Intelligent interconnected switch system based on deep learning and working method

Abstract

The application relates to the technical field of smart grids, in particular to a smart interconnected switch system based on deep learning and a working method thereof, which comprises the following steps: a data capturing module collects time sequence data and state data of equipment operation; a data processing expansion module forms migration data by collecting fault data, generates simulation data by using a GANs model, and constructs a data set in combination with equipment historical data; a diagnosis and prediction module inputs the data set into a DualLSTM-DropNet model, generates equipment state parameter prediction results, performs fault diagnosis, health state evaluation and equipment residual life prediction on high-voltage switch equipment, and generates a diagnosis conclusion; and an intelligent intervention and interconnection module judges whether to intervene in the equipment, and if intervention is needed, a DQN algorithm is used to output a hierarchical intervention scheme. Therefore, the problems of insufficient effective data samples and single diagnosis algorithm in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, specifically to a deep learning-based intelligent interconnected switch system and its operating method. Background Technology

[0002] High-voltage switchgear is crucial in power systems, and its operational stability is affected by multiple factors. Structurally, it encompasses several complex systems, including mechanical, mechanical, transmission, electrical, insulation, and safety protection components. During operation, the internal structure undergoes performance changes over time and with repeated use. For example, frequent opening and closing operations can cause wear on mechanical parts, and electrical connection points may experience overheating and oxidation due to prolonged energization. External environmental factors are equally significant. Voltage and current fluctuations can impact electrical performance, electromagnetic interference can affect control and monitoring systems, and environmental factors such as temperature, humidity, air pressure, and vibration can interact with equipment performance. High humidity environments can reduce equipment insulation performance and increase the risk of short circuits. Early maintenance relied primarily on manual inspections and periodic maintenance. Manual inspections can only detect surface problems and are difficult to detect potential internal faults. Periodic maintenance suffers from the problem of not being able to promptly detect and address faults occurring between maintenance periods, failing to meet the actual needs of ensuring stable equipment operation.

[0003] With technological advancements, power systems have entered an era of automated and intelligent monitoring, enabling real-time data acquisition and status monitoring by installing sensors on equipment. However, existing intelligent diagnostic technologies suffer from several challenges. Regarding data, the available diagnostic data samples are limited, making it difficult for diagnostic models to fully learn the characteristics of normal and fault states. Data validity and correlation are insufficient; the collected data may contain a large amount of redundant or invalid information, such as data affected by environmental noise. As for diagnostic algorithms, most currently employ single algorithms, which are insufficient for handling complex equipment conditions and diverse fault types. They fail to fully exploit the potential characteristics of the data, resulting in low fault diagnosis accuracy and poor practicality, impacting condition-based maintenance, predictive maintenance, and full lifecycle management of high-voltage switchgear. Summary of the Invention

[0004] This application provides a deep learning-based intelligent interconnected switch system and its working method to solve the problems of insufficient effective data samples, low diagnostic accuracy, and single diagnostic algorithm in the prior art.

[0005] The first aspect of this application provides a deep learning-based intelligent interconnected switch system, comprising: a data capture module, a data processing expansion module, a diagnostic prediction module, and an intelligent intervention and interconnection module; wherein, the data capture module is used to collect time-series data and status data of equipment operation; the data processing expansion module is used to collect cross-domain fault data of similar equipment, integrate it to form migration data, use GANs models to generate simulation data, expand the data source, and construct a dataset by combining historical equipment data; the diagnostic prediction module is used to input the dataset into a DualLSTM-DropNet model that integrates multi-scale attention mechanisms to generate equipment status parameter prediction results, perform fault diagnosis, health status assessment, and remaining life prediction of high-voltage switch equipment, and generate diagnostic conclusions; the intelligent intervention and interconnection module determines whether to intervene in the equipment based on the diagnostic conclusions. If intervention is required, it uses the DQN algorithm, combined with real-time equipment operating conditions and historical intervention effect feedback, to dynamically output a graded intervention plan. At the same time, the graded intervention plan is transmitted to the control terminal via the Internet of Things for dynamic intervention.

[0006] Preferably, the data acquisition module includes a sensor selection and configuration unit and a real-time data acquisition unit. The sensor selection and configuration unit is used to select suitable electrical and non-electrical sensors and determine their installation locations and quantities according to non-intrusive monitoring requirements. The real-time data acquisition unit is used to collect time-series data and status data of the device operation in real time.

[0007] Preferably, the data processing expansion module includes a data migration unit, a simulation data generation unit, a data optimization unit, and an integration and construction unit. The data migration unit collects fault data from similar high-voltage switchgear in different regions, under different operating conditions, and at different lifespans, and integrates this data to form migration data. The simulation data generation unit generates simulation data based on the time-series and status data of the equipment operation using a GANs model. The data optimization unit filters out abnormal fluctuations in the simulation data and migration data using IQR; it quantifies the distribution differences between the simulation data and real data using JS divergence, and verifies the authenticity of the simulation data using a logistic regression model and a confusion matrix. The integration and construction unit integrates the migration data, the simulation data, and the time-series and status data of the equipment operation, and constructs a dataset by combining historical equipment data.

[0008] Preferably, the formula for the GANs model is:

[0009] ;

[0010] in, A discriminant for real and simulated data; A generator for simulation data; For generator The minimum value; For discriminator The maximum value; It is a value function; For discriminator For real data The expected logarithm of the judgment result; It is the natural logarithm; For discriminator For input data The probability of being judged as real data; noise distribution Expectations; Represents generator Using noise The generated data.

[0011] Preferably, the diagnostic prediction module includes an extraction prediction unit and an evaluation diagnosis unit. The extraction prediction unit inputs the dataset into the DualLSTM-DropNet model, uses bidirectional LSTM layers to capture temporal dependencies in the data, and integrates a multi-scale attention mechanism to enhance the extraction of features from different periods, outputting equipment state parameter prediction results. The evaluation diagnosis unit evaluates the equipment state parameter prediction results using the RMSE index and R² coefficient, and performs fault diagnosis, health status assessment, and remaining life prediction for the high-voltage switchgear based on the state parameter prediction results, forming a diagnostic conclusion.

[0012] Preferably, the intelligent interconnection and intervention module includes an intervention decision unit and a data transmission unit. The intervention decision unit determines whether to intervene based on the diagnostic conclusion. If intervention is required, it dynamically outputs a graded intervention plan by combining the DQN algorithm with the real-time operating conditions of the equipment and the feedback of historical intervention effects. The data transmission unit transmits the graded intervention plan to the control terminal via the Internet of Things.

[0013] The second aspect of this application provides a method for operating intelligent interconnected switches based on deep learning, comprising: acquiring time-series data and state data of equipment operation; processing the time-series data and state data, inputting them into a GANs model to generate simulation data, and simultaneously collecting cross-domain fault data of similar high-voltage switchgear in different regions, under different operating conditions, and at different lifespans, integrating them to form migration data; using IQR to filter out fluctuation and abnormal data in the simulation data and migration data, and using JS divergence to quantify the distribution differences between the simulation data and the time-series data and the state data; using logistic regression algorithm and confusion matrix to verify the authenticity of the simulation data; and constructing a dataset by combining historical equipment data; inputting the dataset into a DualLSTM-DropNet model with a multi-scale attention mechanism to generate state parameter prediction results, and performing fault diagnosis, health status assessment, and remaining lifespan prediction on the high-voltage switchgear based on the state parameter prediction results to form a diagnostic conclusion; and determining whether to intervene in condition-based maintenance, predictive maintenance, and full life-cycle management of the equipment based on the diagnostic conclusion. If intervention is required, using the DQN algorithm, combined with real-time equipment operating conditions and historical intervention effect feedback, to generate a graded intervention plan.

[0014] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement a deep learning-based intelligent interconnection switch operation method as described in the above embodiments.

[0015] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement a deep learning-based intelligent interconnection switch operation method as described in the above embodiments.

[0016] A fifth aspect of this application provides a computer program product, including a computer program or instructions, for implementing a deep learning-based intelligent interconnection switch operation method as described in the above embodiments.

[0017] Therefore, this application has the following beneficial effects:

[0018] This application's embodiments utilize a data capture module to collect time-series and status data of equipment operation using adapted power and non-power sensors, improving the comprehensiveness and real-time acquisition capability of equipment operation data. A data processing expansion module filters out abnormal data using IQR, combines JS divergence to quantify distribution differences, and verifies data authenticity using a logistic regression model and confusion matrix, generating datasets to provide data support for diagnostic prediction. Simulation data is generated through GANs models, combined with cross-domain fault migration data, expanding data sources and improving data richness and diversity. The diagnostic prediction module inputs the dataset into a DualLSTM-DropNet model incorporating multi-scale attention mechanisms, improving the accuracy of fault diagnosis, health status assessment, and remaining life prediction. The intelligent intervention and interconnection module uses the DQN algorithm combined with real-time operating conditions and historical feedback to generate tiered intervention plans, dynamically intervening via IoT transmission to improve the timeliness and targeted nature of equipment operation and maintenance. This solves the problems of insufficient effective data samples, low diagnostic accuracy, and limited diagnostic algorithms in existing technologies.

[0019] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0020] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein...

[0021] Figure 1 This is a schematic diagram of a deep learning-based intelligent interconnected switch system provided according to an embodiment of this application.

[0022] Figure 2 This is a schematic diagram of a data capture module provided according to an embodiment of this application.

[0023] Figure 3 This is a schematic diagram of high-voltage switchgear data acquisition according to an embodiment of this application.

[0024] Figure 4 This is a schematic diagram of a data processing extension module provided according to an embodiment of this application.

[0025] Figure 5 This is a schematic diagram of data processing for a high-voltage circuit breaker according to an embodiment of this application.

[0026] Figure 6 This is a schematic diagram of a diagnostic prediction module provided according to an embodiment of this application.

[0027] Figure 7 This is a schematic diagram illustrating predictive diagnostics of a high-voltage circuit breaker according to an embodiment of this application.

[0028] Figure 8 This is a schematic diagram of an intelligent intervention and interconnection module provided according to an embodiment of this application.

[0029] Figure 9 This is a schematic diagram of graded intervention of a high-voltage circuit breaker according to an embodiment of this application.

[0030] Figure 10 This is a schematic diagram of a deep learning-based intelligent interconnected switch system provided according to an embodiment of this application.

[0031] Figure 11 This is a flowchart illustrating a deep learning-based intelligent interconnection switch operation method according to an embodiment of this application.

[0032] Figure 12 This is a schematic diagram illustrating the verification of simulation data according to an embodiment of this application.

[0033] Figure 13 This is a schematic diagram illustrating a deep learning-based intelligent interconnection switch operation method according to an embodiment of this application.

[0034] Figure 14 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] The following description, with reference to the accompanying drawings, illustrates an embodiment of a deep learning-based intelligent interconnected switch system and its operating method. Addressing the issue of insufficient effective data samples mentioned in the background section, this application provides a deep learning-based intelligent interconnected switch system. In this system, a data acquisition module utilizes adapted power and non-power sensors to collect time-series and status data of the equipment operation, improving the comprehensiveness and real-time acquisition capability of the equipment operation data. A data processing expansion module filters out abnormal data using IQR, combines JS divergence to quantify distribution differences, and verifies data authenticity using a logistic regression model and confusion matrix, generating datasets to provide data support for diagnostic prediction. Simulation data is generated through GANs models, combined with cross-domain fault migration data, expanding the data source and improving data richness and diversity. The diagnostic prediction module inputs the dataset into a DualLSTM-DropNet model incorporating multi-scale attention mechanisms, improving the accuracy of fault diagnosis, health status assessment, and remaining life prediction. The intelligent intervention and interconnection module uses the DQN algorithm combined with real-time operating conditions and historical feedback to generate tiered intervention plans, dynamically intervening through IoT transmission to improve the timeliness and targeted nature of equipment operation and maintenance. This solves the problems of insufficient effective data samples, low diagnostic accuracy, and single diagnostic algorithm in existing technologies.

[0037] Figure 1 This is a schematic diagram of a deep learning-based intelligent interconnected switch system provided in an embodiment of this application.

[0038] This application provides a deep learning-based intelligent interconnected switch system 10, which includes: a data capture module 100, a data processing expansion module 200, a diagnosis and prediction module 300, and an intelligent intervention and interconnection module 400.

[0039] The data acquisition module 100 is used to collect time-series data and status data of equipment operation; the data processing expansion module 200 is used to collect cross-domain fault data of similar equipment, integrate them to form migration data, use GANs models to generate simulation data, expand data sources, and construct datasets by combining historical equipment data; the diagnosis and prediction module 300 is used to input the dataset into the DualLSTM-DropNet model that integrates multi-scale attention mechanisms to generate equipment status parameter prediction results, perform fault diagnosis, health status assessment and remaining life prediction of high-voltage switchgear, and generate diagnostic conclusions; the intelligent intervention and interconnection module 400 determines whether to intervene in the equipment based on the diagnostic conclusions. If intervention is required, it uses the DQN algorithm, combined with the real-time operating conditions of the equipment and historical intervention effect feedback, to dynamically output a graded intervention plan. At the same time, the graded intervention plan is transmitted to the control terminal through the Internet of Things for dynamic intervention.

[0040] It is understood that, in this embodiment, the data capture module uses adapted power and non-power sensors to collect time-series and status data of equipment operation, improving the comprehensiveness and real-time acquisition capability of equipment operation data. The data processing expansion module uses IQR to filter out abnormal data, combines JS divergence to quantify distribution differences, and uses logistic regression models and confusion matrices to verify data authenticity, generating datasets to provide data support for diagnosis and prediction. Simulation data is generated through GANs models, combined with cross-domain fault migration data, expanding data sources and improving data richness and diversity. The diagnosis and prediction module inputs the dataset into a DualLSTM-DropNet model that integrates multi-scale attention mechanisms, improving the accuracy of fault diagnosis, health status assessment, and remaining life prediction. The intelligent intervention and interconnection module uses the DQN algorithm combined with real-time operating conditions and historical feedback to generate tiered intervention plans, and performs dynamic intervention through IoT transmission, improving the timeliness and targeting of equipment operation and maintenance. Thus, it solves the problems of insufficient effective data samples, low diagnostic accuracy, and single diagnostic algorithms in existing technologies.

[0041] In this embodiment of the application, the data capture module 100 includes: as follows Figure 2 As shown, the sensor selection and configuration unit and the real-time data acquisition unit are shown.

[0042] The sensor selection and configuration unit is used to select suitable electrical and non-electrical sensors and determine their installation location and quantity according to the non-intrusive monitoring requirements; the real-time data acquisition unit is used to collect the time-series data and status data of the equipment operation in real time.

[0043] It is understood that the embodiments of this application use a sensor selection and configuration unit to screen suitable electrical and non-electrical sensors and determine their installation locations and quantities, providing suitable and reasonable hardware support for the collection of equipment operation data, ensuring that the sensors match the monitoring requirements, and avoiding data collection deviations caused by sensor incompatibility or unreasonable locations and quantities; the real-time data acquisition unit collects the time-series data and status data of equipment operation in real time, ensuring that the acquired data is real-time and continuous, reflecting the equipment operating status, and improving the adaptability, real-time performance, and completeness of equipment operation data collection.

[0044] For example, such as Figure 3As shown, when a power company implemented intelligent monitoring for its 10kV high-voltage switchgear, during the sensor selection and configuration phase, based on the requirement of non-intrusive monitoring (avoiding damage to the original structure of the switchgear), the sensor selection and configuration unit selected suitable electrical and non-electrical sensors: Three high-precision Hall current sensors (model CHB-100, range 0-500A, error ≤0.5%) were selected and installed at the incoming, outgoing, and busbar connections of the switchgear to collect the time-series data of the three-phase current; two voltage sensors (model LV28-) were selected... The P sensor (with a range of 0-12kV and linearity ≤0.2%) is installed at the insulating bushing on the incoming line side to collect line voltage data. For non-electrical quantity sensors, four infrared temperature sensors (model MLX90614, temperature range -40~125℃, accuracy ±0.5℃) are configured and mounted on the moving contact, stationary contact, and cable joints on both sides to monitor the temperature of key components. Additionally, one vibration sensor (model SEN-09197, range ±2g, frequency response 10-1000Hz) is installed on the side wall of the cabinet to collect mechanical vibration data. After the real-time data acquisition unit is started, it continuously collects data from the above sensors at a frequency of 100ms / time. Over a continuous 24-hour period, a total of the following time-series data were acquired: 1,728,000 sets of current data (e.g., phase A current fluctuated between 120-135A, phase B between 118-132A, and phase C between 122-136A); 576,000 sets of voltage data (line voltage remained stable between 10.2-10.5kV); and 691,200 sets of status data (contact temperature peaked at 38℃, ambient temperature at 25℃); and 1,728,000 sets of vibration data (vibration amplitude ≤0.1g). All data were stored in real time, completing the data acquisition phase and providing complete and real-time raw data support for subsequent processes.

[0045] In this embodiment of the application, the data processing expansion module 200 includes: Figure 4 As shown, there are a data migration unit, a simulation data generation unit, a data optimization unit, and an integration and construction unit.

[0046] The data migration unit collects fault data from similar high-voltage switchgear in different regions, under different operating conditions, and at different lifespans, and integrates them to form migration data. The simulation data generation unit generates simulation data based on the time-series and status data of the equipment operation using GANs models. The data optimization unit filters out abnormal fluctuation data in simulation data and migration data using IQR. It also quantifies the distribution differences between simulation data and real data by combining JS divergence and verifies the authenticity of simulation data using logistic regression models and confusion matrices. The integration and construction unit integrates migration data, simulation data, and the time-series and status data of the equipment operation, and constructs a dataset by combining historical equipment data.

[0047] It should be noted that IQR refers to the interquartile range, which measures the dispersion of data. The formula is:

[0048] ;

[0049] in, Interquartile range; This is the value located at the 75th percentile of the data. This represents the value for a data center that is in the 25th percentile.

[0050] It is understood that the embodiments of this application collect fault data of similar high-voltage switchgear in different regions, under different operating conditions, and at different lifespans through a data migration unit, integrate it to form migration data, expand the coverage and scenario diversity of the data, supplement fault characteristics that are difficult to obtain from a single device or environment, and avoid the one-sidedness caused by data being limited to specific conditions; through a simulation data generation unit, based on the time-series data and status data of equipment operation, simulation data is generated through a GANs model, supplementing simulation samples that meet the data characteristics when the original samples are insufficient, especially filling the sample gaps of rare fault types, and solving the problem of insufficient sample size; through a data optimization unit... The method utilizes IQR to filter out fluctuating and abnormal data in simulation and migration data, combines JS divergence to quantify the distribution differences between simulation and real data, and uses logistic regression models and confusion matrices to verify the authenticity of simulation data, ensuring data quality, eliminating invalid outliers, and ensuring that simulation data and real data have consistent characteristics, thus avoiding interference from low-quality or distorted data in subsequent analysis. By integrating migration data, simulation data, and time-series and status data of equipment operation through integrated building units, and combining historical equipment data to construct a dataset, the method merges multi-source and multi-type data to form a complete dataset covering different scenarios and states, thereby improving the comprehensiveness of the dataset.

[0051] For example, such as Figure 5As shown, when a power company was processing data on a batch of outdoor 110kV high-voltage circuit breakers, the data migration unit specifically collected cross-domain fault data of similar equipment: 350 porcelain insulator corrosion fault data and 220 terminal oxidation fault data were obtained from coastal salt spray areas; 280 operating mechanism sealing failure data and 190 insulation strength reduction data were obtained from high-altitude low-temperature areas (above 3000m altitude, -20~10℃); 310 transmission component jamming data and 240 sensor signal interference data were obtained from windy and sandy mining areas; and 260 spring fatigue fault data and 200 hydraulic mechanism leakage data were obtained from equipment that had been in operation for about 10 years. These data were integrated to form 1760 migration data, covering fault characteristics of different natural environments, operating conditions, and life stages. The simulation data generation unit, based on 4,000 normal operation timing data (such as opening and closing coil current and contact travel speed) and 900 common fault state data (such as pressure changes during slight leakage) collected from this batch of circuit breakers, generated simulation data using a GANs model: 5,000 normal operation simulation data and 1,200 simulation data each for typical faults such as salt spray corrosion, low temperature sealing failure, and wind and sand jamming, generating a total of 9,600 simulation data, filling the gap in rare fault samples under extreme storms. The data optimization unit first processed the data using the IQR method: removing 160 abnormal fluctuation values ​​(such as distorted records of sudden pressure drops) from 1,760 migration data and removing 630 data deviating from the true characteristics from 9,600 simulation data, retaining 1,600 migration data and 8,970 simulation data; then, using JS divergence quantification, the difference between the simulation data and the real data distribution was 0.18 (<0.3, meeting the consistency requirement); finally, the simulation data was verified using a logistic regression model and confusion matrix, achieving a 91% accuracy rate in terms of realism. The integration building unit combines 1,600 migration data entries and 8,970 simulation data entries with the original 4,900 device runtime sequence and status data entries and 1,500 historical fault data entries to form a dataset of 16,970 entries, providing high-quality data support covering multiple scenarios for the future.

[0052] In this embodiment of the application, the formula for the GANs model is:

[0053] ;

[0054] in, A discriminant for real and simulated data; A generator for simulation data; For generator The minimum value; For discriminator The maximum value; It is a value function; For discriminator For real data The expected logarithm of the judgment result; It is the natural logarithm; For discriminator For input data The probability of being judged as real data; noise distribution Expectations; Represents generator Using noise The generated data.

[0055] It should be noted that the application process of GANs models is as follows: First, a generative adversarial network consisting of a generator and a discriminator is constructed. The generator adopts a multi-layer fully connected structure, combining batch normalization, LeakyReLU activation function, and L2 regularization, with the last layer using the tanh activation function. The discriminator is also a fully connected layer, using LeakyReLU activation function and L2 regularization, introducing a gradient penalty mechanism, and optimizing the quality of generated samples through feature matching loss and trend consistency loss. During training, the Adam optimizer is used, and appropriate strategies are set to ensure training stability, saving weights as the generator loss improves.

[0056] The formula for the LeakyReLU activation function is as follows:

[0057] ;

[0058] in, For the fully connected layer output of the generator or discriminator; A small constant that is set; This is the output of the activation function.

[0059] The normalization formula is:

[0060] ;

[0061] in, Batch-normalized output tensors; This is the input to the fully connected layer; This represents the mean of the data within the current training batch. This represents the variance of the current training batch; This is the scaling factor; It is a small constant; This is the offset coefficient.

[0062] The L2 regularization formula is:

[0063] ;

[0064] in, This is a loss item; The strength coefficient; These are trainable parameters; It is a relational character; It is a set of trainable parameters; for The square of the L2 norm.

[0065] The feature matching loss formula is:

[0066] ;

[0067] in, For feature matching loss; For expectation calculation; This represents a real data sample and its distribution. Noise samples and their distribution; This is the feature extraction function for the intermediate layer of the discriminator; This is a real data sample; Generator noise Generated fake data.

[0068] The formula for trend consistency loss is:

[0069] ;

[0070] in, This represents a loss due to trend consistency. For expectation calculation; This represents the actual data sample and its distribution. Noise samples and their distribution; This is a trend extraction function; This is a real data sample; This is the output of the generator.

[0071] The gradient penalty formula is:

[0072] ;

[0073] in, For the loss function term; This is the penalty coefficient; For expectation calculation; The interpolation samples and their distribution are shown below. For the discriminator to interpolate samples The gradient; This is the output of the discriminator.

[0074] It is understood that the embodiments of this application generate simulation data through GANs models, which can supplement simulation samples that conform to the operating characteristics of the equipment when the original samples are insufficient, fill the sample gaps of rare fault types, enrich the coverage of the dataset in terms of scenarios and state types, enhance the coverage of the data on the diverse operating states of the equipment, improve the integrity of the dataset, and provide data support for the construction of the dataset.

[0075] For example, when processing data from 35kV high-voltage switchgear, GANs models are used to generate simulation data. The original operating data for this batch of switchgear includes 5000 normal time-series data points (dynamic changes in three-phase current from 150-200A and line voltage from 34-36kV) and 800 common state data points (such as contact temperature from 30-45℃ and cabinet vibration amplitude from 0.05-0.1g). However, severe fault data is scarce; there are only 60 samples of "insulation breakdown" and only 45 samples of "mechanical interlocking failure," resulting in a coverage rate of less than 10% for these rare fault types. Based on this raw data, the simulation data generation unit activates the GANs model: the generator learns the distribution of features such as current fluctuation patterns and temperature change trends, and generates simulation data by combining this with noise input; the discriminator continuously distinguishes between real and simulation data, driving the generator to optimize its output. The final generated simulation data includes 8000 normal operation data points and 500 simulation data points each for severe faults such as "insulation breakdown" and "mechanical interlocking failure," generating a total of 9000 simulation data points. It effectively filled the gap in scarce fault samples, expanded the data source size from 5,800 to 14,800, and increased the fault type coverage to 85%, providing sufficient data support for subsequent integration and construction of a comprehensive dataset.

[0076] In this embodiment of the application, the diagnostic prediction module 300 includes: Figure 6 As shown, the extraction prediction unit and the evaluation diagnosis unit are included.

[0077] The extraction and prediction unit inputs the dataset into the DualLSTM-DropNet model, uses bidirectional LSTM layers to capture temporal dependencies in the data, and integrates a multi-scale attention mechanism to enhance the extraction of features from different periods, outputting the prediction results of equipment state parameters. The evaluation and diagnosis unit evaluates the prediction results of equipment state parameters using the RMSE index and R² coefficient, and performs fault diagnosis, health status assessment, and remaining life prediction of high-voltage switchgear based on the state parameter prediction results, forming a diagnostic conclusion.

[0078] It should be noted that the application process of the DualLSTM-DropNet model is as follows: A bidirectional LSTM architecture is adopted, with a two-layer LSTM structure. A Dropout layer is added after each layer to suppress overfitting. Input data is fed in through the InputLayer, sequentially processed by two layers of LSTM feature extraction and Dropout, and finally the prediction result is output by a fully connected layer. The formula is:

[0079] ;

[0080] ;

[0081] ;

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] in, Output for the forget gate; For activation functions; Here is the weight matrix for the forget gate; For the model in Abstract operational characteristics learned from historical data of the equipment in constant progress; for Input characteristics at any given time; For the bias term of the forget gate; For input gate output; This is the weight matrix of the input gate; This is the bias term for the input gate; These are candidate features for early signs of failure; This is the weight matrix of the currently input features; The bias term for the current input feature; For activation functions; This involves integrating the filtered historical key information with current key features; Output gate output; This is the weight matrix of the output gate; This is the bias term for the output gate; This is the current core feature; To learn the positive sequence trend characteristic of "gradually increasing temperature → increasing probability of failure"; Forward LSTM unit; To learn the inverse correlation features of "fault occurrence → temperature drop"; For backward LSTM units; A random mask for Dropout; This represents the probability of being dropped in Dropout. The current core features after Dropout processing; This is the output of the first layer of the DualLSTM; This is a first-layer bidirectional LSTM module; This represents the drop probability of the first layer of Dropout. This is the output of the second layer of DualLSTM; This is a second-layer bidirectional LSTM module; This represents the drop probability of the second-level Dropout. This is the model's predicted output.

[0093] The RMSE (Real Money Result) index measures the average magnitude of the square root of the sum of the squares of the deviations between the predicted and actual values. The formula is:

[0094] ;

[0095] in, For the first The true value of each sample; For the first Predicted values ​​for each sample; The total number of samples.

[0096] The R² coefficient is an indicator used to measure the degree of agreement between model predictions and actual equipment parameters. The formula is:

[0097] ;

[0098] in, For the first The true value of each sample; For the first Predicted values ​​for each sample; The total number of samples; This is the average of the true parameter values ​​for all samples.

[0099] It is understood that, in this embodiment, the dataset is input into the DualLSTM-DropNet model through the extraction prediction unit. The bidirectional LSTM layer is used to capture the temporal dependencies in the data, and a multi-scale attention mechanism is integrated to enhance the extraction of features in different cycles. The output is the prediction result of equipment state parameters, which captures the temporal correlation pattern of equipment operation data and the key features of different cycles, thereby improving the accuracy and pertinence of state parameter prediction. The evaluation and diagnosis unit evaluates the prediction result of equipment state parameters with RMSE index and R² coefficient, performs fault diagnosis, health status assessment and remaining life prediction of high-voltage switchgear, forms a diagnostic conclusion, enhances the reliability of the diagnostic conclusion, and provides decision support for intervention.

[0100] For example, such as Figure 7 As shown, a power operation and maintenance team analyzed a set of 110kV high-voltage circuit breakers using a diagnostic prediction module. The circuit breaker dataset contains 1200 processed migration data entries, 5780 simulation data entries, 2500 original operating data entries, and 1000 historical fault data entries, totaling 10480 data entries. This dataset covers current, temperature, vibration, and other characteristics under different regions and operating conditions. The prediction unit inputs the dataset into the DualLSTM-DropNet model: the bidirectional LSTM layer captures the current fluctuations over time (e.g., peak current of 120-140A from 9:00-17:00 on weekdays, decreasing to 80-100A at night) and the temporal correlation between temperature and load. After fusing a multi-scale attention mechanism, the extraction of features from daily load cycles (24 hours) and weekly maintenance cycles (7 days) is enhanced, outputting prediction results for state parameters such as contact temperature and mechanical vibration amplitude. The predicted contact temperature for the next 72 hours shows a 92% trend agreement with the actual value. The assessment and diagnosis unit verified the reliability of the prediction results through the RMSE index (calculated as 2.1℃, less than the threshold of 3℃) and the R² coefficient (0.95, close to 1). Based on this, further analysis diagnosed that the circuit breaker had a slight contact wear fault, with a health index of 83 (out of 100) and an estimated remaining life of 3.2 years. Finally, a complete diagnostic conclusion was formed to provide a basis for subsequent intervention.

[0101] In this embodiment, the intelligent interconnection and intervention module 400 includes, as follows: Figure 8 As shown, the intervention decision-making unit and the data transmission unit.

[0102] The intervention decision unit determines whether to intervene based on the diagnostic conclusion. If intervention is required, it dynamically outputs a graded intervention plan by combining the DQN algorithm with the real-time operating conditions of the equipment and the feedback of historical intervention effects. The data transmission unit transmits the graded intervention plan to the control terminal through the Internet of Things.

[0103] It should be noted that the DQN algorithm formula is as follows:

[0104] ;

[0105] ;

[0106] ;

[0107] in, This represents the total cumulative rewards that can be obtained in the future. This refers to the real-time operating status of the equipment. This is a tiered intervention approach; These are the parameters of the network; Theoretically optimal value; For the first Step goal value; For the first The gains or losses resulting from the step; Discount factor; The state at the next moment; For all possible intervention actions; These are the parameters of the target network; To maximize the function; The loss function; For expectation calculation; For the first Step-by-step device status; For the first The intervention actions to be performed step by step.

[0108] It is understood that the embodiments of this application determine whether to intervene through an intervention decision unit. If intervention is required, a graded intervention plan is dynamically output through the DQN algorithm combined with the real-time operating conditions of the equipment and feedback from historical intervention effects. The intervention needs are grasped based on the diagnostic results, and the algorithm dynamically adapts to the current state of the equipment and historical experience, making the intervention plan more in line with the actual scenario, avoiding blind intervention, and improving the pertinence and effectiveness of the intervention measures. The graded intervention plan is transmitted to the control terminal through the Internet of Things via the data transmission unit, and the intervention instructions are transmitted efficiently and in real time, ensuring that the control terminal obtains and executes the plan in a timely manner, improving the response speed of equipment intervention, strengthening the remote collaboration capability of the system, and ensuring the smooth implementation of dynamic intervention.

[0109] For example, such as Figure 9As shown, after monitoring a set of 110kV high-voltage circuit breakers in operation at a substation, the diagnostic conclusion was: "The circuit breaker has a 'stuck operating mechanism' fault (medium fault level), health index 65, and the load is expected to rise to 85% in the next 24 hours." Based on this diagnostic conclusion, the pre-decision unit determined that intervention was necessary. Using the DQN algorithm, real-time equipment operating data (current mechanism operating voltage deviation 3%, ambient temperature 32℃) was retrieved, and combined with historical intervention records (in the past three similar stuck faults, the "adding special lubricant + load reduction operation" solution had a success rate of 92%), a "level two intervention" solution was dynamically output: immediately remotely control the addition of lubricant; temporarily reduce the circuit breaker load to 70% and re-check the mechanism operating parameters after 2 hours. The data transmission unit then encrypted and transmitted this level-two intervention solution to the intelligent operation and maintenance system of the substation control center via the Internet of Things. Upon receiving the signal, the control terminal executed the lubricant addition command within 15 minutes and simultaneously adjusted the load distribution, achieving precision and timeliness of the intervention.

[0110] This application proposes a deep learning-based intelligent interconnected switch system. Through a data acquisition module, it collects time-series and status data of equipment operation using adapted power and non-power sensors, improving the comprehensiveness and real-time acquisition capability of equipment operation data. A data processing expansion module filters out abnormal data using IQR, combines JS divergence to quantify distribution differences, and verifies data authenticity using a logistic regression model and confusion matrix, generating datasets to provide data support for diagnosis and prediction. Simulation data is generated through GANs models, combined with cross-domain fault migration data, expanding data sources and improving data richness and diversity. The diagnosis and prediction module inputs the dataset into a DualLSTM-DropNet model incorporating multi-scale attention mechanisms, improving the accuracy of fault diagnosis, health status assessment, and remaining life prediction. The intelligent intervention and interconnection module uses the DQN algorithm combined with real-time operating conditions and historical feedback to generate tiered intervention plans, and performs dynamic intervention via IoT transmission, improving the timeliness and targeted nature of equipment operation and maintenance. This solves the problems of insufficient effective data samples, low diagnostic accuracy, and limited diagnostic algorithms in existing technologies.

[0111] The following will illustrate a specific embodiment of a deep learning-based intelligent interconnected switch system, such as... Figure 10 As shown, it includes:

[0112] A power grid company launched an intelligent operation and maintenance system for a set of 220kV outdoor high-voltage switchgear within its jurisdiction. The initial phase of the operation was completed by the data acquisition module. This module includes a sensor selection and configuration unit and a real-time data acquisition unit, which work together to achieve comprehensive data perception of the equipment. Based on non-intrusive monitoring requirements, the sensor selection and configuration unit selected suitable electrical and non-electrical sensors: three Hall current sensors (range 0-600A) were installed at the switchgear's incoming line end to capture dynamic fluctuations in three-phase current; four infrared temperature sensors (range -20~150℃) were attached to the contact surface to monitor temperature changes at the contact points; two vibration sensors (range 0-1g) were fixed to the side of the cabinet to record the vibration amplitude and frequency during mechanical operation; and one voltage sensor (range 0-300kV) was deployed in the busbar compartment to collect temporal changes in line voltage. All sensors were installed non-intrusively to avoid affecting the original structure of the equipment. The real-time data acquisition unit continuously collects monitoring data from the aforementioned sensors at a frequency of 1 minute per acquisition, forming two types of core data: first, time-series data, such as the periodic fluctuation of current peak of 250-300A from 9:00 to 17:00 on weekdays and dropping to 100-150A at night, as well as the dynamic changes in voltage within the range of 220±5kV; second, status data, such as the gradient change of contact temperature with increasing load (25℃ to 60℃), and the vibration peak value (0.08-0.12g) when the operating mechanism is in motion. After three months of continuous acquisition, a total of 144,000 basic data points were obtained, providing raw materials for subsequent processing.

[0113] After data acquisition, the data processing expansion module's data migration unit collects cross-domain fault data from similar 220kV switchgear: 320 data points on "operating mechanism freezing and jamming" and 280 data points on "insulation component condensation flashover" from frigid northern regions (-20~5℃); 410 data points on "contact oxidation overheating" and 290 data points on "cabinet corrosion grounding" from humid and hot southern regions (25~35℃, humidity >85%); and 350 data points on "mechanical wear and jamming" and 220 data points on "secondary circuit aging" from equipment that has been in operation for more than 15 years. These data are then integrated to form 1870 migration data points, covering fault characteristics across different environments, operating conditions, and lifespan stages. The simulation data generation unit, based on the original time-series and state data, initiates a GANs model for sample expansion. The generator (G) learns the distribution of features such as current fluctuation patterns and temperature change trends, and generates simulation data by combining random noise; the discriminator (D) continuously distinguishes between real and simulation data, driving the generator to optimize its output. Ultimately, 4000 simulation data points for normal operation and 800 simulation data points each for rare faults such as "contact welding" and "insulation breakdown" were generated, totaling 5600 data points, filling the sample gap under extreme operating conditions. The data optimization unit then intervened: using IQR, 120 outlier values ​​(such as distorted records of a sudden temperature rise to 180℃) were removed from the migration data, and 350 data points deviating from the true characteristics were removed from the simulation data, retaining 1750 and 5250 valid data points respectively. Then, using JS divergence quantification, the distribution difference between the simulation data and the real data was found to be 0.21 (<0.3, meeting the consistency requirement). Finally, the simulation data was validated using a logistic regression model and confusion matrix, achieving a 94% accuracy rate in terms of realism. The integrated construction unit combines 1,750 migration data entries and 5,250 simulation data entries with 144,000 original operational data entries and 3,000 historical fault data entries to form a dataset containing 154,000 records. The data covers 12 states, including normal operation, minor faults, and serious faults, providing comprehensive support for subsequent diagnosis and prediction.

[0114] In the diagnostic prediction module, the prediction unit inputs 154,000 dataset entries into a DualLSTM-DropNet model that integrates a multi-scale attention mechanism. The bidirectional LSTM layer captures the temporal dependency between peak current and contact temperature through forward and backward computation (e.g., 15 minutes after the current reaches 300A, the temperature rises to 55℃). The multi-scale attention mechanism enhances feature extraction for daily load cycles (24 hours) and weekly maintenance cycles (7 days), focusing on equipment status changes during peak load periods. The final output is the predicted status parameters for the next 72 hours, with the deviation between the predicted and actual contact temperature values ​​controlled within ±2℃, and the trend agreement for voltage fluctuation prediction reaching 96%. The diagnostic unit evaluates the reliability of the predictions using quantitative indicators: the calculated RMSE of the predicted status parameters is 1.8 (far below the threshold of 5), and the R² coefficient is 0.97 (close to 1), confirming the reliability of the prediction results. Based on this, the unit further analyzed and concluded that: cabinet #3 in this group of switch cabinets has a "slight overheating of contacts" fault (current temperature 62℃, exceeding the normal threshold by 5℃), with a health index of 76 (out of 100) and an estimated remaining lifespan of 3.8 years; the other cabinets are in normal condition, but the vibration amplitude of cabinet #5 needs to be monitored (the recent fluctuation amplitude has increased by 12%).

[0115] Based on the diagnostic conclusions, the intervention decision unit determined that although the "slight overheating of contacts" fault in cabinet #3 did not endanger safety, the health index was below the warning threshold (80), and the load was expected to rise to 90% in the next 72 hours, requiring timely intervention. The unit invoked the DQN algorithm, combining real-time operating conditions (current ambient temperature 33℃, load 280A) and historical intervention records (in the past two similar faults, the success rate of the "remotely increasing the cooling fan speed + re-inspection after 3 days" solution reached 95%), and dynamically output the "Level 1 Intervention" solution: ① Immediately remotely control the cooling fan speed of cabinet #3 to increase to 80% (from 50%); ② Reduce the load of the cabinet to 220A; ③ Automatically collect the contact temperature after 72 hours, and if it still exceeds the threshold, trigger "Level 2 Intervention" (arrange manual maintenance). The data transmission unit transmitted the graded intervention solution to the intelligent operation and maintenance platform of the power grid dispatch and control center in real time through an encrypted IoT channel. After receiving the plan, the control terminal issues instructions for adjusting the cooling fan speed and distributing the load within 10 minutes. After the on-site switchgear executes the instructions, it provides feedback on the execution results (fan speed has reached 80%, load is stable at 220A), forming an intervention closed loop. Re-inspection data three days later shows that the contact temperature of cabinet #3 has dropped to 52℃, returning to normal.

[0116] In summary, this application embodiment acquires time-series and status data of equipment operation through sensor selection, configuration, and real-time acquisition in the data acquisition module, providing comprehensive and real-time raw data for subsequent system analysis. The data processing expansion module collects cross-domain fault data to form migration data, and uses GANs models to generate simulation data to integrate and construct a high-quality dataset, expanding data sources, solving the problem of sample scarcity, and enriching the scenarios and state types covered by the data. The diagnosis and prediction module inputs the dataset into a DualLSTM-DropNet model that integrates multi-scale attention mechanisms to generate state parameter prediction results, and combines quantitative indicators to form diagnostic conclusions, improving the accuracy and reliability of equipment fault diagnosis, health assessment, and lifespan prediction. Based on the diagnostic conclusions, the intelligent intervention and interconnection module dynamically outputs graded intervention plans using the DQN algorithm and transmits them via the Internet of Things, enabling targeted and timely dynamic intervention to ensure the safe and stable operation of equipment and improve the intelligent operation and maintenance level of high-voltage switchgear.

[0117] Next, referring to the accompanying drawings, a method for operating a deep learning-based intelligent interconnected switch according to an embodiment of this application is described.

[0118] like Figure 11 As shown, the working method of this deep learning-based intelligent interconnected switch includes the following steps:

[0119] In step S101, timing data and status data of the device operation are acquired.

[0120] It is understood that the embodiments of this application obtain the timing data and status data of the device operation, record the operating characteristics and dynamic change patterns of the device at different times, provide original and authentic basic information for subsequent stages, ensure the integrity and reliability of the data source, and guarantee the scientificity and effectiveness of the overall decision-making.

[0121] In step S102, the time series data and state data are processed and input into the GANs model to generate simulation data. At the same time, cross-domain fault data of similar high-voltage switchgear in different regions, under different operating conditions and at different life periods are collected and integrated to form migration data. IQR is used to filter out the fluctuation and abnormal data of simulation data and migration data, and the distribution difference between simulation data and time series data and state data is quantified by JS divergence. Logistic regression algorithm and confusion matrix are used to verify the authenticity of simulation data, and a dataset is constructed by combining the historical data of the equipment.

[0122] JS divergence is a symmetric index that measures the difference between two probability distributions; the formula is:

[0123] ;

[0124] in, The predicted state probability distribution; This represents the probability distribution of the actual operating state. Relative entropy; for and The average distribution.

[0125] It should be noted that logistic regression is a classification algorithm that predicts the probability of an event occurring. The formula is:

[0126] ;

[0127] ;

[0128] in, The input feature vector; The feature weight vector; This represents the transpose of the weight vector; For bias terms; This is the result of a linear combination; The probability value predicted by the model; For activation functions; It is a natural constant; It is the negative of the linear result.

[0129] A confusion matrix is ​​a table that displays the comparison between the predicted class and the actual class in matrix form.

[0130] It is understood that the embodiments of this application can intuitively evaluate the feature consistency between simulation data and real data through JS divergence, determine whether the simulation data conforms to the operating rules of real equipment, construct a classification model through logistic regression to determine the authenticity of the data, and the confusion matrix can clearly present the precision, recall and other indicators of the verification results, reflecting the degree of matching between simulation data and real data, ensuring that the simulation data included in the dataset is effective and usable, and improving the overall quality of the dataset.

[0131] For example, when processing simulation data of a 110kV high-voltage switchgear, a power operation and maintenance team first used JS divergence to measure the distribution difference between the simulation data and the real data. The calculated difference value was 0.23 (threshold < 0.3), indicating that the feature distributions of the two types of data were similar, and it was initially judged that the simulation data closely matched the actual operating patterns. Subsequently, a classification model was constructed using a logistic regression algorithm, taking 10 features such as voltage deviation and vibration frequency as input, and dividing the data into "real" and "simulation" categories for training. After the model training was completed, 2000 test data (including 1000 real data and 1000 simulation data) were classified, and the classification results were output. Finally, the effect was verified by confusion matrix analysis: 920 real data were correctly identified (true positives), and 80 were misclassified as simulation data (false negatives); 910 simulation data were correctly identified (true negatives), and 90 were misclassified as real data (false positives). The calculated precision was 91.9% and the recall was 92%, confirming the reliability of the simulation data and its inclusion in the dataset for subsequent analysis.

[0132] In step S103, the dataset is input into the DualLSTM-DropNet model that integrates a multi-scale attention mechanism to generate state parameter prediction results. Based on the state parameter prediction results, fault diagnosis, health status assessment and remaining life prediction are performed on the high-voltage switchgear to form a diagnostic conclusion.

[0133] Among them, the multi-scale attention mechanism refers to the mechanism that adaptively focuses on key information and suppresses secondary information from different time or feature scales of data, and the formula is:

[0134] ;

[0135] ;

[0136] in, For the first Scale-based feature mapping; For the first Scale-based convolution or sliding window operations; The bidirectional hidden states output by the DualLSTM-DropNet model; For time step; To hide dimensions; It is the set of real numbers; This represents the total number of scales.

[0137] It is understood that the embodiments of this application enhance the extraction of different periodic features in equipment operation data by integrating a multi-scale attention mechanism, such as daily load cycle and weekly maintenance cycle, so that the model can accurately capture key features at different time scales; improve the pertinence and accuracy of equipment status parameter prediction results, provide a more reliable basis for subsequent fault diagnosis, health status assessment and remaining life prediction, and ensure the effectiveness of diagnostic conclusions.

[0138] In step S104, based on the diagnostic conclusion, it is determined whether to perform condition-based maintenance, predictive maintenance, and full lifecycle management intervention on the equipment. If intervention is required, the DQN algorithm is used to generate a graded intervention plan by combining the real-time operating conditions of the equipment with historical intervention effect feedback.

[0139] Among them, the tiered intervention program is a plan to develop and implement corresponding differentiated intervention measures based on the different levels and degrees of problems or individuals.

[0140] It is understood that the embodiments of this application provide targeted intervention measures adapted to different fault levels and equipment states through a graded intervention scheme, avoiding blind or insufficient maintenance and ensuring the accuracy of intervention; at the same time, dynamic intervention can be carried out by transmitting data to the control terminal through the Internet of Things, ensuring the timeliness and efficiency of intervention, and improving the stability of equipment operation and the scientific nature of operation and maintenance management.

[0141] For example, during monitoring of a 110kV high-voltage circuit breaker at a substation, the diagnostic conclusion showed that circuit breaker #2 had a "contact overheating (65℃, threshold 55℃)" fault, with a health index of 72, and the load was expected to reach 90% in the next 24 hours. Intervention was deemed necessary. The DQN algorithm retrieved real-time operating conditions (ambient temperature 34℃, current load 75%) and historical data (the "remote fan adjustment + load reduction" solution had a 90% success rate in the last three similar faults), outputting a three-level intervention plan: Level 1 (immediate execution): remotely increase the cooling fan speed to 100% and reduce the load to 60%; Level 2 (after 2 hours): if the temperature does not drop below 55℃, activate the backup cooling system; Level 3 (after 12 hours): if the temperature still exceeds the limit, generate a maintenance work order and arrange for personnel to handle the situation on-site. The data transmission unit encrypted and transmitted the plan to the control center via the Internet of Things, completing the execution of the Level 1 intervention command within 10 minutes. Two hours later, monitoring showed the temperature had dropped to 52℃, and the system automatically terminated subsequent interventions to avoid over-operation. This process demonstrates the dynamic adaptability of the hierarchical scheme, enabling rapid response to faults while reducing unnecessary resource investment.

[0142] According to an embodiment of this application, a deep learning-based intelligent interconnected switch operation method is proposed. Through a data acquisition module, adapted power and non-power sensors are used to collect time-series and status data of the device operation, improving the comprehensiveness and real-time acquisition capability of the device operation data. With the help of a data processing expansion module, abnormal data is filtered out using IQR, and the distribution differences are quantified by combining JS divergence and the authenticity of the data is verified using a logistic regression model and confusion matrix, generating a dataset to provide data support for diagnosis and prediction. Simulation data is generated through a GANs model, combined with cross-domain fault migration data, expanding the data source and improving the richness and diversity of the data. The diagnosis and prediction module inputs the dataset into a DualLSTM-DropNet model that integrates a multi-scale attention mechanism, improving the accuracy of fault diagnosis, health status assessment, and remaining life prediction. The intelligent intervention and interconnection module uses the DQN algorithm combined with real-time operating conditions and historical feedback to generate a graded intervention plan, and performs dynamic intervention through IoT transmission, improving the timeliness and pertinence of equipment operation and maintenance. This solves the problems of insufficient effective data samples, low diagnostic accuracy, and single diagnostic algorithms in existing technologies.

[0143] The following will illustrate a deep learning-based intelligent interconnected switch operation method through a specific embodiment, such as... Figure 13 As shown, it includes:

[0144] A power operation and maintenance team conducted intelligent operation and maintenance on a set of 220kV outdoor high-voltage switchgear, acquiring operational data of the high-voltage switchgear: recording time-series data every 30 minutes, including three-phase contact temperature (peak 65℃ on summer workdays, dropping to 40℃ at night), vibration amplitude of the operating mechanism (0.15-0.2g when opening, dropping to below 0.05g after closing), and circuit resistance (stable at 50-60μΩ); simultaneously collecting status data, such as the aging degree of insulation tie rods (tested monthly, dropping from an initial 90% integrity rate to 78%), sealing pressure (consistently maintained at 0.5-0.6MPa), and number of opening and closing cycles (cumulative 320 times), forming a basic data reserve.

[0145] Subsequently, the data was processed: the aforementioned time-series and state data were input into the GANs model to generate 4000 simulation data points, covering rare fault scenarios such as "contact welding" and "operating mechanism jamming." Simultaneously, cross-domain fault data of similar high-voltage switches were collected from high-altitude areas (280 faults due to low air pressure causing insulation strength degradation), industrial areas (320 faults due to dust pollution causing contact overheating), and equipment that had been in operation for over 20 years (250 faults due to mechanical component wear), integrating them into 850 migration data points. The IQR method was used to filter out 180 outliers in the simulation data where the contact temperature suddenly rose to 180℃, and 90 distorted records in the migration data where the loop resistance was negative. Then, JS divergence calculation showed that the distribution difference between the simulation data and the real data was 0.22 (<0.3), meeting the consistency requirements. Finally, logistic regression and confusion matrix verification were used, and the simulation data's accuracy reached 95%. Ultimately, it was integrated with 5000 historical operating data points to form a dataset of 9670 data points.

[0146] Next, the dataset was input into the DualLSTM-DropNet model, which incorporates a multi-scale attention mechanism. The model captures the temporal correlation of "the contact temperature rises by 8°C one hour after the loop resistance rises to 58μΩ" through a bidirectional LSTM layer. The multi-scale attention mechanism enhances the feature extraction of daily operating peaks (8:00-18:00) and seasonal temperature differences (summer and winter), and outputs the predicted state parameters for the next 45 days—the peak contact temperature is expected to reach 72°C, and the integrity rate of the insulation rod will drop to 75%. The prediction results are reliable based on RMSE (1.9) and R² (0.96) indicators. Based on this, the high-voltage switch is diagnosed with a "contact oxidation" fault, with a health index of 68 and an estimated remaining life of 3.5 years, thus forming a diagnostic conclusion.

[0147] Finally, based on the diagnostic conclusion, intervention was deemed necessary. Using the DQN algorithm, combined with real-time operating conditions (current ambient temperature 36℃, 10 opening and closing operations completed) and historical intervention records (the "contact grinding + enhanced heat dissipation" solution had a 96% success rate in the past three similar faults), a tiered intervention plan was generated: Level 1 (immediate execution): remotely control the automatic contact grinding device and simultaneously increase the cooling fan speed to 90%; Level 2 (after 72 hours): if the loop resistance does not drop below 55μΩ, arrange for the replacement of local seals; Level 3 (after 15 days): if the standard is still not met, formulate a comprehensive overhaul plan. The plan was transmitted to the high-voltage switch control terminal via the Internet of Things, and the terminal completed the execution of the Level 1 intervention command within 20 minutes. After 72 hours, the test showed that the loop resistance had dropped to 54μΩ, eliminating the need for further measures, achieving precise and efficient dynamic intervention.

[0148] In summary, this application's embodiments acquire time-series and status data of equipment operation to provide foundational information for subsequent analysis; enrich data sources and ensure data quality through data processing and expansion; generate status parameter prediction results using the DualLSTM-DropNet model to predict fault diagnosis, health assessment, and equipment lifespan; and dynamically output and transmit tiered intervention plans through intelligent intervention for precise and timely equipment intervention. This enhances the initiative and accuracy of high-voltage switchgear operation and maintenance, ensures stable equipment operation, prevents faults, and optimizes the efficiency of full lifecycle management.

[0149] Figure 14 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0150] The memory 1401, the processor 1402, and the computer program stored on the memory 1401 and executable on the processor 1402.

[0151] When the processor 1402 executes the program, it implements a deep learning-based intelligent interconnection switch operation method provided in the above embodiments.

[0152] Furthermore, electronic devices also include:

[0153] Communication interface 1403 is used for communication between memory 1401 and processor 1402.

[0154] The memory 1401 is used to store computer programs that can run on the processor 1402.

[0155] The memory 1401 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0156] If the memory 1401, processor 1402, and communication interface 1403 are implemented independently, then the communication interface 1403, memory 1401, and processor 1402 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 14 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0157] Optionally, in a specific implementation, if the memory 1401, processor 1402, and communication interface 1403 are integrated on a single chip, then the memory 1401, processor 1402, and communication interface 1403 can communicate with each other through an internal interface.

[0158] The processor 1402 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of this application.

[0159] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for operating a deep learning-based intelligent interconnected switch.

[0160] Furthermore, this application also provides a computer program product, including a computer program or instructions, which, when executed, implement the aforementioned method for operating a deep learning-based intelligent interconnected switch.

[0161] In the description of this specification, the references to "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0162] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0163] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0164] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0165] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0166] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. An intelligent interconnected switch system based on deep learning, characterized in that, Comprise: Data capture module, data processing expansion module, diagnostic prediction module and intelligent intervention and interconnection module; wherein, The data capture module is used to collect the time series data and state data of the equipment operation; The data processing expansion module is used to collect cross-domain fault data of similar equipment, integrate to form migration data, generate simulation data using GANs model, expand data source, and construct data set combining with equipment historical data; wherein, the data processing expansion module comprises a data migration unit, a simulation data generation unit, a data optimization unit and an integration and construction unit, wherein the data migration unit is used to collect fault data of similar high-voltage switchgear in different regions, different working conditions and different life periods, and integrate to form migration data; the simulation data generation unit generates simulation data based on the time series data and state data of the equipment operation through the GANs model; the data optimization unit is used to exclude fluctuation abnormal data of simulation data and migration data through IQR; combining JS divergence, quantifying the distribution difference between simulation data and real data, using logistic regression model and confusion matrix to verify the authenticity of the simulation data; the integration and construction unit is used to integrate the migration data, the simulation data and the time series data and state data of the equipment operation, and construct data set combining with equipment historical data; The diagnostic prediction module is used to input the data set into the DualLSTM-DropNet model with fusion multi-scale attention mechanism to generate equipment state parameter prediction results, and perform fault diagnosis, health state evaluation and equipment remaining life prediction on the high-voltage switchgear, and generate diagnostic conclusion; The intelligent intervention and interconnection module judges whether to intervene the equipment based on the diagnostic conclusion, if intervention is needed, uses DQN algorithm, combines with real-time working condition and historical intervention effect feedback to dynamically output hierarchical intervention scheme, at the same time, transmits the hierarchical intervention scheme to the control end through the Internet of Things for dynamic intervention. 2.The intelligent interconnected switch system based on deep learning of claim 1, wherein, The data capture module comprises a sensor selection and configuration unit and a real-time data acquisition unit, wherein the sensor selection and configuration unit is used to select suitable power and non-power sensors according to non-invasive monitoring requirements and determine their installation position and quantity; the real-time data acquisition unit is used to acquire the time series data and state data of the equipment operation in real time. 3.The intelligent interconnected switch system based on deep learning of claim 1, wherein, The formula of the GANs model is: ； in, A discriminant for real and simulated data; A generator for simulation data; For generator The minimum value; For discriminator The maximum value; Value function; For discriminator For real data The expected logarithm of the judgment result; It is the natural logarithm; For discriminator For input data The probability of being judged as real data; noise distribution Expectations; Represents generator Using noise The generated data. 4.The intelligent interconnected switch system based on deep learning of claim 1, wherein, The diagnostic prediction module comprises an extraction and prediction unit and an evaluation and diagnosis unit, wherein the extraction and prediction unit inputs the data set into the DualLSTM-DropNet model, uses the bidirectional LSTM layer to capture the time series dependence in the data, and fuses the multi-scale attention mechanism to strengthen the extraction of different period features, and outputs the equipment state parameter prediction results; the evaluation and diagnosis unit evaluates the equipment state parameter prediction results through RMSE index and R² coefficient, and performs fault diagnosis, health state evaluation and remaining life prediction on the high-voltage switchgear based on the state parameter prediction results, and forms diagnostic conclusion. 5.The intelligent interconnected switch system based on deep learning of claim 1, wherein, The intelligent intervention and interconnection module comprises an intervention decision unit and a data transmission unit, wherein the intervention decision unit judges whether to intervene based on the diagnosis conclusion, and if intervention is needed, a hierarchical intervention scheme is dynamically output by combining the DQN algorithm, real-time working conditions of the equipment and historical intervention effect feedback; the data transmission unit transmits the hierarchical intervention scheme to the control end through the Internet of Things.

6. A working method applied to the intelligent interconnected switch system based on deep learning according to any one of claims 1-5, characterized in that, The working method comprises: acquiring time series data and state data of equipment operation; processing the time series data and state data, inputting the GANs model to generate simulation data, collecting cross-domain fault data of similar high-voltage switch equipment in different regions, different working conditions and different life periods, integrating to form migration data, using IQR to exclude fluctuation abnormal data of the simulation data and the migration data, quantifying the distribution difference between the simulation data and the time series data and the state data by JS divergence, verifying the authenticity of the simulation data by using a logistic regression algorithm and a confusion matrix, and combining historical data of the equipment to build a data set; inputting the data set into a DualLSTM-DropNet model with a fusion multi-scale attention mechanism to generate a state parameter prediction result, and based on the state parameter prediction result, performing fault diagnosis, health state evaluation and residual life prediction on the high-voltage switch equipment to form a diagnosis conclusion; judging whether to perform state maintenance, predictive maintenance and full life cycle management intervention on the equipment according to the diagnosis conclusion, and if intervention is needed, using the DQN algorithm to combine real-time working conditions of the equipment and historical intervention effect feedback to generate a hierarchical intervention scheme.

7. An electronic device, comprising: It comprises: a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to realize the working method of the intelligent interconnected switch system based on deep learning according to claim 6.

8. A computer readable storage medium having stored thereon a computer program or instructions, characterized in that, The computer program or instructions are executed to realize the working method of the intelligent interconnected switch system based on deep learning according to claim 6.

9. A computer program product comprising computer programs or instructions, characterized in that, The computer program or instructions are executed to realize the working method of the intelligent interconnected switch system based on deep learning according to claim 6.

Images