A comprehensive online monitoring system for power cables

Abstract

This invention provides a comprehensive online monitoring system for power cables, relating to the field of data processing technology. The system includes: extracting a set of partial discharge distribution points from a multi-dimensional feature fusion dataset and calculating the weighting factor for each distribution point; based on the weighting factor, constructing the boundary of the partial discharge source probability distribution using a weighted point set concave hull generation algorithm to obtain a set of boundary feature points; inputting the set of boundary feature points and the multi-dimensional feature fusion dataset into a convolutional-long short-term memory hybrid neural network to generate diagnostic results; generating a three-level early warning signal based on the diagnostic results, pushing the early warning information to remote operation and maintenance terminals and mobile apps via the MQTT protocol, automatically triggering a digital work order service to generate maintenance tasks and perform location verification. This invention improves the accuracy of monitoring and fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a comprehensive online monitoring system for power cables. Background Technology

[0002] As the core carrier of urban power grids and industrial power distribution systems, the operating status of power cables directly affects the safety and stability of the power system. With the increase in power grid load density and the lengthening of cable service life, the risks of abnormal grounding current, joint overheating, and partial discharge faults are increasing. Traditional single-parameter monitoring and manual diagnosis methods are no longer sufficient to meet reliability requirements.

[0003] In addition, cable condition is affected by environmental temperature, humidity and pollution. Some existing systems store environmental parameters as independent data and do not deeply integrate them into the condition feature extraction process, resulting in insufficient model generalization ability. Some traditional systems rely on a single communication method, which is prone to data packet loss in complex electromagnetic environments or network congestion scenarios, affecting the timeliness of monitoring. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a comprehensive online monitoring system for power cables, which improves the accuracy of monitoring and fault diagnosis.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a comprehensive online monitoring system for power cables includes:

[0007] The data acquisition module is used to collect the time-domain waveform of the grounding current of power cables, the temperature field distribution of the joint, the partial discharge pulse sequence and environmental parameters, and obtain multiple types of raw data sets;

[0008] The data processing module is used to input multiple types of raw data sets into the edge computing node, perform data preprocessing, and generate standardized state data packets;

[0009] The data fusion module is used to transmit standardized status data packets to the cloud platform through a dual-channel redundant network of TCP / IP and power line carrier. It uses a dynamic time warping algorithm to align the spatiotemporal dimensions of current, temperature, and partial discharge data, and fuses environmental parameters to generate a multi-dimensional feature fusion dataset.

[0010] The fault diagnosis module is used to extract the partial discharge distribution point set from the multi-dimensional feature fusion dataset and calculate the weight factor of each distribution point. Based on the weight factor, a weighted point set concave hull generation algorithm is used to construct the boundary of the partial discharge source probability distribution and obtain the boundary feature point set. The boundary feature point set and the multi-dimensional feature fusion dataset are input into a convolutional-long short-term memory hybrid neural network to generate diagnostic results.

[0011] The intelligent response module is used to generate a three-level early warning signal based on the diagnostic results, push the early warning information to the remote operation and maintenance terminal and mobile APP via the MQTT protocol, and automatically trigger the digital work order service to generate maintenance tasks and perform location verification.

[0012] Furthermore, the various types of raw data sets are input into the edge computing nodes for data preprocessing to generate standardized state data packets, including:

[0013] The time-domain waveform data of power cable grounding current from multiple types of raw datasets are processed to generate power frequency current waveform data; and temperature field data are generated by Kriging interpolation based on discrete point data of temperature field distribution of power cable joints from multiple types of raw datasets; the temperature field data are then calculated and processed to obtain temperature field analysis results.

[0014] The partial discharge pulse sequence data from multiple types of raw datasets are processed to obtain a partial discharge pulse feature dataset; and the environmental parameters from multiple types of raw datasets are processed to generate effective environmental parameter data.

[0015] The current waveform data after power frequency rejection, temperature field analysis results, partial discharge pulse characteristic dataset, and effective environmental parameter data are collected and processed to generate a multi-source state dataset; the multi-source state dataset is then normalized to obtain a standardized state data package.

[0016] Furthermore, standardized status data packets are transmitted to the cloud platform via a dual-channel redundant network of TCP / IP and power line carrier. A dynamic time warping algorithm is used to align the spatiotemporal dimensions of current, temperature, and partial discharge data, and environmental parameters are fused to generate a multi-dimensional feature fusion dataset, including:

[0017] Standardized status data packets are transmitted to the cloud platform in parallel via TCP / IP network channel and power line carrier communication channel; after the data packets received from both channels are verified to be error-free through redundancy check, a cloud-received verification data packet is generated.

[0018] The received verification data packets from the cloud are processed to obtain current time series, temperature time series and partial discharge pulse time series; and a dynamic time warping algorithm and nonlinear time axis transformation are used to generate spatiotemporally aligned current series, temperature aligned series and partial discharge aligned series.

[0019] Based on the spatiotemporally aligned current sequence, temperature aligned sequence, and partial discharge aligned sequence, normalized environmental parameter data is extracted from the verification data packet received in the cloud. The current aligned sequence, temperature aligned sequence, partial discharge aligned sequence, and environmental parameter data are integrated according to time points to generate a multi-dimensional feature fusion dataset.

[0020] Furthermore, the partial discharge distribution point set is extracted from the multi-dimensional feature fusion dataset, and the weight factor of each distribution point is calculated. Based on the weight factor, a weighted point set concave hull generation algorithm is used to construct the boundary of the partial discharge source probability distribution, obtaining a boundary feature point set. The boundary feature point set and the multi-dimensional feature fusion dataset are input into a convolutional-long short-term memory hybrid neural network to generate diagnostic results, including:

[0021] The spatiotemporal distribution point set of partial discharge pulses is extracted from the multidimensional feature fusion dataset. Based on the spatiotemporal distribution point set of partial discharge pulses, the time aggregation weight factor of the timestamp of each point and the relative intensity weight factor of the amplitude of each point are calculated and the two are fused to generate a joint weight factor, thus obtaining the weighted partial discharge point set.

[0022] The partial discharge distribution point set and the corresponding weight factor of each point are used to generate a set of convex boundary points that can surround all partial discharge distribution points by using a weighted point set concave hull generation algorithm. This process constructs the probability distribution boundary of the high-weight region and obtains the set of boundary feature points.

[0023] The set of boundary feature points is mapped to generate a spatial probability matrix and fused with time series data. The matrix is ​​then input into a convolutional-LSTM hybrid neural network to extract spatiotemporal features and generate diagnostic results.

[0024] Furthermore, based on the diagnostic results, a three-level early warning signal is generated, and the early warning information is pushed to remote operation and maintenance terminals and mobile APP via the MQTT protocol. This automatically triggers the digital work order system to generate maintenance tasks and perform location verification, including:

[0025] The cable insulation condition risk level value in the diagnostic results is analyzed; the risk level value is matched with the preset three-level warning threshold range to obtain the warning signal level; the spatial location coordinates and timestamp data of the partial discharge source in the diagnostic results are extracted; and the warning signal level, location coordinates, and timestamp are encapsulated into a structured MQTT protocol message.

[0026] The structured MQTT protocol messages are pushed to remote operation and maintenance terminals and mobile APP in parallel through the cloud platform to provide early warning information; corresponding maintenance priority identifiers are generated according to the early warning signal level; the location coordinates, early warning signal level, and priority identifier are bound to generate structured maintenance instructions; and the digital work order service is automatically triggered to generate electronic maintenance work orders.

[0027] The spatial positioning service of the on-site verification terminal is driven by the positioning coordinates in the maintenance work order to perform local source positioning and receive the feedback verification results.

[0028] Furthermore, standardized status data packets are transmitted to the cloud platform in parallel via a TCP / IP network channel and a power line carrier communication channel; after the data packets received from both channels are verified for redundancy, a cloud-received verification data packet is generated, including:

[0029] The standardized status data packet is copied twice and transmitted to the cloud platform in parallel through the TCP / IP network channel and the power line carrier communication channel, respectively. The cloud platform receives the data packet copies from the two channels, which are denoted as copy A and copy B, and performs preliminary data integrity checks on each.

[0030] Based on the status of replicas A and B after preliminary verification, a preset time window is started to wait; if both replicas are valid within the window period, valid replicas A and B are obtained; otherwise, exception handling is triggered and the process is terminated.

[0031] Apply the same preset verification algorithm to valid copies A and B respectively to generate verification identifier A and verification identifier B, and compare the two; if they match, generate copies A and B with identical content; otherwise, trigger data inconsistency exception handling and terminate this process.

[0032] Based on the confirmed identical content of copies A and B, one of the copies is selected according to a preset selection strategy, and a cloud-received verification data packet is generated and output.

[0033] Furthermore, the received verification data packets from the cloud are processed to obtain current time series, temperature time series, and partial discharge pulse time series; and a dynamic time warping algorithm and nonlinear time axis transformation are used to generate spatiotemporally aligned current series, temperature aligned series, and partial discharge aligned series, including:

[0034] Based on receiving verification data packets in the cloud, performing predefined protocol parsing operations, and generating the original current time series, temperature time series, and partial discharge pulse time series;

[0035] The temperature time series is designated as the reference series, and a dynamic time warping algorithm is applied between the current time series and the temperature time series. The nonlinear time warping path between the two series is calculated to generate the current-temperature warping mapping relationship.

[0036] A dynamic time warping algorithm is applied between the partial discharge pulse time series and the temperature time series; the curved path is calculated to generate the partial discharge-temperature warping mapping relationship.

[0037] Based on the extracted original time series and mapping relationship, a nonlinear time axis transformation is performed to generate a current-aligned sequence and a partial discharge-aligned sequence that are completely aligned with the temperature sequence time axis, and the original temperature sequence is used as the temperature-aligned sequence.

[0038] Furthermore, using a weighted point set concave hull generation algorithm, the partial discharge distribution point set and the corresponding weight factors of each point are calculated to obtain a set of convex boundary points that can surround all partial discharge distribution points. This constructs the probability distribution boundary of the high-weight region, resulting in a set of boundary feature points, including:

[0039] Based on the weighted partial discharge site set, configure the initial concave radius parameter setting value and the radius adjustment step size;

[0040] Based on the weighted partial discharge site set and the initial radius setting, a concave hull generation algorithm with joint weighting factor as spatial density weight is adopted to calculate and generate a set of candidate boundary points that enclose all points;

[0041] The sum of the joint weight factors of all points within the region enclosed by the concave hull formed by the candidate boundary point set is calculated to obtain the enclosed weight sum; the sum of the joint weight factors of all points in the weighted partial discharge point set is calculated to obtain the total weight sum.

[0042] Calculate the wrapping rate of high-weight regions, dynamically adjust the radius parameter based on the comparison between the wrapping rate and the target threshold, and iteratively perform candidate boundary calculation and evaluation until the wrapping rate meets the requirements, finally obtaining the set of boundary feature points.

[0043] In a second aspect, a computing device includes:

[0044] One or more processors;

[0045] A storage device for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the system.

[0046] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the system.

[0047] The above-described solution of the present invention has at least the following beneficial effects:

[0048] By employing edge computing nodes to preprocess data, the amount of invalid data transmission is reduced, improving processing efficiency. A dual-channel redundant network design using TCP / IP and power line carrier effectively mitigates data loss caused by single transmission link failures, ensuring the real-time nature and continuity of monitoring data. A weighted point set concave hull algorithm is used to construct the probability distribution boundary of partial discharge sources. Combined with a convolutional-long short-term memory hybrid neural network, the spatial characteristics and temporal correlations of the data are fully explored, enhancing the identification accuracy and early warning capability of partial discharge faults, and reducing false alarms and missed alarms.

[0049] Furthermore, through real-time online monitoring, precise fault location, and rapid repair response, potential cable defects can be detected in a timely manner and intervention can be carried out in advance, effectively improving the power outage accident caused by the expansion of faults, enhancing the reliability and safety of power system operation, and reducing economic losses caused by faults. Attached Figure Description

[0050] Figure 1This is a schematic diagram of an integrated online monitoring system for power cables provided by an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram illustrating the process of a comprehensive online monitoring system for power cables provided by an embodiment of the present invention. The system transmits standardized status data packets to a cloud platform via a dual-channel redundant network of TCP / IP and power line carrier. It uses a dynamic time warping algorithm to align the spatiotemporal dimensions of current, temperature, and partial discharge data, and integrates environmental parameters to generate a multi-dimensional feature fusion dataset. Detailed Implementation

[0052] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0053] like Figure 1 As shown, an embodiment of the present invention proposes a comprehensive online monitoring system for power cables, comprising:

[0054] The data acquisition module is used to collect the time-domain waveform of the grounding current of power cables, the temperature field distribution of the joint, the partial discharge pulse sequence and environmental parameters, and obtain multiple types of raw data sets;

[0055] The data processing module is used to input multiple types of raw data sets into the edge computing node, perform data preprocessing, and generate standardized state data packets;

[0056] The data fusion module is used to transmit standardized status data packets to the cloud platform through a dual-channel redundant network of TCP / IP and power line carrier. It uses a dynamic time warping algorithm to align the spatiotemporal dimensions of current, temperature, and partial discharge data, and fuses environmental parameters to generate a multi-dimensional feature fusion dataset.

[0057] The fault diagnosis module is used to extract the partial discharge distribution point set from the multi-dimensional feature fusion dataset and calculate the weight factor of each distribution point. Based on the weight factor, a weighted point set concave hull generation algorithm is used to construct the boundary of the partial discharge source probability distribution and obtain the boundary feature point set. The boundary feature point set and the multi-dimensional feature fusion dataset are input into a convolutional-long short-term memory hybrid neural network to generate diagnostic results.

[0058] The intelligent response module is used to generate a three-level early warning signal based on the diagnostic results, push the early warning information to the remote operation and maintenance terminal and mobile APP via the MQTT protocol, and automatically trigger the digital work order service to generate maintenance tasks and perform location verification.

[0059] In this embodiment of the invention, edge computing nodes are used to preprocess data, reducing the amount of invalid data transmission and improving processing efficiency. A dual-channel redundant network design using TCP / IP and power line carrier effectively mitigates data loss caused by single transmission link failures, ensuring the real-time nature and continuity of monitoring data. A weighted point set concave hull algorithm is used to construct the probability distribution boundary of partial discharge sources. Combined with a convolutional-long short-term memory hybrid neural network, the spatial characteristics and temporal correlations of the data are fully explored, improving the identification accuracy and early warning capability of partial discharge faults, and reducing false alarms and missed alarms.

[0060] Furthermore, through real-time online monitoring, precise fault location, and rapid repair response, potential cable defects can be detected in a timely manner and intervention can be carried out in advance, effectively improving the power outage accident caused by the expansion of faults, enhancing the reliability and safety of power system operation, and reducing economic losses caused by faults.

[0061] In a preferred embodiment of the present invention, multiple types of raw data sets are input into an edge computing node for data preprocessing to generate standardized state data packets, including:

[0062] The time-domain waveform data of power cable grounding current from multiple types of raw datasets are processed to generate power frequency current waveform data; and temperature field data are generated by Kriging interpolation based on discrete point data of temperature field distribution of power cable joints from multiple types of raw datasets; the temperature field data are then calculated and processed to obtain temperature field analysis results.

[0063] The partial discharge pulse sequence data from multiple types of raw datasets are processed to obtain a partial discharge pulse feature dataset; and the environmental parameters from multiple types of raw datasets are processed to generate effective environmental parameter data.

[0064] The current waveform data after power frequency rejection, temperature field analysis results, partial discharge pulse characteristic dataset, and effective environmental parameter data are collected and processed to generate a multi-source state dataset; the multi-source state dataset is then normalized to obtain a standardized state data package.

[0065] In this embodiment of the invention, the sliding average method is used to perform baseline correction on the original current time-domain waveform (sampling frequency 1kHz-10kHz); the window size (5-20 sampling points) is dynamically adjusted according to the sampling frequency. For example, when sampling at 1kHz, the window takes 20 points (corresponding to 20ms), and when sampling at 10kHz, the window takes 5 points (corresponding to 0.5ms). The slow trend caused by sensor drift is eliminated by calculating the average value through sliding, while retaining the high-frequency fluctuation components.

[0066] Based on the 50Hz power frequency of my country's power system, the center frequency of the band-stop filter is set to 50Hz, with a bandwidth of 2Hz (covering the 49-51Hz range) to ensure accurate targeting of the power frequency interference band. To ensure the interference suppression effect, the stopband attenuation is set to ≥40dB (which can reduce the amplitude of the power frequency signal to less than 1% of the original value), and the passband ripple is set to ≤1dB (to reduce the attenuation of useful signals).

[0067] Considering the computing power limitations of edge computing nodes, we select 4th to 8th order Butterworth band-stop filters (4th order can meet the basic attenuation requirements, 8th order is suitable for strong interference scenarios, and the higher the order, the steeper the filter curve, but the computational load increases).

[0068] Based on the above parameters, filter processing is performed at the edge computing node to determine the coefficients of each order. The pre-processed ground current time-domain waveform data (with trend terms removed) is input into the filter, and signal filtering is achieved through iterative calculation (the output value at each moment is obtained by weighted summation of the current and the input values ​​from the previous 3-7 moments, depending on the order). The ratio of the remaining amplitude to the original amplitude of the 50Hz signal is calculated to ensure an attenuation rate ≥90%. Simultaneously, the amplitude loss rate of the useful signal frequency band (such as 1-10kHz related to partial discharge) is checked to be ≤5% to avoid distortion of the useful signal. Value range description: Filter order 4-8, stopband attenuation 40-60dB, passband ripple 0.5-1dB.

[0069] Standard deviation (σ): Takes a value of 0.5-2 sampling intervals, which is adjusted according to the sampling frequency: When sampling at 1kHz (sampling interval 1ms), σ takes 1-2ms (i.e., 1-2 sampling intervals); when sampling at 10kHz (sampling interval 0.1ms), σ takes 0.05-0.1ms (i.e., 0.5-1 sampling intervals), balancing noise removal and feature preservation.

[0070] The window length is 6-8 times σ (covering more than 99% of the energy of the Gaussian distribution). For example, when σ = 1ms, the window size is 6-8ms (corresponding to 6-8 sampling points); when σ = 0.1ms, the window size is 0.6-0.8ms (corresponding to 6-8 sampling points).

[0071] The center of the window has the highest weight (close to 1), which gradually decreases towards both ends (weight of edge points ≤ 0.01), and the total weight is normalized to 1 (to avoid changes in the overall signal amplitude after smoothing).

[0072] Sliding window processing is applied to the current waveform data: the smoothed value of each sampling point is equal to the sum of the products of the original values ​​of all points within the window and their corresponding weights.

[0073] Boundary handling: For regions where the beginning and end of the waveform are less than the length of a window, a mirror filling method is used (copying neighboring data with the boundary point as the axis of symmetry) to avoid boundary point distortion.

[0074] Value range description: σ = 0.5-2 sampling intervals, window size = 6-8 times σ, weight value 0.01-1 (after normalization).

[0075] The original temperature discrete points are from distributed fiber optic or infrared sensors (sampling point spacing 0.05-0.2 meters). Outliers are removed: the effective temperature range is set to -20℃ to 90℃ (covering the normal operating temperature of the cable and extreme environments). Data outside this range is judged as sensor failure and deleted. For missing points (the proportion ≤5%), they are filled by the average of the three nearest valid points to ensure data continuity.

[0076] Construction, training, and implementation of the variogram model (taking the spherical model as an example):

[0077] A spherical model is chosen to quantify the spatial correlation of the temperature field, defining three core parameters: nugget value (C0): reflecting measurement error, with a value of 0.1-2℃. 2 (Based on sensor accuracy, a sensor with ±0.5℃ corresponds to C0≤0.25℃) 2 ); Sill value (C0+C): reflects total spatial variation, ranging from 5-20℃ 2 (Corresponds to a maximum actual temperature difference of 3-5℃); Range (a): Reflects spatial distance, with a value of 0.1-0.5 meters (must be less than the cable joint length by 0.5-2 meters).

[0078] The model logic is as follows: when the distance is less than or equal to the range, the semi-variable value approaches the sill value as the distance increases; when the distance is greater than the range, the semi-variable value stabilizes at the sill value.

[0079] Model training:

[0080] Divide the spatial distance between temperature point pairs (0.05-1 meter) into 10-15 lag distance groups (each group width 0.05-0.1 meter), each group containing ≥30 point pairs, and calculate the semi-variance value (average of the squares of the temperature difference) for each group; parameter initialization: C0 is taken as 10%-20% of the semi-variance value of the smallest lag distance group, the sill value is taken as 90%-100% of the semi-variance value of the largest lag distance group, and the range is taken as 1 / 3-1 / 2 of the total lag distance range.

[0081] Iterative optimization: By adjusting parameters (range ±0.05 meters, C0 ±0.1℃) 2 , sill value ±1℃ 2 The goal is to minimize the root mean square error (RMSE) between the model's predicted and experimental values, iterating until the change in RMSE is ≤0.05℃. 2 Or up to 30 times; Constraints: C0 ≤ base value × 10%, range ≤ 1 / 2 of joint length.

[0082] Model Implementation and Validation:

[0083] Visual verification: Plot the experimental values ​​against the fitted curve, requiring that more than 80% of the points fall within ±0.5℃ of the curve. 2 Within the range; cross-validation: use 80% of the data for training and 20% for validation, ensuring that the RMSE difference between the validation set and the training set is <10%; if the standard is not met, readjust the lag distance grouping or change the model (such as the exponential model).

[0084] The cable joint area (0.5-2 meters in length) is divided into a three-dimensional grid of 0.01-0.05 meters. The temperature value of each grid point is calculated based on the trained variogram model to generate continuous temperature field distribution data.

[0085] Calculate the highest temperature, lowest temperature, average temperature of the temperature field, and the maximum temperature gradient (the ratio of temperature difference between adjacent grids to distance, reflecting the rate of heat diffusion).

[0086] Hotspot identification: Set a temperature threshold of 70-80℃ (based on the long-term allowable operating temperature of cross-linked polyethylene cable of 90℃), mark areas exceeding the threshold as hotspots, and record their spatial coordinates and temperature values.

[0087] Calculate the standard deviation of the temperature field. Under normal operation, the standard deviation should be ≤5℃; otherwise, it is considered that the temperature distribution is uneven. Set the amplitude threshold to 3-5 times the mean of the background noise (to distinguish between effective pulses and noise), and the width threshold to 0.1-10μs (to filter nanosecond-level electromagnetic interference and millisecond-level mechanical noise), retaining pulse signals that conform to the characteristics of partial discharge. For each effective pulse, extract the peak value (0.1-100mV), rise time (10-100ns), fall time (50-500ns), and pulse width (100ns-1μs). Extract three main frequency components (1-100MHz) and their corresponding amplitudes through spectrum analysis. Calculate the pulse repetition rate (1-1000 times / minute), mean, and variance of the amplitude distribution within one minute.

[0088] Through analysis, redundant features with a correlation of ≤0.3 with the type of partial discharge (such as corona discharge or surface discharge) are removed, and 8-15 core features are retained to form a feature dataset.

[0089] Outliers were removed from environmental parameters (temperature, humidity, air pressure, and rainfall), and the effective ranges were set as follows: ambient temperature -40℃ to 50℃, humidity 0 to 100%, air pressure 80 to 110 kPa, and rainfall 0 to 50 mm / h. For missing values ​​of ≤5 minutes, linear interpolation was used to fill in the missing values.

[0090] Missing data type determination: Scan the original environmental parameter sequence (sampling interval 1 minute), mark the location of missing values, and classify them according to the duration of the missing data into short missing data: consecutive missing data points ≤ 5 (corresponding to ≤ 5 minutes, which meets the preprocessing requirements); long missing data: consecutive missing data points > 5 (linear interpolation is not applicable to this type, and it needs to be marked as invalid segment and processed separately).

[0091] For short missing intervals, determine the valid data points before and after them (denoted as point A before and point B after). Points A and B must be normal data that have not been removed (e.g., ambient temperature within the range of -40℃ to 50℃), and there must be no other valid points between A and B (to avoid multiple interpolation).

[0092] Let the timestamp of the preceding point A be t1 (e.g., 10:00), and the corresponding parameter value be v1 (e.g., 25℃); let the timestamp of the following point B be t2 (e.g., 10:03), and the corresponding parameter value be v2 (e.g., 27℃); calculate the time interval Δt between A and B = t2 - t1 (unit: minutes), with a value range of 2-6 minutes (since the missing time is ≤5 minutes, plus one valid point before and after, the minimum total interval is 2 minutes).

[0093] Number of missing points: Suppose there are n missing points (n = 1-5) between A and B, and the timestamp of each missing point is t. 1+1 , t 1+2 , ..., t 1+n (1 minute interval).

[0094] Slope calculation logic: Based on the parameter difference and time interval between points A and B, determine the parameter change rate (slope) per unit time. For example: point A is 25℃ (10:00), point B is 27℃ (10:03), Δt = 3 minutes, the total parameter change Δv = 2℃, then the change rate per minute is 2℃ / 3≈0.67℃ / minute.

[0095] Fill in the gaps point by point:

[0096] First missing point (10:01): Value = v1 + 1 × rate of change (25 + 0.67 ≈ 25.67℃); Second missing point (10:02): Value = v1 + 2 × rate of change (25 + 1.33 ≈ 26.33℃); and so on, until all missing points are filled. The calculated value of the last missing point must be consistent with v2 at point B (error ≤ 0.1℃, to ensure linear continuity).

[0097] Error range control: whether the interpolation result is consistent with the trend of change of the neighboring valid points (e.g., when the temperature rises overall, the interpolation result should increase monotonically), and the deviation of a single interpolation point from the value calculated according to the linear trend should be ≤0.5℃ (humidity ≤2%, air pressure ≤0.5kPa).

[0098] Anomaly correction: If the interpolation result exceeds the reasonable range of the parameter (e.g., temperature interpolation > 50℃), linear interpolation is abandoned, and the historical average value of the parameter for the same period (e.g., the average value of the same period in the last 7 days) is used to fill the gap, ensuring that the physical meaning of the data is reasonable.

[0099] The correlation between each parameter and the cable joint temperature is calculated by correlation analysis. Parameters with a correlation of ≥0.6 (usually ambient temperature and humidity) are retained, while weakly correlated parameters such as air pressure are removed. The retained environmental parameters are converted into dimensionless indices (such as humidity index = actual humidity ÷ 100, temperature index = (actual temperature + 40) ÷ 90) to facilitate cross-parameter fusion analysis.

[0100] The four types of data (current waveform after power frequency removal, temperature field analysis results, partial discharge pulse characteristics, and effective environmental parameters) were aligned by timestamps and a unified sampling interval of 1 minute to form a multi-source state dataset.

[0101] The min-max normalization method is used to map the data to the interval [0, 1]. The minimum and maximum values ​​of each parameter are taken as the extreme values ​​of their historical normal operating range (such as ground current 0-10A, maximum temperature field 20-90℃), and finally a standardized state data packet is generated.

[0102] By eliminating power frequency interference and handling outliers, noise and interference are effectively removed while retaining key features. Normalization and time alignment address the heterogeneity of multi-source data, ensuring that data of different types and magnitudes can be fused and analyzed. Steps such as temperature field hotspot identification and partial discharge pulse feature extraction target and mine key information related to cable faults, improving the input effectiveness of subsequent diagnostic models. Edge node preprocessing reduces the amount of invalid data transmission, saving computing power for deep analysis on the cloud platform and improving system real-time performance.

[0103] In a preferred embodiment of the present invention, standardized state data packets are transmitted to a cloud platform via a dual-channel redundant network of TCP / IP and power line carrier. A dynamic time warping algorithm is used to align the spatiotemporal dimensions of current, temperature, and partial discharge data, and environmental parameters are fused to generate a multi-dimensional feature fusion dataset, including:

[0104] Standardized status data packets are transmitted to the cloud platform in parallel via TCP / IP network channel and power line carrier communication channel; after the data packets received from both channels are verified to be error-free through redundancy check, a cloud-received verification data packet is generated.

[0105] The received verification data packets from the cloud are processed to obtain current time series, temperature time series and partial discharge pulse time series; and a dynamic time warping algorithm and nonlinear time axis transformation are used to generate spatiotemporally aligned current series, temperature aligned series and partial discharge aligned series.

[0106] Based on the spatiotemporally aligned current sequence, temperature aligned sequence, and partial discharge aligned sequence, normalized environmental parameter data is extracted from the verification data packet received in the cloud. The current aligned sequence, temperature aligned sequence, partial discharge aligned sequence, and environmental parameter data are integrated according to time points to generate a multi-dimensional feature fusion dataset.

[0107] In this embodiment of the invention, the TCP / IP channel is configured as follows: the industrial Ethernet protocol is adopted, the transmission rate is 100Mbps-1Gbps, an industrial dedicated port (such as 502, 8080) is selected, the data packet format is JSON (including timestamp, device ID, data type, and payload fields), the sending interval is consistent with the standardized data packet generation frequency (1 minute / time), and the timeout retransmission threshold is set to 3 times (if it exceeds 3 times, it will switch to the backup path).

[0108] Power line carrier channel configuration: Low-voltage power line carrier communication (PLC) is adopted, with an operating frequency of 3-500kHz (avoiding power frequency interference), the modulation method is orthogonal frequency division multiplexing (OFDM), the transmission rate is 500kbps-2Mbps, and the data packets are supplemented with RS code channel coding (error correction capability of 4-8 bytes) to adapt to power line noise environment.

[0109] Edge nodes simultaneously send the same standardized status data packets to both channels, marking the channel identifier ("CH1" for TCP / IP and "CH2" for power line carrier) to ensure that the data packets are transmitted independently at the physical link layer.

[0110] The cloud platform sorts dual-channel data packets by timestamp (with an error margin of ±100ms), and data packets with the same timestamp are considered to be from the same source.

[0111] Calculate the CRC16 check value for each data packet (using a polynomial of 0x8005, with an initial value of 0xFFFF), and compare it with the check field provided in the data packet. If the check passes, the packet is marked as valid. If both channels are valid, select the TCP / IP channel data (with lower transmission latency) as the baseline and the power line carrier data as the backup.

[0112] If a single-channel verification fails, valid data from the other channel is used directly. If both channels fail, a data retransmission request is triggered (limited to data packets within the last 5 minutes). If the number of retransmissions exceeds the limit, the data is marked as missing (the percentage must be ≤0.1%). Finally, a cloud-received verification data packet (containing a complete timestamp and multiple data types) is generated.

[0113] The current time series is extracted from the verification data packets received from the cloud: sampling frequency 1kHz-10kHz, each data point includes a timestamp (accurate to milliseconds) and a normalized current value (0-1); temperature time series: sampling frequency 1 time / minute, including timestamps (accurate to minutes) and temperature field statistical characteristics (maximum temperature, average temperature, etc., all normalized); partial discharge pulse time series: non-equal interval sampling (recorded when the pulse is triggered), including pulse occurrence time (accurate to microseconds) and feature vectors (peak value, frequency, etc., normalized); duplicate timestamp data in the time series is removed (the latest value is retained) to ensure that the timestamp is unique within a single sequence.

[0114] Construction, training, and implementation of the Dynamic Time Warping (DTW) algorithm:

[0115] The core objective is to align sequences with different sampling frequencies and start times (e.g., current sequences are high-frequency and dense, while temperature sequences are low-frequency and sparse), defining three key elements:

[0116] Distance metric: Euclidean distance is used to measure the similarity between two data points (applicable to normalized data, with a value range of 0-√2); Path constraint: Sakoe-Chiba window constraint is used (the window width is 5%-10% of the sequence length, such as 50-100 points for a 1000-point sequence) to limit the search range of the alignment path and avoid physically unreasonable matching (such as aligning a sudden current change with a stable temperature segment); Cumulative distance: The sum of the cumulative distances from the start point to the end point of the sequence, used to evaluate the overall alignment effect (the smaller the value, the higher the alignment).

[0117] Training data: Select 30 days of historical normal operation data (including current, temperature, and partial discharge sequences) to ensure coverage of different operating conditions (such as peak and off-peak loads); Window width optimization: Calculate the physical consistency of the aligned sequences (such as the matching degree of current peak and temperature rise trend) for different window widths (3%-15% of sequence length), select the smallest window with a matching degree ≥85% (balancing accuracy and computational load), and finally determine the window width to be 5%-10%; Distance threshold setting: Statistically calculate the cumulative distance after normal data alignment, and take the 95th percentile value as the threshold (usually 0.5-2.0, after normalization). If the distance exceeds the threshold, it is judged as an alignment anomaly (manual review is required).

[0118] Sequence preprocessing: The current sequence is downsampled to 1 time / minute (consistent with the frequency of the temperature sequence), retaining the peak and average values ​​within each minute as features; Distance matrix construction: The Euclidean distance between each pair of points in the current feature sequence (length N) and the temperature sequence (length M) is calculated to form an N×M matrix; Path search: Starting from the upper left corner (starting point) of the matrix, the path with the smallest cumulative distance is selected within the constraint window, allowing one-to-one, one-to-many, or many-to-one matching (e.g., matching 1 current peak point with 3 temperature rise points); Nonlinear time axis transformation: The timestamps of the temperature sequence are adjusted according to the adaptation path to align each temperature point with the corresponding current feature point in time (error ≤ 10 seconds); Partial discharge sequence alignment: Using the same logic, the partial discharge pulse time is mapped to the aligned time axis (based on the current sequence time) to ensure that the pulse occurrence time matches the trend of current and temperature changes in time.

[0119] Finally, spatiotemporally aligned current, temperature, and partial discharge sequences are generated (timestamps are uniformly accurate to the second and of consistent length).

[0120] Normalized environmental parameter data (ambient temperature index and humidity index, 0-1 range) are extracted from the verification data packets received from the cloud, and their timestamps are consistent with the temperature sequence (1 time / minute). Through timestamp matching, the environmental parameters are associated with the aligned current, temperature, and partial discharge sequences (four data items with the same timestamp form a feature sample).

[0121] Using the aligned timeline (longest sequence length, usually 24×60=1440 points / day) as a benchmark, check whether each time point contains four data items: if environmental parameters are missing at a certain time point, linear interpolation is used to fill them in (error ≤0.05); if partial discharge data is missing, it is marked as "no pulse" (all feature vectors are 0).

[0122] After integration, each sample includes: current-aligned features (3-5 statistics), temperature-aligned features (4-6 statistics), partial discharge-aligned features (8-15 core features), and environmental parameters (2 indices), forming a structured data record.

[0123] 10% of the samples are randomly selected to check the consistency of timestamps (error ≤ 1 second) and data integrity (missing rate ≤ 0.5%), ensuring that the multi-dimensional features correspond one-to-one in time, and finally generating a multi-dimensional feature fusion dataset.

[0124] The dual-channel redundancy design of TCP / IP and power line carrier ensures a data transmission success rate of ≥99.9%, with automatic switching in case of single-channel failure to prevent data loss. The integrated dataset contains multi-dimensional features including current, temperature, partial discharge, and environmental characteristics, with consistent spatiotemporal dimensions, thus improving model recognition accuracy. Algorithm parameters (such as window width and distance threshold) are optimized based on historical data training, balancing real-time performance with computational power requirements.

[0125] In a preferred embodiment of the present invention, a set of partial discharge distribution points is extracted from a multi-dimensional feature fusion dataset, and a weight factor for each distribution point is calculated. Based on the weight factors, a weighted point set concave hull generation algorithm is used to construct the boundary of the partial discharge source probability distribution, resulting in a set of boundary feature points. The set of boundary feature points and the multi-dimensional feature fusion dataset are input into a convolutional-long short-term memory hybrid neural network to generate diagnostic results, including:

[0126] The spatiotemporal distribution point set of partial discharge pulses is extracted from the multidimensional feature fusion dataset. Based on the spatiotemporal distribution point set of partial discharge pulses, the time aggregation weight factor of the timestamp of each point and the relative intensity weight factor of the amplitude of each point are calculated and the two are fused to generate a joint weight factor, thus obtaining the weighted partial discharge point set.

[0127] The partial discharge distribution point set and the corresponding weight factor of each point are used to generate a set of convex boundary points that can surround all partial discharge distribution points by using a weighted point set concave hull generation algorithm. This process constructs the probability distribution boundary of the high-weight region and obtains the set of boundary feature points.

[0128] The set of boundary feature points is mapped to generate a spatial probability matrix and fused with time series data. The matrix is ​​then input into a convolutional-LSTM hybrid neural network to extract spatiotemporal features and generate diagnostic results.

[0129] In this embodiment of the invention, partial discharge pulse data are selected from the multi-dimensional feature fusion dataset, and the spatiotemporal attributes of each pulse are extracted to form a spatiotemporal distribution point set; each point includes: timestamp (accurate to milliseconds, consistent with the aligned time axis); spatial coordinates (based on the relative position of the three-dimensional model of the cable joint, such as axial distance 0-2 meters, radial distance 0-0.5 meters); pulse amplitude (normalized to 0-1, corresponding to the original amplitude 0.1-100mV).

[0130] Duplicate points need to be removed from the point set (the pulse with the largest amplitude within ±1ms of the same spatiotemporal coordinate and timestamp is retained) to ensure that there is no redundant data in a single point set.

[0131] The aggregation degree of statistical points is statistically analyzed using a sliding time window, with the window size set to 5-10 minutes (covering the typical burst cycle of partial discharge pulses). For each point, the total number of pulses within its window is calculated (denoted as N). The time aggregation weight factor = N ÷ the maximum number of pulses within the window (value range 0-1). For example, if there are a maximum of 50 pulses in a certain window and a certain point's window has 30 pulses, then its time weight is 0.6. The window sliding step size is set to 1-2 minutes to ensure that the time resolution matches the overall monitoring frequency (1 minute / time).

[0132] The maximum amplitude of all pulses in the point set is used as the benchmark (denoted as Vmax). The relative intensity weight factor of each point is equal to the amplitude of that point divided by Vmax (the value range is 0-1). To avoid small amplitude pulses being ignored, a minimum intensity threshold of 0.1 is set (the weight of points with amplitudes lower than Vmax×10% is 0.1, not 0), to ensure the reasonable weight of weak signals.

[0133] Joint weighting factor = 0.4 × time aggregation weight + 0.6 × relative intensity weight (time weight accounts for 40%-60%, which can be dynamically adjusted according to the type of partial discharge, such as increasing the time weight for periodic partial discharge).

[0134] After fusion, the joint weight ranges from 0.1 to 1 (since the intensity weight is the lowest at 0.1). The higher the value, the greater the contribution of that point to fault diagnosis. Finally, a weighted partial discharge site set is generated (each point is given an additional joint weight).

[0135] Weight threshold determination: The weight threshold is set based on the noise level in the multi-dimensional feature fusion data, with a value range of 0.3-0.7. If the monitoring environment noise is high (such as complex cable laying environment), the threshold is set to 0.5-0.7 to filter low weight noise points. If the noise is low (such as laboratory or new cable system), the threshold is set to 0.3-0.5 to retain more potential effective signals.

[0136] High-weight key point extraction: Points with joint weight factors higher than the above threshold are selected to form a high-weight key point subset; the number of points in this subset usually accounts for 20%-50% of the total number of points in the original distribution point set (too low a proportion may result in the loss of effective information, while too high a proportion will retain redundant points).

[0137] Detailed process of constructing the boundary of the probability distribution of a partial discharge source using the weighted point set concave hull generation algorithm:

[0138] From the weighted partial discharge site set, low-weighted points with a joint weighting factor < 0.1 are removed (considered as noise or interference signals), retaining the effective point set (denoted as P, containing N points, typically 50-500, depending on monitoring duration and discharge activity). The spatial coordinates of the effective points (e.g., distance values ​​along the cable length) are mapped to a unified coordinate system (e.g., with the cable start point as the origin, the coordinate range is normalized to relative values ​​of 0-100) to avoid the impact of absolute distance differences on boundary calculations. The joint weighting factor of all retained points is ensured to be within the range of 0.1-1.0 (if the original calculation exceeds this range, it is adjusted to this range through linear scaling), providing a unified benchmark for subsequent weighted calculations.

[0139] Initial convex hull generation (basics of boundary calculation): Calculate the minimum convex hull (i.e. the smallest convex polygon that can completely enclose all points) for the effective point set P: First, find the point with the smallest y-coordinate in the point set (if there are multiple points, select the one with the smallest x-coordinate) as the starting point. Sort the remaining points in ascending order of their polar angle with the starting point. Then, use a stack structure to determine the direction of the points in turn (left turns are retained, right turns are divided). Finally, the vertex sequence of the convex hull is formed (denoted as H, containing M vertices, M is usually 5-30, less than the number of points in the point set P).

[0140] Connect the vertices of the convex hull in sequence to form M consecutive edge segments (denoted as L1, L2, ..., L...). m Each line segment is determined by two adjacent vertices, and the length of the line segment is usually between 5 and 50 (corresponding to 0.5-5m of the actual cable length in normalized coordinates).

[0141] Convex hull edge evaluation and candidate concave point identification: For each convex hull edge L i (Connect vertices A and B), traverse all valid point sets P that are not in L i For each point on the line, calculate the distance L from each point to the line segment. i The perpendicular distance (denoted as d) is used to filter out points where d > 0 (i.e., points located at L). i The inner point may become a concave reference point.

[0142] Weighted distance sorting: For the selected inner points, calculate the "weighted distance" = d × the joint weight factor of the point (the higher the weight, the greater the influence of the point on the edge adjustment), and sort them from largest to smallest weighted distance. Take the top 3-5 points as candidate concave points of the edge (these points are the key points that are most likely to "pull" the edge inward).

[0143] Weighted edge concavity adjustment:

[0144] For each edge L i And its candidate concave points, and determine whether they meet the concave condition: the joint weight factor of the candidate points is ≥0.5 (to ensure that the point has sufficient reference value), and the point is within L.i The distance should be ≥ 10% of the edge length (to avoid meaningless fine-tuning due to being too close to the edge).

[0145] If there exists a candidate point that meets the conditions (denoted as C), then the original edge L is... i (AB) is split into two segments: AC and CB. At this time, the "convex angle" of the original convex hull is replaced by the "concave angle" (point C becomes the new boundary vertex).

[0146] After splitting, the total length of the new line segments AC and CB is equal to that of the original line segment L. i The length difference should be ≤30% (to prevent excessive concavity from causing boundary distortion); if the weight of candidate point C is >0.8 (high weight), the concavity should be retained first; if the weight is between 0.5 and 0.8, it is necessary to check whether there are other high weight points (≥0.7) around to support the concavity. If there are, retain it; otherwise, discard it.

[0147] Multi-round iterative optimization: Repeat the above concave operation for all edges, iterating 2-3 times (to avoid excessive iteration leading to boundary fragmentation). After each round, recalculate the convex hull vertex sequence and update the boundary segments.

[0148] After concave adjustment, the boundary may contain too many vertices (e.g., the distance between some adjacent vertices is less than 2 normalized units, corresponding to an actual distance of less than 0.2m). The boundary is simplified by using the Douglas-Peucker algorithm, retaining key vertices (with the error threshold set to 1-2 normalized units) and reducing the number of feature points (ultimately retaining 10-50, balancing accuracy and computational efficiency).

[0149] Douglas-Peucker algorithm steps:

[0150] Let the set of boundary points to be simplified be an ordered sequence S = [Q0, Q1, ..., Q...]. n (Arranged in clockwise / counterclockwise order, with the beginning and end connected to form a closed boundary), set a distance threshold ε (set according to the accuracy requirements, such as 1-2 units in normalized coordinates, corresponding to 0.1-0.2m of the actual cable length); the smaller the threshold, the more points are retained after simplification, and the closer the shape is to the original boundary; the larger the threshold, the higher the degree of simplification and the fewer points.

[0151] Take the first point Q0 and the last point Q in the point set. n Connect to form the initial line segment M = Q0 - Q n .

[0152] Calculate all intermediate points (Q1 to Q) in the point set S. n-1 Find the point Q with the largest perpendicular distance from line segment M to line segment M. k Let the distance be d. k .

[0153] Judgment and recursive processing:

[0154] If d k >ε: Explanation of Q k These are key feature points that affect the shape of the boundary and must be preserved; the point set S is split into two subsets: S1 = [Q0, Q1, ..., Q...]. k ] and S2=[Q k Q k+1 Q n ], repeat the recursive simplification for each of the two subsets; if d k ≤ε: This means that the distance from all intermediate points to line segment M is less than the threshold. These points can be considered redundant and directly removed. The simplified line segment is M, and only Q0 and Q are retained. n After the recursion is completed, all the retained key feature points are spliced ​​together in their original order to form a simplified set of boundary points.

[0155] Example explanation: Assume the original boundary has 100 vertices, and the threshold ε = 1.5 (normalized coordinates):

[0156] Round 1: Take the first and last points Q0 and Q... 99 Calculate the distance from the middle 98 points to line segment Q0-Q. 99 Find the farthest point Q within the given distance. 50 (Distance d=3>1.5), keep Q 50 Split into [Q0-Q 50 ] and [Q 50 -Q 99 ].

[0157] Round 2: Against [Q0-Q] 50] Find the farthest point Q 20 (d=2>1.5), retain and split; for [Q 50 -Q 99 Find the farthest point Q. 70 (d=1.2≤1.5), remove intermediate points and keep only Q. 50 and Q 99 .

[0158] By repeating the recursion, 100 points can eventually be reduced to 15-20 key vertices, which preserves the concave and convex features of the boundary while significantly reducing the amount of data.

[0159] Ensure that the simplified vertex sequence is connected end to end to form a closed concave polygon boundary. This boundary must completely enclose all points with a weight ≥ 0.3 in the effective point set (low-weight points can be excluded from the boundary and regarded as secondary interference). Arrange the vertices of the final boundary in clockwise or counterclockwise order to form a boundary feature point set (each point contains standardized coordinates and corresponding weights), which will serve as one of the input features of the subsequent neural network.

[0160] Network structure construction:

[0161] Input layer: Receives a spatial probability matrix (32×32 or 64×64 in size) mapped from a set of boundary feature points, where the matrix values ​​are the joint weights of the corresponding positions, reflecting the spatial distribution; and a time series of a multi-dimensional feature fusion dataset (containing current, temperature, and environmental parameters, with time segments of 10-30 minutes in length).

[0162] Convolutional layers: Set up 2-3 convolutional layers. The first layer uses 32 3×3 filters (to extract local spatial features), and the second layer uses 64 3×3 filters (to extract combined features). Each layer is followed by 2×2 max pooling (to reduce dimensionality). The activation function is ReLU (to avoid gradient vanishing).

[0163] LSTM layer: Set up 1-2 LSTM units (64-128 units / layer) to receive the spliced ​​data of spatial features and time series output from the convolutional layer and learn the temporal correlation (such as the hysteresis relationship between partial discharge pulse and temperature rise).

[0164] Output layer: Set the number of nodes according to the diagnostic target (e.g., 3 types of faults + 1 type of normal, a total of 4 nodes), and output the probability of each category (range 0-1).

[0165] Network training process:

[0166] 70% of the historical monitoring data is selected as the training set (including normal and fault labels), 20% as the validation set, and 10% as the test set, ensuring that the proportion of fault samples is 30%-40% (balanced class distribution); the optimizer uses Adam (learning rate 0.001-0.01, decaying by 10% every 5 epochs), and the parameters are adjusted through iterative optimization, with 30-50 iterations (until the accuracy of the validation set no longer improves); a Dropout layer (proportion 0.2-0.3) and L2 regularization (coefficient 0.0001) are added to prevent overfitting (the difference between the accuracy of the training set and the validation set should be <5%).

[0167] Input data preprocessing: the spatial probability matrix is ​​normalized to [0, 1], and the time series is spliced ​​into continuous segments at the minute level.

[0168] The convolutional layer outputs a spatial feature map (reflecting the spatial morphology of high-weight regions), the LSTM layer outputs a temporal feature vector (reflecting the trend of parameter changes), and the probability distribution is output after being fused by the fully connected layer.

[0169] The category with the highest probability is taken as the final result. If the highest probability is less than 0.7 (insufficient credibility), it is marked as "to be reviewed" and the manual verification process is triggered.

[0170] Value range description: 32-64 convolutional filters, 64-128 LSTM units, learning rate 0.001-0.01, training iterations 30-50, spatial matrix size 32×32-64×64, time segment length 10-30 minutes.

[0171] By strengthening high-value partial discharge (PD) signals with weighting factors and reducing noise interference, the model focuses on fault-related features, improving diagnostic sensitivity. The weighted concave hull boundary accurately encloses high-weight regions, transforming 3D spatial features into structured boundary points, resolving the ambiguity in PD source localization and reducing spatial localization error. The convolutional-LSTM network simultaneously captures the spatial morphology and temporal evolution of PD (such as changes in pulse frequency and amplitude during fault development), improving diagnostic accuracy compared to single models (pure convolution or pure LSTM).

[0172] In a preferred embodiment of the present invention, a three-level early warning signal is generated based on the diagnostic results, and the early warning information is pushed to a remote operation and maintenance terminal and a mobile APP via the MQTT protocol. This automatically triggers the digital work order system to generate maintenance tasks and perform location verification, including:

[0173] The cable insulation condition risk level value in the diagnostic results is analyzed; the risk level value is matched with the preset three-level warning threshold range to obtain the warning signal level; the spatial location coordinates and timestamp data of the partial discharge source in the diagnostic results are extracted; and the warning signal level, location coordinates, and timestamp are encapsulated into a structured MQTT protocol message.

[0174] The structured MQTT protocol messages are pushed to remote operation and maintenance terminals and mobile APP in parallel through the cloud platform to provide early warning information; corresponding maintenance priority identifiers are generated according to the early warning signal level; the location coordinates, early warning signal level, and priority identifier are bound to generate structured maintenance instructions; and the digital work order service is automatically triggered to generate electronic maintenance work orders.

[0175] The spatial positioning service of the on-site verification terminal is driven by the positioning coordinates in the maintenance work order to perform local source positioning and receive the feedback verification results.

[0176] Extract the cable insulation condition risk level value from the diagnostic results (normalized range 0-100, 0 indicates no risk, 100 indicates extremely high risk).

[0177] Preset three-level warning threshold range:

[0178] Level 1 Warning (highest level): Risk level 70-100 (corresponding to severe insulation deterioration and intense partial discharge); Level 2 Warning: Risk level 30-70 (corresponding to mild insulation aging and moderate partial discharge); Level 3 Warning: Risk level 0-30 (corresponding to slight insulation abnormalities and weak or occasional partial discharge).

[0179] The analyzed risk level values ​​are matched with threshold ranges to determine the warning signal level (e.g., risk level 85 matches a level 1 warning).

[0180] Extract the spatial location coordinates of the partial discharge source from the diagnostic results: a three-dimensional coordinate system based on the cable laying path, including axial distance (0-5km, accurate to ±1m), radial depth (0-1m, accurate to ±0.1m), and elevation (±50m, reference ground elevation); extract timestamp data: accurate to the second (format YYYY-MM-DDHH:MM:SS, consistent with the time of diagnostic result generation, error ≤10s).

[0181] Set the message topic to "PowerCable / Warning / [Device ID]" (the device ID is a unique 10-20 digit code, such as "CBL-20230723-001").

[0182] The payload field contains structured data: warning level ("Level1", "Level2", "Level3"); location coordinates (axial, radial, and elevation values ​​and units); timestamp (the above format string); and risk level value (retaining one decimal place, such as 85.2).

[0183] Set the message QoS (Quality of Service) to 1 (ensure at least one delivery), and set reserved messages to "Enable" (newly accessed terminals can obtain historical warnings).

[0184] The cloud platform pushes messages to remote operation and maintenance terminals (industrial tablets, supporting 4G / 5G) and mobile apps (iOS / Android systems) simultaneously via MQTT broker.

[0185] Remote terminal: Upon receiving the signal, it triggers an audible and visual alarm (volume ≥ 80dB, light flashing frequency 2 times / second) and displays the complete warning information on the screen (font enlarged to 24 points, highlighting the location coordinates).

[0186] Mobile App: Push notification message (title includes "Warning Level + Device ID", content summary displays risk level and time), clicking will redirect to the details page (including map location entry).

[0187] If the first push fails, it will be retried three times at intervals of 10s, 30s, and 60s. If it still fails, it will be logged and a manual intervention reminder will be triggered.

[0188] Prioritize based on warning level:

[0189] Level 1 Warning: Priority "Urgent" (marked as "P0"), requiring a response within 1 hour and arrival at the scene within 4 hours; Level 2 Warning: Priority "High" (marked as "P1"), requiring a response within 2 hours and arrival at the scene within 12 hours; Level 3 Warning: Priority "General" (marked as "P2"), requiring a response within 24 hours and completion of verification within 72 hours.

[0190] The maintenance instruction includes: equipment ID, warning level, priority identifier, location coordinates (format after latitude and longitude conversion, error ±5m), and a summary of diagnostic results (such as "partial discharge source is located at connector A3, amplitude 120mV").

[0191] Work order numbering rules: "WO-last two digits of year-month-day-serial number" (e.g., "WO-25-07-23-005"); includes the following fields: work order status (initially "pending dispatch"), responsible team (automatically assigned based on equipment affiliation), estimated completion time (calculated based on priority), and associated warning ID (bound to the unique identifier of the MQTT message).

[0192] Within 5 minutes of the work order being generated, it is pushed to the responsible team's terminal via the system and simultaneously updated on the operation and maintenance management platform dashboard.

[0193] After maintenance personnel arrive at the approximate area with the on-site verification terminal (supporting GPS + Beidou dual-mode positioning, positioning accuracy ±3m), the terminal automatically activates the spatial positioning service based on the positioning coordinates in the work order: it starts the partial discharge detector (detection frequency band 30kHz-30MHz) and reduces the detection range according to the coordinate information (initial radius 10m, dynamically adjusted according to signal strength: reduced to 3m when the signal is strong, and expanded to 20m when the signal is weak).

[0194] The terminal screen displays the real-time positioning deviation (the straight-line distance between the current location and the target coordinates, in meters). When the deviation is ≤5m, the system prompts "Enter the verification area".

[0195] On-site personnel enter the verification results through the terminal. The result type is: "Confirmed Fault", "Suspected Fault", or "No Fault". Additional information includes: text description (e.g., "Temperature at connector A3 is 38°C, partial discharge signal detected"), on-site photos (≥2 photos, including equipment identification and measuring instrument readings), and positioning correction coordinates (if the actual position deviates from the work order coordinates by more than 5m, manual adjustment is required).

[0196] Feedback information is uploaded to the cloud platform via 4G / 5G, triggering a work order status update (such as "verified"), and is synchronously linked to the warning record to form a closed loop.

[0197] A three-tiered threshold range is bound to priorities to ensure rapid response to high-risk faults and orderly handling of low-risk issues, optimizing resource allocation. The entire process, from early warning location to work order generation and on-site verification, is automated, shortening fault diagnosis time. The MQTT protocol, combined with a retry mechanism and multi-terminal push, improves the delivery rate of early warning information. Dual-mode positioning and dynamic detection range adjustment improve the verification success rate.

[0198] In a preferred embodiment of the present invention, standardized status data packets are transmitted to the cloud platform in parallel via a TCP / IP network channel and a power line carrier communication channel; after the data packets received from both channels are verified to be error-free through redundancy checks, a cloud-received verification data packet is generated, including:

[0199] The standardized status data packet is copied twice and transmitted to the cloud platform in parallel through the TCP / IP network channel and the power line carrier communication channel, respectively. The cloud platform receives the data packet copies from the two channels, which are denoted as copy A and copy B, and performs preliminary data integrity checks on each.

[0200] Based on the status of replicas A and B after preliminary verification, a preset time window is started to wait; if both replicas are valid within the window period, valid replicas A and B are obtained; otherwise, exception handling is triggered and the process is terminated.

[0201] Apply the same preset verification algorithm to valid copies A and B respectively to generate verification identifier A and verification identifier B, and compare the two; if they match, generate copies A and B with identical content; otherwise, trigger data inconsistency exception handling and terminate this process.

[0202] Based on the confirmed identical content of copies A and B, one of the copies is selected according to a preset selection strategy, and a cloud-received verification data packet is generated and output.

[0203] In this embodiment of the invention, a standardized state data packet is completely copied to generate two identical copies (copy A and copy B); copy A is transmitted through a TCP / IP network channel, and copy B is transmitted through a power line carrier communication channel, with both channels starting transmission simultaneously.

[0204] Transmission rate range: TCP / IP channels are typically 10Mbps-10Gbps (limited by network bandwidth); power line carrier channels are typically 10kbps-200Mbps (affected by power line noise and transmission distance).

[0205] The cloud platform receives replicas from two channels respectively and performs a preliminary check on each replica, including data packet length verification: it checks whether the length of the received data packet is within a preset range (e.g., 128 bytes - 2048 bytes, depending on the data content definition). If it exceeds the range, it is determined to be invalid.

[0206] Identifier integrity check: Check whether the start identifier (such as a fixed 2 bytes "0xAA55") in the header and the end identifier (such as "0x55AA") in the tail of the data packet are complete. If they are missing, the packet is deemed invalid.

[0207] Simple checksum check: Calculate the sum of all data bytes in the data packet and compare it with the checksum field carried in the data packet. The two must be completely consistent (the error range is 0), otherwise it is determined to be invalid.

[0208] If the first valid copy (e.g., copy A) arrives, a preset time window is started to wait for the second copy (copy B).

[0209] Time window range: usually 50ms-500ms (set according to the maximum transmission delay difference between the two channels, for example, the window can be appropriately enlarged when the power line carrier channel has a high delay).

[0210] Result determination: If both copies in the window pass the preliminary check (both are valid), the process continues; otherwise, exception handling (such as re-requesting data) is triggered and the process terminates.

[0211] Verification identifier generation and comparison: Apply the same preset verification algorithm (such as CRC32, SHA-1) to valid copy A and copy B respectively to generate verification identifiers; taking CRC32 as an example: calculate the 32-bit CRC value for the complete data content of the copy (excluding the verification field itself) to obtain verification identifier A and verification identifier B, with a value range of 0-0xFFFFFFFF (32-bit unsigned integer).

[0212] Comparison rules: Verification identifier A and verification identifier B must be completely identical (the numerical difference is 0); otherwise, the data is considered inconsistent, and exception handling is triggered and the process terminates.

[0213] Based on a preset strategy, one of the two identical copies, A and B, is selected as the output for receiving verification data packets in the cloud:

[0214] Examples of selection strategies: First-come, first-served (selecting the first arriving replica); Channel priority (e.g., default TCP / IP channel priority); Random selection.

[0215] By employing dual-channel parallel transmission and redundant verification, data loss due to single-channel failure is avoided, improving transmission success rate. Multi-level verification (preliminary integrity + checksum comparison) ensures that received data is consistent with the original data, reducing the risk of transmission errors. An anomaly handling mechanism can promptly address issues such as channel delays and data corruption, ensuring process stability.

[0216] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0217] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0218] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A comprehensive online monitoring system for power cables, characterized in that, include: The data acquisition module is used to collect the time-domain waveform of the grounding current of power cables, the temperature field distribution of the joint, the partial discharge pulse sequence and environmental parameters, and obtain multiple types of raw data sets; The data processing module is used to input multiple types of raw data sets into the edge computing node, perform data preprocessing, and generate standardized state data packets; The data fusion module transmits standardized state data packets to the cloud platform via a dual-channel redundant network of TCP / IP and power line carrier. It uses a dynamic time warping algorithm to align the spatiotemporal dimensions of current, temperature, and partial discharge data, and fuses environmental parameters to generate a multi-dimensional feature fusion dataset. Specifically, this involves: simultaneously copying the standardized state data packets twice and transmitting them to the cloud platform in parallel via the TCP / IP network channel and the power line carrier communication channel, respectively. The cloud platform receives the data packet copies from both channels, denoted as copy A and copy B, and performs preliminary data integrity checks on each. Based on the status of copies A and B after preliminary verification, a preset time window is initiated for waiting. If both copies are valid within the window period, valid copies A and B are obtained; otherwise, exception handling is triggered and the process terminates. The same preset verification algorithm is applied to valid copies A and B to generate verification labels. The system identifies identifier A and verification identifier B, and compares them. If they match, it generates identical copies A and B; otherwise, it triggers data inconsistency handling and terminates the process. Based on the confirmed identical copies A and B, it selects one of the copies according to a preset selection strategy, generates and outputs a cloud-received verification data packet. The cloud-received verification data packet is processed to obtain current time series, temperature time series, and partial discharge pulse time series. A dynamic time warping algorithm and nonlinear time axis transformation are used to generate spatiotemporally aligned current, temperature, and partial discharge sequences. Based on the spatiotemporally aligned current, temperature, and partial discharge sequences, normalized environmental parameter data is extracted from the cloud-received verification data packet. The current, temperature, and partial discharge sequences and environmental parameter data are integrated by time point to generate a multi-dimensional feature fusion dataset. The fault diagnosis module is used to extract the partial discharge distribution point set from the multi-dimensional feature fusion dataset and calculate the weight factor of each distribution point. Based on the weight factor, a weighted point set concave hull generation algorithm is used to construct the boundary of the partial discharge source probability distribution and obtain the boundary feature point set. The boundary feature point set and the multi-dimensional feature fusion dataset are input into a convolutional-long short-term memory hybrid neural network to generate diagnostic results. The intelligent response module is used to generate a three-level early warning signal based on the diagnostic results, push the early warning information to the remote operation and maintenance terminal and mobile APP via the MQTT protocol, and automatically trigger the digital work order service to generate maintenance tasks and perform location verification.

2. The integrated online monitoring system for power cables according to claim 1, characterized in that, Multiple types of raw data sets are input into the edge computing nodes for data preprocessing to generate standardized state data packets, including: The time-domain waveform data of power cable grounding current from multiple types of raw datasets are processed to generate power frequency current waveform data; and temperature field data are generated by Kriging interpolation based on discrete point data of temperature field distribution of power cable joints from multiple types of raw datasets; the temperature field data are then calculated and processed to obtain temperature field analysis results. The partial discharge pulse sequence data from multiple types of raw datasets are processed to obtain a partial discharge pulse feature dataset; and the environmental parameters from multiple types of raw datasets are processed to generate effective environmental parameter data. The current waveform data after power frequency rejection, temperature field analysis results, partial discharge pulse characteristic dataset, and effective environmental parameter data are collected and processed to generate a multi-source state dataset; the multi-source state dataset is then normalized to obtain a standardized state data package.

3. The integrated online monitoring system for power cables according to claim 2, characterized in that, A set of partial discharge distribution points is extracted from the multi-dimensional feature fusion dataset, and the weight factor of each distribution point is calculated. Based on the weight factor, a weighted point set concave hull generation algorithm is used to construct the boundary of the partial discharge source probability distribution, obtaining a set of boundary feature points. The set of boundary feature points and the multi-dimensional feature fusion dataset are input into a convolutional-long short-term memory hybrid neural network to generate diagnostic results, including: The spatiotemporal distribution point set of partial discharge pulses is extracted from the multidimensional feature fusion dataset. Based on the spatiotemporal distribution point set of partial discharge pulses, the time aggregation weight factor of the timestamp of each point and the relative intensity weight factor of the amplitude of each point are calculated and the two are fused to generate a joint weight factor, thus obtaining the weighted partial discharge point set. The partial discharge distribution point set and the corresponding weight factor of each point are used to generate a set of convex boundary points that can surround all partial discharge distribution points by using a weighted point set concave hull generation algorithm. This process constructs the probability distribution boundary of the high-weight region and obtains the set of boundary feature points. The set of boundary feature points is mapped to generate a spatial probability matrix and fused with time series data. The matrix is ​​then input into a convolutional-LSTM hybrid neural network to extract spatiotemporal features and generate diagnostic results.

4. The integrated online monitoring system for power cables according to claim 3, characterized in that, Based on the diagnostic results, a three-level early warning signal is generated. The warning information is pushed to remote maintenance terminals and mobile apps via the MQTT protocol, automatically triggering the digital work order system to generate maintenance tasks and perform location verification, including: The cable insulation condition risk level value in the diagnostic results is analyzed; the risk level value is matched with the preset three-level warning threshold range to obtain the warning signal level; the spatial location coordinates and timestamp data of the partial discharge source in the diagnostic results are extracted; and the warning signal level, location coordinates, and timestamp are encapsulated into a structured MQTT protocol message. The structured MQTT protocol messages are pushed to remote operation and maintenance terminals and mobile APP in parallel through the cloud platform to provide early warning information; corresponding maintenance priority identifiers are generated according to the early warning signal level; the location coordinates, early warning signal level, and priority identifier are bound to generate structured maintenance instructions; and the digital work order service is automatically triggered to generate electronic maintenance work orders. The spatial positioning service of the on-site verification terminal is driven by the positioning coordinates in the maintenance work order to perform local source positioning and receive the feedback verification results.

5. The integrated online monitoring system for power cables according to claim 4, characterized in that, The received verification data packets from the cloud are processed to obtain current time series, temperature time series and partial discharge pulse time series; Furthermore, a dynamic time warping algorithm and nonlinear time axis transformation are employed to generate spatiotemporally aligned current sequences, temperature-aligned sequences, and partial discharge-aligned sequences, including: Based on receiving verification data packets in the cloud, performing predefined protocol parsing operations, and generating the original current time series, temperature time series, and partial discharge pulse time series; The temperature time series is designated as the reference series, and a dynamic time warping algorithm is applied between the current time series and the temperature time series. The nonlinear time warping path between the two series is calculated to generate the current-temperature warping mapping relationship. A dynamic time warping algorithm is applied between the partial discharge pulse time series and the temperature time series; the curved path is calculated to generate the partial discharge-temperature warping mapping relationship. Based on the extracted original time series and mapping relationship, a nonlinear time axis transformation is performed to generate a current-aligned sequence and a partial discharge-aligned sequence that are completely aligned with the temperature sequence time axis, and the original temperature sequence is used as the temperature-aligned sequence.

6. The integrated online monitoring system for power cables according to claim 5, characterized in that, The partial discharge distribution point set and its corresponding weight factor are used to generate a set of convex boundary points that can enclose all partial discharge distribution points using a weighted point set concave hull generation algorithm. This constructs the probability distribution boundary of the high-weight region, resulting in a set of boundary feature points, including: Receive the partial discharge distribution point set and the joint weight factor corresponding to each point; perform a descending sorting operation on the entire point set according to the weight factor value of each point; after sorting, the point with the largest weight factor is located at the beginning of the sequence, select the first three non-collinear points with the largest weight factors, calculate and construct an initial convex hull; Traverse the remaining point set in descending order of weight, determine the positional relationship between each point and the convex hull, and if it is outside, dynamically reconstruct the convex hull and update the boundary vertex set to complete the ordered iteration; After the ordered iteration is completed, the set of convex hull vertices formed by high-weight points is processed first based on the weight descending mechanism, and the set of boundary feature points representing the boundary of the probability distribution of the local discharge source is output.

7. A computing device, characterized in that, include: one or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the system as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the system as described in any one of claims 1 to 6.

Images