An edge autonomy and dynamic load perception-based distribution box state monitoring method and system

Abstract

The application discloses a kind of based on edge autonomy and dynamic load perception's distribution box state monitoring method and system, edge gateway obtains the operating condition data collected by sensor, according to the historical load time series data stored locally, call the prediction model independently executed inference that solidification deployment is locally, output load variation trend, and according to the trend hierarchical adjustment sensor's collection frequency;Edge gateway also executes the comprehensive abnormality judgment of three dimensions of current collection value, historical baseline value and load variation trend to real-time collection data, when determining abnormality, execute protection action locally;Edge gateway monitors the communication link state with cloud, when link interruption, data containing prediction result and collection frequency scheduling record are buffered locally, and are transmitted after link recovery.The application realizes the complete autonomous decision closed loop of edge side in communication interruption scene, and is suitable for communication signal unstable distribution place.

Description

Technical Field

[0001] This invention relates to the field of online condition monitoring technology for low-voltage power distribution equipment, specifically to a method and system for monitoring the condition of distribution boxes based on edge autonomy and dynamic load perception. Background Technology

[0002] Electrical connection points within low-voltage distribution boxes, such as busbar joints, circuit breaker input / output terminals, and cable joints, are prone to failure due to long-term exposure to current heat and mechanical vibration. Statistics show that over 60% of electrical faults in distribution boxes originate from insulation aging or burning caused by overheating at connection points. Therefore, continuous monitoring of the temperature and circuit electrical parameters at these critical points is a fundamental method for assessing the condition of distribution boxes. Early methods relied on regular manual inspections, using infrared thermometers to measure each point on-site. This approach resulted in long inspection intervals, making it difficult to detect anomalies between inspections, and the distribution boxes were often deployed in remote locations such as poles and underground spaces, leading to high labor costs.

[0003] Remote monitoring systems based on a two-layer "end-cloud" architecture have achieved a shift from "manual inspection" to "remote online operation." However, in actual deployments, once the communication link is interrupted (unstable 4G signals, broken fiber optic cables, etc., are not uncommon in actual operation and maintenance), the cloud cannot obtain on-site data, and the entire monitoring system becomes ineffective. To address this, the industry has introduced edge computing technology, forming a three-tier collaborative architecture of "end-edge-cloud." Chinese invention patent CN117375002B (Southern Power Grid Science Research Institute, authorization announcement date January 6, 2026) discloses a cloud-edge fusion load regulation method and system based on an improved LSTM, which essentially involves "cloud training and edge inference." Its model parameters are synchronized from the cloud and cannot operate independently after a network outage. Chinese invention patent CN119225260A discloses an intelligent data acquisition method and system for industrial edge gateways, using LSTM to predict the data growth rate and adjusting the acquisition frequency based on acquisition pressure. However, its prediction object is the data growth rate rather than the physical trend of the load, and the adjustment purpose is to reduce the gateway burden rather than capture transient load characteristics. Gui Yuan et al. verified the feasibility of deploying lightweight LSTMs for temperature prediction on edge devices in the journal Sensors (2025), but their prediction object was temperature rather than load trend, and the prediction results were not linked to the acquisition strategy.

[0004] It is evident that existing technologies still face a fundamental problem: the lack of independent closed-loop decision-making capabilities at the edge. In existing solutions, model parameters rely on cloud synchronization (CN117375002A), anomaly detection is mostly based on fixed threshold comparisons (CN113177646A), and frequency adjustment is based on the data acquisition burden rather than the physical state of the load (CN119225260A). Furthermore, none of these solutions can independently complete the entire link from prediction to execution after a network outage. Even if some solutions achieve localized optimization at the edge, they only partially delegate cloud capabilities, rather than building complete autonomous logic at the edge. Therefore, it is necessary to propose an edge-autonomous distribution box status monitoring solution that can independently complete load trend perception, adaptive adjustment of data acquisition strategies, and comprehensive anomaly detection without relying on the cloud. Summary of the Invention

[0005] (a) Technical issues This invention addresses the fundamental deficiency in existing distribution box status monitoring systems where the edge side lacks independent closed-loop decision-making capabilities. It proposes a distribution box status monitoring method and system based on edge autonomy and dynamic load perception to solve the technical problem that edge gateways cannot independently complete trend prediction, data acquisition scheduling, and anomaly response in communication interruption scenarios.

[0006] (II) Technical Solution To address the aforementioned technical problems, this invention provides a method for monitoring the status of a distribution box based on edge autonomy and dynamic load perception, comprising: a data acquisition step, acquiring operational status data collected by sensors deployed at various monitoring points within the distribution box; a trend prediction step, where an edge gateway, based on locally stored historical load time-series data, calls a locally fixed prediction model to perform inference and outputs the load change trend for future periods; the model parameters of the prediction model are stored in a fixed form in the edge gateway, allowing it to complete inference independently without relying on cloud synchronization; a data acquisition and scheduling step, where the edge gateway determines the current load change state based on the load change trend and adjusts the data acquisition frequency of the sensors in stages according to the load change state; and anomaly detection. The steps include: the edge gateway performing a comprehensive anomaly judgment on the real-time collected data, comprising at least three dimensions: the current collected value dimension, the historical baseline value dimension, and the load change trend dimension; when the current collected value dimension is triggered, it is determined as an alarm; when the historical baseline value dimension and the load change trend dimension are triggered simultaneously but the current collected value dimension is not triggered, it is determined as a warning; the autonomous execution step involves the edge gateway performing protection actions locally without waiting for instructions from the cloud; and the data synchronization step involves the edge gateway monitoring the communication link status with the cloud, and caching data including the load change trend and collection frequency scheduling records locally when the link is interrupted, and retransmitting the data after the link is restored.

[0007] Furthermore, in the data acquisition and scheduling step, the tiered adjustment specifically includes: when the load change trend is characterized as a steady state, reducing the acquisition frequency to an energy-saving frequency; when the load change trend is characterized as a load ramp-up or transitional state, increasing the acquisition frequency to a medium-frequency frequency; when the load change trend is characterized as a rapid change or abrupt change, increasing the acquisition frequency to a high-frequency frequency; the energy-saving frequency is lower than the medium-frequency frequency, and the medium-frequency frequency is lower than the high-frequency frequency.

[0008] Furthermore, the prediction model is a lightweight LSTM model with no more than 2 network layers and no more than 64 hidden units, and the model parameters are compressed and stored in the edge gateway after being quantized by INT8 integers.

[0009] Furthermore, the sensors include passive wireless temperature sensors deployed at key electrical connection points within the distribution box. Each sensor acquires operating energy through a radio frequency field and has a unique ID code. The edge gateway sends frequency adjustment commands to the sensors via wireless radio frequency communication.

[0010] Furthermore, the passive wireless temperature sensor is deployed as follows: the reader antenna is installed in the non-metallic area of ​​the inner wall of the distribution box; all sensors are divided into several communication groups according to the internal space size of the distribution box and the position of the metal partition; each sensor in the same communication group sends data to the reader in different time slots; a relay forwarding node is added to the sensor located in the signal attenuation area to receive and forward the radio frequency signal of the sensor.

[0011] Furthermore, the edge gateway continuously monitors the communication conflict rate of each communication packet during operation. When the communication conflict rate exceeds a preset threshold, it automatically splits the corresponding communication packet into smaller sub-packets.

[0012] Furthermore, in the anomaly judgment step: the trigger condition for the current collected value dimension is: the real-time collected value exceeds a preset fixed alarm threshold; the trigger condition for the historical baseline value dimension is: the deviation between the real-time collected value and the historical baseline value under the same working conditions exceeds a set proportion; the trigger condition for the load change trend dimension is: the trend indication state output by the prediction model continues to deteriorate.

[0013] Furthermore, in the data synchronization step, the data retransmitted after the link is restored includes the prediction results output by the prediction model during the link interruption, as well as the time points and adjustment basis records of each acquisition frequency adjustment in the acquisition scheduling step.

[0014] A power distribution box status monitoring system based on edge autonomy and dynamic load perception includes: a sensor array deployed at various monitoring points within the power distribution box for collecting operational status data; an edge gateway deployed inside or adjacent to the power distribution box, comprising: a data aggregation and preprocessing unit for receiving data collected by the sensor array and performing verification and format unification; a local storage unit for storing historical time-series data; a prediction unit with a fixed prediction model for independently performing inference based on historical load time-series data in the local storage unit and outputting load change trends for future periods; a data acquisition scheduling unit for adjusting the data acquisition frequency of the sensor array according to the load change trends; an anomaly judgment unit for performing comprehensive anomaly judgment on real-time acquired data, including current acquisition value dimensions, historical baseline value dimensions, and load change trend dimensions; an autonomous execution unit for performing protection actions locally when an alarm or warning is detected; a communication and synchronization unit for monitoring the communication link status with the cloud and caching data locally when the link is interrupted and retransmitting it after the link is restored; and a cloud platform for receiving data uploaded by the edge gateway.

[0015] Furthermore, the sensor array includes passive wireless temperature sensors deployed at key electrical connection points within the distribution box, and electrical parameter acquisition modules installed on the incoming side and outgoing circuits of the distribution box; the edge gateway is connected to the passive wireless temperature sensors via wireless radio frequency communication and to the electrical parameter acquisition modules via a bus.

[0016] (III) Beneficial Effects Compared with existing technologies, this invention has the following advantages: 1. It realizes a complete autonomous decision-making closed loop at the edge in the event of communication interruption. This invention constructs a complete logical chain of "prediction → scheduling → judgment → execution" that does not rely on the cloud by fully localizing and solidifying the prediction model, autonomously scheduling the acquisition frequency, performing three-dimensional comprehensive judgment, and implementing local autonomous execution. Even if communication is completely interrupted, the edge gateway can still independently complete all monitoring and decision-making functions. The completeness of this logical chain is not disclosed in any single document in the existing technology, and its technical concept is fundamentally different from conventional edge caching or fixed threshold judgment in this field.

[0017] 2. This invention implements a feedforward-based data acquisition and scheduling mechanism based on load trend prediction. By predicting load change trends in advance, the data acquisition and scheduling mechanism switches frequencies before load changes occur. In steady-state conditions, it actively reduces the frequency to minimize data redundancy, and actively increases the frequency to capture transient characteristics during load changes. This "feedforward" mechanism changes the timing of data acquisition strategy adjustments from "after-the-fact" to "before-the-fact," fundamentally avoiding the risk of transient data loss. Its working principle is substantially different from the feedback-based or load-driven adjustments in existing technologies.

[0018] 3. This invention achieves proactive anomaly identification and false alarm suppression through three-dimensional fusion of "current collected values ​​+ historical baseline + predicted trend". It integrates three-dimensional information into a single judgment logic: when equipment slowly deteriorates but has not yet exceeded a fixed threshold, the historical deviation dimension and the trend deterioration dimension can trigger an early warning; conversely, when occasional interference causes a single-point surge but the historical baseline and predicted trend are normal, the system does not trigger an alarm. This three-dimensional collaborative mechanism simultaneously achieves proactive early warning and false alarm suppression within the same logical framework. Existing solutions typically require separate modules or reliance on cloud-based judgment to achieve both; this invention achieves this effect through logical reorganization at the edge.

[0019] 4. This invention enables traceable data retransmission during network outages. The retransmitted data includes prediction results and data acquisition frequency scheduling records, allowing the cloud to fully reconstruct the edge gateway's decision-making logic during the outage, providing complete decision chain data support for post-outage fault analysis. This effect is based on the autonomous logic chain of "prediction → scheduling → judgment → execution," and is a natural extension of this logic chain in the data synchronization stage. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall architecture of the distribution box status monitoring system of the present invention.

[0021] Figure 2 This is a functional block diagram of the edge gateway of the present invention.

[0022] Figure 3 This is a flowchart of the dynamic acquisition frequency hierarchical adjustment and three-dimensional anomaly judgment of the present invention.

[0023] Figure 4 This is a schematic diagram of the deployment of the passive wireless temperature sensor of the present invention in a power distribution box.

[0024] Figure 5 This is a graph showing the relationship between load changes and sampling frequency adjustments in an embodiment of the present invention.

[0025] Explanation of reference numerals in the attached diagram: 1—Distribution box enclosure; 2—Busbar connection point; 3—Circuit breaker inlet / outlet terminals (wiring terminals); 4—Cable connector; 5—Passive wireless temperature sensor; 6—Reader antenna; 7—Insulating board; 8—Relay relay node; 9—Metal partition; 10—Edge gateway. Detailed Implementation Example

[0026] This embodiment uses the deployment of a monitoring system for 12 low-voltage distribution boxes in an industrial park as an example. The park is located in a suburban area with unstable 4G signals, and had previously experienced two unplanned power outages due to overheating of cable joints.

[0027] System deployment and sensor installation, such as Figure 1 As shown, one edge gateway 10 is installed in each distribution box. Three main considerations were taken into account during the selection process: first, power consumption needed to be controlled below 3W to avoid excessive heat generation within the enclosed space of the distribution box; second, the processor needed to support fixed-point arithmetic acceleration to adapt to the inference of the quantized predictive model; and third, local storage needed to meet at least 7 days of data caching requirements during network outages. This embodiment uses an ARM Cortex-A7 dual-core processor (1.2GHz) equipped with 2GB of eMMC local storage. This configuration is common in industrial-grade edge gateways, offering controllable cost and sufficient computing power to support lightweight model inference (the measured inference time for a single iteration is approximately 15ms).

[0028] A three-phase electrical parameter acquisition module is installed at the incoming end of the distribution box, with the range selected according to the rated capacity of the distribution box (rated current 400A, transformer ratio 800:5 in this embodiment). Passive wireless temperature sensors 5 are installed at the inlet and outlet terminals 3 of the circuit breakers and cable joints 4 of each outgoing circuit. These sensors are radio frequency power-generating sensors based on the surface acoustic wave (SAW) principle, with a temperature measurement range of -25℃ to +150℃ and an accuracy of ±1℃.

[0029] It should be noted that the distribution box, as a closed metal cavity, suffers from severe shielding and multipath reflection effects on radio frequency signals. Communication coverage between the sensors and the reader is the primary engineering challenge during deployment. To address this issue, this embodiment adopts the following deployment strategy: The reader operates at the 433MHz industrial frequency band. This band has better penetration capability within the metal cavity than 2.4GHz, but the communication rate is lower, requiring a trade-off between penetration capability and data throughput. In this embodiment, the number of sensors in each distribution box does not exceed 16, and the 433MHz frequency band is sufficient to support this.

[0030] Sensor array deployment in a metallic cavity environment, such as Figure 4 As shown, the distribution box measures 800mm wide × 600mm deep × 1800mm high, with an internal metal partition 9 separating the incoming and outgoing cable areas. During deployment, the reader antenna 6 is installed on the insulating plate 7 inside the distribution box door. This location is made of non-metallic material, resulting in minimal RF signal attenuation. The entire box contains 16 sensor temperature measurement points, divided into two communication groups based on spatial location: eight on the upper layer (incoming busbar 2 and upper-layer outgoing circuits) and eight on the lower layer (lower-layer outgoing circuits and cable connectors 4). Each of the eight sensors within the same group is assigned a different time slot offset (from 0ms to 70ms, with a step size of 10ms) to ensure that data is not simultaneously sent to the reader. For the two cable connector sensors located deep at the bottom of the distribution box (whose signal attenuation exceeded 20dB in actual measurements), a relay node 8 is added at the cable tray between them and the reader to relay the sensor signals to the reader.

[0031] It should be noted that the division of communication groups is not a one-time process and remains unchanged throughout the entire process. During debugging, it was found that when multiple sensors within a group are simultaneously scheduled to high-frequency acquisition mode (e.g., every 30 seconds), the communication conflict rate increases from the normal less than 1% to approximately 8% to 12%. The root cause of this problem is that the wake-up and response frequency of each sensor increases in high-frequency acquisition mode. In the multipath reflection environment of the metal cavity, signal delay jitter is aggravated, and the time slot allocation scheme that works reliably in low-frequency mode is no longer applicable in high-frequency mode. To address this, the edge gateway 10 continuously monitors the communication conflict rate of each group during operation. When the conflict rate exceeds 5%, it automatically splits the group into smaller subgroups or activates a backup relay channel. This dynamic matching mechanism between communication resources and acquisition frequency is key to ensuring communication reliability in variable-frequency acquisition mode and is a necessary design determined after repeated debugging in engineering deployment.

[0032] The parameter configuration and prediction model deployment are based on a base acquisition frequency of 5 minutes / time, an energy-saving frequency of 20 minutes / time, a medium-frequency frequency of 2 minutes / time, and a high-frequency frequency of 30 seconds / time. These values ​​are determined based on industry norms (environmental sensors typically operate at 15 to 30 minutes / time) and the design philosophy of this solution: a higher base frequency and a wider dynamic adjustment range. The higher base frequency ensures sufficient data density under normal conditions, while the wider adjustment range provides ample room for energy saving and data encryption. In actual deployment, adjustments can be made according to specific scenarios. For example, for distribution boxes with frequent load fluctuations, the energy-saving frequency can be appropriately increased to 15 minutes / time to shorten response latency.

[0033] It's important to note that in conventional fixed-frequency acquisition schemes, designers face a dilemma: setting the acquisition frequency to a high frequency (e.g., 30 seconds / time) leads to a surge in sensor power consumption, congested wireless communication channels, and rapid saturation of edge gateway data storage; setting it to a low frequency (e.g., 20 minutes / time) inevitably results in the loss of transient data during load surges. This contradiction is irreconcilable in traditional schemes, requiring a compromise based on experience. In this embodiment, however, this contradiction is resolved through a tiered adjustment mechanism in the acquisition scheduling steps: under normal conditions, low-frequency operation conserves resources; once the prediction model identifies a load change trend, the system proactively increases the frequency before the actual change occurs.

[0034] The following thresholds are used to determine load change states: steady state corresponds to a load change rate of less than 5% per minute; transient state corresponds to a load change rate between 5% and 15% per minute; and abrupt change state corresponds to a load change rate greater than 15% per minute. These thresholds are set based on the following: normal load fluctuations in low-voltage distribution boxes generally do not exceed ±3% per minute (referring to the recommended values ​​for statistical characteristics of low-voltage distribution loads in IEC 60364 standard); fluctuations exceeding 5% per minute indicate significant load change events; and fluctuations exceeding 15% per minute indicate impulsive loads or abnormal operating conditions, requiring maximum monitoring density. These two thresholds can be fine-tuned during the commissioning phase based on specific load characteristics. For example, in circuits primarily loaded with motors, the first threshold can be appropriately relaxed to 8% per minute to avoid unnecessary frequency increases during normal motor startup.

[0035] The prediction model employs a lightweight LSTM with a single-layer structure and 32 hidden units. Before deployment, the model was trained offline in the development environment using historical load data (current and power data for 30 consecutive days with a 1-minute sampling interval) from the park's power distribution boxes. After training, the 32-bit floating-point model parameters were compressed using INT8 integer quantization, resulting in a model file size of approximately 120KB, which was then stored in the local flash memory of the edge gateway. The prediction window was set to 10 minutes, with an input sequence length of 30 time points (corresponding to 150 minutes of historical data). The 10-minute window was chosen based on the following trade-offs: a window that is too short (e.g., 2 to 3 minutes) leaves insufficient time for adjusting the sampling frequency; a window that is too long (e.g., 30 to 60 minutes) reduces prediction accuracy and increases the computational power consumption of the edge gateway. Actual testing shows that, with a 10-minute window, the model's inference error (MAPE) on the edge gateway is approximately 6.8%, meeting the requirements for engineering applications. The loss of accuracy caused by quantization is within an acceptable range (the inference error increased from 6.2% in FP32 to 6.8%), but the model size was reduced by about 75% and the inference speed was increased by about 4 times, which is a necessary trade-off for the limited computing power of the edge gateway.

[0036] After the system is put into operation, it will run in a cyclical manner according to the following process. For example... Figure 3 As shown, the sensor collects data at a base frequency of 5 minutes and uploads it to the edge gateway 10. After each data collection, the edge gateway 10 performs a trend prediction step, retrieving the load data of the past 150 minutes from local storage, inputting it into the prediction model for inference, and outputting the load change trend for the next 10 minutes.

[0037] In a concentrated equipment startup event in the workshop (the load increased from 40% to 85% of rated capacity within 3 minutes, a change rate of approximately 15% / minute), the predictive model identified the rapid load increase trend in the second minute. Based on this, the data acquisition and scheduling unit increased the acquisition frequency from 5 minutes / time to 2 minutes / time when the load change rate exceeded the first threshold (in the third minute), and further increased it to 30 seconds / time when the load change rate exceeded the second threshold (in the fourth minute). This proactive scheduling enabled the system to fully record the current waveform and temperature change data throughout the entire load increase process, totaling approximately 180 valid data points. If the frequency were fixed at 5 minutes / time, only about 8 data points could be recorded. Figure 5 As shown, the sampling frequency is adjusted as soon as the predictive model identifies the upward trend, rather than waiting until the load change rate actually exceeds the threshold. This is the essential characteristic that distinguishes "feedforward" scheduling from "feedback" scheduling.

[0038] During the incident, temperature sensor readings showed that the temperature at the outgoing terminal of a circuit breaker in a certain outgoing circuit rose from 42℃ to 67℃ (the rated temperature rise threshold for this circuit is 75℃). The judgment logic of the anomaly detection steps is as follows: the first dimension (current value 67℃ < 75℃) was not triggered; the second dimension (current value 67℃ deviates from the historical baseline of 48℃ by approximately 40%, exceeding the set deviation ratio of 30%) was triggered; the third dimension (the trend output by the prediction model indicates that it will continue to rise in the next 10 minutes) was triggered. The second and third dimensions were triggered simultaneously, while the first dimension was not triggered, which was judged as a "warning". The edge gateway 10 then generated a warning signal locally and activated the audible and visual alarm (autonomous execution step). After receiving the warning, the maintenance personnel intervened in advance and found that the wiring terminals of the circuit were loose, resulting in increased contact resistance. After timely tightening, the temperature returned to the normal range.

[0039] Here, it's worth noting a side effect of the three-dimensional collaborative mechanism: During the initial commissioning phase, there were several instances where the temperature of high-power equipment briefly spiked upon startup but then quickly dropped back down. In these situations, the first dimension might trigger an alarm, but the second dimension (no deviation from the historical baseline) and the third dimension (no indication of a continuously deteriorating trend) remained normal, preventing the system from outputting an alarm and effectively avoiding false alarms. Conversely, when the contact resistance increases slowly, the temperature may not exceed a fixed threshold, but the second dimension (deviation from the historical baseline) and the third dimension (deteriorating trend) can identify potential problems in advance. This three-dimensional collaborative mechanism improves early warning sensitivity while maintaining the reliability of the judgment, two things that are usually difficult to achieve simultaneously. This solution achieves this goal through the combination of three-dimensional logic.

[0040] During the three-month trial operation, the system experienced four communication outages, the longest lasting approximately 45 minutes. During each outage, Edge Gateway 10 independently and cyclically executed trend prediction, data collection and scheduling, anomaly detection, and autonomous execution steps. A total of 12 warning records were generated during the outages, two of which corresponded to temperature anomalies that were subsequently verified. All raw data collected during the outages, the prediction results output by the prediction model, and the data collection frequency scheduling records (including the time points and basis for each frequency adjustment) were completely written to local storage and automatically re-uploaded to the cloud after network recovery. The completeness of the re-uploaded data allows maintenance personnel to fully reconstruct the edge gateway's decision-making process during the outages, including when and why the data collection frequency was adjusted, when the warnings were triggered, and which dimensions were used for judgment. This is of significant value for root cause analysis. Example

[0041] This embodiment uses the deployment of a monitoring system in an underground power distribution room (8 low-voltage distribution cabinets) of a commercial complex as an example. The power distribution room is located on the second basement level, where mobile communication signals are extremely weak (only a weak 2G signal), making traditional cloud-dependent monitoring systems basically unusable.

[0042] The main difference from Embodiment 1 lies in the communication strategy: due to the extremely weak communication signal, the edge gateway 10 is configured in "timed upload" mode, which does not upload data by default. It only connects the Wi-Fi module via a timed switch between 2:00 AM and 3:00 AM daily (during the off-peak period after the mall closes) to upload the previous day's data summary to the cloud. At other times, the edge gateway operates completely independently. Furthermore, the power distribution room is small, with power distribution cabinets arranged back-to-back with a gap of less than 300mm between them. Some temperature measurement points are located deep within the cabinets (more than 600mm from the cabinet door), further exacerbating the signal coverage difficulties in the aforementioned metal cavity environment. During deployment, one reader antenna 6 (a total of 8) is installed on the inner insulating plate of each power distribution cabinet door. The number of sensors in each communication group is controlled to within 6, and the time slot interval is extended to 70ms to overcome signal interference caused by multipath reflections within the metal cavity.

Claims

1. A method for monitoring the status of a distribution box based on edge autonomy and dynamic load sensing, characterized in that, include: The data acquisition step involves acquiring the operating status data collected by sensors deployed at various monitoring points within the distribution box; In the trend prediction step, the edge gateway uses locally stored historical load time-series data to call a locally deployed prediction model to perform inference and output the load change trend in the future period. The model parameters of the prediction model are stored in the edge gateway in a fixed form and can complete inference independently without relying on cloud synchronization. In the data acquisition and scheduling step, the edge gateway determines the current load change status based on the load change trend, and adjusts the data acquisition frequency of the sensor in stages according to the load change status. The anomaly detection step involves the edge gateway performing a comprehensive anomaly detection on the real-time collected data, which includes at least three dimensions: the current collected value dimension, the historical baseline value dimension, and the load change trend dimension. When the current collected value dimension is triggered, it is determined as an alarm; when the historical baseline value dimension and the load change trend dimension are triggered simultaneously but the current collected value dimension is not triggered, it is determined as a warning. In the autonomous execution steps, when an alarm or warning is detected, the edge gateway does not wait for instructions from the cloud and directly executes protection actions locally; In the data synchronization step, the edge gateway monitors the status of the communication link with the cloud. When the link is interrupted, it caches the data, including the load change trend and the collection frequency scheduling record, locally and retransmits it after the link is restored.

2. The method according to claim 1, characterized in that, In the data acquisition and scheduling step, the hierarchical adjustment specifically includes: When the load change trend is characterized as a steady state, the sampling frequency is reduced to an energy-saving frequency; When the load change trend is characterized as a load ramp-up or transition state, the sampling frequency is increased to the intermediate frequency. When the load change trend is characterized as a rapid change or abrupt change, the sampling frequency is increased to a high frequency. The energy-saving frequency is lower than the intermediate frequency, and the intermediate frequency is lower than the high frequency.

3. The method according to claim 1, characterized in that, The prediction model is a lightweight LSTM model with no more than 2 network layers and no more than 64 hidden units. The model parameters are compressed using INT8 integer quantization and then stored in the edge gateway.

4. The method according to claim 1, characterized in that, The sensors include passive wireless temperature sensors deployed at key electrical connection points within the distribution box. Each sensor acquires operating energy through a radio frequency field and has a unique ID code. The edge gateway sends frequency adjustment commands to the sensors via wireless radio frequency communication.

5. The method according to claim 4, characterized in that, The passive wireless temperature sensor is deployed as follows: Install the reader antenna in the non-metallic area of ​​the inner wall of the distribution box; Based on the internal space dimensions of the distribution box and the position of the metal partition, all sensors are divided into several communication groups. Each sensor in the same communication group sends data to the reader in different time slots. For sensors located in signal attenuation areas, a relay forwarding node is added to receive and forward the sensor's radio frequency signals.

6. The method according to claim 5, characterized in that, During operation, the edge gateway continuously monitors the communication conflict rate of each communication group. When the communication conflict rate exceeds a preset threshold, it automatically splits the corresponding communication group into smaller subgroups.

7. The method according to claim 1, characterized in that, In the anomaly detection step: The trigger condition for the current collected value dimension is: the real-time collected value exceeds a preset fixed alarm threshold; The trigger condition for the historical baseline value dimension is: the deviation between the real-time collected value and the historical baseline value under the same working conditions exceeds a set proportion; The trigger condition for the load change trend dimension is that the trend indication state output by the prediction model continues to deteriorate.

8. The method according to claim 1, characterized in that, In the data synchronization step, the data retransmitted after the link is restored includes the prediction results output by the prediction model during the link interruption, as well as the time point and adjustment basis record of each acquisition frequency adjustment in the acquisition scheduling step.

9. A distribution box status monitoring system based on edge autonomy and dynamic load sensing, characterized in that, include: Sensor arrays are deployed at various monitoring points within the distribution box to collect operational status data; Edge gateways, deployed inside or adjacent to distribution boxes, include: The data aggregation and preprocessing unit is used to receive the data collected by the sensor array and perform verification and format unification. Local memory is used to store historical time-series data; The prediction unit is equipped with a prediction model, which is used to independently perform inference based on the historical load time series data in the local memory and output the load change trend in the future period. A data acquisition and scheduling unit is used to adjust the data acquisition frequency of the sensor array in stages according to the load change trend. The anomaly detection unit is used to perform comprehensive anomaly detection on real-time acquired data, including the dimensions of current acquired value, historical baseline value, and load change trend. Autonomous execution unit, used to perform protective actions locally when an alarm or warning is detected; The communication and synchronization unit is used to monitor the status of the communication link with the cloud, and to cache data locally when the link is interrupted and retransmit it after the link is restored. A cloud platform is used to receive data uploaded by the edge gateway.

Images