the HI value, the mean squared error loss function is used to measure the difference between the model's predicted value and the true label. For classification tasks, such as determining the fault type, cross-entropy loss is commonly used. The Adam and SGD optimization algorithms are used, and all parameters in the model are continuously adjusted through backpropagation to minimize the loss function. After training, the attention weights are visualized. For example, when analyzing cases of contact overheating, it may be found that the model assigns extremely high attention weights to contact temperature and the current characteristics of that contact.

[0174] The deep learning model based on the attention mechanism consists of an input layer, a one-dimensional convolutional layer, a long short-term memory network layer, an attention layer, and a fully connected output layer. The attention layer is used to calculate the attention weights at each time step in the temporal feature matrix and inputs the weighted summed context vector into the fully connected layer for fault classification and health index prediction.

[0175] The specific method for establishing the fault mode knowledge base in step IV is as follows:

[0176] The switchgear system is considered as a whole, and then decomposed layer by layer into subsystems, assemblies and individual components. For example, it can be decomposed into subsystems such as "circuit breaker mechanism", "insulation system" and "contact system". The "contact system" can be further decomposed into "moving contact" and "stationary contact". This hierarchical decomposition helps to accurately identify the specific location of the fault.

[0177] Furthermore, based on importance and failure frequency, identify the key components and subsystems that need to be focused on in the initial stage of the knowledge base to avoid making the scope too broad and difficult to start.

[0178] By collecting and organizing historical work orders, maintenance records, and accident reports, and injecting fault modes into digital twin models, we can obtain fault data that is difficult to obtain in actual applications, especially slow-developing and early-stage latent fault data, which greatly enriches the breadth and depth of the knowledge base.

[0179] Establish the association between fault modes and combinations of characteristic parameters and change patterns. For example, a fault mode usually triggers changes in multiple characteristic parameters. "Poor contact" may simultaneously lead to "abnormal increase in contact temperature", "temperature rise exceeding normal value when passing the same current", and "increased activity of partial discharge signal". The knowledge base needs to record these combinations of characteristic parameters.

[0180] Then, an SQL database was used for storage, and multiple related data tables were designed. Offline confirmed fault cases and real feature data were fed back into the knowledge base and continuously evolved.

[0181] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent operating system for low-voltage switchgear, characterized in that: It includes a data acquisition module, a storage module, a prediction module, an early warning module, and a diagnostic module; Data acquisition module: The data acquisition module is used to collect multi-source operating data of the low-voltage switchgear in real time. The multi-source operating data includes operating parameters and status information. Storage module: The storage module is used to store the multi-source operating data of the switchgear collected by the data acquisition module, as well as the configuration parameters and control program of the switchgear; Prediction Module: The prediction module predicts potential faults of the switchgear itself based on multi-source operational data acquired by the data acquisition module, data from the storage module, and its own pre-trained digital twin model, and outputs the prediction results to the early warning module. Early warning module: The early warning module outputs an early warning signal based on the predicted fault results output by the prediction module, and transmits the early warning signal to the diagnostic module; Diagnostic module: The diagnostic module analyzes the faults of the switchgear itself based on the early warning signal output by the early warning module and generates a diagnostic report.

2. The intelligent operating system for low-voltage switchgear according to claim 1, characterized in that: The prediction module predicts potential faults in the switchgear based on a digital twin model, specifically as follows: Step S1: Construct a joint geometric-physical model of the switchgear; Step S2: Construct a physical information neural network model based on the switch cabinet's geometric-physical joint model, and generate a digital twin model through the physical information neural network model; Step S3: Inject the pre-designed fault modes into the digital twin model to simulate the entire process of the switchgear gradually deteriorating from a healthy state under various potential operating conditions. Step S4: Continuously input the real-time collected sensor data into the digital twin model and calculate the overall health index of the switchgear.

3. The intelligent operating system for low-voltage switchgear according to claim 2, characterized in that: In step S1, constructing the geometric-physical joint model of the switchgear specifically involves the following steps: Step S1.1: First, perform digital modeling and data acquisition for the switchgear: The process involves using computer-aided design drawings of switchgear, combined with 3D laser scanning technology, to obtain high-precision geometric models of the switchgear and key components, and then embedding physical laws based on these high-precision geometric models. Step S1.2: Deploy the sensor array and establish a data channel: Multiple sensor arrays, including temperature sensors, current transformers, voltage sensors, and partial discharge sensors, are deployed in key areas inside the switch cabinet. In step S2, constructing the physical information neural network model specifically involves the following steps: Step S2.1: The network architecture directly embeds the physical laws from step S1.1 above as constraints into the loss function of the neural network. The input layer of the network receives real-time sensing data, the hidden layer learns the complex mapping between physical laws and observation data, and the output layer predicts key state parameters. After training, a digital twin model is obtained. Step S2.2: Next, model parameter calibration and verification are performed. Using the switchgear factory test data, historical normal operation data, and some known minor anomaly data, the constructed physical information neural network model is trained and calibrated. The internal parameters of the model are adjusted using the Simulink Design Optimization tool to minimize the data error of the digital twin simulation output and ensure that the virtual model can reflect the real-time status of a specific switchgear with high fidelity.

4. The intelligent operating system for low-voltage switchgear according to claim 3, characterized in that: In step S3, injecting the fault mode into the digital twin model specifically involves: In the digital twin model calibrated in step S2.2 above, specific failure modes are introduced into the digital twin model; In the virtual environment of the digital twin model, by accelerating simulation, a digital twin injected with fault modes is run to simulate the entire process of the switchgear gradually deteriorating from a healthy state under various potential operating conditions.

5. The intelligent operating system for low-voltage switchgear according to claim 3, characterized in that: In step S4, the real-time collected sensor data is continuously input into the digital twin model, and the comprehensive health index of the switchgear is calculated, including the following steps: Step S4.1: Extract key health feature parameters; From the synchronously updated digital twin model, extract key feature parameters that can quantify equipment performance degradation, including instantaneous deviation features and trend evolution features; Instantaneous deviation characteristics: reflect the absolute difference between the current state and the ideal health benchmark; Trend evolution characteristics: Reflects the long-term rate of change of equipment status parameters.

6. The intelligent operating system for low-voltage switchgear according to claim 5, characterized in that: Step S4.1, extracting key health feature parameters, further includes: Step S4.1.1: Preprocess the key feature parameters and assign weights to the feature data; Z-score normalization is used to transform all key feature parameters to the same dimension, as shown by the formula x. std = (x-μ) / σ, where X is the original feature value, μ is the historical mean of the feature, and σ is the standard deviation; Step S4.1.2, Feature weight configuration: Based on the degree of influence of each feature parameter on the overall health status of the switchgear, assign weights to them; Step S4.1.3: Based on the preprocessed key feature parameters and their feature weight configuration, calculate the comprehensive health index HI using a weighted average model; HI=∑(ω i ×F i (P i )); where P i Let be the i-th preprocessed feature parameter value; F i () is a feature function for the i-th parameter, used to map the actual value of the parameter to a sub-score or health degree that reflects its health level; ω i Let ω be the weight coefficient of the i-th feature, satisfying ∑ω i =1; The calculated comprehensive health index HI is compared with a preset threshold to determine the current health status and decide whether to trigger an alert. If the comprehensive health index HI ≥ 80, the system is operating normally; If 50≤HI<80, the system experiences performance degradation and the signal is transmitted to the diagnostic module. If the comprehensive health index HI < 50, the system's health status has seriously deteriorated, and the signal will be transmitted to the diagnostic module.

7. The intelligent operating system for low-voltage switchgear according to claim 6, characterized in that: When the diagnostic module receives an alarm signal, it analyzes the reasons for the changes in the Comprehensive Health Index (HI), including the following steps: Step 1: The diagnostic module acquires all raw sensor data of the comprehensive health index HI within a time window before and after the change, where the time window is 10 seconds before and after the change. Step II: The diagnostic module then extracts time-domain and frequency-domain features; and constructs a high-dimensional time-series feature matrix by combining all extracted time-domain and frequency-domain features with the raw values ​​directly measured by the sensor. Step III: Then, the temporal feature matrix is ​​input into a pre-trained deep learning model based on an attention mechanism. This model can automatically learn and assign the contribution of different feature parameters to the HI change under a specific fault mode. Step IV, the diagnostic module also includes a fault mode knowledge base; The current high time-domain and frequency-domain features obtained in steps II and III, along with the contribution of feature parameters to HI changes under specific fault modes, are compared with the patterns in the fault knowledge base. Through retrieval and matching, the diagnostic module finally outputs the fault diagnosis conclusion. Step V: The diagnostic module generates a diagnostic report.

8. The intelligent operating system for low-voltage switchgear according to claim 7, characterized in that: In step II, the construction of the time-series feature matrix includes the following steps: Step II-I: First, clean the raw data of the multi-sensor array to remove noise and obvious outliers; handle missing values, fill them with interpolation, and normalize the data. Step II-II: Then, the sliding window technique is used to divide the continuous time series data into multiple time segments, forming a series of continuous, overlapping data windows; Furthermore, by using a sliding window segmentation method, the data within each window constitutes the basic unit for feature extraction; Steps II-III: Calculate the mean, standard deviation, maximum, minimum, and quantile of the data within each window to describe the basic distribution characteristics of the data over that time period; Steps II-IV: Introduce historical data as features; concatenate all types of features extracted from the same time window along the feature dimension to form a wide vector representing the overall state of the window; stack the feature vectors corresponding to all time windows in chronological order to form a complete time-series feature matrix, where each row of the matrix corresponds to a time window and each column corresponds to a feature variable.

9. The intelligent operating system for low-voltage switchgear according to claim 8, characterized in that: The deep learning model establishment in step III includes the following steps: Step III-I: Data preparation and preprocessing, using a one-dimensional convolutional neural network or a long short-term memory network as the backbone network; Step III-II: After the feature extractor, an attention layer is connected. This attention layer will calculate an attention weight for each feature parameter at each time step in the temporal feature matrix. Step III-III: Multiply the learned attention weights with the original features and perform a weighted sum to obtain a "context vector" that represents the key information of the entire time window. Finally, this vector is fed into a fully connected layer to output the predicted value of the health index. Steps III-IV: Model training and optimization are then performed.

10. The intelligent operating system for low-voltage switchgear according to claim 9, characterized in that: The specific method for establishing the fault mode knowledge base in step IV is as follows: Treat the switchgear system as a whole, and then break it down layer by layer into subsystems, assemblies and individual components; Collect and organize historical work orders, maintenance records, and accident reports, and inject fault modes into the digital twin model to establish the association between fault modes and characteristic parameter combinations and change patterns. Then, use an SQL database for storage, design multiple related data tables, and incorporate offline confirmed fault cases and real characteristic data.

Images