A power facility remote monitoring method based on internet of things

Abstract

This invention relates to the field of Internet of Things (IoT) technology, and in particular to an IoT-based remote monitoring method for power facilities, comprising the following steps: S101: multi-dimensional heterogeneous sensor data acquisition and timestamping; S102: edge data filtering, normalization, and feature extraction; S103: data encryption and transmission protocol encapsulation; S104: cloud data reception, verification, and time-series database storage; S105: real-time status assessment and threshold alarm judgment; S106: historical data analysis and AI prediction model iterative training; S107: equipment remaining life prediction and health index generation; S108: fault warning level classification and maintenance instruction issuance. This invention solves the problems of poor real-time performance, single data dimension, high deployment costs, and lack of predictive maintenance capabilities in traditional monitoring methods. It upgrades the maintenance mode to predictive maintenance based on the actual condition of the equipment, significantly improving monitoring efficiency, reducing operation and maintenance costs, and extending equipment lifespan.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) technology, and in particular to a method for remote monitoring of power facilities based on IoT. Background Technology

[0002] As a core component of the power system, the stability of the operation of power facilities (such as transformers, circuit breakers, and transmission lines) directly affects the safe and reliable power supply of the power system.

[0003] Traditional power facility condition monitoring mainly relies on two methods: First, manual periodic inspections, where maintenance personnel periodically visit the site to check equipment operation through visual inspection, infrared thermography, etc. This method is inefficient, costly, and cannot capture instantaneous or sudden fault precursor data, resulting in significant monitoring lag. Second, monitoring systems based on SCADA or wired networks, while capable of real-time data acquisition, typically require complex wiring, especially in geographically dispersed, harsh environments, or high-voltage isolation areas, leading to extremely high wiring and maintenance costs and poor deployment flexibility. Therefore, a remote monitoring method for power facilities based on the Internet of Things (IoT) is needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a remote monitoring method for power facilities based on the Internet of Things.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a remote monitoring method for power facilities based on the Internet of Things, comprising the following steps: S101: Multidimensional heterogeneous sensor data acquisition and timestamp marking: Key operating parameters of power facilities are synchronously collected using multi-dimensional heterogeneous sensors, and all collected data are precisely timestamped. The key operating parameters include temperature, humidity, vibration, noise, partial discharge, and oil chromatography parameters. The synchronous acquisition of data by the sensors is achieved through a four-layer processing approach: hardware-level clock synchronization, unified sampling frequency, synchronous triggering mechanism, interface adaptation, and delay compensation, with a timing deviation of ≤1ms. S102: Edge data filtering, normalization, and feature extraction: Using IoT terminals and gateways as edge computing nodes, the raw data marked in step S101 is subjected to quality checks, outliers are removed, and then Min-Max normalization is performed. Time-domain features and frequency-domain features are extracted. The time-domain features include the root mean square of the vibration signal, and the frequency-domain features include the core numerical features of the PRPD spectrum of partial discharge. The extracted features are used for initial local anomaly judgment at the edge and input to the cloud model. S103: Data encryption and transmission protocol encapsulation: The IoT terminal and gateway use the AES-256 algorithm to encrypt the feature data processed in step S102 with high strength, and encapsulate it into a standard IoT protocol packet, which is then transmitted to the cloud platform through a low-power wide area network or a 5G cellular network; the standard IoT protocol includes the MQTT protocol or the CoAP protocol. S104: Cloud-based data reception, verification, and time-series database storage: The cloud platform's data receiving service receives the data transmitted in step S103, performs integrity verification and decryption, and then stores it in a high-throughput time-series database. The time-series database is indexed by device ID, sensor ID, and timestamp, retaining complete time-series information and sensor identifiers to provide full data for model training. S105: Real-time Status Assessment and Threshold Alarm Judgment: The cloud platform performs real-time analysis on the data stored in step S104 and determines whether the device's operating status is abnormal based on a preset safety threshold. If the data exceeds the preset safety threshold, it is determined to be in an abnormal state, and step S108 is executed. If the data does not exceed the preset safety threshold, it is determined to be in a normal state, and step S106 is executed. S106: Historical Data Analysis and Iterative Training of AI Prediction Models The big data processing module and AI prediction and analysis module of the cloud platform collaborate to perform offline incremental iterative training on the Long Short-Term Memory (LSTM) network model using historical data. The LTM network model adopts a four-layer architecture: an input layer, a two-layer LTM network feature extraction layer, a two-layer fully connected feature fusion layer, and an output layer. Specifically: The input layer receives time-series data, wherein the time step length of the time-series data is set to 1008 and the feature dimension is set to 20. In the two-layer LSTM feature extraction layer, the first layer of the Long Short-Term Memory network has 128 neurons, the activation function is tanh, the return sequence is set to True, the Dropout rate is 0.2, and the L2 regularization coefficient is 0.001. The second layer of the Long Short-Term Memory network has 64 neurons, the activation function is tanh, the return sequence is set to False, the Dropout rate is 0.2, and the L2 regularization coefficient is 0.001. In the two-layer fully connected feature fusion layer, the first fully connected layer has 32 neurons, the activation function is ReLU, and the L2 regularization coefficient is 0.001; the second fully connected layer has 16 neurons and the activation function is ReLU. The output layer has two neurons, which use the Linear activation function and the Sigmoid activation function respectively. The training process uses the Adam optimizer and a weighted loss function, and performs incremental iterative training at a preset cycle of 1-7 days. S107: Equipment Remaining Life Prediction and Health Index Generation: The AI ​​prediction and analysis module calls the long short-term memory network model trained in step S106, combines it with the current seven-day operating data stored in step S104, predicts the remaining lifespan of the device, and generates a comprehensive health index. S108: Fault Warning Level Classification and Maintenance Instruction Issuance: Based on the anomaly judgment result of step S105 and the prediction result of step S107, the fault warning level is divided, and maintenance instructions containing specific repair content and time limits are issued to maintenance personnel through the remote monitoring center; the fault warning level includes Level I emergency shutdown, Level II recent maintenance, and Level III observation.

[0006] As a further description of the above technical solution: The multidimensional heterogeneous sensor in step S101 includes a temperature sensor, a humidity sensor, a vibration sensor, a noise sensor, a partial discharge sensor, and an oil chromatography sensor.

[0007] As a further description of the above technical solution: In step S103, the low-power wide-area network includes either a LoRa network or an NB-IoT network.

[0008] As a further description of the above technical solution: In step S104, the time-series database adopts a high-throughput design, which supports the fast writing and querying of massive monitoring data. During the data storage process, complete timestamp information and sensor identification information are retained, providing full time-series data including normal operation, slight degradation, and historical fault precursor states for the training of the long short-term memory network model in step S106.

[0009] As a further description of the above technical solution: In step S106, the training parameters of the Long Short-Term Memory (LSTM) network model are configured as follows: the initial learning rate is 0.001, and a learning rate decay strategy is adopted, which reduces the learning rate to 0.9 times the original value every 10 rounds; the number of training rounds is 500, the batch size is 64, and an early stopping strategy is enabled, which stops training if the validation set loss does not decrease for 10 consecutive rounds; in incremental training mode, the parameters of the two-layer LSM network feature extraction layer are frozen, and only the parameters of the fully connected layer are fine-tuned; the model after training is named and stored with device ID and training timestamp, retains the last 5 versions and supports automatic rollback.

[0010] As a further description of the above technical solution: The constraint rules for the value of the comprehensive health index in step S107 include: The input features are mapped to the [0,1] interval through Min-Max normalization; The output of the Long Short-Term Memory Network model is activated by the Sigmoid function and scaled by 100 times to obtain the basic comprehensive health index in the (0,100) range. The expert rule calibration adopts quantitative correction, with a fixed total weight of 1 and a single correction range of ≤ ±15 points; Forced boundary correction is applied before the final output: values ​​below 0 are taken as 0, values ​​above 100 are taken as 100, and integers are retained. Through trend verification of the results over the past three times, abnormal fluctuations were corrected using a weighted moving average to avoid random values.

[0011] As a further description of the above technical solution: In step S108, the rules for classifying fault warning levels are as follows: Level I Emergency Shutdown: Real-time core parameters exceed the standard by ≥30%, comprehensive health index <60 points, and remaining life expectancy ≤3 months; Level II Recent Maintenance: Real-time parameters exceed the standard by 10%-30%, comprehensive health index is 60-80 points, and remaining lifespan is 3-12 months; Level III observation: Real-time parameters exceed the standard by ≤10%, comprehensive health index ≥80, and remaining life expectancy >12 months.

[0012] As a further description of the above technical solution: Steps S101-S103 are executed at the edge layer, and steps S104-S108 are executed at the cloud platform layer; the edge layer includes multi-dimensional heterogeneous sensors and IoT terminals, and gateways; the cloud platform layer includes data receiving and storage services, big data processing modules, and AI predictive analysis modules.

[0013] As a further description of the above technical solution: The cloud platform constructs a digital twin model of the power facility, and presents the real-time status assessment results of step S105, the remaining life prediction results of step S107, and the equipment health index in a visual form. The visualization content includes the internal temperature distribution and stress change simulation of the equipment.

[0014] The present invention has the following beneficial effects: 1. Compared with existing technologies, this invention adopts a transmission method that combines wireless IoT terminals with low-power wide area networks and 5G cellular networks. It does not require complex wiring and can be flexibly deployed in remote, high-voltage or difficult-to-wire areas to achieve comprehensive coverage monitoring of power facilities. At the same time, multi-dimensional heterogeneous sensors collect data synchronously, and the edge layer and cloud platform process the data collaboratively. The data transmission latency is low, which can capture instantaneous or sudden fault precursor data, completely solving the problem of lag in traditional manual inspection and realizing real-time monitoring.

[0015] 2. Compared with existing technologies, this invention overcomes the limitations of traditional monitoring systems' single-dimensional threshold alarms. Through multi-dimensional heterogeneous data fusion and acquisition, it provides comprehensive input features for the AI ​​model. Utilizing a Long Short-Term Memory (LSTM) network model for in-depth data mining, it can accurately predict the remaining lifespan of equipment and generate a health index. This upgrades the maintenance model from passive repair after a failure or fixed-time preventative maintenance to predictive maintenance based on the actual condition of the equipment. This model avoids unnecessary maintenance operations, reduces resource waste, and ensures timely maintenance at critical points of equipment degradation, significantly extending equipment lifespan and reducing the failure rate.

[0016] 3. Compared with existing technologies, the wireless deployment method of this invention eliminates the complex wiring construction and subsequent maintenance costs of wired systems, making it particularly suitable for geographically dispersed and harsh environments. Furthermore, AI-driven predictive maintenance reduces the frequency of manual inspections, lowers labor costs, and avoids significant economic losses caused by downtime due to malfunctions.

[0017] 4. Compared with existing technologies, this invention preprocesses and encrypts data at the edge gateway before transmission, which reduces transmission bandwidth pressure and effectively prevents data from being stolen or tampered with during transmission, thus ensuring data security. At the same time, the combination of feature extraction at the edge layer and iterative training of AI models in the cloud improves the efficiency and accuracy of data analysis. The high throughput design of the time-series database supports the rapid storage and retrieval of massive amounts of data, providing efficient data support for real-time status assessment and predictive maintenance. Attached Figure Description

[0018] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] Reference Figure 1-2 The present invention provides a remote monitoring method for power facilities based on the Internet of Things, which consists of system deployment and detailed implementation of method steps.

[0020] (I) System Deployment The monitoring system in this embodiment consists of three parts: edge layer devices, transmission network, and cloud platform. The specific deployment is as follows: Edge layer device deployment: Multi-dimensional heterogeneous sensor deployment: Temperature, humidity, and partial discharge sensors are deployed on the top of the transformer tank; vibration and noise sensors are deployed on the sidewalls of the tank; and an oil chromatography sensor (detecting components: H2, CH4, C2H6, C2H4, C2H2, detection limit ≤1μL / L) is deployed at the transformer oil conservator. All sensors are connected to IoT terminals and gateways via waterproof and explosion-proof junction boxes, and their deployment locations avoid high-voltage energized areas, meeting power safety protection requirements.

[0021] IoT terminal and gateway configuration: Select an industrial-grade RTU gateway (supporting LPWAN / NB-IoT / 5G multi-mode communication), integrate a lightweight TensorFlowLite model, configure 2GB of local storage (caching key data within 1 week), and support AES-256 encryption algorithm.

[0022] Transmission network selection: This embodiment is deployed in a substation in a remote suburb, and NB-IoT low power wide area network is selected. The data transmission rate is configured to 128kbps to meet the real-time monitoring requirements; the communication protocol between the gateway and the cloud platform adopts the MQTT protocol.

[0023] Cloud platform setup: The cloud platform is deployed on the power system's private cloud, with the following configuration: Data reception and storage services: A Kafka cluster (3-node deployment) is used to receive data, and InfluxDB is selected as the time-series database; Big data processing module: Based on the Spark Streaming framework, it can process data in real time with a throughput of ≥50MB / s and supports operations such as data cleaning and feature fusion; AI Predictive Analysis Module: Deploys a Long Short-Term Memory Network model (input feature dimension: 20 dimensions, including mean temperature, root mean square of vibration, partial discharge PRPD spectral features, etc.; output is RUL predicted value (unit: day) and health index HI). The model is trained using the Adam optimizer with a learning rate of 0.001 and 500 iterations. Digital Twin Model: Based on Unity3D, a digital twin of the transformer is constructed. The transformer CAD design model is imported, and parameters such as internal temperature distribution, oil flow status, vibration and displacement are mapped through real-time data. The visualization delay is ≤1 second.

[0024] (II) Detailed Implementation of Methods and Procedures S101: Multidimensional Heterogeneous Sensor Data Acquisition and Timestamp Marking. It selects multidimensional heterogeneous sensors such as temperature / humidity, vibration / noise, partial discharge / oil chromatography, covering all environmental, mechanical, electrical, and chemical parameters of equipment operation. Sensor synchronization is achieved through a four-layer processing: the gateway has a built-in high-precision RTC module, calibrating the master clock via NTPv4 / PTPv2 protocol (annual drift ≤31.5 seconds), and then aligning all sensor clocks through the communication interface (synchronization accuracy ≤1ms); a mandatory unified sampling period of 100ms / time is enforced, with high-frequency sensors downsampling and low-frequency sensors interpolating for completion; a hardware pulse-first synchronization triggering mechanism is adopted (trigger delay ≤1μs); a multi-protocol conversion module adapts to different sensor interfaces, measures and compensates for transmission delays, ensuring synchronized data reception. All acquired data is marked with millisecond-level timestamps and sensor IDs to ensure timing consistency.

[0025] S102: Edge Data Filtering, Normalization, and Feature Extraction. As an edge computing node, the IoT gateway first performs quality checks on the raw data, eliminating outliers (such as abnormal peak values ​​from vibration sensors caused by electromagnetic interference) using the 3σ criterion. Then, Min-Max normalization is used to map the data to the [0,1] interval, eliminating dimensional differences. Subsequently, time-domain features (root mean square of vibration, peak factor, etc.) and frequency-domain features (average discharge quantity, phase concentration, discharge repetition frequency, etc. of the PRPD spectrum) are extracted. The gateway integrates a lightweight MobileNet model, comparing the extracted features with a preset template to complete initial local anomaly detection, uploading only abnormal data or key features to reduce invalid data transmission.

[0026] S103: Data Encryption and Transmission Protocol Encapsulation. The gateway uses the AES-256 high-strength encryption algorithm to encrypt the processed feature data and encapsulates it into MQTT or CoAP standard IoT protocol packets. The transmission network is selected based on the deployment scenario: LoRa / NB-IoT low-power wide area network is used in remote areas, while 5G cellular networks can be used in urban / factory scenarios to ensure secure, low-latency data transmission to the cloud platform.

[0027] S104: Cloud-based data reception, verification, and time-series database storage. The cloud-based Kafka cluster receives transmitted data, verifies data integrity using a CRC32 checksum, and decrypts it using a negotiated key after confirmation. The decrypted valid data is then written to the InfluxDB time-series database according to an index structure of device ID, sensor ID, and timestamp. This database supports high-throughput writing and fast querying of massive amounts of data, preserving complete time-series information and sensor identifiers, providing high-quality data support for subsequent model training and prediction.

[0028] S105: Real-time status assessment and threshold alarm judgment. The cloud platform calls preset safety thresholds (such as the upper limit of 85℃ for 110kV transformer temperature and 50pC for partial discharge) to perform real-time analysis on the current data. If the data exceeds the threshold, it is judged as an equipment abnormality, and step S108 is executed directly; if the data does not exceed the limit, it is judged as a normal state, and the model training stage is entered.

[0029] S106: Historical Data Analysis and Offline Iterative Training of the Long Short-Term Memory Network Model. The big data processing module extracts historical data (including normal operation, slight degradation, and pre-fault states) of the target device from the time-series database for the past three months, and divides it into training, validation, and test sets in a 7:2:1 ratio. Offline incremental training is performed on the Long Short-Term Memory network model: the model input is a time-series sequence with 1008 time steps × 20-dimensional features. The first LSTM layer captures short-term fluctuations at the hourly level, and the second LSTM layer captures long-term degradation trends at the daily / weekly level. After feature fusion through a fully connected layer, the remaining lifespan and health index prediction results are output. Training uses the Adam optimizer (initial learning rate 0.001, decaying to 0.9 times every 10 epochs), a weighted loss function (RUL weight 0.6, HI weight 0.4), 500 training epochs, a batch size of 64, and an early stopping strategy to avoid overfitting. In incremental training mode, the LSTM layer parameters are frozen and only the fully connected layers are fine-tuned. The training cycle is 1-7 days / cycle. The trained model is stored by device ID and training timestamp, and the last 5 versions are retained to support automatic rollback.

[0030] S107: Equipment Remaining Life Prediction and Health Index Generation. The AI ​​prediction and analysis module loads the latest version of the Long Short-Term Memory network model, inputs the current 7 days of time-series data (preprocessed and aligned with the training data format), and obtains the raw prediction results. Remaining life is calibrated through outlier correction (<0 is taken as 0, exceeding the design life is taken as the upper limit), trend smoothing (weighted moving average of the last 3 times), and unit conversion (day → year / month). The health index is constrained by five layers: input normalization, model output constraints (Sigmoid + 100x scaling), expert rule quantification calibration (fixed weights + limited correction range), boundary forced correction (0-100 as a safety net), and trend verification (abnormal fluctuation smoothing), ensuring that the value is within the range of 0-100, truly reflecting the health status of the equipment.

[0031] S108: Fault Warning Level Classification and Maintenance Instruction Issuance. Based on the real-time anomaly results of S105 and the prediction results of S107, three warning levels are classified according to quantitative rules: Level I Emergency Shutdown (requires response within 12 hours), Level II Near-Term Maintenance (repair within 7-30 days), and Level III Observation (enhanced monitoring, no shutdown required). The remote monitoring center pushes maintenance instructions via SMS and APP, specifying the repair content (such as fault location, parts to be replaced), time nodes, and priorities; the digital twin model synchronously visualizes the fault area and equipment status, assisting maintenance personnel in accurate repair.

[0032] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for remote monitoring of power facilities based on the Internet of Things, characterized in that, Includes the following steps: S101: Multidimensional heterogeneous sensor data acquisition and timestamp marking: Key operating parameters of power facilities are synchronously collected using multi-dimensional heterogeneous sensors, and all collected data are precisely timestamped. The key operating parameters include temperature, humidity, vibration, noise, partial discharge, and oil chromatography parameters. The synchronous acquisition of data by the sensors is achieved through a four-layer processing approach: hardware-level clock synchronization, unified sampling frequency, synchronous triggering mechanism, interface adaptation, and delay compensation, with a timing deviation of ≤1ms. S102: Edge data filtering, normalization, and feature extraction: Using IoT terminals and gateways as edge computing nodes, the raw data marked in step S101 is subjected to quality checks, outliers are removed, and then Min-Max normalization is performed. Time-domain features and frequency-domain features are extracted. The time-domain features include the root mean square of the vibration signal, and the frequency-domain features include the core numerical features of the PRPD spectrum of partial discharge. The extracted features are used for initial local anomaly judgment at the edge and input to the cloud model. S103: Data encryption and transmission protocol encapsulation: The IoT terminal and gateway use the AES-256 algorithm to encrypt the feature data processed in step S102 with high strength, and encapsulate it into a standard IoT protocol packet, which is then transmitted to the cloud platform through a low-power wide area network or a 5G cellular network; the standard IoT protocol includes the MQTT protocol or the CoAP protocol. S104: Cloud-based data reception, verification, and time-series database storage: The cloud platform's data receiving service receives the data transmitted in step S103, performs integrity verification and decryption, and then stores it in a high-throughput time-series database. The time-series database is indexed by device ID, sensor ID, and timestamp, retaining complete time-series information and sensor identifiers to provide full data for model training. S105: Real-time Status Assessment and Threshold Alarm Judgment: The cloud platform performs real-time analysis on the data stored in step S104 and determines whether the device's operating status is abnormal based on a preset safety threshold. If the data exceeds the preset safety threshold, it is determined to be in an abnormal state, and step S108 is executed. If the data does not exceed the preset safety threshold, it is determined to be in a normal state, and step S106 is executed. S106: Historical Data Analysis and Iterative Training of AI Prediction Models The big data processing module and AI prediction and analysis module of the cloud platform collaborate to perform offline incremental iterative training on the Long Short-Term Memory (LSTM) network model using historical data. The LTM network model adopts a four-layer architecture: an input layer, a two-layer LTM network feature extraction layer, a two-layer fully connected feature fusion layer, and an output layer. Specifically: The input layer receives time-series data, wherein the time step length of the time-series data is set to 1008 and the feature dimension is set to 20. In the two-layer LSTM feature extraction layer, the first layer of the Long Short-Term Memory network has 128 neurons, the activation function is tanh, the return sequence is set to True, the Dropout rate is 0.2, and the L2 regularization coefficient is 0.

001. The second layer of the Long Short-Term Memory network has 64 neurons, the activation function is tanh, the return sequence is set to False, the Dropout rate is 0.2, and the L2 regularization coefficient is 0.

001. In the two-layer fully connected feature fusion layer, the first fully connected layer has 32 neurons, the activation function is ReLU, and the L2 regularization coefficient is 0.001; the second fully connected layer has 16 neurons and the activation function is ReLU. The output layer has two neurons, which use the Linear activation function and the Sigmoid activation function respectively. The training process uses the Adam optimizer and a weighted loss function, and performs incremental iterative training at a preset cycle of 1-7 days. S107: Equipment Remaining Life Prediction and Health Index Generation: The AI ​​prediction and analysis module calls the long short-term memory network model trained in step S106, combines it with the current seven days of operating data stored in step S104, predicts the remaining lifespan of the device, and generates a comprehensive health index. S108: Fault Warning Level Classification and Maintenance Instruction Issuance: Based on the anomaly judgment result of step S105 and the prediction result of step S107, the fault warning level is divided, and maintenance instructions containing specific repair content and time limits are issued to maintenance personnel through the remote monitoring center; the fault warning level includes Level I emergency shutdown, Level II recent maintenance, and Level III observation.

2. The method for remote monitoring of power facilities based on the Internet of Things according to claim 1, characterized in that: The multidimensional heterogeneous sensor in step S101 includes a temperature sensor, a humidity sensor, a vibration sensor, a noise sensor, a partial discharge sensor, and an oil chromatography sensor.

3. The method for remote monitoring of power facilities based on the Internet of Things according to claim 1, characterized in that: In step S103, the low-power wide-area network includes either a LoRa network or an NB-IoT network.

4. The method for remote monitoring of power facilities based on the Internet of Things according to claim 1, characterized in that: In step S104, the time-series database adopts a high-throughput design, which supports the fast writing and querying of massive monitoring data. During the data storage process, complete timestamp information and sensor identification information are retained, providing full time-series data including normal operation, slight degradation, and historical fault precursor states for the training of the long short-term memory network model in step S106.

5. The method for remote monitoring of power facilities based on the Internet of Things according to claim 1, characterized in that: In step S106, the training parameters of the long short-term memory network model are configured as follows: the initial learning rate is 0.001, and the learning rate decays to 0.9 times the original value every 10 rounds; the number of training rounds is 500, the batch size is 64, and an early stopping strategy is enabled, which stops training if the validation set loss does not decrease for 10 consecutive rounds. In incremental training mode, the parameters of the feature extraction layer of the two-layer long short-term memory network are frozen, and only the parameters of the fully connected layer are fine-tuned. The trained model is stored with the device ID and training timestamp, retaining the last 5 versions and supporting automatic rollback.

6. The method for remote monitoring of power facilities based on the Internet of Things according to claim 1, characterized in that: The constraint rules for the value of the comprehensive health index in step S107 include: The input features are mapped to the [0,1] interval through Min-Max normalization; The output of the Long Short-Term Memory Network model is activated by the Sigmoid function and scaled by 100 times to obtain the basic comprehensive health index in the (0,100) range. The expert rule calibration adopts quantitative correction, with a fixed total weight of 1 and a single correction range of ≤ ±15 points; Forced boundary correction is applied before the final output: values ​​below 0 are taken as 0, values ​​above 100 are taken as 100, and integers are retained. Through trend verification of the results over the past three times, abnormal fluctuations were corrected using a weighted moving average to avoid random values.

7. The method for remote monitoring of power facilities based on the Internet of Things according to claim 1, characterized in that: In step S108, the rules for classifying fault warning levels are as follows: Level I Emergency Shutdown: Real-time core parameters exceed the standard by ≥30%, comprehensive health index <60 points, and remaining life expectancy ≤3 months; Level II Recent Maintenance: Real-time parameters exceed the standard by 10%-30%, comprehensive health index is 60-80 points, and remaining lifespan is 3-12 months; Level III observation: Real-time parameters exceed the standard by ≤10%, comprehensive health index ≥80, and remaining life expectancy >12 months.

8. The method for remote monitoring of power facilities based on the Internet of Things according to claim 1, characterized in that: Steps S101-S103 are executed at the edge layer, and steps S104-S108 are executed at the cloud platform layer; the edge layer includes multi-dimensional heterogeneous sensors and IoT terminals, and gateways; the cloud platform layer includes data receiving and storage services, big data processing modules, and AI predictive analysis modules.

9. The method for remote monitoring of power facilities based on the Internet of Things according to claim 1, characterized in that: The cloud platform constructs a digital twin model of the power facility, and presents the real-time status assessment results of step S105, the remaining life prediction results of step S107, and the equipment health index in a visual form. The visualization content includes the internal temperature distribution and stress change simulation of the equipment.

Images