A cloud-edge collaborative integrated approach to leakage current protection and remote operation and maintenance of lighting systems.

Abstract

This invention discloses an integrated method for leakage current protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration, belonging to the field of cloud-edge collaboration technology. Through a cloud-edge collaboration architecture, edge computing nodes locally collect and preprocess data in real time, achieving real-time leakage current risk quantification and dynamic feature extraction through dual-path parallel analysis, balancing real-time protection with the integrity of diagnostic data. Multi-level collaborative diagnosis in the cloud combines historical case pattern matching and deep temporal network model analysis, generating accurate comprehensive diagnostic decision factors through nonlinear fusion. Furthermore, risk mapping incorporates three-dimensional data, accurately determining risk levels and fault types. The collaborative response strategy includes adaptive threshold adjustment, operation and maintenance decision suggestions, and model update instructions. After executing protection, edge nodes also evaluate the effectiveness of the actions, improving the timeliness and accuracy of leakage current protection, reducing safety hazards, optimizing remote operation and maintenance efficiency, and enhancing the long-term adaptability and performance stability of the system.

Description

Technical Field

[0001] This invention relates to the field of cloud-edge collaboration technology, specifically to an integrated method for leakage current protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration. Background Technology

[0002] During long-term operation, lighting systems are prone to leakage current due to factors such as aging insulation, loose wiring, and humid environments (e.g., underground garages, restrooms). If leakage protection is not timely, it can lead to safety accidents such as electric shock, wiring burnout, or even fire. However, existing leakage protection technologies have significant shortcomings:

[0003] Insufficient real-time and accuracy of protection: Traditional leakage protection relies on leakage current protectors with fixed thresholds (such as a preset leakage current threshold of 10mA for tripping), without considering the impact of load fluctuations in lighting circuits (such as a temporary increase in leakage current when LED lights are turned on) on leakage current. This can easily trigger invalid alarms or fail to detect real leakage risks due to excessively high thresholds. Moreover, most of these are passive protections that "trip locally," lacking real-time quantitative analysis of leakage current change trends and failing to provide early warnings of potential hazards.

[0004] Disconnection between protection and operation and maintenance: In existing technologies, leakage protection (such as circuit breaker tripping) and remote operation and maintenance are separate links. The protection action can only cut off the fault circuit and cannot synchronize the fault data (such as leakage current waveform and load status) to the operation and maintenance platform. Operation and maintenance personnel need to conduct on-site inspections to determine the fault type and location, resulting in delayed operation and maintenance response and low efficiency.

[0005] Lack of long-term health management capabilities: Focusing only on handling instantaneous leakage faults without monitoring and evaluating the long-term health status of lighting circuits (such as the slow increase in leakage current due to insulation aging), it is difficult to intervene in chronic faults in advance, increasing the risk of sudden system failures.

[0006] To address the pain points of leakage current protection and maintenance in lighting systems, existing technologies cannot achieve integrated management and control encompassing "real-time risk quantification, precise protection, remote collaborative maintenance, and continuous optimization." On one hand, local nodes have limited computing power, making it difficult to extract fine-grained features of leakage current and perform complex risk analysis. On the other hand, while cloud platforms possess strong computing power, they lack real-time collaboration with local nodes, hindering rapid response to instantaneous high-risk faults, and diagnostic models and maintenance rules are difficult to dynamically optimize based on actual operating conditions. Therefore, a cloud-edge collaborative technology solution is urgently needed to combine the "real-time data acquisition and rapid protection" capabilities of edge nodes with the "big data analysis, precise diagnosis, and global maintenance optimization" capabilities of cloud platforms. This would achieve integrated leakage current protection and remote maintenance for lighting systems, resolving the issues of poor real-time performance, low accuracy, and insufficient collaboration inherent in existing technologies. Summary of the Invention

[0007] The technical problem to be solved by this invention is: how to solve the problems of poor real-time performance, low accuracy and insufficient collaboration in the existing technology, and to provide an integrated method for leakage protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration.

[0008] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0009] S1: By deploying edge computing nodes in the lighting distribution box, the voltage, current and leakage current waveform data of each lighting circuit are collected synchronously, the waveform data are preprocessed, the load power data is calculated, and the leakage current time sequence data is extracted and standardized.

[0010] S2: Edge computing nodes use local analysis models to perform dual-path parallel analysis. The first path calculates the real-time leakage risk quantification index of the lighting circuit based on leakage current time series data and load power data, and records the real-time leakage risk quantification index and its timestamp as local risk time series data. The second path performs online feature extraction on the current waveform to generate dynamic feature vectors.

[0011] S3: The edge computing node compares the real-time leakage risk quantification index with the local early warning threshold. When the threshold is exceeded, an alarm is triggered and the dynamic feature vector, the current local risk time series data and the related time period are uploaded to the cloud operation and maintenance platform.

[0012] S4: After receiving the data, the cloud-based operations and maintenance platform initiates multi-level collaborative diagnosis. The diagnosis process includes: first, performing pattern matching based on historical cases on the received dynamic feature vector to generate a first reference coefficient; second, analyzing the received complete data sequence using a cloud-based evaluation model to generate a second reference coefficient; and finally, calculating a comprehensive diagnostic decision factor based on the confidence levels of the first and second reference coefficients using a nonlinear fusion function. The complete data sequence includes the dynamic feature vector and local risk time-series data uploaded by edge computing nodes.

[0013] S5: The cloud-based operations and maintenance platform performs risk mapping based on comprehensive diagnostic decision factors and preliminary fault type identification results. Based on the final risk level determined by the risk mapping and the preliminary fault type identification results, it generates a collaborative response strategy and distributes the strategy to edge computing nodes and mobile operations and maintenance terminals. The collaborative response strategy includes adaptive threshold adjustment parameters for adjusting edge detection sensitivity, operations and maintenance decision suggestions corresponding to the final risk level and fault type, and model update instructions for optimizing the performance of the local analysis model.

[0014] S6: The edge computing node receives and parses the collaborative response strategy, adjusts the parameters to update the local early warning threshold according to the adaptive threshold, executes the model update instruction to optimize the local analysis model, and performs corresponding protection actions; at the same time, the mobile operation and maintenance terminal generates operation and maintenance work orders according to the operation and maintenance decision suggestions, guides the operation and maintenance personnel to handle the situation on-site, and feeds back the handling results to the cloud operation and maintenance platform.

[0015] S7: The cloud-based operations and maintenance platform optimizes multi-level collaborative diagnosis and risk mapping rules based on feedback processing results data using an incremental learning mechanism.

[0016] Furthermore, in step S2, a real-time leakage current risk quantification index is calculated. The detailed process is as follows:

[0017] S211: A sliding window-based differential accumulation algorithm is used to process the standardized leakage current timing data. Process the data and calculate the cumulative change within each window. ;

[0018] S212: Combine the load power data of the corresponding lighting circuit at the same time. Calculate the sensitivity of the correlation between load and leakage current. ;

[0019] S213: Real-time Leakage Risk Quantification Indicator Calculated using the following formula:

[0020] ;

[0021] Among them, cumulative change The calculation method is as follows: sum the absolute differences in leakage current between adjacent sampling points within the sliding window; correlation sensitivity The calculation method is as follows: by calculating the load power data With standardized leakage current timing data The absolute value of the correlation coefficient within the sliding window is obtained by linear scaling; t represents time t.

[0022] Furthermore, in step S2, the specific process of online feature extraction for the second path is as follows:

[0023] S221: After resampling and aligning the current waveform for each power frequency cycle, perform synchronous compressed wavelet transform to obtain the time spectrum; extract three dimensions of features from the time spectrum: calculate the ratio of the energy of the set high-frequency subband to the fundamental energy as the high-frequency transient energy coefficient; extract the derivative statistical features of the zero-sequence current waveform near the zero crossing point, the zero-sequence current waveform is the leakage current waveform; calculate the percentage offset of the harmonic distortion rate of the current waveform relative to the historical baseline value;

[0024] S222: Concatenate the features of the above three dimensions to form a dynamic feature vector.

[0025] Furthermore, in step S4, the first reference coefficient The generation process is as follows: The cloud-based operations and maintenance platform stores a historical case feature library. Each historical case contains a historical feature vector and a final fault label, where the historical feature vector and the dynamic feature vector have the same feature dimension; the received dynamic feature vector... The k-nearest neighbor algorithm is used to retrieve the most similar feature from the historical case feature database. Calculate the dynamic feature vector from historical cases, i.e., neighbor cases. Historical feature vectors for each historical case cosine similarity and with The reciprocal of the distance is used as the weight to obtain the similarity weight. ,in Sort by distance; calculate Cosine similarity The weighted average of the neighboring cases is used as the average similarity. First reference coefficient Calculated using the following formula:

[0026] ;

[0027] in, Fault labels for neighboring cases. For indicator functions, when the fault label belongs to a predefined set of emergency fault types. The value is 1 when the time is right, and 0 otherwise. for Information entropy of fault labels in a neighboring case This represents the total number of fault label types.

[0028] Furthermore, in step S4, the cloud-based evaluation model is a model based on a deep temporal network; the second reference coefficient... The generation process is as follows: The cloud-based assessment model takes the received data sequence containing dynamic feature vectors and historical risk quantification indicators as input, extracts time-dependent features through a long short-term memory network layer, and then assigns weights to features at different time steps through an attention layer; the output of the cloud-based assessment model is mapped to a real-time risk probability through a fully connected layer. Second reference coefficient By analyzing real-time risk probabilities and the confidence level of the model in predicting the results. The specific formula is obtained by combining the results:

[0029] ;

[0030] Among them, confidence estimation It is calculated from the concentration of the attention weight distribution.

[0031] Furthermore, in step S4, the nonlinear fusion function generates a comprehensive diagnostic decision factor. The specific process is as follows: First, based on the average similarity of neighboring cases... Calculate the confidence factor of the first reference coefficient Based on the second reference coefficient Confidence level in the calculation process Calculate its credibility factor Then, a confidence-weighted sigmoid fusion function is used to calculate... :

[0032] ;

[0033] in, For the Sigmoid function, For positive numbers used to prevent division by zero.

[0034] Furthermore, after step S4, the method further includes: calculating the slope of the health decay trend of the corresponding lighting circuit within a preset historical period. Slope of the health decline trend It is the slope of the trend line obtained by linear fitting based on the historical risk quantification index sequence; the risk mapping process in step S5 is as follows: using comprehensive diagnostic decision factors The first dimension is the risk level of the initially identified fault types, and the second dimension is the slope of the health decay trend. The third dimension is used to normalize the values ​​of the three dimensions and input them into the risk space partitioning model. The risk space partitioning model maps the input vector to a risk space composed of subspaces with different risk levels, thereby determining the final risk level.

[0035] Furthermore, in step S5, the calculation process for the adaptive threshold adjustment parameter is as follows: based on the final risk level and comprehensive diagnostic decision factors... Determine the required adjustment of the warning threshold offset for edge computing nodes. The specific calculation formula is as follows:

[0036] ;

[0037] in, The basic warning threshold for the corresponding lighting circuit, The steepness coefficient is related to the final risk level.

[0038] Furthermore, in step S6, after the edge computing node performs the protection action, it starts a high-frequency monitoring window of a set duration to capture the transient response waveforms of voltage, current and leakage current after the action. Each transient response waveform is compared with the standard response waveform generated based on the theoretical model or historical normal action data, the action effectiveness evaluation score is calculated, and the score and the response waveform summary are fed back to the cloud operation and maintenance platform.

[0039] Furthermore, in step S7, the specific process of the incremental learning mechanism is as follows: the cloud operation and maintenance platform records the complete process from uploading data from the edge computing node to generating a collaborative response strategy, executing edge protection, and finally receiving feedback on the handling results, as a closed-loop event sample, and trains and adjusts the cloud evaluation model online; at the same time, it optimizes the risk space partitioning model or the final risk level determination rule by utilizing the correspondence between the risk mapping results recorded in the closed-loop event sample and the handling effect; and based on the accuracy of historical case matching and the success rate of handling results, it maintains and dynamically updates a long-term utility weight for each historical case in the feature historical case library, which is used to represent the priority of the corresponding historical case being retrieved in subsequent pattern matching.

[0040] Compared with existing technologies, this invention has the following advantages: This cloud-edge collaborative integrated method for leakage current protection and remote operation and maintenance of lighting systems utilizes a cloud-edge collaborative architecture. Edge computing nodes collect and preprocess data locally in real time, and dual-path parallel analysis achieves real-time leakage current risk quantification and dynamic feature extraction, balancing the real-time nature of protection with the integrity of diagnostic data. Cloud-based multi-level collaborative diagnosis combines historical case pattern matching and deep temporal network model analysis, generating accurate comprehensive diagnostic decision factors through nonlinear fusion. Risk mapping incorporates three-dimensional data, accurately determining risk levels and fault types. The collaborative response strategy includes adaptive threshold adjustment, operation and maintenance decision suggestions, and model update instructions. After the edge node executes protection, it also evaluates the effectiveness of the action. The cloud continuously optimizes diagnostic and risk mapping rules through incremental learning, while mobile operation and maintenance terminals form a closed-loop handling process. This not only improves the timeliness and accuracy of leakage current protection and reduces safety hazards, but also optimizes remote operation and maintenance efficiency and enhances the long-term adaptability and performance stability of the system. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the integrated method for leakage protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration, as proposed in this invention. Detailed Implementation

[0042] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0043] like Figure 1 As shown, this embodiment provides a technical solution: an integrated method for leakage current protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration, including the following steps:

[0044] Step 1: Data Acquisition and Preprocessing;

[0045] By deploying edge computing nodes in the lighting distribution box, the voltage, current and leakage current waveform data of each lighting circuit are collected synchronously. The waveform data is preprocessed to calculate the load power data, and the leakage current time sequence data is extracted and standardized.

[0046] In this embodiment, the leakage current waveform data is the zero-sequence current waveform data used for leakage current monitoring.

[0047] In this embodiment, preprocessing mainly includes two types of operations: signal cleaning and data conversion.

[0048] Signal cleaning: The raw waveform data of voltage, current and leakage current directly acquired are filtered and denoised to eliminate high-frequency interference and measurement noise, providing a stable signal source for subsequent analysis.

[0049] Data transformation: involves two parallel computation processes:

[0050] Calculate load power data: Using the cleaned voltage and current waveforms, power data characterizing the circuit load status is obtained through instantaneous power calculation and periodic averaging.

[0051] Standardized leakage current time series data: Using the cleaned leakage current waveform data as input, amplitude calibration (such as converting to milliampere value) and baseline calibration (eliminating inherent offset) are performed to generate a standardized leakage current data sequence that can be used for threshold comparison and trend analysis.

[0052] Step 2: Generate dynamic feature vectors;

[0053] Edge computing nodes utilize local analysis models to perform dual-path parallel analysis. The first path calculates the real-time leakage risk quantification index of the lighting circuit based on leakage current time-series data and load power data, and records the real-time leakage risk quantification index and its timestamp as local risk time-series data. The second path performs online feature extraction on the current waveform to generate dynamic feature vectors for fine-grained diagnosis in the cloud.

[0054] In this embodiment, in step two, the first path calculates the real-time leakage current risk quantification index. The detailed process is as follows: First, a differential accumulation algorithm based on a sliding window is used to process the standardized leakage current time series data. Process the data and calculate the cumulative change within each window. Then, combine the load power data of the corresponding lighting circuit at the same time. Calculate the sensitivity of the correlation between load and leakage current. Ultimately, the real-time leakage risk quantification index Calculated using the following formula:

[0055] ;

[0056] Among them, cumulative change The calculation method is as follows: take the sum of the absolute differences in leakage current between adjacent sampling points within the sliding window, that is... ; Association sensitivity The calculation method is as follows: by calculating the load power data With standardized leakage current timing data The absolute value of the correlation coefficient within the sliding window is then obtained by linear scaling.

[0057] Specifically, taking the lighting circuit of the corridor on the third floor of an office building as an example:

[0058] The loop contains 10 8W LED lights (rated total power 80W), the edge computing node sampling interval is 1 second, and the sliding window width is... , The warning threshold is 0.6, and the single leakage current threshold is 1.2mA.

[0059] Operating condition 1: Load fluctuation (no leakage current);

[0060] Calculate the standardized leakage current as an LED light gradually starts up within a 15-second window. (Unit: mA) in order According to the formula Summing the absolute values ​​of adjacent differences yields It accurately captures the gradual change in leakage current as the load starts.

[0061] Calculate: Simultaneous load power (Unit: W) in order , and Correlation coefficient Linear scaling have to This indicates that the leakage current is highly correlated with the load and is a normal fluctuation.

[0062] Calculation: Substituting into the formula yields If the value is below the threshold of 0.6, no alarm will be triggered.

[0063] Operating Condition 2: Stable load (actual leakage current);

[0064] Calculate the standardized leakage current within a 15-second window when all LEDs are lit (stable power 80W). (Unit: mA) in order Calculate the absolute values ​​of adjacent differences using the formula and sum them to obtain... This clearly reflects the continuous upward trend of leakage current.

[0065] Calculation: Power is constant at 80W. and Correlation coefficient After linear scaling, This indicates that the leakage current is unrelated to the load and represents a genuine leakage risk.

[0066] Calculation: Substituting into the formula yields If the current exceeds the threshold of 0.6, the edge node will trigger an alarm to accurately identify the actual leakage current.

[0067] It should be noted that the Pearson correlation coefficient The Pearson Correlation Coefficient is a statistical indicator that measures the degree of linear correlation between two continuous variables. It can accurately quantify the strength of the linear relationship where both variables increase or decrease simultaneously, or where one increases while the other decreases. Its value ranges from [value missing]. When the coefficient is close to When the time is right, it indicates a strong positive linear correlation between the two; when it is close to the time it is right, it indicates a strong positive linear correlation between the two. It shows a strong negative linear correlation; close to At times, the linear correlation is extremely weak or non-linear, a characteristic that perfectly suits the need to "distinguish between normal changes in leakage current caused by load fluctuations (strong correlation) and actual leakage current (weak correlation)".

[0068] The specific calculation method for the Pearson correlation coefficient is as follows:

[0069] Suppose there is a sliding window One sampling point (in this embodiment) Corresponding window width ,variable For load power sequence ,variable For standardized leakage current sequence Then the Pearson correlation coefficient The calculation formula is:

[0070] ;

[0071] in:

[0072] The sum of the products of "power × leakage current" at all sampling points;

[0073] , : These are the sums of the power sequence and the leakage current sequence, respectively;

[0074] , : These are the sum of squares of the power sequence and the leakage current sequence, respectively;

[0075] Number of sampling points within the sliding window (in this embodiment) ).

[0076] In this invention, the cumulative change in leakage current is calculated using a sliding window differential accumulation algorithm. It can effectively capture the gradual or abrupt trend of leakage current over a period of time, avoiding risk misjudgment caused by the randomness of data from a single sampling point; and calculate the correlation sensitivity by combining load power data. It can distinguish between normal changes in leakage current caused by load power fluctuations and actual leakage risks caused by line insulation damage, reducing invalid alarms; ultimately, through formula... The leakage current variation trend is integrated with the load correlation to form a real-time leakage risk quantification index. This indicator can accurately quantify the degree of leakage risk under different operating conditions, providing an objective and reliable basis for subsequent threshold comparison and alarm triggering of edge computing nodes, and significantly improving the accuracy of leakage risk identification and anti-interference capability.

[0077] In this embodiment, the specific process of online feature extraction for the second path in step two includes: resampling and aligning the current waveform for each power frequency cycle, performing synchronous compressed wavelet transform to obtain a time-spectrum image; extracting features in three dimensions from the time-spectrum image: first, calculating the ratio of the energy of the set high-frequency subband to the fundamental energy as the high-frequency transient energy coefficient; second, extracting the derivative statistical features of the zero-sequence current waveform near the zero-crossing point; third, calculating the percentage offset of the harmonic distortion rate of the current waveform relative to the historical baseline value; and concatenating the above three-dimensional features to form a dynamic feature vector.

[0078] Specifically, the lighting circuit scenario for the corridor on the third floor of the office building is continued.

[0079] This circuit is a three-phase four-wire lighting circuit (containing 10 8W LED lights, rated total power 80W, power frequency 50Hz, single power frequency cycle 0.02s), and the edge computing nodes adopt a "dual sampling mode":

[0080] Low-frequency sampling (1Hz, sampling interval 1 second): used for acquiring leakage current timing data and load power data, adapted for risk trend analysis;

[0081] High-frequency sampling (originally 5kHz, resampled to 10kHz): only used for fine-grained feature extraction of current waveforms (started within a single fault analysis period, non-continuous high-frequency sampling, balancing computing power and feature accuracy), with 200 sampling points per power frequency cycle.

[0082] Historical baseline harmonic distortion rate (Typical values ​​for a normal lighting circuit), the current simulation is a "partial insulation breakdown" leakage fault (this fault will generate high-frequency transient signals, zero-sequence current fluctuations, and aggravated harmonic distortion). The specific feature extraction process is as follows:

[0083] Step 1: Current waveform resampling and alignment;

[0084] Resampling: Under the original 5kHz high-frequency sampling, there are only 100 sampling points per power frequency cycle, and some transient details are easily lost; by resampling to 10kHz (sampling interval 0.0001s) through linear interpolation, 200 sampling points are obtained per power frequency cycle, ensuring that the current change signal at the fault moment (such as t=125 seconds) is accurately captured.

[0085] Alignment: Using the zero-crossing point of the grid voltage as a reference, the starting point of the current waveform of each power frequency cycle is aligned to the zero-crossing point of the voltage (e.g., t=125.00s, 125.02s, 125.04s...) to eliminate the phase shift of waveforms of different cycles and ensure that subsequent time-frequency analysis is performed under a unified time reference.

[0086] Step 2: Synchronous compressed wavelet transform and time-spectrum graph acquisition;

[0087] Using Morlet wavelets as basis functions (suitable for transient signal analysis), a synchronous compressed wavelet transform is performed on the aligned current waveform:

[0088] The transformed image is a time-frequency spectrum (horizontal axis: time, unit s; vertical axis: frequency, unit Hz; color intensity: energy intensity).

[0089] Fault characteristics: During the insulation breakdown period of t=125.03s~125.05s (corresponding to one complete power frequency cycle), a significant energy concentration area appears in the 1~5kHz high frequency band of the time spectrum (the energy in this frequency band is <0.05J under normal operating conditions), the energy in the 50Hz fundamental frequency band fluctuates slightly, and an energy anomaly appears near 0Hz (DC component) corresponding to the zero sequence current.

[0090] Step 3: 3D feature extraction and dynamic feature vector generation;

[0091] Feature 1: High-frequency transient energy coefficient;

[0092] Definition: Set the high-frequency subband to 1-5kHz (the main frequency band of transient signals from insulation breakdown faults), and calculate the total energy of this subband. Energy in the 50Hz fundamental frequency band The ratio is given by the formula:

[0093] ;

[0094] Calculation process: The result is obtained by integrating the time-frequency spectrum. , Substituting, we get:

[0095] ;

[0096] Under normal operating conditions The current value indicates the presence of a significant high-frequency transient, consistent with insulation breakdown characteristics.

[0097] Feature 2: Statistical characteristics of the zero-crossing derivative of the zero-sequence current;

[0098] Definition: Zero-sequence current (In case of failure) Significantly increased); found Calculate the rate of change (derivative) of the current within 0.0001s before and after the waveform crosses zero. ), and the statistical mean μ and standard deviation σ.

[0099] Calculation process: within t=125.03s~125.05s, The zero-crossing points are t = 125.035s, 125.042s, and 125.048s, corresponding to the derivatives. , , The calculation yields:

[0100] mean ;

[0101] Standard deviation ;

[0102] Under normal operating conditions The current value indicates that the zero-sequence current changes drastically, which corroborates grounding faults.

[0103] Feature 3: Harmonic distortion rate offset percentage;

[0104] Definition: First calculate the current total harmonic distortion rate. (formula , For fundamental current, (for harmonic currents), then calculate relative to Offset percentage:

[0105] ;

[0106] Calculation process: Actual measurement , , Substituting, we get:

[0107] ;

[0108] Known ,but:

[0109]

[0110] A deviation exceeding 100% indicates abnormal line harmonics, further confirming the fault.

[0111] Step 4: Dynamic feature vector concatenation;

[0112] The three dimensions of characteristics (high-frequency transient energy coefficient 0.02, mean value of zero-crossing derivative of zero-sequence current) are used to determine the three dimensions of characteristics. Standard deviation Harmonic distortion rate offset percentage By concatenating these features, a dynamic feature vector is obtained. After receiving the data, the cloud-based operations and maintenance platform compares it with historical "partial insulation breakdown" fault characteristic vectors (typical values). Highly matched, quickly and accurately determining the type of fault.

[0113] In this invention, the power frequency periodic current waveform is first resampled and aligned to eliminate the interference of sampling frequency deviation and waveform phase shift on subsequent analysis, ensuring the consistency of time-frequency feature extraction. By obtaining the time-frequency spectrum through synchronous compressed wavelet transform, the subtle features such as transients and harmonics hidden in the current waveform can be accurately captured (traditional Fourier transform is difficult to distinguish time-varying signals). The extracted high-frequency transient energy coefficient, zero-sequence current zero-crossing point derivative statistical features, and harmonic distortion rate offset percentage are three dimensions of features, which correspond to different physical manifestations of leakage faults (such as insulation breakdown and grounding faults). The features are comprehensive and highly identifiable. The final spliced ​​dynamic feature vector can provide high-dimensional and targeted data support for fine-grained diagnosis in the cloud, avoiding fault misjudgment caused by a single feature, and significantly improving the accuracy of cloud-based positioning of leakage fault type and fault degree.

[0114] Step 3: Compare and upload relevant data for alerts;

[0115] Edge computing nodes compare real-time leakage risk quantification indicators with local early warning thresholds. When the threshold is exceeded, an alarm is triggered and dynamic feature vectors, local risk time series data for the current and related periods are uploaded to the cloud operation and maintenance platform.

[0116] Specifically, the lighting circuit scenario for the corridor on the third floor of the office building is continued.

[0117] The parameters of the circuit edge node are consistent with those described above: 10 8W LED lights (rated total power 80W), and 1Hz low-frequency sampling for calculating real-time leakage risk quantification indicators. Local warning threshold The associated time period is "the current moment and the previous 5 low-frequency sampling points (6 seconds in total)," and 10kHz high-frequency sampling is used to generate dynamic feature vectors.

[0118] Operating Condition 1: Not exceeding the threshold (no alarms, no uploads);

[0119] Edge nodes (edge ​​computing nodes) are sampled at 1Hz and continuously compute from t=120s to 125s. The values ​​are 0.292, 0.315, 0.338, 0.351, 0.367, and 0.382, respectively, all less than the threshold of 0.6. At this point, the edge nodes only store these values ​​locally. The corresponding timestamp and load power (both are 80W, and the load is stable) do not trigger alarms or transmit data to the cloud, thus avoiding invalid data from occupying bandwidth (a single dynamic feature vector is about 128 bytes, and 6 seconds of time-series data is about 144 bytes, so this operation directly saves about 272 bytes of transmission).

[0120] Operating Condition 2: Exceeding the threshold (triggers alarms and uploads data);

[0121] Subsequently, between t=126s and 128s, the line experienced insulation breakdown and leakage. It rose to 0.523 (t=126s), 0.557 (t=127s), and 0.618 (t=128s).

[0122] Alarm triggered at t=128s The edge node immediately triggers a local audible and visual alarm and simultaneously starts data uploading;

[0123] Related time series data processing: Extracting 6 seconds of data from t=123s to 128s – timestamp 123s ( (Power 80W), 124s (0.367, 80W), 125s (0.382, 80W), 126s (0.523, 80W), 127s (0.557, 80W), 128s (0.618, 80W);

[0124] Dynamic feature vector acquisition: Start 10kHz high-frequency sampling and generate vectors according to steps 1 to 4 above. (Including high-frequency transient energy coefficient, mean / standard deviation of zero-sequence current derivative, and percentage shift of harmonic distortion rate);

[0125] Data Upload: The above vectors and 6-second time-series data are packaged and uploaded to the cloud. The total data volume is approximately 272 bytes, with a transmission latency of <0.5 seconds. This provides complete data support for multi-level collaborative diagnosis (such as historical case matching and model analysis) and ensures accurate diagnosis.

[0126] In this invention, edge computing nodes can quickly trigger alarms by directly comparing local real-time leakage risk quantification indicators with local early warning thresholds, avoiding protection delays caused by uploading all data to the cloud before judgment, and ensuring the real-time performance of leakage protection. At the same time, dynamic feature vectors and local risk time-series data for the associated period are uploaded only when the indicators exceed the threshold, rather than continuously uploading all data, which greatly reduces the amount of data transmitted between the edge and cloud, and reduces bandwidth consumption and energy consumption. Moreover, the uploaded data includes leakage current risk trends (local risk time-series data) and fine-grained waveform features (dynamic feature vectors), which can provide a complete and accurate data source for subsequent multi-level collaborative diagnosis in the cloud, realizing the collaborative connection of "local rapid response + cloud accurate diagnosis".

[0127] Step 4: Calculation of comprehensive diagnostic decision factors;

[0128] After receiving the data, the cloud-based operations and maintenance platform initiates multi-level collaborative diagnosis. The diagnosis process includes: first, performing pattern matching based on historical cases on the received dynamic feature vector to generate a first reference coefficient; second, analyzing the received complete data sequence using a cloud-based evaluation model to generate a second reference coefficient; and finally, calculating a comprehensive diagnostic decision factor based on the confidence levels of the first and second reference coefficients using a nonlinear fusion function. The complete data sequence includes the dynamic feature vector and local risk time-series data uploaded by edge computing nodes.

[0129] In this embodiment, in step four, the first reference coefficient The generation process is as follows: The cloud-based operations and maintenance platform stores a historical case feature library. Each historical case contains a historical feature vector and a final fault label, where the historical feature vector and the dynamic feature vector have the same feature dimension; the received dynamic feature vector... The k-nearest neighbor algorithm is used to retrieve the most similar feature from the historical case feature database. Calculate the dynamic feature vector from historical cases, i.e., neighbor cases. Historical feature vectors for each historical case cosine similarity and with The reciprocal of the distance is used as the weight to obtain the similarity weight. ,in Sort by distance; calculate Cosine similarity The weighted average of the neighboring cases is used as the average similarity. First reference coefficient Calculated using the following formula:

[0130] ;

[0131] in, Fault labels for neighboring cases. For indicator functions, when the fault label belongs to a predefined set of emergency fault types. The value is 1 when the time is right, and 0 otherwise. for Information entropy of fault labels in a neighboring case This represents the total number of fault label types.

[0132] In this embodiment, in step four, the cloud-based evaluation model is a model based on a deep temporal network; the second reference coefficient... The generation process is as follows: The cloud-based assessment model takes the received data sequence containing dynamic feature vectors and historical risk quantification indicators as input, extracts time-dependent features through a long short-term memory network layer, and then assigns weights to features at different time steps through an attention layer; the output of the cloud-based assessment model is mapped to a real-time risk probability through a fully connected layer. Second reference coefficient By analyzing real-time risk probabilities and the confidence level of the model in predicting the results. The specific formula is obtained by combining the results:

[0133] ;

[0134] Among them, confidence estimation It is calculated from the concentration of the attention weight distribution.

[0135] In this embodiment, in step four, the nonlinear fusion function generates a comprehensive diagnostic decision factor. The specific process is as follows: First, based on the average similarity of neighboring cases... Calculate the confidence factor of the first reference coefficient Based on the second reference coefficient Confidence level in the calculation process Calculate its credibility factor Then, a confidence-weighted sigmoid fusion function is used to calculate... :

[0136] ;

[0137] in, For the Sigmoid function, It is a very small positive number.

[0138] In this embodiment, a specific example of step four is as follows: continuing the scenario of the lighting circuit in the corridor on the third floor of the office building, the fault is a partial breakdown of the line insulation.

[0139] It is known that the cloud-based operations and maintenance platform receives data uploaded by edge nodes: dynamic feature vectors. (High-frequency transient energy coefficient, mean / standard deviation of zero-sequence current zero-crossing derivative, percentage offset of harmonic distortion rate), local risk time series data (real-time leakage risk quantification indicators from t=123s to 128s). Preset parameters: k-nearest neighbor algorithm (5 most similar historical cases) (Distance weighting coefficient), a predefined set of emergency fault types. Includes "partial insulation breakdown of the line" (fault label) ), total number of fault tag types (Insulation breakdown, poor contact, normal), extremely small positive number .

[0140] 1. Calculate the first reference coefficient ;

[0141] (1) Historical case retrieval and similarity calculation: In the cloud-based historical case database, cases similar to those found in the historical case database are retrieved. The historical feature vectors of the 5 most similar neighbor cases are... The cosine similarity is in order , , , , Sort by similarity in descending order, and by distance from the sorting index. The numbers are 1, 2, 3, 4, and 5 respectively.

[0142] (2) Similarity weight Calculation: According to the formula Substituting, we get:

[0143] , , , , .

[0144] (3) Average similarity of neighboring cases calculate: .

[0145] (4) Information entropy Calculation: The fault labels for all 5 neighbor cases are (belonging to) ), probability distribution ( Percentage) ( ), ( According to the entropy formula ,have to .

[0146] (5) Calculation: According to the formula ,because Therefore Substituting, we get:

[0147] ;

[0148] (Since the weighted sum is 3.892, the numerator equals the denominator).

[0149] 2. Calculate the second reference coefficient ;

[0150] (1) Cloud-based evaluation model input and feature extraction: "Dynamic feature vector" +Local risk time series data The data sequence is input into the LSTM + attention layer model, and the LSTM extracts... The temporal trend increased from 0.351 to 0.618, with the attention layer at t=126s~128s. The rapid upward phase is assigned a high weight (weight ratio 0.85), and the attention weight distribution is highly concentrated.

[0151] (2) Calculation of real-time risk probability and confidence level: The model outputs the real-time risk probability. (Probability of determining it as an insulation breakdown fault), confidence level The confidence level is 0.92, calculated from the concentration of attention weight distribution (the more concentrated the weights, the higher the confidence level).

[0152] (3) Calculation: According to the formula Substituting into .

[0153] It should be noted that the attention layer of the cloud-based evaluation model assigns an attention weight to each time step (assuming a total of n time steps) of the input data sequence (including dynamic feature vectors and historical risk quantification indicators), denoted as . And all weights satisfy the normalization constraint. (The total weight is 1 to ensure that the allocation logic is reasonable).

[0154] Attention weight distribution concentration is used to measure the model’s attention to “critical time steps” (such as periods of sudden increase in leakage risk). The more concentrated the weights are in a few critical time steps (non-uniform distribution), the higher the concentration, indicating that the model can accurately capture core features and the greater the corresponding confidence (Conf). Conversely, the more dispersed the weights are (uniform distribution), the lower the concentration and the smaller the Conf.

[0155] In this embodiment, the concentration of attention weight distribution is quantified based on information entropy, and the specific process is as follows:

[0156] First, calculate the information entropy of the attention weight distribution. (The smaller the entropy value, the more concentrated the distribution);

[0157] Normalize the entropy value (map to the 0-1 range): ( It is the maximum entropy when the weights are uniformly distributed over n time steps.

[0158] Confidence calculation Conf: (The smaller the entropy normalization value, the closer Conf is to 1, and the higher the concentration).

[0159] In short, its core is to determine the accuracy of the model's attention to key features by quantifying the uniformity of the attention weight distribution, and ultimately convert this into the confidence level of the prediction result, supporting the second reference coefficient. The calculation.

[0160] 3. Calculate comprehensive diagnostic decision factors ;

[0161] (1) Calculation of credibility factor: (Confidence level of the first reference coefficient) (Reliability of the second reference coefficient).

[0162] (2) Weighted fusion and Sigmoid transform: according to the formula First, calculate the weighted sum:

[0163] ;

[0164] Then through the Sigmoid function calculate: .

[0165] (3) Significance of the results: (generally (Judging as high risk) accurately reflects the current high-risk status of insulation breakdown faults, avoiding the problem of relying solely on historical cases (which may lead to misjudgment if there are no similar cases in the case library) or models (which may lead to misjudgment if there is a lot of data noise).

[0166] In this invention, the cloud first generates a first reference coefficient through k-nearest neighbor pattern matching based on historical cases, which can quickly locate similar faults based on existing fault experience and reduce diagnostic delays in new scenarios. Then, a second reference coefficient is generated through a cloud evaluation model based on deep temporal networks, which can accurately capture the temporal dependence features in the data and improve the risk identification capability under complex dynamic conditions. Finally, a comprehensive diagnostic decision factor (DDF) is generated by fusing the credibility of the two reference coefficients (average similarity of neighbor cases and model confidence) as weights through a nonlinear fusion function. This effectively avoids the limitations of a single diagnostic method (such as edge scenarios not covered by historical cases or model misjudgment of sudden faults), allowing diagnostic decisions to take into account both experience reliability and data-driven accuracy, and significantly improving the comprehensiveness and credibility of leakage fault diagnosis.

[0167] Step 5: Generate a collaborative response strategy;

[0168] Based on comprehensive diagnostic decision factors and preliminary fault type identification results, the cloud-based operation and maintenance platform performs risk mapping. Based on the final risk level determined by the risk mapping and the preliminary fault type identification results, it generates a collaborative response strategy and distributes the strategy to edge computing nodes and mobile operation and maintenance terminals. The collaborative response strategy includes adaptive threshold adjustment parameters for adjusting edge detection sensitivity, operation and maintenance decision suggestions corresponding to the final risk level and fault type, and model update instructions for optimizing the performance of local analysis models.

[0169] In this embodiment, after step four, the method of the present invention further includes: calculating the slope of the health decay trend of the corresponding lighting circuit within a preset historical period. Slope of the health decline trend It is the slope of the trend line obtained by linear fitting based on the historical risk quantification index sequence; the risk mapping process in step five is as follows: using comprehensive diagnostic decision factors The first dimension is the risk level of the initially identified fault types, and the second dimension is the slope of the health decay trend. The third dimension is used as the input vector. After normalizing the values ​​of these three dimensions, the vector is input into a risk space partitioning model pre-trained based on a clustering algorithm. The risk space partitioning model maps the input vector to a risk space composed of subspaces with different risk levels, thereby determining the final risk level.

[0170] Specifically, continuing with the lighting circuit scenario on the third floor of the office building corridor, the fault was a partial breakdown of the line insulation.

[0171] Given premise: Step four has already generated comprehensive diagnostic decision factors. (High risk tendency, value range 0-1); The preliminary fault type identification result is "partial insulation breakdown of the line", and the preset risk level of this fault type is 0.9 (set according to the fault severity, value range 0-1, 0 is no risk, 1 is extremely high risk); The preset historical period is 72 hours (based on 1 hour). The risk space partitioning model is a three-level model pre-trained with K-means clustering (low risk: all three normalized values ​​<0.5; medium risk: at least one dimension 0.5-0.7, the rest <0.8; high risk: at least two dimensions >0.7).

[0172] 1. Calculate the slope of the health decline trend. ;

[0173] Slope of the health decline trend Based on 72-hour historical risk quantification indicators ( The mean sequence, fitted by a linear fitting formula ( This is a time variable, in hours, with values ​​ranging from 0 to 71. For the corresponding time Mean; Calculation of the intercept:

[0174] The first 60 hours ( ): Circuit health, The mean value is stable between 0.3 and 0.35. hour , hour , hour , hour ;

[0175] The next 12 hours ( Insulation begins to break down. The mean gradually increases, such as hour , hour , hour (Compared to t=128s in step three) Consistent).

[0176] The least squares method was used to analyze 72... Linear fitting of data points yields the fitting equation. Therefore, the slope of the health decline trend (unit: Mean / hour, positive slope indicates that health decays over time; the steeper the slope, the faster the decay. (Normal loop) (Usually <0.001).

[0177] 2. Normalization of three-dimensional risk indicators;

[0178] Using min-max normalization (mapping the indicator to the 0-1 range), the formula is:

[0179] ;

[0180] First Dimension (DDF): , preset (No risk) (Extremely high risk), then ;

[0181] Second dimension (fault type risk level): , preset , ,but ;

[0182] The third dimension ( : , preset (No attenuation) (Limited decay rate), then .

[0183] 3. Input the risk space partitioning model to determine the final risk level;

[0184] Normalized three-dimensional vector Input the pre-trained K-means clustering risk space partitioning model:

[0185] Model determination logic: In this vector, , , It meets the high-risk judgment condition of "at least two dimensions > 0.7";

[0186] Final risk level: High risk, perfectly consistent with the actual scenario of "partial insulation breakdown requiring emergency handling," avoiding reliance solely on... Hesitation or neglect that may arise when approaching a high-risk threshold The intervention was delayed due to the continued deterioration of the displayed health trend.

[0187] In this invention, the slope of the health decay trend of the lighting circuit within a preset historical period is calculated. It can capture the changing trend of the long-term health status of the circuit, avoiding short-term misjudgments caused by relying solely on the transient comprehensive diagnostic decision factor (DDF) and the preliminary identification results of the fault type (e.g., a high transient DDF but a stable long-term health trend may be an occasional fluctuation, or a moderate transient DDF but...). (Indicating continued health deterioration requiring early intervention); simultaneously, using DDF, fault type risk level, The risk assessment system is constructed using three dimensions, rather than a single dimension. After normalization, it is input into a pre-trained clustering risk space partitioning model, which can more comprehensively reflect the current risk and long-term health hazards of the circuit. This allows the final risk level determination to take into account the instantaneous risk intensity, the inherent hazard of the fault, and the rate of health decay, significantly improving the accuracy and foresight of the risk level classification and providing a more scientific basis for the subsequent generation of targeted collaborative response strategies.

[0188] In this embodiment, the calculation process of the adaptive threshold adjustment parameter in step five is as follows: based on the final risk level and comprehensive diagnostic decision factors... Determine the required adjustment of the warning threshold offset for edge computing nodes. The specific calculation formula is as follows:

[0189] ;

[0190] in, The basic warning threshold for the corresponding lighting circuit, The steepness coefficient is related to the final risk level.

[0191] Specific examples are as follows (continuing the lighting circuit scenario of the 3rd floor corridor of the office building):

[0192] The basic early warning threshold for this circuit is known. (Based on the leakage current fluctuation range under normal operating conditions, corresponding to) Based on the fundamental judgment criteria, the final risk level and DDF for the two scenarios have been determined through the above method: Scenario 1 is "partial breakdown of line insulation" (final risk level: high risk). ); Scenario 2 is "LED light startup load fluctuation" (final risk level: low risk). The preset steepness coefficient γ rule is: high risk γ=2 (larger threshold adjustment range), low risk γ=1 (smaller adjustment range).

[0193] 1. Scenario 1: High risk (partial breakdown of line insulation);

[0194] (1) Parameter determination: , Ultimately, the risk level is high → γ=2;

[0195] (2) Calculate the adaptive threshold offset ΔT: according to the formula Substituting the data, we get:

[0196] ;

[0197] (3) Adjusted local early warning threshold: High risk requires increased detection sensitivity, therefore the adjusted threshold = ;

[0198] (4) Actual effect: The adjusted threshold of 0.458 is lower than the original basic threshold of 0.5. When the threshold gradually increases from 0.35 (normal) to 0.46 (initial risk stage), an alarm can be triggered, avoiding the problem of "alarms only being triggered after the risk has spread" that may have been caused by the original fixed threshold of 0.5, and gaining 3-5 minutes of response time for handling serious faults.

[0199] 2. Scenario 2: Low risk (LED light startup load fluctuation);

[0200] (1) Parameter determination: , Ultimately, the risk level is low → γ=1;

[0201] (2) Calculate the adaptive threshold offset ΔT: Substituting into the formula, we get:

[0202] ;

[0203] (3) Adjusted local early warning threshold: Low-risk areas need to reduce invalid alarms, therefore the adjusted threshold = ;

[0204] (4) Actual effect: When the LED light is turned on It briefly rose to 0.6 (which would have triggered false alarms at the original fixed threshold of 0.5), but because it did not reach the adjusted threshold of 0.85, it did not trigger an alarm. This effectively avoids such low-risk fluctuations from consuming operation and maintenance resources, and meets the actual need for "accurate alarms only in high-risk scenarios".

[0205] In this invention, by combining the final risk level and the comprehensive diagnostic decision factor (DDF) to calculate the adaptive threshold offset ΔT, the local warning threshold of the edge computing node can be dynamically adjusted according to the real-time risk status of the lighting circuit, instead of using a fixed threshold. When the risk is high, ΔT is amplified by matching the steepness coefficient γ of the risk level to reduce the warning threshold and improve detection sensitivity (avoiding missed detection of serious faults). When the risk is low, ΔT is reduced and the warning threshold is increased to reduce sensitivity (reducing invalid alarms caused by occasional fluctuations). This ensures that the threshold adjustment range is accurately matched with the severity of the fault, significantly enhancing the flexibility and scenario adaptability of edge node leakage detection, and avoiding the problem of misjudgment or missed detection under different risk scenarios with a fixed threshold.

[0206] Step Six: Local Updates and On-site Handling;

[0207] Edge computing nodes receive and parse collaborative response strategies, adjust parameters to update local early warning thresholds based on adaptive thresholds, execute model update instructions to optimize local analysis models, and perform corresponding protective actions. Meanwhile, mobile operation and maintenance terminals generate operation and maintenance work orders based on operation and maintenance decision suggestions, guide operation and maintenance personnel to handle matters on-site, and feed back the handling results to the cloud operation and maintenance platform.

[0208] In this embodiment, in step six, after the edge computing node performs the protection action, it starts a brief high-frequency monitoring window to capture the transient response waveforms of voltage, current and leakage current after the action. Each transient response waveform is compared with the standard response waveform generated based on the theoretical model or historical normal action data, the action effectiveness evaluation score is calculated, and the score and the response waveform summary are fed back to the cloud operation and maintenance platform.

[0209] Specific examples are as follows (continuing the scenario of the lighting circuit in the corridor on the 3rd floor of the office building, the protective action is "disconnecting the circuit breaker"):

[0210] Known scenario: The circuit triggers the alarm in step three due to partial insulation breakdown. The cloud sends out a collaborative response strategy, and the edge node executes the protection action of "disconnecting the circuit breaker". Preset high-frequency monitoring window parameters: sampling frequency 10kHz (consistent with high-frequency sampling to ensure transient detail capture), monitoring duration 0.1 seconds (covering the complete transient process of voltage, current and leakage current after the circuit breaker action); standard response waveform is generated based on historical normal circuit breaker action data: the voltage waveform drops from 220V (RMS) to 0V within 0.02 seconds, the current waveform drops from 5A (fundamental current) to 0A synchronously, and the leakage current waveform drops from 2.38mA (fault leakage current) to 0mA within 0.03 seconds, with no abnormal spike signals.

[0211] 1. Start the high-frequency monitoring window and transient waveform capture;

[0212] While the edge node performs the circuit breaker disconnection action, it immediately initiates 10kHz high-frequency sampling for 0.1 seconds (a total of 1000 sampling points) to capture three types of transient response waveforms:

[0213] Voltage waveform: 220V before operation, drops to 0V at 0.021 seconds (slightly later than the standard 0.02 seconds due to the mechanical delay of the circuit breaker), and remains at 0V for the next 0.021-0.1 seconds;

[0214] Current waveform: 5A before operation, then drops to 0A synchronously after 0.021 seconds, with no residual current;

[0215] Leakage current waveform: 2.38mA before operation, dropping to 0mA at 0.032 seconds (slightly later than the standard 0.03 seconds), with no abnormal fluctuations.

[0216] 2. Calculate the effectiveness assessment score for the action;

[0217] The "waveform similarity percentage" is used as the validity score. The comprehensive similarity is calculated based on the Pearson correlation coefficients of three types of waveforms: voltage, current, and leakage current. The formula is as follows:

[0218] (1) Single-class waveform similarity: ( To capture the Pearson correlation coefficient between the waveform and the standard waveform, the absolute value is taken to ensure a positive correlation, and then multiplied by 100% to convert it to a percentage.

[0219] (2) Overall validity score: (The three types of waveforms are weighted equally, which is consistent with the overall impact of the protective action on the electrical parameters).

[0220] Specific calculations:

[0221] Voltage waveform: Correlation between captured waveform and standard waveform (The action delay is only 0.001 seconds, and the overall trend is highly consistent), therefore ;

[0222] Current waveform: Correlation coefficient between the captured waveform and the standard waveform (The current synchronously drops to 0A without deviation), therefore ;

[0223] Leakage current waveform: Correlation coefficient between the captured waveform and the standard waveform (Action delay 0.002 seconds, no abnormality), therefore ;

[0224] Overall validity score: (generally (The protective action was deemed effective).

[0225] 3. Data feedback and cloud support;

[0226] The edge node packages the "Action Effectiveness Evaluation Score 98%" and the "Voltage / Current / Leakage Current Transient Response Waveform Summary (including key parameters: voltage drops to 0V in 0.021 seconds, current drops to 0A in 0.021 seconds, leakage current drops to 0mA in 0.032 seconds)" and feeds it back to the cloud-based operations and maintenance platform.

[0227] Cloud-based assessment: A validity score of 98% indicates that the circuit breaker has operated effectively, the faulty circuit has been successfully disconnected, and no secondary protective action is required.

[0228] Subsequent optimization: The cloud records the waveform summary and score to the closed-loop event sample. If similar loops occur subsequently... (For example, if the circuit breaker disconnects after a delay of 0.05 seconds), it will trigger the operation and maintenance recommendation of circuit breaker parameter calibration, reflecting its value in supporting the long-term reliability of the system.

[0229] In this invention, after the edge computing node executes a protective action, it captures the transient response waveform by activating a high-frequency monitoring window. This allows for real-time verification of the actual effectiveness of the protective action (such as circuit breaker tripping or residual current device (RCD) tripping), avoiding the risk of undetected protective action failures (such as a circuit breaker jamming and failing to trip, or an RCD malfunctioning and failing to disconnect the faulty circuit). Simultaneously, the transient response waveform is compared with the standard response waveform to calculate an effectiveness score, which can quantitatively evaluate the execution quality of the protective action. Furthermore, this score and waveform summary are fed back to the cloud, providing real-world data support for optimizing protection strategies (such as adjusting circuit breaker action parameters) and updating risk mapping rules. This achieves a closed loop of "protective action - effectiveness verification - strategy optimization," improving the reliability and self-optimization capability of the leakage current protection system.

[0230] Step 7: Incremental learning optimization;

[0231] Based on the feedback of the handling results, the cloud-based operation and maintenance platform uses an incremental learning mechanism to optimize the multi-level collaborative diagnosis and risk mapping rules.

[0232] In this embodiment, in step seven, the specific process of the incremental learning mechanism is as follows: the cloud operation and maintenance platform records the complete process from uploading data from the edge computing node to generating a collaborative response strategy, executing edge protection, and finally receiving feedback on the handling results, as a closed-loop event sample, and trains and fine-tunes the cloud evaluation model online; at the same time, it optimizes the risk space partitioning model or the final risk level determination rule by utilizing the correspondence between the risk mapping results recorded in the closed-loop event sample and the handling effect; and based on the accuracy of historical case matching and the success rate of handling results, it maintains and dynamically updates a long-term utility weight for each historical case in the feature historical case library, which is used to represent the priority of the corresponding historical case being retrieved in subsequent pattern matching.

[0233] Specific examples are as follows (continuing the scenario of the lighting circuit in the corridor on the 3rd floor of the office building, the closed-loop event is "handling of partial insulation breakdown fault in the line"):

[0234] It is known that the sample of this closed-loop event contains a complete data chain: dynamic feature vectors uploaded by edge nodes. Local risk time series data from t=123s to 128s ( → Cloud generation The final risk level is "high risk". The coordinated response strategy (circuit breaker disconnection + threshold adjustment to 0.458) → edge execution of protection actions (effectiveness score 98%) → maintenance personnel replace the insulated cable on site. The result of the handling is "fault eliminated and circuit restored to normal" (marked as "handling successful").

[0235] 1. Closed-loop event samples are used for cloud-based evaluation and model fine-tuning;

[0236] The cloud-based evaluation model is an LSTM + attention layer model. It uses the "dynamic feature vector + local risk time-series data" from the aforementioned closed-loop event samples as new training samples, and then fine-tunes them online after fusing them with the original training set (1000 historical fault samples).

[0237] Fine-tuning objective: Optimize the LSTM layer for " Rapid ascent phase (t=126s~128s, The feature extraction weights increased from 0.523 to 0.618.

[0238] Fine-tuning effect: The model before fine-tuning for this class The feature capture loss function for trends was initially set at 0.12, which was fine-tuned to 0.05. This improved performance when encountering similar trends later. When a fault occurs with a "small jump in a short period of time", the model outputs the real-time risk probability. The accuracy rate has increased from 92% to 96%, reducing misjudgments caused by insufficient feature capture.

[0239] 2. Optimize risk mapping rules;

[0240] Analyze the matching between "risk level assessment result (high risk)" and "response effect (successfully eliminated fault)" in this closed-loop event sample:

[0241] In the original risk mapping rules, the determination of "high risk" required that "at least two normalized dimensions > 0.7" (e.g., , ), in this sample It complies with the rules and the handling is effective;

[0242] Further analysis of nearly 10 similar closed-loop samples of "partial insulation breakdown" revealed that when (Close to 0.7) but , In practice, high-risk responses are also required, so the rules are optimized based on incremental learning: "..." and Even at that time "It is still judged as high risk" to avoid underestimating the risk due to a single dimension approaching the threshold.

[0243] 3. Dynamically update the long-term utility weights of historical cases;

[0244] Define the formula for calculating long-term utility weights: ( Case matching accuracy is the rate at which the diagnostic results of a case match the actual fault after it is retrieved. The success rate is defined as the success rate of handling maintenance recommendations generated based on the matching results of this case. The value ranges from 0 to 1, with higher weights indicating higher search priority. Two related historical cases are selected to update the weight:

[0245] Case 1 (Original Weight) The fault was matched with "partial insulation breakdown" three times. (All three diagnoses were correct). (All three treatments were successful), substituting into the formula yields... Its weight is increased to 1.0, and it will be prioritized for matching in subsequent searches;

[0246] Case 2 (Original Weight) It was searched twice, and misdiagnosed once (misclassified as "poor contact"). One out of two attempts was successful. Substituting into the formula, we get The weight is reduced to 0.5, and the priority is lowered in subsequent searches to avoid interfering with accurate diagnosis.

[0247] In this invention, the entire process from edge data uploading, collaborative response strategy generation, protection execution to handling feedback is recorded as a closed-loop event sample. An incremental learning mechanism can be used to fine-tune the cloud-based evaluation model online, allowing it to continuously adapt to changes in actual operating conditions (such as fault characteristic shifts caused by line aging). At the same time, risk mapping rules or level judgment criteria are optimized based on closed-loop samples to improve the accuracy of risk level classification. Furthermore, the long-term utility weight of cases can be dynamically updated based on the historical case matching accuracy and handling success rate, ensuring that subsequent pattern matching prioritizes the retrieval of high-value cases and avoids outdated or low-accuracy cases interfering with diagnosis. This achieves continuous iterative optimization of the system's diagnostic capabilities, the rationality of risk judgment, and the effectiveness of the case library, extending the reliability of the system's long-term service life.

[0248] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A cloud-edge collaborative integrated method for leakage current protection and remote operation and maintenance of lighting systems, characterized in that, Includes the following steps: S1: By deploying edge computing nodes in the lighting distribution box, the voltage, current and leakage current waveform data of each lighting circuit are collected synchronously, the waveform data are preprocessed, the load power data is calculated, and the leakage current time sequence data is extracted and standardized. S2: Edge computing nodes use local analysis models to perform dual-path parallel analysis. The first path calculates the real-time leakage risk quantification index of the lighting circuit based on leakage current time series data and load power data, and records the real-time leakage risk quantification index and its timestamp as local risk time series data. The second path performs online feature extraction on the current waveform to generate dynamic feature vectors. S3: Edge computing nodes compare real-time leakage risk quantification indicators with local early warning thresholds. When the threshold is exceeded, an alarm is triggered and dynamic feature vectors, local risk time series data for the current and related periods are uploaded to the cloud operation and maintenance platform. S4: After receiving the data, the cloud-based operations and maintenance platform initiates multi-level collaborative diagnosis. The diagnosis process includes: first, performing pattern matching based on historical cases on the received dynamic feature vector to generate a first reference coefficient; second, analyzing the received complete data sequence using a cloud-based evaluation model to generate a second reference coefficient; and finally, calculating a comprehensive diagnostic decision factor based on the confidence levels of the first and second reference coefficients using a nonlinear fusion function. The complete data sequence includes the dynamic feature vector and local risk time-series data uploaded by edge computing nodes. S5: The cloud-based operations and maintenance platform performs risk mapping based on comprehensive diagnostic decision factors and preliminary fault type identification results. Based on the final risk level determined by the risk mapping and the preliminary fault type identification results, it generates a collaborative response strategy and distributes the strategy to edge computing nodes and mobile operations and maintenance terminals. The collaborative response strategy includes adaptive threshold adjustment parameters for adjusting edge detection sensitivity, operations and maintenance decision suggestions corresponding to the final risk level and fault type, and model update instructions for optimizing the performance of the local analysis model. S6: The edge computing node receives and parses the collaborative response strategy, adjusts the parameters to update the local early warning threshold according to the adaptive threshold, executes the model update instruction to optimize the local analysis model, and performs corresponding protection actions; at the same time, the mobile operation and maintenance terminal generates operation and maintenance work orders according to the operation and maintenance decision suggestions, guides the operation and maintenance personnel to handle the situation on-site, and feeds back the handling results to the cloud operation and maintenance platform. S7: The cloud-based operations and maintenance platform optimizes multi-level collaborative diagnosis and risk mapping rules based on feedback processing results data using an incremental learning mechanism.

2. The integrated method for leakage protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration as described in claim 1, characterized in that, In step S2, a real-time leakage risk quantification index is calculated. The detailed process is as follows: S211: A sliding window-based differential accumulation algorithm is used to process the standardized leakage current timing data. Process the data and calculate the cumulative change within each window. ; S212: Combine the load power data of the corresponding lighting circuit at the same time. Calculate the sensitivity of the correlation between load and leakage current. ; S213: Real-time Leakage Risk Quantification Indicator Calculated using the following formula: ； Among them, cumulative change The calculation method is as follows: sum the absolute differences in leakage current between adjacent sampling points within the sliding window; correlation sensitivity The calculation method is as follows: by calculating the load power data With standardized leakage current timing data The absolute value of the correlation coefficient within the sliding window is obtained by linear scaling; t represents time t.

3. The integrated method for leakage current protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration as described in claim 1, characterized in that, In step S2, the specific process of online feature extraction for the second path is as follows: S221: After resampling and aligning the current waveform for each power frequency cycle, perform synchronous compressed wavelet transform to obtain the time spectrum; extract three-dimensional features from the time spectrum: calculate the ratio of the energy of the set high-frequency subband to the fundamental wave energy, as the high-frequency transient energy coefficient; Extract the statistical characteristics of the derivative of the zero-sequence current waveform near the zero-crossing point. The zero-sequence current waveform is the leakage current waveform. Calculate the percentage deviation of the harmonic distortion rate of the current waveform relative to the historical baseline value. S222: Concatenate the features of the above three dimensions to form a dynamic feature vector.

4. The integrated method for leakage protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration as described in claim 1, characterized in that, In step S4, the first reference coefficient The generation process is as follows: The cloud-based operations and maintenance platform stores a historical case feature library. Each historical case contains a historical feature vector and a final fault label, where the historical feature vector and the dynamic feature vector have the same feature dimension; the received dynamic feature vector... The k-nearest neighbor algorithm is used to retrieve the most similar feature from the historical case feature database. Calculate the dynamic feature vector from historical cases, i.e., neighbor cases. Historical feature vectors for each historical case cosine similarity and with The reciprocal of the distance is used as the weight to obtain the similarity weight. ,in The sorting number is based on distance. , where is the distance weighting coefficient; Calculation Cosine similarity The weighted average of the neighboring cases is used as the average similarity. First reference coefficient Calculated using the following formula: ； in, Fault labels for neighboring cases. For indicator functions, when the fault label belongs to a predefined set of emergency fault types. The value is 1 when the time is right, and 0 otherwise. for Information entropy of fault labels in a neighboring case This represents the total number of fault label types.

5. The integrated method for leakage current protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration as described in claim 4, characterized in that, In step S4, the cloud-based evaluation model is a deep temporal network-based model; the second reference coefficient... The generation process is as follows: The cloud-based evaluation model takes the received data sequence containing dynamic feature vectors and historical risk quantification indicators as input, extracts time-dependent features through the long short-term memory network layer, and then assigns weights to the features at different time steps through the attention layer. The output of the cloud-based assessment model is mapped to a real-time risk probability through a fully connected layer. Second reference coefficient By analyzing real-time risk probabilities and the confidence level of the model in predicting the results. The specific formula is obtained by combining the results: ； Among them, confidence estimation It is calculated from the concentration of the attention weight distribution.

6. The integrated method for leakage current protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration as described in claim 5, characterized in that, In step S4, the nonlinear fusion function generates a comprehensive diagnostic decision factor. The specific process is as follows: First, based on the average similarity of neighboring cases... Calculate the confidence factor of the first reference coefficient Based on the second reference coefficient Confidence level in the calculation process Calculate its credibility factor Then, a confidence-weighted sigmoid fusion function is used to calculate... : ； in, For the Sigmoid function, For positive numbers used to prevent division by zero.

7. The integrated method for leakage protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration as described in claim 6, characterized in that, After step S4, the method further includes: calculating the slope of the health decay trend of the corresponding lighting circuit within a preset historical period. Slope of the health decline trend It is the slope of the trend line obtained by linear fitting based on the historical risk quantification index sequence; the risk mapping process in step S5 is as follows: using comprehensive diagnostic decision factors The first dimension is the risk level of the initially identified fault types, and the second dimension is the slope of the health decay trend. The third dimension is used to normalize the values ​​of the three dimensions and input them into the risk space partitioning model. The risk space partitioning model maps the input vector to a risk space composed of subspaces with different risk levels, thereby determining the final risk level.

8. The integrated method for leakage current protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration as described in claim 1, characterized in that, In step S5, the calculation process for the adaptive threshold adjustment parameter is as follows: based on the final risk level and comprehensive diagnostic decision factors... Determine the required adjustment of the warning threshold offset for edge computing nodes. The specific calculation formula is as follows: ； in, The basic warning threshold for the corresponding lighting circuit, The steepness coefficient is related to the final risk level.

9. The integrated method for leakage current protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration as described in claim 1, characterized in that, In step S6, after the edge computing node performs the protection action, it starts a high-frequency monitoring window of a set duration to capture the transient response waveforms of voltage, current and leakage current after the action. Each transient response waveform is compared with the standard response waveform generated based on the theoretical model or historical normal action data, the action effectiveness evaluation score is calculated, and the score and response waveform summary are fed back to the cloud operation and maintenance platform.

10. The integrated method for leakage protection and remote operation and maintenance of lighting systems based on cloud-edge collaboration as described in claim 1, characterized in that, In step S7, the specific process of the incremental learning mechanism is as follows: The cloud operation and maintenance platform records the complete process from uploading data from the edge computing node to generating a collaborative response strategy, executing edge protection, and finally receiving feedback on the handling results, as a closed-loop event sample, and trains and adjusts the cloud evaluation model online; at the same time, it optimizes the risk space partitioning model or the final risk level determination rule by using the correspondence between the risk mapping results recorded in the closed-loop event sample and the handling effect; and based on the accuracy of historical case matching and the success rate of handling results, it maintains and dynamically updates a long-term utility weight for each historical case in the feature historical case library. The long-term utility weight is used to represent the priority of the corresponding historical case being retrieved in subsequent pattern matching.

Images