A cable branch box online monitoring method and device

Abstract

The application relates to the technical field of power equipment monitoring, in particular to a cable branch box online monitoring method and device, and aims to solve the problem that in the prior art, long-term exposure of equipment to outdoor installation environments results in that current, voltage and local discharge signals and other multi-source monitoring data generally have distortion phenomena, the local discharge signals are particularly susceptible to noise interference and are difficult to effectively identify, and thus fault feature extraction is difficult. The application obtains multi-source monitoring data and performs environment self-adaptive calibration, and then inputs a light mixed model to perform fault identification, solves the problem of data distortion caused by outdoor environment interference and the imbalance between cost and performance, and has the effects of improving monitoring accuracy, reducing deployment cost and enhancing real-time performance.

Description

Technical Field

[0001] This application relates to the field of power equipment monitoring technology, and more specifically, to an online monitoring method and device for cable branch boxes. Background Technology

[0002] Cable distribution boxes, as core facilities in the power distribution networks of small and medium-sized enterprises (SMEs), play a crucial role in power distribution and transmission. However, in practical applications, SMEs encounter multiple technical bottlenecks in the monitoring of cable distribution boxes. The outdoor installation environment exposes equipment to drastic fluctuations in temperature and humidity, as well as complex electromagnetic interference, leading to widespread distortion of multi-source monitoring data such as current, voltage, and partial discharge signals. Partial discharge signals are particularly susceptible to noise interference, making them difficult to identify effectively and consequently hindering fault feature extraction.

[0003] Meanwhile, monitoring systems face a severe imbalance between cost and performance: high-precision monitoring solutions rely on complex algorithms and high-performance hardware, resulting in high overall deployment costs that far exceed the economic affordability of small and medium-sized enterprises (SMEs); while economical solutions suffer from limited data processing capabilities, leading to significantly insufficient accuracy in fault warnings, frequent false alarms, and an inability to promptly mitigate the risk of sudden failures. Furthermore, SMEs generally lack professional algorithm maintenance teams, and existing monitoring systems are complex to operate in areas such as model updates and parameter debugging. Additionally, traditional processing models exhibit slow inference responses, making it difficult to meet the timeliness requirements of real-time warnings.

[0004] In current technical practices, although multi-source data fusion and lightweight models have their own applications, simply introducing multi-source data while ignoring dynamic environmental interference can lead to insufficient calibration and misjudgments. On the other hand, using only lightweight models while ignoring data quality can weaken feature recognition capabilities. Conventional combination methods, which involve simply splicing data and directly inputting it into the model, fail to effectively coordinate the contradiction between dynamic environmental changes and limited computing resources. They also lack a collaborative optimization mechanism for data calibration and model design, thus failing to meet the comprehensive requirements of low cost, high accuracy, and easy operation and maintenance.

[0005] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0006] (a) Technical problems to be solved The purpose of this application is to provide an online monitoring method and device for cable branch boxes, which has the effects of improving monitoring accuracy, reducing deployment costs, and enhancing real-time performance.

[0007] (II) Technical Solution This application provides an online monitoring method for cable branch boxes, the technical solution of which is as follows: Acquire multi-source monitoring data from the cable branch box. The multi-source monitoring data should include at least current, voltage, partial discharge signal, and environmental parameters. Environmental adaptive calibration is performed on multi-source monitoring data to obtain calibrated electrical parameter data. Environmental adaptive calibration includes first-level calibration based on dynamic adjustment of filter parameters based on environmental parameters, and second-level calibration based on the correlation of multi-source data. The calibrated electrical parameter data is input into the fault identification model for fault identification. The fault identification model is a lightweight hybrid model that combines attention mechanism and extreme learning machine. Based on the identification results, the early warning information is uploaded to the operation and maintenance platform via low-power communication.

[0008] Furthermore, this application also proposes a first-stage calibration method based on dynamically adjusting filter parameters according to environmental parameters, specifically including: The gain parameters of the Kalman filter are dynamically adjusted based on the real-time collected environmental parameters to filter current, voltage, and partial discharge signals.

[0009] Furthermore, this application proposes a second-level calibration based on the correlation of multi-source data, specifically including: The reliability of the partial discharge signal is determined based on the stability of current and voltage data, and the partial discharge signal is weighted and corrected based on the reliability weight.

[0010] Furthermore, this application proposes that in the fault identification model, an attention mechanism is used to assign weights to the input features, and an extreme learning machine is used to classify faults based on the weighted features.

[0011] Furthermore, this application proposes that the weights of the attention mechanism be used to initialize the input layer weight matrix of the extreme learning machine.

[0012] Furthermore, this application proposes that multi-source monitoring data be collected through the following sensors: current sensor, voltage sensor, partial discharge sensor, and temperature and humidity sensor, with the sensors clock-synchronized by a microcontroller.

[0013] Furthermore, this application also proposes that the low-power communication method is LoRa wireless communication, and the communication strategy includes timed data upload and fault-triggered real-time upload.

[0014] Furthermore, this application also proposes an online monitoring device for cable branch boxes, used to perform the above-mentioned method, comprising: The data acquisition module is used to acquire multi-source monitoring data of the cable branch box; The calibration module is used to perform environmental adaptive calibration on multi-source monitoring data to obtain calibrated electrical parameter data. The fault identification module includes a lightweight hybrid model that combines attention mechanism and extreme learning machine, used to identify faults in calibrated electrical parameter data; The communication module is used to upload early warning information based on the identification results; The operations and maintenance platform module is used to receive and display early warning information and provide model update functions.

[0015] Furthermore, this application also proposes that the calibration module includes: The first-level calibration unit is used to dynamically adjust the filter parameters based on environmental parameters; The second-level calibration unit is used to perform confidence-weighted correction on the partial discharge signal based on the stability of current and voltage data.

[0016] (III) Beneficial Effects Compared with the prior art, the beneficial effects of the present invention are as follows: This invention solves the problems of data distortion and cost-performance imbalance caused by outdoor environmental interference by acquiring multi-source monitoring data, performing environmental adaptive calibration, and then inputting it into a lightweight hybrid model for fault identification. It has the effects of improving monitoring accuracy, reducing deployment costs, and enhancing real-time performance. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a logical diagram of the online monitoring method for cable branch boxes. Detailed Implementation

[0019] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0020] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance. Example

[0021] In traditional online monitoring of cable branch boxes, changes in environmental parameters and electromagnetic noise interference lead to distortion of multi-source monitoring data, making it difficult to accurately extract fault characteristics. Furthermore, current, voltage, and partial discharge signals collected by single-source sensors are susceptible to interference, affecting the reliability of fault diagnosis. Moreover, existing monitoring systems suffer from significant problems: complex solutions have high hardware costs and require high-performance computing resources, while low-cost solutions have low fault warning accuracy and high false alarm rates. Simultaneously, the maintenance of complex models requires professional personnel for model updates and parameter adjustments, resulting in large inference delays and failing to meet real-time early warning requirements.

[0022] For example, in a monitoring scenario of outdoor cable distribution boxes in an industrial park, when the ambient temperature and humidity fluctuate, the partial discharge signal is distorted by electromagnetic noise interference, and traditional single-source sensors cannot distinguish between real fault signals and noise interference. Furthermore, in existing monitoring systems, complex solutions cannot be fully deployed due to high hardware costs, while low-cost solutions generate frequent false alarms due to insufficient fault identification accuracy, requiring maintenance personnel to conduct multiple on-site verifications, increasing the operational burden. In addition, the model inference delay is large, making it unable to respond promptly to sudden faults, affecting the system's real-time performance and stability.

[0023] If these issues are not addressed, the monitoring system will be unable to effectively identify cable distribution box faults, leading to increased operational risks in the power distribution network. Small and medium-sized enterprises may face equipment damage and production interruptions. Consequently, high maintenance costs and complex operations will limit the widespread application of monitoring technology, impacting the safety and reliability of the power distribution network.

[0024] In this regard, such as Figure 1 As shown, an online monitoring method for cable branch boxes is proposed, including the following steps: S100. Acquire multi-source monitoring data of the cable branch box. The multi-source monitoring data shall include at least current, voltage, partial discharge signal and environmental parameters. S200. Perform environmental adaptive calibration on multi-source monitoring data to obtain calibrated electrical parameter data. Environmental adaptive calibration includes first-level calibration based on dynamic adjustment of filter parameters based on environmental parameters, and second-level calibration based on the correlation of multi-source data. S300. Input the calibrated electrical parameter data into the fault identification model for fault identification. The fault identification model is a lightweight hybrid model that combines attention mechanism and extreme learning machine. Based on the identification results, the S400 uploads the early warning information to the operation and maintenance platform via low-power communication.

[0025] For ease of understanding, the following explains some key terms in this embodiment: A cable distribution box is a power distribution device used for power distribution and transmission. It is usually installed outdoors and is susceptible to environmental factors. Internally, it contains functional units for cable connection, protection, and monitoring.

[0026] Multi-source monitoring data refers to data on various physical quantities acquired from different types of sensors, such as current, voltage, partial discharge signals, and environmental parameters. This data is used together to comprehensively assess the operational status of the cable distribution box.

[0027] Environmental parameters refer to physical quantities that affect the operating environment of a cable distribution box, such as temperature and humidity. Changes in these parameters may cause fluctuations or distortions in the electrical signals inside the cable distribution box.

[0028] Environmental adaptive calibration is a data processing method that dynamically adjusts the calibration strategy based on real-time environmental conditions to eliminate or reduce the impact of environmental factors on monitoring data, thereby improving the accuracy and reliability of the data.

[0029] The first-level calibration is a stage of environmental adaptive calibration, which mainly adjusts the parameters of data filtering dynamically based on changes in environmental parameters to adapt to instantaneous environmental interference.

[0030] The second-level calibration is another stage of environmental adaptive calibration. It utilizes the inherent correlation between different monitoring data to cross-validate and correct the data, further improving the reliability of the data.

[0031] The calibrated electrical parameter data refers to the electrical quantity data such as current, voltage, and partial discharge after environmental adaptive calibration. These data have higher accuracy and anti-interference ability, and are more suitable for fault identification.

[0032] The fault identification model is a machine learning model that is trained to analyze calibrated electrical parameter data and identify different types of cable branch box fault modes.

[0033] Attention mechanisms are neural network components that assign different weights to different parts of the input data, allowing the model to focus on more important features when processing the data.

[0034] Extreme Learning Machine (ELM) is a type of feedforward neural network that features fast learning speed and good generalization performance. It is often used for classification and regression tasks. Its input layer weights and hidden layer biases are randomly generated, and the output layer weights are calculated in one step.

[0035] Lightweight hybrid models refer to fault identification models that combine multiple technologies (such as attention mechanisms and extreme learning machines) and have low computational resource requirements. They aim to reduce the requirements for hardware computing power while ensuring identification performance.

[0036] Low-power communication refers to communication technologies that consume less power during data transmission, such as LoRa and NB-IoT, which are suitable for battery-powered applications or applications with strict energy consumption requirements.

[0037] The operation and maintenance platform is a software system that integrates data reception, storage, analysis, display and management. It is used to remotely monitor the operating status of cable branch boxes, receive early warning information, and support the management and maintenance of the monitoring system.

[0038] This embodiment provides an online monitoring method for cable branch boxes, which is characterized by the following aspects: First, acquire multi-source monitoring data from the cable distribution box. In practical applications, data can be collected by deploying various sensors inside or near the cable distribution box. For example, current transformers and voltage transformers can be installed to acquire current and voltage data; ultrasonic or ultra-high frequency sensors can be installed to acquire partial discharge signals; and temperature and humidity sensors can be installed to acquire environmental parameters. These sensors can operate independently, transmitting their collected data to a central processing unit. While this method can acquire multi-dimensional data, it may suffer from inconsistencies in data acquisition times and data formats between different sensors, requiring subsequent processing to ensure data consistency.

[0039] Secondly, environmental adaptive calibration is performed on the multi-source monitoring data to obtain calibrated electrical parameter data. The calibration process includes two stages: first-level calibration and second-level calibration. In the first-level calibration, a basic digital filter, such as a moving average filter or a Butterworth filter, can be used, and its filtering parameters can be adjusted according to preset environmental parameter thresholds. For example, when the ambient temperature exceeds a certain threshold, the filter's cutoff frequency can be adjusted to filter out noise within a specific frequency range.

[0040] This method of adjusting filter parameters based on environmental parameters can initially address interference caused by environmental changes, but its adjustment strategy may be relatively simple and unable to finely adapt to complex dynamic environments. In the second-level calibration, simple cross-comparisons can be performed on different types of data. For example, when a local discharge signal becomes abnormal, the simultaneous current and voltage data can be checked for drastic fluctuations. If the current and voltage data also show drastic fluctuations, the anomaly in the partial discharge signal may be related to external power grid disturbances rather than internal equipment faults. This simple comparison can provide a preliminary assessment of the validity of the partial discharge signal, but this method may lack quantitative evidence and make it difficult to accurately evaluate the signal's reliability.

[0041] Next, the calibrated electrical parameter data is input into a fault identification model for fault identification. This fault identification model is designed as a lightweight hybrid model combining an attention mechanism and an extreme learning machine (ELM). Specifically, an attention mechanism is first used to extract features from the input calibrated electrical parameter data. The attention mechanism assigns weights to different features based on the importance of the data; for example, when identifying partial discharge faults, features of partial discharge signals may be given higher weights. Then, the weighted features processed by the attention mechanism are directly input into an ELM for fault classification. The ELM can quickly learn and output the identification results. This combination can improve identification efficiency and accuracy to a certain extent while maintaining the model's lightweight nature. However, the information transfer and fusion between the attention mechanism and the ELM may be too direct, failing to fully leverage the potential of their synergy.

[0042] Finally, based on the identification results, the warning information is uploaded to the operation and maintenance platform via low-power communication. When the fault identification model determines that a potential fault or anomaly exists, the system generates corresponding warning information. This warning information can be transmitted through a low-power wireless communication module, such as a GPRS or 2G network-based communication module, which can achieve low power consumption even when transmitting small amounts of data. The warning information is sent to the remote operation and maintenance platform for operation and maintenance personnel to view and handle. This communication method can meet the basic requirements for warning information transmission, but its communication efficiency and real-time performance may be limited by network coverage and bandwidth, and in some extreme cases, there may still be room for improvement in power consumption optimization.

[0043] The following example will provide a more detailed explanation of the above technical solution: Suppose that an early partial discharge fault is occurring inside an outdoor cable branch box at location A. However, due to the complex environment of the area, large fluctuations in temperature and humidity, and strong electromagnetic interference, traditional monitoring methods are unable to accurately identify the fault.

[0044] First, the monitoring system acquires multi-source monitoring data from the cable distribution box. Specifically, current sensors, voltage sensors, partial discharge sensors, and temperature and humidity sensors collect data in real time. For example, at a certain point in time, the temperature and humidity sensor might detect an ambient temperature of -10°C and humidity of 90%, while simultaneously collecting current, voltage, and partial discharge signals. This multi-source data is transmitted to the processing unit, providing comprehensive information for subsequent calibration and identification.

[0045] Next, environmental adaptive calibration is performed on these multi-source monitoring data. In the first-level calibration, the processing unit dynamically adjusts the parameters of the preset digital filter based on the real-time collected temperature and humidity parameters of -10℃ and 90%. For example, when the temperature is low and the humidity is high, the system will determine that the environmental noise may be enhanced, and therefore adjust the cutoff frequency or gain of the filter to more effectively filter out transient noise caused by the low temperature and high humidity environment, such as slight drift of voltage signals or background noise of partial discharge signals.

[0046] Subsequently, in the second-level calibration, the processing unit analyzes the stability of the calibrated current and voltage data. If the current and voltage data fluctuate within the normal range, but the partial discharge signal shows abnormal pulses, the system considers the partial discharge signal to be highly reliable because it is unlikely to be caused by external power grid fluctuations. Through this two-level calibration, the distortion caused by environmental interference and noise in the original data is significantly reduced, resulting in cleaner and more reliable electrical parameter data.

[0047] Then, the calibrated electrical parameter data are input into the fault identification model for fault identification. This model is a lightweight hybrid model that combines an attention mechanism and an extreme learning machine. The calibrated current, voltage, and partial discharge data are input into the model.

[0048] The attention mechanism automatically identifies and focuses on the features most relevant to partial discharge faults based on the characteristics of these data, such as the amplitude and frequency of partial discharge pulses and their relationship with voltage phase, and assigns higher weights to these key features.

[0049] Subsequently, the weighted features are input into the Extreme Learning Machine (ELM). Leveraging its rapid learning capabilities, the ELM performs efficient fault classification based on these weighted features, quickly identifying an "early partial discharge fault" in the current cable branch box.

[0050] Finally, based on the identification results, the early warning information is uploaded to the operation and maintenance platform via low-power communication. Once the model identifies an "early partial discharge fault," the system immediately generates an early warning message, such as "An early partial discharge fault has been detected in cable branch box A at location A; immediate investigation is recommended." This warning message is sent to the remote operation and maintenance platform via a low-power wireless communication module (e.g., a 2G network-based module). Upon receiving the information, the operation and maintenance platform immediately displays the warning on its interface and may notify operation and maintenance personnel via SMS or email. In this way, operation and maintenance personnel can obtain fault information in a timely manner, arrange maintenance, and thus prevent the fault from worsening and reduce potential economic losses.

[0051] Based on the above examples, the technical concept of this embodiment demonstrates a significant technical contribution. Traditional cable distribution box monitoring methods often suffer from low fault identification accuracy due to data distortion in complex outdoor environments like Location A. For example, if only single current or voltage monitoring is used, drastic fluctuations in ambient temperature and humidity can directly cause signal drift, completely masking weak fault signals such as early partial discharges, making them impossible to effectively detect. This embodiment, by acquiring multi-source monitoring data, particularly incorporating environmental parameters into the monitoring scope, provides crucial basis for subsequent data calibration, fundamentally improving the comprehensiveness of data input.

[0052] Furthermore, the environmental adaptive calibration mechanism employed in this embodiment, particularly its two-stage calibration design, represents a significant advancement in the prior art. Traditional monitoring systems may use filters with fixed parameters, which cannot dynamically adapt to environmental changes. For example, in the above example, if a fixed-parameter filter is used, it may fail to effectively filter out the transient noise generated when the ambient temperature drops sharply, resulting in the calibrated data still containing a large amount of interference.

[0053] The first-level calibration in this embodiment can dynamically adjust the filtering strategy according to real-time environmental parameters, effectively dealing with transient environmental interference. The second-level calibration utilizes the correlation between multi-source data for cross-validation, such as judging the reliability of partial discharge signals by the stability of current and voltage, which is lacking in existing technologies. This two-level collaborative calibration results in higher purity and reliability of the calibrated electrical parameter data, significantly reducing data distortion and providing high-quality input for subsequent fault identification.

[0054] Furthermore, the fault identification model in this embodiment, namely the lightweight hybrid model combining attention mechanisms and extreme learning machines, demonstrates superiority in addressing the problem of "imbalanced cost-effectiveness of monitoring systems" in existing technologies. While traditional complex deep learning models (such as CNN and GNN) offer high recognition accuracy, they require significant hardware computing power, are expensive, and have long inference latency, making them unsuitable for small and medium-sized enterprises. On the other hand, a single extreme learning machine model may suffer from low accuracy when faced with complex fault characteristics due to insufficient feature capture.

[0055] This embodiment uses an attention mechanism to assign weights to key features, enabling the Extreme Learning Machine to more effectively focus on fault-related information. This significantly improves the accuracy and response speed of fault identification while maintaining a lightweight model. In the example above, the model can quickly and accurately identify early partial discharge faults, with a response time far lower than the latency of traditional complex models, meeting the requirements for real-time early warning.

[0056] Finally, this embodiment employs a low-power communication method to upload early warning information to the operation and maintenance platform, effectively solving the problem of "mismatch between operation and maintenance difficulty and computing power" in existing technologies. Traditional communication methods may consume high power, increasing operation and maintenance costs, and the operation and maintenance platform may require complex operations by professional personnel. The low-power communication design of this embodiment reduces system energy consumption and extends device battery life. Simultaneously, early warning information is directly uploaded to the operation and maintenance platform, simplifying the information transmission process and enabling operation and maintenance personnel to promptly obtain and process fault information through an easy-to-use platform, eliminating the need for professional algorithm operation and maintenance personnel to perform complex model updates or parameter debugging. This end-to-end collaborative design enables the entire monitoring system to achieve high accuracy and fast response while also possessing low cost and easy operation and maintenance characteristics, providing a practical solution for small and medium-sized enterprises.

[0057] In some of the embodiments described above in this application, a first-level calibration based on dynamically adjusting the filter parameters according to environmental parameters is proposed to reduce the impact of environmental interference on monitoring data. However, in this process, due to the drastic fluctuations in environmental parameters, the adjustment of filter parameters lacks a specific real-time mechanism and adaptive optimization, which may lead to unstable filtering effect, incomplete data distortion problem, and affect the accuracy of subsequent calibration.

[0058] In response, this application further proposes a specific implementation method for the first-level calibration, including dynamically adjusting the gain parameters of the Kalman filter based on the real-time acquired environmental parameters, and filtering the current, voltage, and partial discharge signals.

[0059] Among them, the environmental parameters collected in real time refer to the external conditions that affect the accuracy of the monitoring data of the cable branch box, such as temperature, humidity, and air pressure. The real-time acquisition of these parameters can be achieved by various environmental sensors deployed near the cable branch box. For example, high-precision temperature and humidity sensors can be used to obtain ambient temperature and relative humidity data in real time, or air pressure sensors can be used to obtain atmospheric pressure data.

[0060] These sensors are typically connected to microcontrollers to sample data periodically or in an event-driven manner, ensuring the immediacy of environmental information. Dynamically adjusting the Kalman filter gain parameter means that this parameter is not fixed but modified in real time according to changes in external conditions or system state. For example, the gain parameter can be calculated and updated based on the trend or magnitude of changes in environmental parameters using a pre-defined lookup table, empirical formula, or adaptive algorithm.

[0061] Another approach is to dynamically optimize the gain parameters by outputting corresponding gain adjustment factors through a fuzzy logic controller or neural network model based on the real-time values ​​of environmental parameters. The purpose of filtering current, voltage, and partial discharge signals is to remove noise components from these original signals and extract the true, useful signal features. Besides Kalman filtering, other adaptive filtering algorithms can be used, such as the Least Mean Square (LMS) algorithm or the Recursive Least Squares (RLS) algorithm. These algorithms can also dynamically adjust the filtering coefficients according to signal characteristics and noise levels to achieve better noise reduction.

[0062] This solution acquires real-time environmental parameters of the cable distribution box and dynamically adjusts the gain parameters of the Kalman filter accordingly, enabling the filtering process to adaptively respond to environmental changes. Specifically, when environmental parameters (such as temperature and humidity) fluctuate, the system precisely adjusts the gain parameters of the Kalman filter based on the changing trends of these real-time parameters or a preset mapping relationship.

[0063] This dynamic adjustment ensures that the filter maintains optimal filtering performance under different environmental conditions, avoiding the over-filtering or under-filtering problems that may occur with fixed-parameter filters when the environment changes drastically. Subsequently, the dynamically gain-adjusted Kalman filter is applied to current, voltage, and partial discharge signals, effectively suppressing environmental noise interference with these key electrical parameters, thereby obtaining cleaner and more reliable monitoring data. This first-level calibration mechanism provides high-quality input data for subsequent second-level calibration (based on the correlation of multi-source data) and fault identification models, significantly improving the accuracy and robustness of the entire online monitoring method for cable branch boxes, and effectively solving the problems of data distortion and unstable filtering effect caused by environmental interference.

[0064] As a specific implementation method, real-time environmental parameters, such as ambient temperature and relative humidity, can be continuously acquired using temperature and humidity sensors integrated within the cable distribution box. These acquired environmental parameters can be used as inputs to dynamically adjust the gain parameters of the Kalman filter through a preset lookup table or piecewise function. For example, an environmental correction function can be set, which outputs an adjustment factor based on different ranges of real-time temperature and relative humidity.

[0065] When the ambient temperature is below 0℃ and the relative humidity is below 70%, the adjustment factor may be smaller, making the Kalman filter more inclined to trust the predicted state in order to cope with sensor drift that may be caused by low temperature. When the ambient temperature is above 30℃ and the relative humidity is above 80%, the adjustment factor may be larger, making the filter more focused on the measurement data, while enhancing the ability to suppress electromagnetic noise in partial discharge signals under high humidity conditions.

[0066] This adjustment factor is then multiplied or added to the Kalman filter's reference gain parameter to form a real-time updated gain parameter. This dynamic gain parameter is used to perform Kalman filtering on the raw signals acquired by current sensors, voltage sensors, and partial discharge sensors, thereby effectively removing noise and obtaining high-quality electrical parameter data under different environmental conditions.

[0067] The above technical solution significantly improves the stability of the filtering effect, ensuring that monitoring data remains highly accurate even under drastic environmental fluctuations. Simultaneously, this solution can accurately eliminate noise for different types of environmental interference, effectively reducing the data distortion rate of current, voltage, and partial discharge signals. This provides a more reliable and accurate data foundation for subsequent fault identification, thereby enhancing the overall performance and early warning capabilities of the cable branch box online monitoring system.

[0068] In some of the solutions described above in this application, a second-level calibration based on the correlation of multi-source data is proposed to improve data quality. However, in this process, there is a lack of an effective evaluation mechanism for the reliability of partial discharge signals, resulting in unsatisfactory calibration results, inability to accurately suppress noise interference, and consequently affecting the accuracy of fault identification.

[0069] In response, this application further proposes a second-level calibration based on the correlation of multi-source data, specifically including: judging the credibility of the partial discharge signal based on the stability of current and voltage data, and performing weighted correction on the partial discharge signal based on credibility weight.

[0070] The reliability of partial discharge signals is assessed based on the stability of current and voltage data. This aims to evaluate the reliability of partial discharge signals by utilizing the inherent stability characteristics of current and voltage data. Environmental interference typically affects multiple electrical parameters simultaneously, but current and voltage, as fundamental parameters, can indirectly reflect the intensity of environmental interference and its impact on the integrity of partial discharge signals through their stability changes. By evaluating the stability of these fundamental parameters, it is possible to effectively distinguish between genuine partial discharge events and spurious signals caused by noise.

[0071] One approach is to calculate the statistical dispersion indices of current and voltage data within a specific time window, such as standard deviation, variance, or coefficient of variation. When these indices are below a preset threshold, the current and voltage data are considered to be in a stable state, thus indicating a high degree of reliability for the partial discharge signal; conversely, the reliability is considered low. Another approach is to use frequency analysis methods to perform Fourier transforms on the current and voltage data and analyze their spectral characteristics. When the main frequency components are concentrated and their amplitudes are stable, it indicates good data stability and a high degree of reliability for the partial discharge signal; if the spectral distribution is broad or abnormal harmonics are present, it indicates data instability and a low degree of reliability for the partial discharge signal.

[0072] The partial discharge signal is weighted and corrected based on confidence level, aiming to dynamically adjust its influence in data processing according to the assessed confidence level of the partial discharge signal. This adaptive weighting mechanism is crucial for improving the signal-to-noise ratio. It ensures that high-reliability signals dominate subsequent analysis, while low-reliability signals contaminated by noise are effectively attenuated or corrected, thereby providing clearer and more accurate data for fault identification.

[0073] One approach is to pre-set different fixed weight values ​​based on the determined confidence level (e.g., high, medium, low). For example, high confidence might correspond to a weight of 0.9, medium confidence to 0.6, and low confidence to 0.3. The partial discharge signal is then multiplied by the corresponding weight for correction. Another approach is to construct a weight generator based on fuzzy logic or a neural network. This generator takes the stability index of current and voltage data as input and outputs a continuous confidence weight value. This weight value can be smoothly adjusted according to subtle changes in the input data, thus achieving more refined weighted correction.

[0074] The solution of this application first acquires multi-source monitoring data from the cable branch box, which includes at least current, voltage, partial discharge signals, and environmental parameters. When performing environmental adaptive calibration on this data, a first-level calibration is first performed, i.e., the filtering parameters are dynamically adjusted based on the environmental parameters to filter the current, voltage, and partial discharge signals.

[0075] Building upon this foundation, this application further introduces a second-level calibration, the core of which lies in the refined processing of partial discharge signals. Because partial discharge signals are highly sensitive to environmental noise, their raw data may exhibit significant distortion. To address this issue, this scheme indirectly assesses the reliability of partial discharge signals under the current environment by analyzing the stability of relatively stable current and voltage data. When the current and voltage data exhibit high stability, it indicates minimal environmental interference, and the reliability of the partial discharge signal is high; conversely, if the current and voltage data fluctuate significantly, the reliability of the partial discharge signal is low.

[0076] Based on this credibility assessment, the system dynamically generates a credibility weight and uses this weight to perform weighted correction on the partial discharge signal. This correction mechanism ensures that the original characteristics of the partial discharge signal are preserved to the greatest extent when the data quality is good; while when the data quality is poor, noise interference is effectively suppressed by reducing its weight and combining it with other reference information. After two levels of calibration, the quality of the electrical parameter data is significantly improved, which can more accurately reflect the operating status of the cable branch box, thereby providing high-quality input for the subsequent fault identification model and ultimately improving the accuracy and reliability of the entire online monitoring method.

[0077] The following is a specific example. As a concrete implementation, during online monitoring of cable branch boxes, the system can continuously collect environmental parameters such as current, voltage, partial discharge signals, and temperature and humidity. During the second-level calibration, the system can set a fixed time window, such as 100 sampling points, and calculate the relative fluctuation coefficients of the current and voltage data within this window. For example, when the calculated When the value is less than or equal to 5%, the system can determine that the current and voltage data are in a "highly stable" state; when Values ​​between 5% and 15% are considered "moderately stable"; when... A value greater than 15% is considered "unstable". Based on this stability assessment, the system can assign dynamic confidence weights to partial discharge signals. For example, in a "highly stable" state, It can be set to 0.9 to preserve the original characteristics of the partial discharge signal to the greatest extent; in the "moderately stable" state, Linear adjustment can be made based on the S value, for example, using the formula. Perform the calculation.

[0078] In an "unstable" state, This can be set to 0.3 to significantly reduce the weight of the partial discharge signal and rely more on other reference information for correction. Subsequently, the system can employ a Bayesian weighted fusion model to process the partial discharge signal after the first-stage filtering. With a reference discharge value derived from current or voltage. To achieve fusion, for example, the reference discharge value can be determined using an empirical formula. The calculation yielded, where This represents the peak voltage. Finally, the corrected partial discharge signal... It can be represented as .

[0079] In this way, the system can adaptively adjust the degree of correction of the partial discharge signal based on the real-time stability of current and voltage data, ensuring that the output partial discharge data has higher reliability and accuracy under different environmental interferences.

[0080] Through the above technical solution, this application effectively solves the problem of insufficient reliability assessment of partial discharge signals in traditional monitoring methods. By introducing the stability of current and voltage data as the basis for evaluating the reliability of partial discharge signals, and performing dynamic weighted correction based on this, the calibration accuracy and anti-interference capability of partial discharge signals are significantly improved. This enables the system to more accurately identify and retain true partial discharge characteristics in complex and variable environments, while effectively suppressing noise and spurious signals. The data after this refined calibration provides a more reliable and high-quality input for subsequent fault identification models, thereby significantly improving the accuracy of fault identification, reducing the risk of false alarms and missed alarms, and providing more stable and reliable technical support for online monitoring of cable branch boxes.

[0081] In some of the solutions described above in this application, a fault identification model is proposed for fault identification. However, in this process, due to the lack of weight allocation for the input features, the model may not be able to effectively distinguish between key fault features and environmental noise features, resulting in insufficient feature capture and thus reducing classification accuracy. To address this, this application proposes a fault identification model in which an attention mechanism is used to allocate weights to the input features, and an extreme learning machine is used to classify faults based on the weighted features.

[0082] Specifically, the attention mechanism is a mechanism that simulates human cognitive attention. Its core lies in dynamically focusing on the more important parts of the input data and assigning them higher weights. One implementation is additive attention, which uses a feedforward neural network to calculate the compatibility between the query and the key, generating an attention score. Another implementation is multiplicative attention, which calculates the similarity between the query and the key through a dot product, generating an attention score. Its function is to identify and highlight key information in the input features while suppressing irrelevant or noisy information. Weight assignment refers to assigning a numerical value to each feature based on its importance to the target task (such as fault identification), reflecting its influence in the decision-making process. One implementation is to normalize the raw scores calculated by the attention mechanism to a probability distribution between 0 and 1 using the softmax function, using these values ​​as weights. Another implementation is to use the sigmoid function or other normalization methods to map the attention scores to a specific range, using these values ​​as weights. Through weight assignment, the model can focus on features that contribute more to fault identification and reduce noise interference.

[0083] Extreme Learning Machine (ELM) is a single-hidden-layer feedforward neural network where the connection weights from the input layer to the hidden layer and the hidden layer biases are randomly generated and do not need to be adjusted during training; only the output layer weights need to be calculated. One implementation approach is to use the sigmoid function as the hidden layer activation function.

[0084] As another implementation, other nonlinear activation functions such as ReLU and Tanh can also be used as hidden layer activation functions. Extreme Learning Machines provide an efficient and fast classifier, especially suitable for processing high-quality preprocessed feature data. Fault classification refers to dividing the operating status of cable branch boxes into different categories based on input features, such as normal, insulation fault, poor contact, and overload.

[0085] One implementation involves mapping weighted features to predefined fault categories using the output layer of an Extreme Learning Machine (ELM). Another approach combines an ELM with a multi-classifier (such as a one-to-many or one-to-one strategy). Its purpose is to identify the specific fault type of the cable branch box, providing a basis for subsequent early warning and maintenance.

[0086] The proposed solution combines an attention mechanism with an extreme learning machine to construct an efficient and accurate fault identification process.

[0087] First, the multi-source monitoring data acquired from the cable branch box undergoes environmental adaptive calibration to obtain calibrated electrical parameter data, which serves as input to the fault identification model. An attention mechanism receives this calibrated electrical parameter data and dynamically evaluates the importance of each feature (such as current, voltage, partial discharge signal, and environmental parameters) to the current fault identification task.

[0088] The attention mechanism assigns higher weights to key fault features while reducing or suppressing the weights of environmental noise and redundant features, thus forming a set of weighted features. These weighted features are then fed into an Extreme Learning Machine (ELM). Due to its fast training speed and strong generalization ability, the ELM can efficiently learn the complex mapping relationships between these attention-optimized weighted features and various fault types. Finally, based on the learned patterns, the ELM outputs the fault classification results for the cable branch box. This approach allows the ELM to receive high-quality, high-value input features, thereby significantly improving the accuracy and robustness of fault identification while maintaining its lightweight and high efficiency. This effectively solves the technical problem of how to effectively distinguish key fault features from environmental noise features and improve classification accuracy under limited computing power.

[0089] As a specific implementation method, in the online monitoring method for cable branch boxes, the calibrated electrical parameter data can include calibrated current amplitude, effective voltage value, partial discharge pulse count, partial discharge energy, and calibrated temperature and humidity features. The attention mechanism can be implemented using an additive attention module based on a multilayer perceptron (MLP). This module receives the aforementioned calibrated electrical parameter data as input and dynamically assigns an attention weight to each input feature by learning a weight vector. For example, when the system detects an abnormal partial discharge signal, the attention mechanism can assign higher weights to features related to partial discharge, while when the ambient temperature changes drastically, the weights of environmental parameters may be appropriately reduced to avoid misjudgment.

[0090] The Extreme Learning Machine (ELM) takes the weighted feature vectors, processed by an attention mechanism, as input. The number of hidden layer nodes in the ELM can be preset, for example, to 500, and the sigmoid function can be used as the hidden layer activation function. The ELM achieves rapid classification of fault types by randomly initializing the input layer weights and hidden layer biases, and then calculating the output layer weights using a generalized inverse matrix. The output layer of the ELM can be configured to correspond to different fault categories, such as "normal operation," "insulation aging," "poor contact," and "overload." After processing the weighted features, the ELM's output indicates the most likely operating state or fault type of the current cable branch box.

[0091] The above technical solution solves the problem that in the fault identification process, the model cannot effectively distinguish between key fault features and environmental noise features due to the lack of weight allocation for input features, resulting in insufficient feature capture and reduced classification accuracy.

[0092] By introducing an attention mechanism to weight input features, the importance of features can be dynamically evaluated, assigning higher weights to key fault features and suppressing noisy features, thereby optimizing feature representation. Extreme Learning Machines (ELMs) classify faults based on these weighted features, enabling more accurate learning of fault patterns. This is because the weighted features reduce redundant interference, improving classification accuracy while maintaining the model's lightweight nature. Combined with environmental adaptive calibration of multi-source monitoring data, the data input to the fault identification model inherently possesses high quality and reliability. Furthermore, the attention mechanism further refines this data, ensuring that the ELM receives the most representative fault features after dual optimization. This combination significantly improves the accuracy and robustness of fault identification, especially in complex outdoor environments and variable fault scenarios, effectively avoiding false alarms and missed alarms. It provides a high-precision, high-efficiency, and easy-to-deploy solution for online monitoring of cable distribution boxes.

[0093] In some of the solutions described above in this application, an attention mechanism is used to assign weights to input features in the fault identification model, and then an Extreme Learning Machine (ELM) performs fault classification based on the weighted features. However, in this process, the input layer weight matrix of the ELM is usually randomly initialized, which may lead to unstable model training, slow convergence speed, and affect the accuracy and efficiency of fault identification. To address this, this application proposes an improved solution: the weights of the attention mechanism are used to initialize the input layer weight matrix of the ELM.

[0094] Specifically, the weights in an attention mechanism are numerical representations that measure the importance or relevance of input features. These weights are typically learned through neural network layers and reflect the degree to which different features contribute to the model's output. For example, these weights can be a vector, where each element corresponds to the importance score of an input feature; or they can be a matrix representing the relationships between different feature dimensions. In one implementation, the attention mechanism generates weights by calculating the similarity between the input features and the query vector, and then normalizes them using a softmax function.

[0095] In another implementation, the attention mechanism can employ multi-head attention, computing weights in parallel across multiple attention subspaces to capture feature correlations at different levels. Initialization refers to the process of assigning initial values ​​to the model's parameters before training begins. Proper initialization helps the model converge to the optimal solution faster and avoids gradient vanishing or exploding problems during training. One initialization method is random initialization, such as randomly sampling values ​​from a uniform or Gaussian distribution.

[0096] Another initialization method is to initialize using the weights of a pre-trained model, i.e., transfer learning. Additionally, heuristic initialization methods, such as Xavier initialization or He initialization, can be used, which adjust the variance of the weights based on the input-output dimensions of the network layers. The input layer weight matrix of an Extreme Learning Machine (ELM) is the set of weights connecting input layer neurons to hidden layer neurons. Each element of this matrix represents the connection strength between an input feature and a hidden layer neuron. In ELMs, the input layer weight matrix is ​​typically randomly generated and remains unchanged during training. The dimension of this matrix is ​​determined by the number of input features and the number of hidden layer neurons. For example, this matrix can be a fully connected matrix, where every input feature is connected to all hidden layer neurons; or, in some variants, it can be a sparse matrix, where only a subset of input features are connected to hidden layer neurons.

[0097] The fault identification model in this application combines an attention mechanism and an extreme learning machine (ELM). The attention mechanism assigns weights to the input electrical parameter data features to highlight those important for fault identification, while the ELM classifies faults based on these weighted features. To address the instability and slow convergence that may result from random initialization of the ELM's input layer weight matrix, this application proposes an improved initialization strategy. Specifically, in the early stages of model training, the importance of the input electrical parameter data features is first evaluated using the attention mechanism, generating attention weights that reflect the contribution of each feature.

[0098] These attention weights do not directly replace the input layer weight matrix of the Extreme Learning Machine (ELM), but are cleverly used to guide the initialization process of that matrix. By transforming the attention weights into a diagonal matrix and fusing them with a randomly generated weight matrix, the initial input layer weight matrix of the ELM is formed. This fusion initialization method imprints the ELM with the importance of features from the very beginning of training; that is, the connection weights corresponding to key fault features are given higher influence in the initial stage, while the connection weights of noisy or unimportant features are suppressed. This mechanism avoids the blindness of traditional random initialization, allowing the model to focus on the most effective features for fault identification in the early stages of training, thus significantly reducing the time and number of iterations required for the model to explore the optimal weights and accelerating the convergence process. Simultaneously, since elements of the random weight matrix are still retained in the fusion initialization, this helps maintain the model's generalization ability, enabling it not only to identify known fault patterns but also to adapt to new and unforeseen fault scenarios. Therefore, this scheme effectively improves the training efficiency, stability, and accuracy of the fault identification model while maintaining the model's lightweight characteristics.

[0099] The following is a concrete example. In the online monitoring method for cable branch boxes, when calibrated electrical parameter data (e.g., features including current, voltage, amplitude and frequency of partial discharge signals) are input into the fault identification model, these features are first processed by an attention mechanism to output a feature weight vector. For example, this weight vector can be represented as... ,in The importance of corresponding current characteristics The importance of corresponding voltage characteristics The importance of corresponding partial discharge amplitude characteristics The importance of the corresponding partial discharge frequency characteristics. These weight values ​​are typically between 0 and 1, and their sum is 1.

[0100] Subsequently, in order to initialize the input layer weight matrix of the extreme learning machine, the attention weight vector can be transformed into a diagonal matrix. The diagonal elements of this diagonal matrix are the elements of the attention weight vector, while the off-diagonal elements are zero. Simultaneously, a random weight matrix uniformly distributed within a specific interval (e.g., [-1, 1]) can be generated. Its dimension is consistent with the dimension required for the input layer weight matrix of the Extreme Learning Machine. Finally, the input layer weight matrix of the Extreme Learning Machine... By and We obtain this by performing matrix multiplication, i.e. This initialization method allows the Extreme Learning Machine to incorporate prior knowledge of feature importance into its input layer weight matrix at the start of training. For example, if the amplitude of partial discharge is a key fault feature, its corresponding... If the value is high, then in The weights associated with this feature will gain a greater initial influence, thus guiding the model to learn effective fault identification patterns more quickly.

[0101] By using the weights of the attention mechanism to initialize the input layer weight matrix of the Extreme Learning Machine (ELM), this application effectively solves the problems of unstable training and slow convergence speed of ELM under random initialization. This scheme enables the fault identification model to focus on the most critical features for fault identification in the early stages of training, significantly improving the training efficiency of the model and reducing the number of iterations required for the model to reach stable performance.

[0102] Meanwhile, this initialization strategy enhances the model's training stability, ensuring consistency and reliability of model performance across different training batches or initialization conditions, avoiding performance fluctuations caused by randomness. Furthermore, because the model can capture key fault characteristics earlier and more accurately, the accuracy of fault identification is further optimized, especially demonstrating stronger capabilities in identifying early, minor faults, thus effectively reducing the risk of missed fault detection. More importantly, this improvement introduces only a small amount of computation during the model training initialization phase, without adding any extra burden during the model inference phase. Therefore, it fully retains the advantages of the lightweight hybrid model combining attention mechanisms and extreme learning machines, enabling it to still meet the practical needs of SMEs for low cost, high accuracy, and easy maintenance without sacrificing the model's lightweight characteristics.

[0103] In some of the solutions described above in this application, multi-source monitoring data acquisition is proposed to obtain current, voltage, partial discharge signals, and environmental parameters. However, in this process, the lack of time consistency in data acquisition from different sensors can lead to time deviations during data fusion, affecting the accuracy of subsequent environmental adaptive calibration and fault identification. To address this, this application proposes a scheme where multi-source monitoring data is acquired through the following sensors: a current sensor, a voltage sensor, a partial discharge sensor, and a temperature and humidity sensor, with the sensors synchronized by a microcontroller.

[0104] Among them, current sensors, voltage sensors, partial discharge sensors, and temperature and humidity sensors are key devices used to collect data on the operating status and environmental conditions of cable branch boxes. Current sensors are used to monitor the magnitude and waveform of the current in the cable in real time; voltage sensors are used to monitor the voltage level and stability of the cable; and partial discharge sensors are used to detect partial discharge signals that may exist inside the cable insulation, which is an important basis for judging insulation degradation and potential faults. Temperature and humidity sensors are used to obtain temperature and humidity information of the environment in which the cable branch box is located; these environmental parameters have a significant impact on the performance and failure modes of electrical equipment.

[0105] In practical applications, current sensors can be Hall effect sensors or current transformers, such as Rogowski coil-based current sensors. Voltage sensors can be resistive voltage divider sensors or voltage transformers, such as high-voltage probes. Partial discharge sensors can be ultra-high frequency (UHF) sensors or ultrasonic sensors. Temperature and humidity sensors can be digital temperature and humidity sensors, such as the BME280. A microcontroller is an integrated circuit chip that integrates a central processing unit (CPU), memory, and various peripheral interfaces. It can execute preset programs to control and process data from external devices. In this solution, the microcontroller is mainly responsible for coordinating and managing the data acquisition process of multiple sensors and synchronizing their clocks. Various models of microcontrollers can be selected, such as the ESP32 series microcontrollers based on the ARM Cortex-M core.

[0106] This application's solution comprehensively acquires electrical and environmental data of the cable branch box through the coordinated operation of current sensors, voltage sensors, partial discharge sensors, and temperature and humidity sensors. The microcontroller, as the core control unit, establishes a unified timing reference and coordinates the sampling actions of all sensors. Specifically, the microcontroller generates a synchronization pulse signal at a preset maximum sampling frequency and sends this signal to all connected sensors via a standard communication interface.

[0107] Upon receiving the synchronization pulse, each sensor initiates data acquisition according to its preset sampling frequency and internal logic. After data acquisition is complete, the microcontroller immediately assigns a high-precision timestamp to the data collected by each sensor. These timestamps are crucial for achieving precise timeline alignment of multi-source data. Through these unified timestamps, even if different sensors sample at different frequencies, their data can be accurately correlated in the time dimension. For example, in the subsequent environmental adaptive calibration stage, the microcontroller can match temperature and humidity data within a specific time window with current, voltage, and partial discharge signals within the same time window based on the timestamps, thereby achieving dynamic filtering adjustments based on environmental parameters.

[0108] Similarly, in the second-level calibration, determining the reliability of partial discharge signals also requires precise correlation of the stability of current and voltage data within the same time period. In the fault identification model, the capture of time-series correlated fault features such as "partial discharge pulses and voltage peaks" also highly relies on this precise alignment of the time axis. This two-layer time-series collaborative logic of "synchronous triggering for reference + timestamp calibration correlation" ensures high consistency of multi-source data in the time dimension, providing a reliable data foundation for subsequent environmental adaptive calibration and fault identification. It cleverly solves the problem of multi-frequency sensors being unable to force synchronous acquisition, avoids data redundancy or loss that may result from traditional "one-size-fits-all" synchronization, and ensures time consistency during data fusion, thereby significantly improving the accuracy and reliability of the overall monitoring system.

[0109] The following example illustrates this, using an STM32F103 series microcontroller as the core processing unit. This microcontroller connects to an ACS712-5A current sensor, an LV25-P voltage sensor, an HFCT-01 partial discharge sensor, and a DHT22 temperature and humidity sensor via an I2C bus. In operation, the STM32F103 microcontroller's internal timer module is configured to generate interrupts at a frequency of 10kHz. Each time an interrupt is triggered, the microcontroller sends a synchronization acquisition command to all sensors via the I2C bus. Upon receiving the command, the HFCT-01 partial discharge sensor immediately performs a data sample. The ACS712-5A current sensor and the LV25-P voltage sensor are configured to sample every 10 synchronization commands (i.e., every 1ms).

[0110] The DHT22 temperature and humidity sensor is configured to sample once every 1000 synchronization commands (i.e., every 10 seconds). After each sensor completes data sampling, the STM32F103 microcontroller immediately reads the sensor data and uses its internal high-precision real-time clock (RTC) module to add a timestamp accurate to the microsecond level. In this way, even if the sampling frequencies of different sensors are different, all collected data have precise timestamps. Therefore, during data processing, data from different sources and at different frequencies can be precisely aligned on the timeline based on these timestamps. For example, during the first-level environmental adaptive calibration, the system can correlate temperature and humidity data at a certain moment with current, voltage, and partial discharge signals at the same moment based on the timestamps, dynamically adjusting the filtering parameters. This microcontroller-based synchronization triggering and timestamp calibration mechanism ensures the temporal consistency of multi-source data, providing a solid foundation for subsequent data calibration and fault identification.

[0111] Through the above technical solution, this application effectively solves the data fusion deviation problem caused by the lack of time consistency in data acquisition from different sensors during multi-source monitoring data acquisition. By combining current sensors, voltage sensors, partial discharge sensors, and temperature and humidity sensors, comprehensive coverage of the electrical and environmental parameters of the cable branch box is achieved. More importantly, the microcontroller-led "synchronous triggering and benchmark setting + timestamp calibration association" mechanism ensures precise alignment of all sensor data in the time dimension, reducing the time deviation of multi-source data to the microsecond level.

[0112] This completely eliminates the time misalignment problem between environmental and electrical parameters caused by traditional asynchronous acquisition, enabling subsequent adaptive environmental calibration to be based on accurate time correspondence and significantly improving the effectiveness of filter parameter adjustment. Simultaneously, this high-precision time synchronization also ensures the accuracy of the timing correlation judgment between current and voltage data stability and partial discharge signals in the second-level cross-source calibration, thereby improving the reliability of partial discharge signal credibility assessment. Furthermore, for the lightweight hybrid model combining attention mechanisms and extreme learning machines, precisely aligned time-series data allows for more accurate capture of key fault characteristics such as "partial discharge pulses and voltage peaks," significantly improving fault identification accuracy. This solution achieves high-precision time synchronization while effectively controlling hardware costs, providing SMEs with a low-cost, high-accuracy, and easy-to-maintain online monitoring solution for cable distribution boxes.

[0113] In some of the solutions mentioned above in this application, a low-power communication method is proposed for uploading early warning information. However, in this process, the communication latency may be too high, which cannot meet the real-time early warning requirements, and the power consumption control may not be precise, resulting in insufficient overall system energy efficiency.

[0114] In this regard, this application further proposes that the low-power communication method is LoRa wireless communication, and the communication strategy includes timed data upload and fault-triggered real-time upload.

[0115] Specifically, the low-power communication method is LoRa wireless communication. LoRa (Long Range) is a long-range wireless communication technology based on spread spectrum technology, belonging to the category of Low Power Wide Area Network (LPWAN) Internet of Things (IoT) communication technologies. Its main characteristics are long-distance transmission, low-power operation, and strong anti-interference capabilities. This technology effectively solves the problems of high power consumption, limited coverage, and high deployment costs faced by traditional wireless communication technologies in cable branch box online monitoring scenarios, providing an efficient and reliable transmission channel for monitoring data. As one possible implementation, a LoRaWAN protocol stack can be used to transmit data to a cloud platform via a LoRaWAN gateway; alternatively, a point-to-point LoRa communication module can be used to directly transmit data to a local receiving device.

[0116] The communication strategy includes timed data upload, which means the system automatically uploads monitoring data from the cable distribution box to the maintenance platform at preset time intervals. This strategy aims to balance the timeliness of data updates with system power consumption. During normal operation of the cable distribution box, data is transmitted at a lower frequency, thereby effectively reducing the power consumption of the communication module, extending the battery life of the equipment, and reducing network bandwidth usage. As one possible implementation, a timer interrupt can be set internally in the microcontroller to trigger data packaging and LoRa module transmission operations when the preset time interval is reached; alternatively, the task scheduling mechanism of the operating system (such as FreeRTOS) can be used to create a timed task to periodically execute the data upload function.

[0117] The communication strategy also includes fault-triggered real-time upload, which means that when the system detects a potential fault or abnormal situation, it immediately interrupts the current task and prioritizes uploading the early warning information to the operation and maintenance platform. The core of this strategy is to ensure that early warning information can be delivered to operation and maintenance personnel in a timely manner during emergencies, meeting the need for real-time early warning and preventing the escalation of faults or losses due to communication delays. As one possible implementation, after the fault identification model outputs an early warning signal, it can notify the communication module to immediately start the data upload process through an interrupt or event flag; alternatively, the communication module can continuously monitor the output of the fault identification module, and once a fault trigger condition is detected, it immediately initiates a high-priority data upload task.

[0118] This application's solution combines LoRa wireless communication technology with a dual communication strategy to construct an efficient, reliable, and low-power online monitoring data transmission mechanism for cable branch boxes. During online monitoring of the cable branch box, multi-source monitoring data acquired by the data acquisition module is processed by the calibration module and then used by the fault identification module for fault identification. When the fault identification module does not detect any abnormalities, the communication module will upload key operating parameters of the cable branch box via LoRa wireless communication at preset time intervals, according to a timed data upload strategy, to maintain low-power operation of the system under normal conditions. Once the fault identification module detects a potential fault and outputs a warning message, the communication module will immediately switch to a fault-triggered real-time upload strategy, prioritizing the upload of the warning message to the operation and maintenance platform via LoRa wireless communication.

[0119] The synergistic operation of this dual strategy enables the system to minimize communication power consumption while ensuring real-time early warning capabilities. LoRa technology's inherent long-range, low-power, and strong anti-interference capabilities make it particularly suitable for monitoring outdoor distributed cable distribution boxes, effectively covering a wide area and resisting complex electromagnetic interference. In this way, the proposed solution not only resolves the contradiction between real-time performance and power consumption in traditional communication methods but also ensures timely delivery of early warning information at critical moments, thereby improving the operational efficiency and reliability of the entire cable distribution box online monitoring system.

[0120] The following is a specific example. In an online monitoring system for cable branch boxes, a LoRa communication module, model SX1278, can be used. Its operating frequency is set to 433MHz to utilize its excellent penetration and anti-interference capabilities. Under normal operating conditions, the system is configured to upload data every 5 minutes. The amount of data uploaded each time is optimized to less than 1KB, containing only the average values ​​of current, voltage, partial discharge signals, environmental parameters, and equipment status indicators.

[0121] When the fault identification model (a lightweight hybrid model combining attention mechanism and extreme learning machine) identifies faults in the calibrated electrical parameter data and determines that the fault prediction probability reaches or exceeds 80%, the LoRa communication module will immediately interrupt the current timed upload task and prioritize initiating the fault-triggered real-time upload process. At this time, the system will immediately package the warning information, which includes the device number, fault type, key abnormal parameters, and a precise timestamp, with the data size controlled within 2KB. To ensure reliable delivery of the warning information, the system will also adopt an acknowledgment and retransmission mechanism. If the receiving end does not return acknowledgment information within 10 milliseconds, the LoRa module will immediately retransmit. In addition, to further optimize system energy efficiency, during non-transmission and non-sampling periods, the LoRa module, front-end sensors, and microcontroller will synchronously enter sleep mode, reducing the sleep current to below 10μA.

[0122] Through the aforementioned technical solutions, the online monitoring system for cable distribution boxes can significantly reduce the transmission delay of early warning information, ensuring that maintenance personnel receive early warnings quickly after fault symptoms appear, thus gaining valuable time for emergency response. Simultaneously, the system's energy efficiency is significantly optimized, effectively solving the problem of inconvenient power supply for outdoor cable distribution boxes and extending the equipment's lifespan. Furthermore, the long-range coverage and strong anti-interference capabilities of LoRa wireless communication ensure reliable communication in complex industrial environments, preventing the loss of early warning information due to communication interruptions. This combination of dual communication strategies ensures timely delivery of critical early warning information while effectively controlling the power consumption and bandwidth usage of daily data transmission, improving the overall efficiency of data transmission. Example

[0123] In some of the solutions described above in this application, a device for performing online monitoring of cable branch boxes has been proposed. However, in this process, the implementation cost of the device is high and the operation and maintenance complexity is large, making it difficult to adapt to the actual needs of small and medium-sized enterprises for low cost, easy deployment and low maintenance, which makes it difficult for the existing system to operate efficiently in outdoor variable environments.

[0124] In response, this application proposes an online monitoring device for cable branch boxes, which includes a data acquisition module, a calibration module, a fault identification module, a communication module, and an operation and maintenance platform module.

[0125] The data acquisition module is the sensing front end of the device, and its function is to acquire necessary raw data from the operating environment of the cable distribution box. The data acquisition module can be a standalone hardware unit, such as a dedicated circuit board integrating multiple sensors, or it can be a software driver layer embedded in the main controller, responsible for managing the read and write operations of external sensors. Its implementation can include: converting analog sensor signals into digital signals via an analog-to-digital converter (ADC), or directly reading digital sensor data through a digital interface (such as I2C, SPI, or UART).

[0126] This module is used to acquire multi-source monitoring data from cable branch boxes, including at least current, voltage, partial discharge signals, and environmental parameters. This data can be acquired using various sensors. For example, current data can be obtained using Hall effect current sensors or shunts; voltage data can be obtained using voltage transformers or resistive voltage dividers; partial discharge signals can be obtained using ultra-high frequency (UHF) sensors or high-frequency current transformers (HFCTs); and environmental parameters (such as temperature and humidity) can be obtained using thermistors, humidity sensors, or integrated temperature and humidity sensors.

[0127] The calibration module is responsible for preprocessing the raw, multi-source monitoring data to eliminate or reduce the impact of environmental factors on data accuracy. The calibration module can be a collection of software algorithms running on a microprocessor, or a dedicated hardware module containing specific filtering circuits and signal processing units. Its implementation can include using a digital signal processor (DSP) to execute complex filtering and calibration algorithms, or utilizing the computing power of a general-purpose microcontroller (MCU) to achieve software-level data calibration. This module is used to perform environmental adaptive calibration on the multi-source monitoring data to obtain calibrated electrical parameter data.

[0128] Environmental adaptive calibration means that the calibration process can be dynamically adjusted according to real-time environmental conditions to adapt to different interference situations. Implementation methods may include: dynamically adjusting the cutoff frequency or gain of a digital filter based on parameters such as ambient temperature and humidity to suppress noise in specific frequency bands; or correcting or eliminating abnormal data by analyzing the correlation between data from different sensors.

[0129] The fault identification module is the core intelligent component of the device, used to analyze calibrated data and determine whether a fault exists in the cable distribution box. The fault identification module can be a machine learning model running on an embedded processor or an inference engine integrated into a dedicated AI chip. Its implementation can include deploying a lightweight neural network model on a microcontroller for inference, or utilizing a field-programmable gate array (FPGA) for parallel computing to accelerate the fault identification process.

[0130] This module includes a lightweight hybrid model combining an attention mechanism and an Extreme Learning Machine (ELM). The attention mechanism can be used to evaluate and weight the importance of input features, thereby highlighting key features; the Extreme Learning Machine (ELM) is a single-hidden-layer feedforward neural network known for its fast learning and good generalization ability. Lightweight design means a simple model structure and few parameters, suitable for running on low-computing-power hardware. Implementation methods can include: the attention mechanism can use simple fully connected layers or convolutional layers to assign feature weights; the Extreme Learning Machine can randomly generate input layer weights and biases and solve for the output layer weights using the least squares method. This module is used for fault identification based on calibrated electrical parameter data. Implementation methods can include: inputting calibrated feature vectors such as current, voltage, and partial discharge into the hybrid model, with the model outputting a fault category (e.g., normal, partial discharge, overload, etc.) or fault probability; or setting a threshold, where an abnormal index output by the model is considered a fault.

[0131] The communication module is responsible for transmitting fault identification results or early warning information to a remote platform, enabling remote monitoring and management. The communication module can be a wireless transceiver, such as a module based on cellular networks (e.g., 2G / 4G / 5G), Wi-Fi, Bluetooth, or LoRa technologies, or it can be an interface based on wired networks (e.g., Ethernet, RS485). Its implementation can include: integrating a wireless communication chip and antenna to transmit data via radio waves; or connecting to a local area network (LAN) or wide area network (WAN) via a network cable for data transmission. This module is used to upload early warning information based on the identification results. Implementation can include: when the fault identification module detects a fault, the communication module immediately encapsulates an early warning data packet and sends it to the early warning server; or, under normal operating conditions, the communication module periodically uploads device status reports at preset time intervals.

[0132] The operations and maintenance (O&M) platform module serves as the backend management center of the entire monitoring system. It receives, processes, and displays data and early warning information uploaded by front-end devices, and provides system maintenance functions. The O&M platform module can be a web application deployed on a cloud server or a desktop application installed on a local computer. Its implementation can include: development based on a B / S (Browser / Server) architecture, accessible through a browser; or development based on a C / S (Client / Server) architecture, where users operate through client software. This module receives and displays early warning information and provides model update functionality. Implementation can include: the O&M platform receiving early warning information through a data interface and displaying it in real-time on the dashboard; simultaneously, O&M personnel can upload new model files to the platform, which then distributes these model files to the front-end devices for updates via a communication module.

[0133] This device, through its modular design, constructs a complete closed-loop system encompassing data acquisition, fault identification, information uploading, and maintenance, aiming to address the pain points of existing devices in terms of cost, maintenance, and environmental adaptability. Its overall operational logic is as follows: The data acquisition module, acting as the system's perception layer, is responsible for acquiring multi-source monitoring data from the cable distribution box in real-time and comprehensively. This data covers at least current, voltage, partial discharge signals, and environmental parameters. Precise sensor configuration and clock synchronization mechanisms ensure the temporal consistency and integrity of the multi-source data, laying the foundation for subsequent data processing. Subsequently, the calibration module receives the raw multi-source monitoring data from the data acquisition module. Given that cable distribution boxes are often located in complex and variable outdoor environments, the raw data is highly susceptible to distortion due to environmental factors (such as temperature and humidity fluctuations and electromagnetic noise). The core function of the calibration module is to perform environmentally adaptive calibration on this data. By dynamically adjusting filter parameters and utilizing the correlation between multi-source data for correction, noise is effectively filtered out and deviations are corrected, thereby obtaining high-quality, highly reliable electrical parameter data.

[0134] This step is crucial for ensuring the accuracy of subsequent fault identification. It avoids model misjudgments due to data quality issues and reduces the complexity requirements of the fault identification model. Next, the calibrated electrical parameter data is input into the fault identification module. This module uses a lightweight hybrid model combining attention mechanisms and extreme learning machines for fault identification.

[0135] Attention mechanisms play a crucial role here, automatically assigning different weights to different features based on the characteristics of the data, highlighting information that is key to fault identification, thereby improving the model's feature extraction capabilities. Extreme Learning Machines, with their advantages of rapid learning and efficient classification, classify faults based on weighted features, quickly and accurately determining the operating status of cable branch boxes and identifying potential fault types. This lightweight hybrid model design significantly reduces the demand for hardware computing power while ensuring recognition accuracy, enabling the device to operate efficiently on low-cost microcontrollers. Once the fault identification module detects an anomaly or fault, the communication module uploads the warning information to the remote operation and maintenance platform via low-power communication based on the identification results.

[0136] This communication method choice fully considers the inconvenience of power supply to outdoor cable branch boxes, achieving system-level energy consumption optimization. Finally, the operation and maintenance platform module receives and displays these early warning messages, enabling maintenance personnel to understand the equipment status in a timely manner. More importantly, the operation and maintenance platform also provides a model update function, allowing remote optimization and iteration of the fault identification model without on-site manual intervention, greatly reducing the difficulty and technical threshold of operation and maintenance. Through the close collaboration of the above modules, the entire device forms a highly efficient, low-cost, easy-to-maintain, and environmentally adaptable online monitoring system. The comprehensiveness of data acquisition, the adaptability of calibration, the lightweight and high-precision fault identification, the low power consumption of communication, and the convenience of operation and maintenance constitute the core advantages of this device, enabling it to effectively cope with challenges such as data distortion, high system costs, and difficult operation and maintenance in complex outdoor environments.

[0137] As a specific implementation, this online monitoring device for cable branch boxes can be configured as follows: The data acquisition module can consist of a series of low-cost, industrial-grade sensors. For example, the current sensor can be an ACS712-5A Hall effect current sensor, the voltage sensor can be an LV25-P voltage sensor, the partial discharge sensor can be an HFCT-01 high-frequency current transformer, and environmental parameters (such as temperature and humidity) can be acquired by a DHT22 temperature and humidity sensor. These sensors are all connected to a main microcontroller, such as an STM32F103 series microcontroller, via a standard interface (such as I2C or SPI). This microcontroller is responsible for clock synchronization and data reading of all sensors. The calibration module can run as software on the aforementioned STM32F103 microcontroller. This program can implement environmental adaptive calibration, for example, dynamically adjusting the parameters of a digital filter (such as a Butterworth filter) based on the real-time temperature and humidity data acquired by the DHT22 sensor to adapt to noise characteristics under different environments.

[0138] Simultaneously, this module can assess the reliability of partial discharge signals by analyzing the stability of current and voltage data, and perform weighted corrections on the partial discharge signals to improve their accuracy. The fault identification module can also be deployed as software on the STM32F103 microcontroller. This program integrates a lightweight hybrid model combining attention mechanisms and extreme learning machines.

[0139] The attention mechanism can be implemented by a simple fully connected layer to assign weights to features such as current, voltage, and partial discharge signals. The extreme learning machine can use a single hidden layer network structure, where the input layer weights and biases are randomly generated, and the output layer weights are quickly solved by the least squares method, thereby achieving fault classification of calibrated electrical parameter data.

[0140] The entire model has been optimized to ensure that its memory footprint and computational complexity are within the microcontroller's tolerance. A low-power LoRa wireless communication module, such as one based on the SX1278 chip, can be used for the communication module. This module connects to the STM32F103 microcontroller via an SPI interface and is responsible for sending fault identification results and warning information through the LoRa protocol.

[0141] The communication strategy can be configured to periodically upload device status data and trigger real-time uploads immediately upon detecting a fault. The operations and maintenance platform module can be a desktop application or a web application deployed on a regular desktop computer. The platform receives alert information from the front-end devices via a LoRa gateway and displays it in intuitive charts and lists on the user interface. Furthermore, the operations and maintenance platform provides a model update interface, allowing operations and maintenance personnel to upload new model parameter files and transmit these parameters to the STM32F103 microcontroller of the front-end device via a LoRa communication link, enabling remote model updates.

[0142] This device, through its unique modular design and collaborative working mechanism, effectively solves the prominent problems of existing online monitoring devices for cable branch boxes in terms of cost, operation and maintenance, and environmental adaptability. Specifically, by adopting a low-cost data acquisition module, investment costs are controlled from the hardware source, making it affordable for small and medium-sized enterprises. The environmental adaptive calibration function of the calibration module significantly improves the accuracy and reliability of data acquisition in complex outdoor environments, effectively avoiding misjudgments caused by data distortion, thereby improving the accuracy of fault identification. The fault identification module adopts a lightweight hybrid model combining attention mechanisms and extreme learning machines, which significantly reduces the hardware computing power requirements while ensuring high identification accuracy, enabling the device to operate efficiently on a low-cost microcontroller, further reducing the overall hardware cost.

[0143] The low-power design of the communication module, combined with the remote model update function of the operation and maintenance platform, greatly simplifies the deployment and subsequent maintenance of the device, reduces reliance on professional technicians, and enables non-professionals to easily manage and operate it. Overall, this device achieves a balance between low cost, high accuracy, easy operation and maintenance, and strong environmental adaptability, overcoming the contradictions between high precision and high cost, and low cost and low precision in existing technologies. It not only provides reliable fault early warning and effectively avoids losses from sudden failures, but its low deployment and maintenance costs, as well as its good adaptability to harsh outdoor environments, make it an ideal solution for online monitoring of cable branch boxes for small and medium-sized enterprises, filling a market gap and possessing significant practical value and industrial promotion significance.

[0144] Traditional online monitoring methods for cable distribution boxes are susceptible to data distortion in harsh outdoor environments due to drastic fluctuations in temperature and humidity, as well as electromagnetic noise interference. Partial discharge signals are easily masked by noise, thus affecting the accuracy of fault feature extraction and identification. While some solutions in this application propose calibration modules for environmentally adaptive calibration, the calibration process may lack a targeted implementation mechanism, failing to effectively handle noise interference caused by dynamic changes in environmental parameters and the reduced reliability of partial discharge signals due to environmental fluctuations. This exacerbates data distortion and impacts the accuracy and reliability of fault identification.

[0145] To address this, this application further proposes a calibration module, which includes a first-level calibration unit and a second-level calibration unit. The first-level calibration unit is used to dynamically adjust the filter parameters based on environmental parameters; the second-level calibration unit is used to perform a confidence-weighted correction of the partial discharge signal based on the stability of current and voltage data.

[0146] The calibration module is a core component of the online monitoring device for cable branch boxes. Its main function is to process multi-source monitoring data obtained from the cable branch boxes to eliminate or reduce errors and noise caused by environmental factors and the characteristics of the sensors themselves, thereby obtaining high-quality and highly reliable electrical parameter data.

[0147] The calibration module can be a standalone hardware unit, such as an application-specific integrated circuit (ASIC) or a field-programmable gate array (FPGA), or a collection of software algorithms running on a microcontroller or embedded processor. The first-level calibration unit is a functional component within the calibration module, specifically responsible for dynamically adjusting filter parameters based on environmental parameters. Its role is to perform preliminary noise suppression and data optimization on the raw current, voltage, and partial discharge signals according to real-time changing environmental conditions. The first-level calibration unit can be implemented using a digital signal processor (DSP) or by running a filtering algorithm on a general-purpose microcontroller. The technical feature of dynamically adjusting filter parameters based on environmental parameters means that the calibration process does not employ a fixed filtering strategy, but rather adjusts the filter parameter settings in real time according to changes in the external environment (such as temperature and humidity). For example, when the ambient noise level is high, the filtering intensity can be increased; when the environment is relatively stable, the filtering intensity can be appropriately reduced to retain more signal details. This dynamic adjustment can be based on preset lookup tables, fuzzy logic control rules, or adaptive filtering algorithms (such as adaptive Kalman filtering and adaptive Wiener filtering).

[0148] The second-level calibration unit is another functional component of the calibration module. Its main task is to further refine the partial discharge signal by utilizing the inherent correlation between multi-source data. It operates after the first-level calibration and aims to address the potential for misjudgment or reduced reliability of partial discharge signals under complex environments. The second-level calibration unit can be a standalone signal processing chip or integrated into the software module of the first-level calibration unit.

[0149] The technique of using stability-based reliability weighting to correct partial discharge signals leverages the fact that current and voltage signals in cable distribution boxes are generally more stable and less susceptible to environmental interference than partial discharge signals. By evaluating the volatility or anomalies in current and voltage data, the reliability of the partial discharge signal under the current environment can be indirectly determined. For example, when current and voltage data exhibit high stability, the reliability weight of the partial discharge signal can be set higher; conversely, when current and voltage data show drastic fluctuations or anomalies, the reliability weight of the partial discharge signal can be appropriately reduced and weighted to minimize interference from unreliable signals in fault identification.

[0150] This correction can be achieved by setting thresholds, using statistical analysis methods (such as ANOVA or correlation analysis) or machine learning models to determine weights and perform operations such as weighted averaging or weighted fusion.

[0151] This calibration module, by introducing a two-stage calibration unit working in tandem, aims to address environmental interference and signal reliability issues, thereby improving the accuracy of data calibration. Specifically, the module first performs preliminary processing on the raw, multi-source monitoring data acquired through the first-stage calibration unit. The core of this first-stage calibration unit lies in dynamically adjusting filter parameters based on environmental parameters. This means that it does not employ a fixed filtering strategy, but rather dynamically adjusts the parameter settings of its internal filters according to real-time acquired environmental parameters (such as temperature and humidity). For example, when environmental parameters indicate significant external interference, the filter parameters are adjusted to enhance the filtering effect and effectively suppress noise; when environmental parameters indicate a relatively stable environment, the filter parameters may be adjusted to retain more signal details and avoid over-filtering. This adaptive filtering mechanism ensures that current, voltage, and partial discharge signals can be effectively suppressed under different environmental conditions, providing a relatively clean data foundation for subsequent processing.

[0152] After the first-level calibration, the calibration module further refines the partial discharge signal through a second-level calibration unit. The key to this second-level calibration unit is to perform a confidence-weighted correction of the partial discharge signal based on the stability of the current and voltage data.

[0153] Since current and voltage data are generally less sensitive to environmental noise than partial discharge signals, their stability can serve as an important criterion for assessing the reliability of partial discharge signals. The second-level calibration unit analyzes the fluctuations or anomalies in the current and voltage data and determines the reliability of the current partial discharge signal accordingly. For example, if the current and voltage data exhibit high stability, the partial discharge signal is considered to have high reliability and is given a larger weight; conversely, if the current and voltage data show drastic fluctuations or anomalies, the partial discharge signal is considered to be subject to significant interference, has low reliability, is given a smaller weight, and is then weighted and corrected based on this weight.

[0154] This correction mechanism effectively reduces the misleading influence of unreliable partial discharge signals on fault identification, ensuring more reliable calibrated data. Through the synergistic effect of the two-stage calibration units, the calibration module improves the quality of multi-source monitoring data from the source. The first-stage calibration unit addresses the common noise interference caused by dynamic environmental changes, while the second-stage calibration unit, considering the specific characteristics of partial discharge signals, further enhances their reliability by utilizing the correlation between multi-source data. This hierarchical, adaptive, and mutually verifying calibration strategy enables the calibrated electrical parameter data to more accurately reflect the actual operating status of the cable branch box, providing high-quality input for subsequent fault identification models. This significantly improves the accuracy and reliability of fault identification, effectively avoiding misjudgments caused by data distortion.

[0155] In one specific implementation, the calibration module can be integrated into an embedded system containing a high-performance microcontroller (such as an ARM Cortex-M series processor) and necessary memory. The first-stage calibration unit can be implemented as follows: the microcontroller reads environmental parameters collected by temperature and humidity sensors in real time. Based on a pre-stored lookup table or fuzzy control rules, it dynamically adjusts the gain parameters of the Kalman filter. For example, when the ambient temperature is below -10°C or above 40°C, or the relative humidity exceeds 85%, the parameters of the process noise covariance matrix and measurement noise covariance matrix of the Kalman filter are adjusted to increase the filter's smoothing capability, thereby providing stronger noise suppression for current, voltage, and partial discharge signals.

[0156] Conversely, under suitable environmental conditions, the filtering parameters will be adjusted to reduce the filtering intensity in order to better preserve the detailed features of the signal. The second-stage calibration unit can be specifically implemented as follows: in the current and voltage data processed by the first-stage calibration, the microcontroller calculates the standard deviation or coefficient of variation within a certain time window to evaluate its stability.

[0157] Simultaneously, preliminary feature extraction is performed on the partial discharge signal, such as calculating its pulse repetition rate and amplitude. If the stability indicators (such as standard deviation) of the current and voltage data are below a preset threshold, indicating stable grid operation, the confidence weight of the partial discharge signal can be set to 0.9; if the stability indicators are above the threshold, indicating potential fluctuations or interference in the grid, the confidence weight of the partial discharge signal can be reduced to 0.5. Subsequently, the amplitude of the partial discharge signal is corrected by multiplying this confidence weight; for example, the corrected partial discharge signal amplitude = original partial discharge signal amplitude × confidence weight. Alternatively, a wavelet transform-based denoising method can be used, with the wavelet threshold adjusted according to the stability of the current and voltage, to further optimize the partial discharge signal.

[0158] Through the aforementioned technical solution, the calibration module, by introducing a first-level calibration unit and a second-level calibration unit, achieves refined and adaptive calibration of multi-source monitoring data from cable branch boxes. The first-level calibration unit dynamically adjusts filter parameters based on environmental parameters, effectively addressing the complex and variability of outdoor environments. This significantly reduces the interference of environmental noise on current, voltage, and partial discharge signals, avoiding the problem of poor performance of traditional fixed filter parameters in harsh environments. The second-level calibration unit utilizes the stability of current and voltage data to perform reliability-weighted correction on partial discharge signals, solving the problems of susceptibility to interference and low reliability of partial discharge signals, ensuring the accuracy of key fault characteristics. This two-level collaborative calibration mechanism fundamentally improves the quality and reliability of monitoring data, providing a cleaner and more accurate input for subsequent fault identification models. This significantly improves the accuracy of fault identification, reduces the false alarm rate, and effectively avoids misjudgments caused by data distortion, providing a solid data foundation for online monitoring of cable branch boxes.

[0159] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for online monitoring of cable branch boxes, characterized in that, Includes the following steps: Acquire multi-source monitoring data of the cable branch box, wherein the multi-source monitoring data includes at least current, voltage, partial discharge signal and environmental parameters; Environmental adaptive calibration is performed on the multi-source monitoring data to obtain calibrated electrical parameter data. The environmental adaptive calibration includes a first-level calibration based on dynamically adjusting the filter parameters according to environmental parameters, and a second-level calibration based on the correlation of multi-source data. The calibrated electrical parameter data is input into the fault identification model for fault identification. The fault identification model is a lightweight hybrid model that combines attention mechanism and extreme learning machine. Based on the identification results, the early warning information is uploaded to the operation and maintenance platform via low-power communication.

2. The method according to claim 1, characterized in that, The first-stage calibration, which involves dynamically adjusting filter parameters based on environmental parameters, specifically includes: The gain parameters of the Kalman filter are dynamically adjusted based on the real-time collected environmental parameters to filter current, voltage, and partial discharge signals.

3. The method according to claim 2, characterized in that, The second-level calibration based on the correlation of multi-source data specifically includes: The reliability of the partial discharge signal is determined based on the stability of current and voltage data, and the partial discharge signal is weighted and corrected based on the reliability weight.

4. The method according to claim 1, characterized in that, In the fault identification model, the attention mechanism is used to assign weights to the input features, and the extreme learning machine is used to classify faults based on the weighted features.

5. The method according to claim 4, characterized in that, The weights of the attention mechanism are used to initialize the input layer weight matrix of the extreme learning machine.

6. The method according to claim 1, characterized in that, The multi-source monitoring data is collected through the following sensors: current sensor, voltage sensor, partial discharge sensor and temperature and humidity sensor, and the sensors are clock-synchronized by a microcontroller.

7. The method according to claim 1, characterized in that, The low-power communication method is LoRa wireless communication, and the communication strategy includes timed data upload and fault-triggered real-time upload.

8. An online monitoring device for cable branch boxes, characterized in that, For performing the method according to any one of claims 1-7, comprising: The data acquisition module is used to acquire multi-source monitoring data of the cable branch box; The calibration module is used to perform environmental adaptive calibration on the multi-source monitoring data to obtain calibrated electrical parameter data; The fault identification module includes a lightweight hybrid model that combines attention mechanism and extreme learning machine, used to identify faults in calibrated electrical parameter data; The communication module is used to upload early warning information based on the identification results; The operations and maintenance platform module is used to receive and display early warning information and provide model update functions.

9. The apparatus according to claim 8, characterized in that, The calibration module includes: The first-level calibration unit is used to dynamically adjust the filter parameters based on environmental parameters; The second-level calibration unit is used to perform confidence-weighted correction on the partial discharge signal based on the stability of current and voltage data.

Images