Rotor lead fracture early warning method, device and equipment and storage medium

Abstract

The invention provides a rotor lead breakage early warning method, device and equipment and a storage medium. According to the method, strain data are acquired by using distributed optical fiber strain sensors arranged along the axial direction of a rotor lead, and discharge data are acquired by arranging a partial discharge sensor in a shielding case. And synchronously collecting the two types of data to form synchronous data. Characteristics such as discharge phase distribution entropy are calculated from synchronous data, and feature vectors are constructed and input into a pre-trained deep learning diagnosis model. The model adopts a hybrid structure combining a convolutional neural network and a long-short-term memory network, and outputs a diagnosis result representing the health state of the rotor lead. And triggering a corresponding grading early warning action according to a diagnosis result so as to realize differentiated timely response. According to the method, the accuracy and timeliness of early warning are effectively improved through multi-parameter synchronous fusion and deep feature extraction in combination with the diagnosis capability of the model, and the problems of high false alarm and slow response caused by single parameter and simple model in a traditional method are solved.

Description

Technical Field

[0001] This application belongs to the field of rotor lead monitoring, and particularly relates to a rotor lead breakage early warning method, device, equipment and storage medium. Background Technology

[0002] An online monitoring system for the rotor leads of a synchronous condenser is used to detect the risk of lead breakage and ensure the safe operation of the synchronous condenser under conditions of high-speed rotation, high temperature, and strong electromagnetic interference. This type of system mainly uses sensors to collect key parameters, enabling fault warnings and preventing equipment downtime or accidents due to lead breakage.

[0003] Existing technologies specifically involve using a single sensor (such as a fiber optic strain sensor or a partial discharge sensor) for data acquisition and implementing simple alarms through threshold comparison. For example, fiber optic strain sensors are deployed with monitoring points at fixed intervals to capture sudden strain changes; partial discharge sensors are installed near joints to detect discharge pulses. When the acquired data exceeds a preset threshold, the system triggers an alarm, achieving basic fault identification.

[0004] However, in the existing technology, the process of "single sensor acquisition and threshold comparison" is hampered by the single monitoring parameter, lack of multi-parameter synchronous acquisition, incomplete feature extraction, and simple diagnostic model. This results in the inability to accurately capture early fault signs, a high false alarm rate, untimely early warning response, and difficulty in effectively preventing rotor lead wire breakage. Summary of the Invention

[0005] The purpose of this application is to overcome the defects in the prior art and provide a rotor lead wire breakage early warning method, device, equipment and storage medium.

[0006] This application provides a rotor lead wire breakage early warning method, including: Strain data is obtained from a distributed optical fiber strain sensor, wherein multiple monitoring points are arranged along the rotor lead axis at preset intervals. Discharge data is acquired from a partial discharge sensor, which is housed within a shielding cover between the rotor leads and the slip ring connector. The data aggregation module synchronously collects the strain data and the discharge data to generate synchronous data; Based on the synchronization data, a feature vector is extracted by the feature extraction module, and the feature vector contains the discharge phase distribution entropy. The feature vector is input into a pre-trained deep learning diagnostic model, which outputs the health status of the rotor leads. The deep learning diagnostic model is a hybrid structure of convolutional neural network and long short-term memory network. The health status includes normal, mild, moderate and severe. Based on the health status, a tiered warning action is triggered, with each tiered warning action corresponding to a different level of response.

[0007] Optionally, the data aggregation module synchronously acquires the strain data and the discharge data to generate synchronous data, including: The data aggregation module uses a field-programmable gate array (FPGA) chip to time-align the strain data and the discharge data, and performs wavelet threshold denoising to remove noise.

[0008] Optionally, based on the synchronization data, a feature vector is extracted by the feature extraction module. The feature vector includes the discharge phase distribution entropy, comprising: The feature extraction module also extracts the discharge pulse frequency and the maximum discharge amount from the synchronization data, and combines them with the discharge phase distribution entropy to form a multi-dimensional feature vector.

[0009] Optionally, the deep learning diagnostic model is a hybrid structure of convolutional neural networks and long short-term memory networks, wherein: The convolutional neural network layer is used to extract local spatial features, and the long short-term memory network layer is used to extract time series trends.

[0010] Optionally, the deep learning diagnostic model is trained based on historical fault data through transfer learning, and the transfer learning uses incremental learning to update the model parameters quarterly.

[0011] Optionally, based on the health status, a tiered early warning action is triggered, including: When the health status is moderate, a remote alarm is sent to the maintenance personnel via mobile application and SMS, and it is recommended to shut down the system for inspection within a preset time period.

[0012] Optionally, the partial discharge sensor is disposed within the shielding cover of the rotor lead and slip ring connector, comprising: The metal shield is located near the rotor leads and the slip ring joint, and is used for electromagnetic shielding.

[0013] This application also provides a rotor lead wire breakage early warning device, comprising: The strain module acquires strain data from a distributed optical fiber strain sensor, which has multiple monitoring points arranged at preset intervals along the rotor lead axis. The discharge module acquires discharge data from a partial discharge sensor, which is housed inside a shielding cover between the rotor leads and the slip ring connector. The synchronization module, through the data aggregation module, synchronously acquires the strain data and the discharge data to generate synchronization data; The extraction module extracts a feature vector based on the synchronization data, and the feature vector contains the discharge phase distribution entropy. The model module inputs the feature vector into a pre-trained deep learning diagnostic model and outputs the health status of the rotor leads; the deep learning diagnostic model is a hybrid structure of convolutional neural network and long short-term memory network; the health status includes normal, mild, moderate and severe. The early warning module triggers tiered early warning actions based on the health status, with each tiered early warning action corresponding to a different level of response.

[0014] This application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0015] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the above-described method.

[0016] The beneficial effects of this application are: This application provides a rotor lead wire fracture early warning method, comprising: acquiring strain data from a distributed optical fiber strain sensor, wherein the distributed optical fiber strain sensor has multiple monitoring points arranged at preset intervals along the rotor lead wire axis; acquiring discharge data from a partial discharge sensor, wherein the partial discharge sensor is disposed within a shielding cover of the rotor lead wire and the slip ring joint; synchronously acquiring the strain data and the discharge data by a data aggregation module to generate synchronous data; extracting feature vectors by a feature extraction module based on the synchronous data, wherein the feature vectors contain the discharge phase distribution entropy; inputting the feature vectors into a pre-trained deep learning diagnostic model to output the health status of the rotor lead wire; wherein the deep learning diagnostic model is a hybrid structure of convolutional neural network and long short-term memory network; wherein the health status includes normal, slight, moderate and severe; and triggering a graded early warning action based on the health status, wherein the graded early warning action corresponds to different levels of response. This application achieves accurate and timely rotor lead wire breakage early warning by simultaneously acquiring and extracting features such as discharge phase distribution entropy through multi-parameter fusion, and using a CNN-LSTM deep learning model for state diagnosis and hierarchical early warning. It overcomes the shortcomings of existing technologies, such as high false alarm rate and untimely response due to single parameters and simple models. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the rotor lead wire breakage early warning process in this application. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is to be understood that various forms of implementation of the present disclosure are intended and should not be limited to the embodiments set forth herein. Rather, the embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] Please refer to Figure 1 This application provides a rotor lead wire breakage early warning method, applied in the field of online monitoring of synchronous condenser rotor leads, to solve the problem of rotor lead wire breakage early warning. The method includes: S101. Obtain strain data from a distributed optical fiber strain sensor, wherein the distributed optical fiber strain sensor has multiple monitoring points arranged at preset intervals along the rotor lead axis.

[0020] The distributed optical fiber strain sensor is part of the optical fiber strain monitoring subunit. The sensor type is a distributed optical fiber strain sensor, with one monitoring point arranged every 50mm along the rotor lead axis, which can accurately capture the strain change when the conductor microcrack occurs.

[0021] Key parameters include the use of polyimide high-temperature resistant material for the encapsulation layer (temperature resistance 200℃), strain resolution ±0.5μ, sampling frequency 1kHz, and acquisition of real-time strain data through optical signal demodulation.

[0022] By deploying multiple monitoring points, strain changes along the entire length of the rotor leads are ensured to be covered, thereby enabling early identification of microcracks. For example, during the operation of the synchronous condenser, the rotor rotates at high speed (3000 r / min and above), and the sensors monitor strain data in real time. When a sudden change in strain occurs, it indicates that there may be a potential crack.

[0023] Furthermore, the insulation layer temperature monitoring subunit is used to acquire temperature data from the miniature patch PT1000 platinum resistance sensor. Three monitoring points are set up on each lead branch (the two end joints + the middle section) to focus on monitoring local overheating of the insulation layer.

[0024] Key parameters include a measurement range of 40℃~200℃, a temperature accuracy of ±0.05℃, and a sampling frequency of 0.1Hz, which can accurately identify temperature anomalies caused by insulation aging.

[0025] The temperature monitoring subunit is equipped with sensors during the system deployment phase. Specifically, with the synchronous condenser off, the rotor lead surface is pre-treated by grinding and cleaning. Then, a PT1000 temperature sensor is attached to the outer surface of the insulation layer to ensure that the sensor is installed firmly and reliably.

[0026] For example, under high-temperature conditions, temperature sensors can detect localized overheating, helping to identify loose connections or aging insulation, and providing supplementary data for subsequent diagnosis.

[0027] S102. Obtain discharge data from a partial discharge sensor, wherein the partial discharge sensor is disposed inside the shielding cover of the rotor lead and the slip ring joint.

[0028] The partial discharge sensor is part of the partial discharge monitoring subunit. The sensor type is an ultra-high frequency (UHF) antenna sensor, which is installed in a metal shield near the rotor lead and the slip ring joint. It is specifically designed to capture partial discharge pulses generated by insulation aging.

[0029] Key parameters include a detection frequency range of 300MHz to 2GHz, sensitivity ≤0.5pC, and sampling frequency of 10MHz, which can effectively identify the pre-discharge signs before the insulation layer carbonizes.

[0030] By monitoring partial discharge, the insulation condition can be assessed to prevent lead failures caused by insulation aging. For example, under high temperature (≤200℃) and strong electromagnetic interference (≥1.5T) conditions, the sensor detects discharge pulses, providing a data basis for subsequent diagnosis.

[0031] The metal shield, located near the rotor leads and slip ring joint, provides electromagnetic shielding with an effectiveness ≥80dB (10kHz~1GHz), eliminating interference from the rotor's strong magnetic field (≥1.5T). During system deployment, sensor installation includes installing a UHF partial discharge sensor within the metal shield near the joint, ensuring accurate sensor positioning. This shield structure effectively reduces external interference and improves data reliability.

[0032] S103. The data aggregation module synchronously collects the strain data and the discharge data to generate synchronous data.

[0033] The data aggregation module integrates an FPGA chip to achieve synchronous acquisition and preliminary filtering of data from multiple sensors, thus avoiding the impact of data timing deviations on subsequent analysis.

[0034] The structural requirements include a volume of ≤100mm x 80mm x 50mm and a weight of ≤200g, and must meet the rotor dynamic balance requirements to prevent additional vibration during rotation.

[0035] FPGA chips are used to perform time alignment and preliminary processing on multi-source data to ensure data consistency. For example, during data acquisition, strain data is sampled at a frequency of 1kHz and discharge data at a frequency of 10MHz. The FPGA chip synchronously acquires these data to generate a unified synchronous data stream, laying the foundation for subsequent analysis.

[0036] Meanwhile, the data aggregation module performs time alignment of the strain data and the discharge data through a field-programmable gate array chip, and performs wavelet threshold denoising processing to remove noise.

[0037] The wavelet thresholding denoising algorithm is based on the db4 wavelet basis and has a decomposition level of 5. It is used to remove electromagnetic interference noise and improve data quality.

[0038] This noise reduction process can effectively eliminate noise introduced by strong electromagnetic interference (≥1.5T), ensuring data accuracy.

[0039] Data is stably transmitted to the ground center via a signal transmission unit, which includes a hybrid transmission channel and an electromagnetic shielding module. The main channel of the hybrid transmission channel is a high-frequency conductive slip ring with a contact resistance ≤3mΩ and a transmission bandwidth ≥100MHz, primarily transmitting high-precision strain and temperature data to ensure the stability of core parameter transmission. The backup channel is an industrial-grade 5G millimeter-wave module with a transmission rate ≥500Mbps and a latency ≤10ms, automatically switching in case of slip ring failure to avoid data interruption.

[0040] The transmission cable of the electromagnetic shielding module adopts a double-layer copper mesh + aluminum foil shielding structure with a shielding effectiveness of ≥80dB (10kHz~1GHz), which can eliminate the interference of the rotor's strong magnetic field (≥1.5T) on the signal and ensure the accuracy of data transmission.

[0041] During the system deployment phase, link debugging is carried out, connecting the signal lines of each sensor to the rotor-end data aggregation module and interfacing with the ground transmission unit through a high-frequency conductive slip ring; at the same time, the 5G millimeter-wave backup channel is debugged to ensure normal switching between the primary and backup channels.

[0042] In addition, the data preprocessing module of the ground processing and early warning center uses Kalman filtering to achieve time synchronization of multi-sensor data, which solves the timing deviation caused by the difference in sampling frequency of different sensors. For example, the sampling frequency of strain data is 1kHz, the sampling frequency of temperature data is 0.1Hz, and the sampling frequency of partial discharge data is 10MHz. The Kalman filter aligns the timestamps to improve data consistency.

[0043] Furthermore, this application employs a dynamic sampling control mechanism: The FPGA chip embeds a real-time pulse detection unit that continuously monitors the raw discharge data stream from the partial discharge sensor. This unit is configured to identify sudden discharge pulses with amplitudes exceeding a preset threshold.

[0044] Under stable operating conditions where no sudden pulses are detected, the distributed fiber optic strain sensor operates at the first sampling frequency. When the pulse detection unit identifies a sudden discharge pulse, the FPGA generates a control signal. This signal triggers the strain sensor's acquisition circuit, switching its sampling frequency to the second sampling frequency.

[0045] The second sampling frequency is higher than the first sampling frequency and is used to capture transient and weak strain changes that may occur synchronously with the discharge event.

[0046] In this application, the duration of the second sampling frequency is related to the duration of the burst pulse that triggers it. Specifically, the FPGA's internal timer begins timing from the rise of the pulse, and the second sampling frequency continues until the pulse ends, followed by a configurable window period to ensure complete coverage of the discharge event and its possible mechanical effects.

[0047] In this application, the discharge pulse signal of the FPGA serves as the trigger signal for the real-time adjustment of the strain data acquisition strategy through coordinated control. This mechanism enables the system to automatically acquire higher-precision mechanical condition data at potential fault critical points, thereby enabling early detection of mechanical damage signs that co-develop with insulation degradation, building upon conventional strain monitoring.

[0048] S104. Based on the synchronization data, the feature extraction module extracts a feature vector, which includes the discharge phase distribution entropy.

[0049] The feature extraction module also extracts the discharge pulse frequency and the maximum discharge amount from the synchronization data, and combines them with the discharge phase distribution entropy to form a multi-dimensional feature vector.

[0050] The frequency of discharge pulses is measured in pulses / min, the maximum discharge quantity is measured in pC, and the discharge phase distribution entropy is used to quantify the complexity of the phase distribution of the discharge pulses and reflect the aging state of the insulation.

[0051] By extracting multi-dimensional features, the lead condition is comprehensively characterized. For example, a 288-dimensional feature vector is extracted from the preprocessed data, including strain features, temperature features, and partial discharge features. However, this step focuses on discharge-related features, such as the discharge phase distribution entropy, which can identify insulation degradation trends.

[0052] The feature extraction module also extracts strain and temperature features from the synchronous data. The strain features include the peak strain value, standard deviation of strain fluctuation, and slope of strain trend within a unit time (1 min), reflecting the degree of conductor fatigue damage. The temperature features include the maximum temperature value, temperature change rate (°C / min), and temperature difference of the three-phase leads, which can identify insulation overheating and loose joints.

[0053] These features, together with the discharge features, form a 288-dimensional feature vector (based on 3 types of parameters x 8 features x 12 monitoring points), providing complete input for deep learning-based diagnostics. For example, in the real-time monitoring phase, the feature extraction module extracts all features from the preprocessed data to form a real-time feature parameter set, ensuring comprehensive diagnostics.

[0054] Furthermore, before extracting the discharge phase distribution entropy, the following steps are also included: The feature extraction module performs spatial analysis on the data from the distributed fiber optic strain sensor and calculates the strain distribution gradient between multiple monitoring points arranged along the rotor lead axis.

[0055] When the strain difference between adjacent monitoring points exceeds a preset threshold, it is marked as a "high strain gradient region".

[0056] Since the strain data and discharge data have been precisely synchronized in time by the data aggregation module, a correlation can be established between "spatial high-risk areas" and "temporal discharge events" to pinpoint the specific time period in which the high strain gradient region occurs.

[0057] When calculating the discharge phase distribution entropy, which characterizes the statistical properties of discharge, the weight of the time window used to calculate the PPDE is dynamically adjusted based on the time period in which the high strain gradient region appears. Discharge pulse data collected within the time period corresponding to the "high strain gradient region" are assigned a higher weight coefficient (e.g., weight = 1.5). Discharge data collected during the strain-stable period are assigned a standard weight (weight = 1.0).

[0058] Using this weighted discharge data to calculate the final discharge phase distribution entropy, the PPCE value is no longer just a reflection of the statistical characteristics of the discharge signal, but contains key information about the mechanical stress state.

[0059] The discharge phase distribution entropy can reflect the composite index of the coupling effect of "mechanical stress concentration" and "electrical insulation degradation". Thus, a forced deep fusion is performed at the feature level, so that the feature vector input to the final deep learning model naturally contains strong correlation information on "where (spatial location) mechanical risks are causing or aggravating what electrical characteristic changes".

[0060] S105. Input the feature vector into a pre-trained deep learning diagnostic model and output the health status of the rotor lead; the deep learning diagnostic model is a hybrid structure of convolutional neural network and long short-term memory network; the health status includes normal, mild, moderate and severe.

[0061] The deep learning diagnostic model is a CNN-LSTM hybrid neural network. The input layer is a 288-dimensional feature vector (based on 3 types of parameters x 8 features x 12 monitoring points). The hidden layer contains 2 CNN convolutional layers for extracting local spatial features and 2 LSTM layers for capturing time series trends. The output layer has 4 types of states.

[0062] The training method adopts transfer learning, and the model is optimized based on 1,000 sets of historical fault data to improve the diagnostic accuracy in small sample scenarios.

[0063] Intelligent diagnosis can be achieved through deep learning models. For example, after analyzing feature vectors, the model outputs the health status to help maintenance personnel issue early warnings.

[0064] The deep learning diagnostic model is trained on historical fault data using transfer learning, with incremental learning used to update model parameters quarterly. During the model optimization phase, newly added normal operation data and fault data (including pre- and post-maintenance status data) are input into the deep learning diagnostic model each quarter. Incremental learning optimizes model parameters, continuously improving the model's ability to identify new fault modes, reducing the false alarm rate (target ≤5%), and ensuring long-term monitoring accuracy. For example, historical data is used to optimize the model, improving its adaptability to new faults and ensuring continuously improving early warning accuracy.

[0065] S106. Based on the health status, trigger a graded warning action, wherein the graded warning action corresponds to different levels of response.

[0066] The warning levels are divided into Level I (normal, only data is recorded), Level II (minor, local audio and visual alerts + maintenance suggestions), Level III (moderate, remote APP + SMS alerts, it is recommended to shut down the system for inspection within 48 hours), and Level IV (severe, triggering emergency shutdown + uploading to the dispatch center).

[0067] When the health status is moderate, a remote alarm is sent to the maintenance personnel via mobile application and SMS, and it is recommended to shut down the system for inspection within a preset period of 48 hours.

[0068] The system ensures timely response through tiered early warnings. For example, if the diagnosis result is moderate, the system will automatically push an alarm to prompt maintenance personnel to take action, thus forming a closed loop of maintenance.

[0069] The tiered early warning and operation and maintenance management module also provides operation and maintenance functions, including automatic generation of monitoring reports, fault history tracing, and maintenance record archiving, supporting the entire lifecycle operation and maintenance of synchronous condensers. The operation and maintenance closed loop includes maintenance personnel taking maintenance measures based on early warning information (such as replacing leads, reinforcing connectors, etc.); after the fault is resolved, the system re-enters monitoring mode, records the post-maintenance status recovery, forming the operation and maintenance closed loop. For example, in the status diagnosis and early warning phase, the system records all operations and generates reports for subsequent analysis, ensuring traceability of operation and maintenance.

[0070] Furthermore, the deep learning diagnostic model is trained on historical fault data using transfer learning, which updates the model parameters quarterly using incremental learning. Each quarter, newly added normal operation data and fault data (including pre- and post-maintenance status data) are input into the model, and parameters are optimized through incremental learning to continuously improve the model's ability to identify new fault modes, reduce the false alarm rate (target ≤5%), and ensure long-term monitoring accuracy.

[0071] The partial discharge sensor is housed within the shielding cover of the rotor leads and slip ring connector. The metal shielding cover, located near the rotor leads and slip ring connector, serves for electromagnetic shielding. The shielding effectiveness is ≥80dB (10kHz~1GHz), eliminating interference from the rotor's strong magnetic field (≥1.5T) and ensuring accurate data transmission.

[0072] This application also provides a rotor lead wire breakage early warning device, comprising: The strain module acquires strain data from a distributed optical fiber strain sensor, which has multiple monitoring points arranged at preset intervals along the rotor lead axis. The discharge module acquires discharge data from a partial discharge sensor, which is housed inside a shielding cover between the rotor leads and the slip ring connector. The synchronization module, through the data aggregation module, synchronously acquires the strain data and the discharge data to generate synchronization data; The extraction module extracts a feature vector based on the synchronization data, and the feature vector contains the discharge phase distribution entropy. The model module inputs the feature vector into a pre-trained deep learning diagnostic model and outputs the health status of the rotor leads; the deep learning diagnostic model is a hybrid structure of convolutional neural network and long short-term memory network; the health status includes normal, mild, moderate and severe. The early warning module triggers tiered early warning actions based on the health status, with each tiered early warning action corresponding to a different level of response.

[0073] This application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0074] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the above-described method.

[0075] The above description of the embodiments is provided to enable those skilled in the art to understand and apply this application. Those skilled in the art will readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without inventive effort. Therefore, this application is not limited to the above embodiments, and any improvements and modifications made to this application based on the disclosure thereof should be within the scope of protection of this application.

Claims

1. A rotor lead breakage early warning method, characterized by, The method comprises the following steps: obtaining strain data from a distributed optical fiber strain sensor, the distributed optical fiber strain sensor being arranged with multiple monitoring points at preset intervals along the axial direction of the rotor lead; obtaining discharge data from a partial discharge sensor, the partial discharge sensor being arranged in a shield cover of a joint between the rotor lead and the collector ring; synchronously collecting the strain data and the discharge data by a data aggregation module to generate synchronous data; extracting a feature vector from the synchronous data by a feature extraction module, the feature vector including discharge phase distribution entropy; inputting the feature vector into a pre-trained deep learning diagnosis model to output a health status of the rotor lead; the deep learning diagnosis model is a hybrid structure of a convolutional neural network and a long short-term memory network; the health status includes normal, slight, moderate and severe; triggering a hierarchical early warning action according to the health status, the hierarchical early warning action corresponding to different levels of response.

2. The method of claim 1, wherein, The method further comprises the following steps: The data aggregation module performs time alignment on the strain data and the discharge data through a field programmable gate array chip, and performs wavelet threshold denoising processing to remove noise.

3. The method of claim 1, wherein, The method further comprises the following steps: The feature extraction module further extracts discharge pulse frequency and discharge amount maximum from the synchronous data, and combines them with the discharge phase distribution entropy to form a multi-dimensional feature vector.

4. The method of claim 1, wherein, The deep learning diagnosis model is a hybrid structure of a convolutional neural network and a long short-term memory network, wherein: The convolutional neural network layer is used to extract spatial local features, and the long short-term memory network layer is used to extract time series trends.

5. The method of claim 1, wherein, The deep learning diagnosis model is trained based on historical fault data through transfer learning, and the transfer learning updates model parameters every quarter using incremental learning.

6. The method of claim 1, wherein, The method further comprises the following steps: When the health status is moderate, a remote alarm is sent to an operation and maintenance personnel through a mobile application and a short message, and it is suggested to stop and check within a preset period.

7. The method of claim 1, wherein, The partial discharge sensor is arranged in a shield cover of a joint between the rotor lead and the collector ring, and the method further comprises the following steps: The metal shield cover is located near the joint between the rotor lead and the collector ring, and is used for electromagnetic shielding.

8. A rotor lead breakage early warning device characterized by, The method comprises the following steps: a strain module, for obtaining strain data from a distributed optical fiber strain sensor, the distributed optical fiber strain sensor being arranged with multiple monitoring points at preset intervals along the axial direction of the rotor lead; a discharge module, for obtaining discharge data from a partial discharge sensor, the partial discharge sensor being arranged in a shield cover of a joint between the rotor lead and the collector ring; a synchronization module, for synchronously collecting the strain data and the discharge data by a data aggregation module to generate synchronous data; a feature extraction module, for extracting a feature vector from the synchronous data, the feature vector including discharge phase distribution entropy; a feature extraction module, for extracting a feature vector from the synchronous data, the feature vector including discharge phase distribution entropy; The model module inputs the feature vector into a pre-trained deep learning diagnosis model, and outputs a health state of the rotor lead; the deep learning diagnosis model is a hybrid structure of a convolutional neural network and a long short-term memory network; the health state includes normal, slight, moderate, and severe; The early warning module triggers a hierarchical early warning action according to the health state, and the hierarchical early warning action corresponds to different levels of response.

9. An electronic device, comprising: A computer readable storage medium, having stored thereon a computer program, the computer program, when executed in a computer, causing the computer to perform the method of claims 1-7.

10. A computer-readable storage medium, characterized in that, A computer readable storage medium, having stored thereon a computer program, the computer program, when executed in a computer, causing the computer to perform the method of claims 1-7.

Images