Lightning arrester leakage current monitoring system and method based on multi-modal data fusion

Abstract

The invention discloses a lightning arrester leakage current monitoring system based on multi-modal data fusion and a method thereof. The system comprises a multi-modal sensing unit, an edge calculation unit and a cloud analysis platform. The multi-mode sensing unit realizes data interaction with the edge computing unit through Lora and 5G dual-mode communication; and the edge computing unit realizes data interaction with the cloud analysis platform through HTTP and MQTT communication. The method has the following beneficial effects: 1, a deep coupling model of multi-dimensional environmental parameters and electrical characteristics is constructed, and the current powerful reinforcement learning algorithm with a memory enhancement mechanism is combined, so that the monitoring reliability under a complex working condition is remarkably improved; 2, the system has the functions of dynamic threshold adjustment, environmental interference compensation and equipment degradation prediction, the false alarm rate can be effectively reduced, the service life of equipment can be prolonged, and a new solution thought is provided for anti-interference of an online monitoring system;

Description

Technical Field

[0001] The present invention relates to the technical field of real-time diagnosis and life prediction of lightning arrester status, and in particular to a lightning arrester leakage current monitoring system and method based on multi-modal data fusion. Background Art

[0002] Existing technology defects

[0003] Single data dimension: For example, CN202110123456.7 only uses leakage current and temperature data, and does not consider changes in surface conductivity caused by rainfall;

[0004] Insufficient compensation for environmental interference: For example, although CN202210654321.1 introduced humidity correction, it did not establish a multi-parameter coupling impact model;

[0005] Poor dynamic adaptability: Most of the existing threshold settings are fixed values, which cannot adapt to complex working conditions such as day and night temperature differences and seasonal changes;

[0006] Difficulty in tracing the source of faults: For example, the CN202023135961.3 device cannot distinguish between leakage anomalies caused by internal moisture and external flashover. Summary of the invention

[0007] The present invention relates to a system and method for accurately monitoring the leakage current of lightning arresters by integrating multi-dimensional sensor data such as temperature and humidity, partial discharge, rainfall, and line electric field, combined with a deep reinforcement learning algorithm. It is suitable for real-time diagnosis and life prediction of lightning arrester status in power transmission and transformation systems with voltage levels of 35kV and above.

[0008] According to an embodiment of the present invention, a lightning arrester leakage current monitoring system based on multimodal data fusion is provided, comprising: a multimodal sensing unit, an edge computing unit and a cloud analysis platform;

[0009] The multimodal perception unit realizes data interaction with the edge computing unit through Lora and 5G dual-mode communication;

[0010] The edge computing unit communicates with the cloud analysis platform through HTTP and MQTT to achieve data interaction;

[0011] The multimodal sensing unit consists of a leakage current detection module, a partial discharge detection module, an environmental parameter acquisition module, and an electric field detection module, and is used to obtain leakage current, partial discharge, temperature and humidity, rainfall, and electric field strength data, and perform pre-filtering on the acquired data;

[0012] The edge computing unit is built-in with a spatio-temporal alignment engine algorithm module, a feature decoupling model, and a dynamic decision-making algorithm module, which are used to perform spatio-temporal alignment and feature extraction on the data processed by the multi-modal sensing unit, run an environmental coupling compensation algorithm, and finally generate a preliminary diagnosis result;

[0013] The cloud analysis platform deploys a digital twin model, a device degradation prediction algorithm module, and a dynamic threshold generator. Based on the preliminary diagnosis result generated by the edge computing unit, it updates the parameters of the digital twin model, performs device degradation prediction, and issues dynamic threshold parameters.

[0014] Furthermore,

[0015] The leakage current detection module consists of a TMR sensor array and is installed at the lightning arrester grounding downlead;

[0016] The partial discharge detection module adopts a combined structure of a UHF sensor and a high-frequency current transformer;

[0017] The environmental parameter acquisition module includes a temperature and humidity sensor, a piezoelectric rain gauge, and a laser scattering pollution degree sensor;

[0018] The electric field detection module includes a three-axis capacitive electric field induction sensor.

[0019] Furthermore,

[0020] The spatio-temporal alignment engine algorithm module uses a fusion algorithm of dynamic time warping and Kalman filtering to achieve spatio-temporal alignment of multi-modal data. The specific steps include:

[0021] Mark the time stamps of each sensor data stream;

[0022] Calculate the phase difference Δτ of different sensor data streams:

[0023]

[0024] Use a Kalman filter to compensate for the spatial differences caused by the physical distances of different sensors;

[0025] The feature decoupling model is constructed based on a neural network with physical constraints, and the loss function satisfies:

[0026]

[0027] The first part , is the weight coefficient, and represent the predicted environmental features and the actual environmental features respectively;

[0028] The second part , where Represents the derivative (rate of change) of environmental characteristics with respect to time. is a predefined environmental coupling equation, associated with temperature T, humidity H, rainfall R, electric field strength E, contamination degree F, partial discharge quantity PD, and redundant environmental parameter N. is another weight coefficient used to measure the difference between the feature rate of change and the f function.

[0029] The dynamic decision-making algorithm module adopts a memory-augmented deep Q-network. The network structure includes: an input layer; a 128-dimensional feature vector; an LSTM memory unit with a hidden layer dimension of 64; and a double Q-network that outputs a threshold adjustment strategy and an alarm level respectively.

[0030] Furthermore,

[0031] The training method of the memory-augmented deep Q-network is as follows:

[0032] Construct a training set containing 20,000 groups of samples, covering: 6 typical fault modes, such as moisture absorption, valve disc aging, and external insulation damage; 12 types of environmental condition combinations, such as heavy rain, high humidity, and salt fog.

[0033] Design a multi-objective reward function:

[0034]

[0035] The weight coefficient has a value of 0.5, has a value of 0.3, has a value of 0.2;

[0036] Adopt a prioritized experience replay mechanism, and the priority calculation is:

[0037]

[0038] where δi is the TD error and the ε value is 0.01;

[0039] The network is updated hard every 1000 steps.

[0040] Furthermore,

[0041] The spatio-temporal alignment engine algorithm module has an abnormal data cleaning mechanism. For temperature data, sliding window median filtering is used, and the window width W satisfies:

[0042]

[0043] Perform outlier detection based on Mahalanobis distance on the electric field strength data:

[0044]

[0045] Among them, is a statistic, is the electric field strength data, is the population mean, is the weighted deviation of the electric field strength in different directions;

[0046] A credibility evaluation matrix is established for the data:

[0047] +

[0048] wherein, is the credibility score, is the signal-to-noise ratio weight coefficient, = 0.6, is the consistency weight coefficient = 0.4, the signal-to-noise ratio weight accounts for 60%, is the consistency value.

[0049] Furthermore,

[0050] The digital twin model includes:

[0051] A three-dimensional electromagnetic field simulation module that calculates the surface potential distribution of the arrester based on the finite element method;

[0052] A thermodynamics coupling module that simulates the temperature rise of the zinc oxide varistor at different ambient temperatures;

[0053] An equipment degradation prediction algorithm module that adopts an improved Transformer architecture:

[0054] Input sequence: {fundamental amplitude of leakage current, third harmonic content, partial discharge times, environmental characteristic parameters};

[0055] The attention mechanism calculates the weights:

[0056]

[0057] wherein, is the input feature representation, and the output prediction results are: remaining useful life estimate (RUL) and health index (HI);

[0058] A dynamic threshold generator that generates an adaptive alarm threshold according to the historical state of the equipment:

[0059]

[0060] wherein, k is the degradation rate coefficient, which is updated every 6 hours.

[0061] According to another embodiment of the present invention, a method for monitoring the leakage current of an arrester based on multi-modal data fusion is provided, including the following steps:

[0062] Modal data acquisition stage: Synchronously acquire leakage current, partial discharge, temperature and humidity, rainfall, and electric field intensity data, and perform pre-filtering processing on the acquired data;

[0063] Edge-side data processing stage: Perform spatio-temporal alignment and feature extraction on the filtered data, run the environmental coupling compensation algorithm, and finally generate a preliminary diagnosis result;

[0064] Cloud deep analysis stage: Based on the preliminary diagnosis result, update the parameters of the digital twin model, perform equipment degradation prediction, and issue dynamic threshold parameters;

[0065] Closed-loop control stage: Adjust the local monitoring strategy according to the latest acquired threshold parameters, and trigger equipment maintenance suggestions or protection actions.

[0066] Furthermore,

[0067] The specific steps of the environmental coupling compensation algorithm are as follows:

[0068] Establish a rainfall impact correction model:

[0069]

[0070] Where, is the original rainfall, The value is 0.15, R is the rainfall intensity (mm / h), H is the relative humidity (%), is an exponential function;

[0071] Temperature drift compensation:

[0072]

[0073] Where, the α value is 0.12 mA / °C, the β value is 0.002 mA / °C², The value is 25°C;

[0074] Calculation of pollution degree influence factor:

[0075]

[0076] Where, is the particle diameter, is the concentration, .

[0077] Furthermore,

[0078] The specific steps of equipment degradation prediction are as follows:

[0079] Feature extraction:

[0080] Time-domain features: Current mean, variance, kurtosis;

[0081] Frequency domain features: proportion of the third harmonic, resonance frequency offset;

[0082] Degradation state classification:

[0083] Level 1 (healthy): HI ≥ 0.9;

[0084] Level 2 (warning): 0.7 ≤ HI < 0.9;

[0085] Level 3 (fault): HI < 0.7;

[0086] Remaining useful life prediction:

[0087]

[0088] wherein, The value is 0.6, is the change rate of the health index.

[0089] Furthermore,

[0090] The specific steps for adjusting the local monitoring strategy according to the latest obtained threshold parameters are as follows:

[0091] Fundamental current threshold:

[0092] +0.03 HI

[0093] wherein, is the average current, is the standard deviation, and HI is the health index.

[0094] HI is adjusted according to the harmonic content:

[0095]

[0096] wherein, Content of the third harmonic.

[0097]

[0098] wherein, is the current change rate, is the reference current.

[0099] According to an arrester leakage current monitoring system and method based on multimodal data fusion according to an embodiment of the present invention, the following beneficial effects are achieved:

[0100] 1. It is proposed to significantly improve the monitoring reliability under complex working conditions by constructing a deep coupling model of multi-dimensional environmental parameters and electrical characteristics and combining with the current powerful reinforcement learning algorithm with a memory enhancement mechanism.

[0101] 2. The system has functions of dynamic threshold adjustment, environmental interference compensation and equipment degradation prediction, which can effectively reduce the false alarm rate and extend the service life of the equipment, providing a new solution idea for anti-interference of the online monitoring system.

[0102] 3. The monitoring accuracy harmonic separation is increased to 42 dB (traditional method ≤ 35 dB); the recognition rate of moisture-related faults is 98.7% (the highest in the comparative literature is 93.2%).

[0103] 4. Adaptive ability: the threshold adjustment response time < 3 s; the false alarm rate is reduced by 76% under heavy rain conditions.

[0104] 5. Energy efficiency ratio: adopting event-driven data transmission, the traffic is saved by 82%; the standby power consumption of the edge node is 0.8 mW (≥ 5 mW for similar products).

[0105] It should be understood that both the foregoing general description and the following detailed description are exemplary and are intended to provide further explanation of the claimed technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] Figure 1 FIG. is a three-level architecture and data flow diagram of a lightning arrester leakage current monitoring system based on multi-modal data fusion according to an embodiment of the present invention.

[0107] Figure 2 FIG. is a flowchart of a lightning arrester leakage current monitoring method based on multi-modal data fusion according to an embodiment of the present invention.

[0108] Figure 3 FIG. is a structural diagram of the ME-DQN network in a lightning arrester leakage current monitoring system based on multi-modal data fusion according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0109] The following will describe in detail the preferred embodiments of the present invention with reference to the accompanying drawings and further elaborate on the present invention.

[0110] First, in combination with Figure 1 、 3 A lightning arrester leakage current monitoring system based on multi-modal data fusion according to an embodiment of the present invention will be described for real-time diagnosis and life prediction of the lightning arrester status, and its application scenarios are very wide.

[0111] Such as Figure 1 、 3As shown in the figure, a lightning arrester leakage current monitoring system based on multi-modal data fusion according to an embodiment of the present invention includes a multi-modal sensing unit, an edge computing unit, and a cloud analysis platform;

[0112] As Figure 1 shown, the multi-modal sensing unit realizes data interaction with the edge computing unit through Lora and 5G dual-mode communication; the edge computing unit realizes data interaction with the cloud analysis platform through HTTP and MQTT communication.

[0113] Specifically, in this embodiment, the multi-modal sensing unit is composed of a leakage current detection module, a partial discharge detection module, an environmental parameter acquisition module, and an electric field detection module, which is used to acquire leakage current, partial discharge, temperature and humidity, rainfall, and electric field intensity data, and perform pre-filtering processing on the acquired data.

[0114] Furthermore, in this embodiment:

[0115] The leakage current detection module is composed of a TMR sensor array and is installed at the grounding downlead of the lightning arrester; it includes: a fundamental wave detection channel (50Hz ± 5Hz) with a resolution of 0.1mA; a harmonic detection channel (3rd, 5th, 7th harmonics) with a signal-to-noise ratio ≥ 60dB.

[0116] The partial discharge detection module adopts a combined structure of a ultra-high frequency (UHF) sensor and a high-frequency current transformer (HFCT); among them: the UHF sensor planar antenna has a working bandwidth of 300MHz - 1.5GHz, and the HFCT sensor adopts a Rogowski coil structure with a working bandwidth of 10kHz - 30MHz.

[0117] The environmental parameter acquisition module includes a temperature and humidity sensor: the measurement range is -40°C - 125°C, and the accuracy is ±0.3°C; a piezoelectric rain gauge: the resolution is 0.01mm / min, and the anti-wind interference level ≥ 8 levels; a laser scattering type pollution degree sensor: it can detect particles with a diameter of 2 - 200um.

[0118] The electric field detection module includes a three-axis capacitive electric field induction sensor with a range of 0 - 50kV / cm and an anisotropy error ≤ 3%; the response time of the electric field change rate detection circuit ≤ 10us.

[0119] Specifically, in this embodiment, the edge computing unit is built-in with a spatio-temporal alignment engine algorithm module, a feature decoupling model, and a dynamic decision algorithm module, which are used to perform spatio-temporal alignment and feature extraction on the data after filtering by the multi-modal sensing unit, run the environmental coupling compensation algorithm, and finally generate a preliminary diagnosis result. The edge computing unit also has a multi-source data fusion gateway (equipped with an NVIDIA Jetson Orin module) and a data processing module.

[0120] Furthermore, in this embodiment:

[0121] The spatio-temporal alignment engine algorithm module realizes the spatio-temporal alignment of multi-modal data by using the fusion algorithm of dynamic time warping (DTW) and Kalman filtering. The specific steps include:

[0122] Mark the time stamps of the data streams of each sensor, with an error not greater than 1 ms;

[0123] Calculate the phase difference Δτ of the data streams of different sensors:

[0124]

[0125] Use a Kalman filter to compensate for the spatial differences caused by the physical distances of different sensors;

[0126] The feature decoupling model is constructed based on a physics-informed neural network (PINN), and the loss function satisfies:

[0127]

[0128] The first part , is the weight coefficient, and represent the predicted environmental features and the features of the actual environment respectively;

[0129] The second part , where represents the derivative (rate of change) of the environmental feature with respect to time, is a predefined environmental coupling equation, which is associated with temperature T, humidity H, rainfall R, electric field strength E, contamination degree F, partial discharge amount PD, and redundant environmental parameter N, is another weight coefficient used to measure the difference between the feature rate of change and the f function;

[0130] The dynamic decision-making algorithm module uses a memory-augmented deep Q-network (ME-DQN). The network structure includes: an input layer; a 128-dimensional feature vector (including 16-dimensional environmental parameters and 32-dimensional electrical features); an LSTM memory unit: the hidden layer dimension is 64; a double Q-network: respectively output the threshold adjustment strategy and the alarm level.

[0131] Furthermore, in this embodiment:

[0132] The training method of the memory-augmented deep Q-network includes the following:

[0133] Construct a training set containing 20,000 groups of samples, covering: 6 typical fault modes, such as moisture absorption, valve disc aging, external insulation damage, etc.; 12 types of environmental condition combinations, such as heavy rain, high humidity, salt fog, etc.;

[0134] Design a multi-objective reward function:

[0135]

[0136] Weight coefficient The value is 0.5, The value is 0.3, The value is 0.2;

[0137] Adopt the prioritized experience replay mechanism, and the priority calculation is:

[0138]

[0139] where δi is the TD error and the ε value is 0.01;

[0140] The network is updated hard every 1000 steps.

[0141] Furthermore, in this embodiment:

[0142] The spatio-temporal alignment engine algorithm module has an abnormal data cleaning mechanism. For temperature data, median filtering in a sliding window is used, and the window width W satisfies:

[0143]

[0144] Perform outlier detection based on Mahalanobis distance on the electric field strength data:

[0145]

[0146] where, is the statistic, is the electric field strength data, is the overall mean, is the weighted deviation of the electric field strength in different directions;

[0147] Establish a credibility evaluation matrix for the data:

[0148] +

[0149] where, is the credibility score, is the signal-to-noise ratio weight coefficient, = 0.6, is the consistency weight coefficient = 0.4, the signal-to-noise ratio weight accounts for 60%, is the consistency value.

[0150] Specifically, in this embodiment, a digital twin model, a device degradation prediction algorithm module, and a dynamic threshold generator are deployed inside the cloud analysis platform. Based on the preliminary diagnosis results generated by the edge computing unit, the parameters of the digital twin model are updated, device degradation prediction is performed, and dynamic threshold parameters are sent down.

[0151] Furthermore, in this embodiment:

[0152] The digital twin model includes:

[0153] A three-dimensional electromagnetic field simulation module that calculates the surface potential distribution of the lightning arrester based on the finite element method;

[0154] A thermodynamics coupling module that simulates the temperature rise of the zinc oxide varistor under different ambient temperatures;

[0155] The device degradation prediction algorithm module adopts an improved Transformer architecture:

[0156] Input sequence: {fundamental amplitude of leakage current, third harmonic content, partial discharge times, environmental characteristic parameters};

[0157] The attention mechanism calculates weights:

[0158]

[0159] Among them, is the input feature representation, and the output prediction results are: remaining useful life estimate (RUL) and health index (HI);

[0160] The dynamic threshold generator generates an adaptive alarm threshold according to the historical state of the device:

[0161]

[0162] Among them, k is the degradation rate coefficient, which is updated every 6 hours.

[0163] The implementation parameters of the system on the 220 kV lightning arrester are:

[0164] Sensor layout: The leakage current sensor is ≤0.5 m from the grounding end; the electric field induction sensor forms an angle of 45° ± 5° with the central axis of the lightning arrester;

[0165] Sampling frequency: Normal mode: current signal 1 kHz, environmental parameters 1 Hz; Lightning activity period: current signal 10 kHz, environmental parameters 10 Hz;

[0166] Alarm delay: General alarm ≤3 s; Emergency alarm ≤500 ms; Local storage capacity: Can store 30 days of original data (sampled at 1 Hz); or 180 days of compressed feature data.

[0167] Such asFigure 2 As shown, according to another embodiment of the present invention, a method for monitoring the leakage current of a lightning arrester based on multi-modal data fusion is provided, which includes the following steps:

[0168] Modal data acquisition stage: Synchronously obtain leakage current, partial discharge, temperature and humidity, rainfall, and electric field intensity data, and perform pre-filtering processing on the acquired data;

[0169] Edge-side data processing stage: Perform spatio-temporal alignment and feature extraction on the filtered data, run the environmental coupling compensation algorithm, and finally generate a preliminary diagnosis result;

[0170] Cloud deep analysis stage: Based on the preliminary diagnosis result, update the parameters of the digital twin model, perform equipment degradation prediction, and issue dynamic threshold parameters;

[0171] Closed-loop control stage: Adjust the local monitoring strategy according to the latest obtained threshold parameters, and trigger equipment maintenance suggestions or protection actions.

[0172] Furthermore, in this embodiment:

[0173] The specific steps of the environmental coupling compensation algorithm are as follows:

[0174] Establish a rainfall influence correction model:

[0175]

[0176] Among them, is the original rainfall, The value is 0.15, R is the rainfall intensity (mm / h), H is the relative humidity (%), is an exponential function;

[0177] Temperature drift compensation:

[0178]

[0179] Among them, the α value is 0.12 mA / °C, the β value is 0.002 mA / °C², The value is 25 °C;

[0180] Calculation of the contamination degree influence factor:

[0181]

[0182] Among them, is the particle diameter, is the concentration, .

[0183] Furthermore, in this embodiment:

[0184] The specific steps for equipment degradation prediction are as follows:

[0185] Feature extraction:

[0186] Time-domain features: mean current, variance, kurtosis;

[0187] Frequency-domain features: proportion of the third harmonic, resonance frequency offset;

[0188] Degradation state classification:

[0189] Level 1 (healthy): HI ≥ 0.9;

[0190] Level 2 (warning): 0.7 ≤ HI < 0.9;

[0191] Level 3 (fault): HI < 0.7;

[0192] Remaining life prediction:

[0193]

[0194] Among them, the value is 0.6, and is the change rate of the health index.

[0195] Furthermore, in this embodiment:

[0196] The specific steps for adjusting the local monitoring strategy according to the latest obtained threshold parameters are as follows:

[0197] Fundamental current threshold:

[0198] +0.03 HI

[0199] Among them, is the average current, is the standard deviation, and HI is the health index.

[0200] HI is adjusted according to the harmonic content:

[0201]

[0202] Among them, is the third harmonic content.

[0203]

[0204] Among them, is the current change rate, is the reference current.

[0205] Above, refer to Figures 1 - 3A lightning arrester leakage current monitoring system and method based on multi-modal data fusion according to an embodiment of the present invention are described, having the following beneficial effects:

[0206] 1. By constructing a deep coupling model of multi-dimensional environmental parameters and electrical characteristics and combining the current powerful reinforcement learning algorithm with a memory enhancement mechanism, the monitoring reliability under complex working conditions is significantly improved.

[0207] 2. The system has functions of dynamic threshold adjustment, environmental interference compensation and equipment degradation prediction, which can effectively reduce the false alarm rate and extend the service life of the equipment, providing a new solution idea for anti-interference of the online monitoring system.

[0208] 3. The monitoring accuracy harmonic separation is increased to 42 dB (traditional method ≤ 35 dB); the recognition rate of moisture damage faults is 98.7% (the highest in the comparative literature is 93.2%).

[0209] 4. Adaptive ability: the threshold adjustment response time < 3 s; the false alarm rate is reduced by 76% under rainstorm working conditions.

[0210] 5. Energy efficiency ratio: adopting event-driven data transmission, the traffic is saved by 82%; the standby power consumption of the edge node is 0.8 mW (similar products ≥ 5 mW).

[0211] Specific application examples:

[0212] Application example 1: Application in a 500 kV substation

[0213] Hardware deployment:

[0214] The spacing of the sensing units ≤ 1.5 m, sampling frequency:

[0215] Leakage current: 10 kHz (during lightning activity period) / 1 kHz (normal)

[0216] Environmental parameters: 1 Hz

[0217] Software configuration:

[0218] Edge layer: Run a lightweight ME-DQN model (parameter quantity 2.3M);

[0219] Cloud: The digital twin engine updates the equipment degradation model every 6 hours.

[0220] Test results (continuous operation for 180 days):

[0221] Index The present invention Traditional method Fault undetected reporting rate 0.9% 6.7% Average daily false alarm times 0.2 1.8

[0222] Application example 2: Comparative test in a high-humidity coastal environment

[0223] Under the condition that the relative humidity > 90%, the detection accuracy rate of the system of the present invention still remains at 92.4%, while that of the comparative device drops to 67.3%.

[0224] It should be noted that in this specification, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising..." does not exclude the presence of additional identical elements in the process, method, article or device comprising said element.

[0225] Although the content of the present invention has been introduced in detail through the above preferred embodiments, it should be recognized that the above description should not be considered as a limitation of the present invention. After those skilled in the art have read the above content, various modifications and alternatives to the present invention will be obvious. Therefore, the protection scope of the present invention should be defined by the appended claims.

Claims

1. A lightning arrester leakage current monitoring system based on multi-modal data fusion, characterized in that, It includes: a multi-modal perception unit, an edge computing unit, and a cloud analysis platform; The multi-modal perception unit realizes data interaction with the edge computing unit through Lora and 5G dual-mode communication; The edge computing unit realizes data interaction with the cloud analysis platform through HTTP and MQTT communication; The multi-modal perception unit is composed of a leakage current detection module, a partial discharge detection module, an environmental parameter acquisition module, and an electric field detection module, which is used to obtain leakage current, partial discharge, temperature and humidity, rainfall, and electric field intensity data, and perform pre-filtering processing on the acquired data; The edge computing unit is built with a spatio-temporal alignment engine algorithm module, a feature decoupling model, and a dynamic decision-making algorithm module, which are used to perform spatio-temporal alignment and feature extraction on the data processed by the multi-modal perception unit, run an environmental coupling compensation algorithm, and finally generate a preliminary diagnosis result; The cloud analysis platform deploys a digital twin model, a device degradation prediction algorithm module, and a dynamic threshold generator. Based on the preliminary diagnosis result generated by the edge computing unit, it updates the parameters of the digital twin model, performs device degradation prediction, and issues dynamic threshold parameters.

2. The lightning arrester leakage current monitoring system based on multi-modal data fusion according to claim 1, characterized in that The leakage current detection module is composed of a TMR sensor array and is installed at the grounding lead of the lightning arrester; The partial discharge detection module adopts a combined structure of a UHF sensor and a high-frequency current transformer; The environmental parameter acquisition module includes a temperature and humidity sensor, a piezoelectric rain gauge, and a laser scattering type pollution degree sensor; The electric field detection module includes a three-axis capacitive electric field induction sensor.

3. The lightning arrester leakage current monitoring system based on multi-modal data fusion according to claim 1, characterized in that The spatio-temporal alignment engine algorithm module adopts a dynamic time warping and Kalman filter fusion algorithm to realize spatio-temporal alignment of multi-modal data. The specific steps include: Mark the time stamps of each sensor data stream; Calculate the phase difference Δτ of different sensor data streams: ， Use a Kalman filter to compensate for the spatial differences caused by the physical distances of different sensors; The feature decoupling model is constructed based on a neural network with physical constraints, and the loss function satisfies: ， The first part , is the weight coefficient, and represent the predicted environmental characteristics and the characteristics of the actual environment respectively; The second part , where represents the derivative (rate of change) of the environmental characteristic with respect to time, is a predefined environmental coupling equation, associated with temperature T, humidity H, rainfall R, electric field strength E, contamination degree F, partial discharge amount PD, and redundant environmental parameter N, is another weight coefficient, used to measure the difference between the rate of change of the characteristic and the f function; The dynamic decision-making algorithm module adopts a memory-augmented deep Q network. The network structure includes: an input layer; a 128-dimensional feature vector; an LSTM memory unit: the hidden layer dimension is 64; a double Q network: which respectively outputs a threshold adjustment strategy and an alarm level.

4. The lightning arrester leakage current monitoring system based on multi-modal data fusion according to claim 3, characterized in that, The training method of the memory-augmented deep Q network includes the following: Construct a training set containing 20,000 groups of samples, covering: 6 typical fault modes, such as moisture absorption, valve disc aging, and external insulation damage; 12 types of environmental condition combinations, such as heavy rain, high humidity, and salt fog; Design a multi-objective reward function: ， Weight coefficient The value is 0.5, The value is 0.3, The value is 0.2; Adopt a prioritized experience replay mechanism, and the priority calculation is: ， Where δi is the TD error and the ε value is 0.01; The network is updated every 1000 steps with a hard update.

5. The lightning arrester leakage current monitoring system based on multi-modal data fusion according to claim 3, characterized in that, The spatio-temporal alignment engine algorithm module has an abnormal data cleaning mechanism, and uses moving window median filtering for temperature data. The window width W satisfies: ， Perform outlier detection based on Mahalanobis distance on the electric field strength data: ， Among them, is a statistic,[[]]END]] is the electric field strength data,[[]]END]] is the overall mean,[[]]END]] is the weighted deviation of the electric field strength in different directions.[[]]END]] Establish a credibility evaluation matrix for the data: + ， Among them, is the credibility score, is the signal-to-noise ratio weight coefficient, = 0.6, is the consistency weight coefficient = 0.4, the signal-to-noise ratio weight accounts for 60%, is the consistency value.

6. The lightning arrester leakage current monitoring system based on multi-modal data fusion according to claim 1, characterized in that The digital twin model includes: A three-dimensional electromagnetic field simulation module that calculates the surface potential distribution of the arrester based on the finite element method; A thermodynamics coupling module that simulates the temperature rise of zinc oxide varistors at different ambient temperatures; The device degradation prediction algorithm module adopts an improved Transformer architecture: Input sequence: {fundamental amplitude of leakage current, third harmonic content, number of partial discharges, environmental characteristic parameters}; The attention mechanism calculates weights: ， Among them, is the input feature representation, and the output prediction results are: the remaining useful life estimate (RUL) and the health index (HI); The dynamic threshold generator generates an adaptive alarm threshold according to the historical state of the device: ， Among them, k is the degradation rate coefficient, which is updated every 6 hours.

7. A method for monitoring the leakage current of a lightning arrester based on multi-modal data fusion, characterized in that, It includes the following steps: Modal data acquisition stage: Synchronously obtain leakage current, partial discharge, temperature and humidity, rainfall, and electric field strength data, and perform pre-filtering processing on the obtained data; Edge-side data processing stage: Perform spatio-temporal alignment and feature extraction on the filtered data, run the environmental coupling compensation algorithm, and finally generate a preliminary diagnosis result; Cloud deep analysis stage: Based on the preliminary diagnosis result, update the parameters of the digital twin model, perform device degradation prediction, and issue dynamic threshold parameters; Closed-loop control stage: Adjust the local monitoring strategy according to the latest obtained threshold parameters, and trigger device maintenance suggestions or protection actions.

8. The method for monitoring the leakage current of a lightning arrester based on multi-modal data fusion according to claim 7, wherein, The specific steps of the environmental coupling compensation algorithm are as follows: Establish a rainfall impact correction model: ， Among them, is the original rainfall,[[]]END The value is 0.15, R is the rainfall intensity (mm / h), and H is the relative humidity (%). is an exponential function; Temperature drift compensation: ， Among them, the value of α is 0.12 mA / °C, and the value of β is 0.002 mA / °C², The value is 25 °C; Calculation of the influence factor of pollution degree: ， Among them, is the particle diameter, is the concentration, .

9. The method for monitoring the leakage current of a lightning arrester based on multimodal data fusion according to claim 7, characterized in that, The specific steps of the device degradation prediction are as follows: Feature extraction: Time-domain features: current mean, variance, kurtosis; Frequency-domain features: third harmonic ratio, resonance frequency offset; Degradation state classification: Level 1 (healthy): HI≥0.9; Level 2 (warning): 0.7≤HI<0.9; Level 3 (fault): HI<0.7; Remaining life prediction: ， Among them, The value is 0.6, which is the change rate of the health index.

10. The method for monitoring the leakage current of a lightning arrester based on multi-modal data fusion according to claim 7, wherein The specific steps of adjusting the local monitoring strategy according to the latest obtained threshold parameters are as follows: Fundamental current threshold: +0.03 HI, Among them, is the average current, is the standard deviation, and HI is the health index; HI adjusts the threshold according to the harmonic content: ， Among them, Third harmonic content; ， Among them, is the current change rate, is the reference current.

Images