Edge computing-based power inspection image recognition computing power scheduling method and system

Abstract

The application relates to the technical field of electric power inspection, and discloses an electric power inspection image recognition computing power scheduling method and system based on edge calculation. The method comprises the following steps: an edge collection hub obtains image data and sensing data, obtains computing power use data of the edge collection hub, and calculates a computing power occupation value; if the computing power occupation value is greater than a set occupation value, the computing power occupation values of other edge collection hubs are obtained, and the other edge collection hubs are marked as distribution nodes; a distribution node with the smallest computing power occupation value is selected, the current edge collection hub sends image data and corresponding algorithm codes to the distribution node; the distribution node processes the image data to obtain processing data; after the processing is completed, a completion signal is returned, and the current edge collection hub reads the processing data returned by the distribution node in response to the completion signal; and the computing power utilization efficiency of the edge calculation equipment is improved.

Description

Technical Field

[0001] This application relates to the technical field of power line inspection, and in particular to a power line inspection image recognition computing power scheduling method and system based on edge computing. Background Technology

[0002] As the core infrastructure for energy supply, the power grid is becoming increasingly complex and its scale expands. Automated inspection has become an essential means to ensure the safety of the power grid. It can achieve all-weather, wide-coverage equipment status monitoring, promptly detect potential faults, reduce the risk of power outages, and provide key support for the reliable operation of the power system.

[0003] Current power line inspections primarily rely on high-definition image recognition technology to detect equipment defects, such as insulator damage and broken conductor strands. However, single visual data is easily affected by factors such as lighting and weather, limiting diagnostic accuracy. To further improve the accuracy and reliability of defect identification, the industry trend is shifting towards multimodal data comprehensive analysis. This involves fusing visual data such as visible light and infrared images collected by inspection equipment with physical parameters such as temperature, humidity, and vibration obtained in real time from sensors on the power facilities themselves to construct a multi-dimensional equipment condition assessment model, enabling more comprehensive fault diagnosis.

[0004] Multimodal data comprehensive analysis requires simultaneous feature extraction from image data and real-time calculation of physical parameters, placing extremely high demands on computing power. However, edge inspection devices are limited by size, power consumption, and cost, resulting in limited computing resources that can only perform simple image preprocessing or parameter acquisition. While the cloud possesses powerful computing capabilities, it relies on network transmission of massive amounts of multimodal data, which is prone to transmission delays and excessive bandwidth consumption. Furthermore, it cannot respond to inspection needs in real time when the network is interrupted, creating a huge gap between edge computing power and cloud computing power, thus hindering the practical application of multimodal inspection technology. Summary of the Invention

[0005] To improve the computing power utilization efficiency of edge computing devices, this application provides a computing power scheduling method and system for power inspection image recognition based on edge computing.

[0006] Firstly, this application provides a power grid inspection image recognition computing power scheduling method based on edge computing, which adopts the following technical solution:

[0007] A power inspection image recognition computing power scheduling method based on edge computing includes: edge acquisition hubs acquiring and preprocessing image data; acquiring computing power usage data of the edge acquisition hubs and calculating computing power occupancy values ​​based on set usage data; if the computing power occupancy value is greater than the set occupancy value, then selecting other edge acquisition hubs with computing power occupancy values ​​less than the set occupancy value as distribution nodes; sending image data and corresponding algorithm codes to the distribution nodes; calculating computing power demand values ​​based on the computing power occupancy value and the set occupancy value, and adjusting the amount of image data distributed based on the positive correlation with the computing power demand value.

[0008] The distributed nodes retrieve the image data from the image algorithm according to the algorithm code, process the image data, and return the processed data; obtain the label calculation time required for the distributed nodes to process the image data, match the label reference time according to the amount of image data, if the label calculation time is greater than the label reference time, calculate the label computing power comparison value according to the label calculation time and the label reference time, and amplify the computing power occupancy value according to the label computing power comparison value.

[0009] By adopting the above technical solution, the edge acquisition hub acquires image data and sensor data, processes these two types of data using a preset preprocessing algorithm, obtains its own computing power usage data, and calculates the computing power occupancy value based on the preset usage data, providing a precise initial basis for computing power scheduling. When the computing power occupancy value exceeds the preset occupancy value, the edge acquisition hub promptly obtains the computing power occupancy values ​​of other edge acquisition hubs, marks other edge acquisition hubs with computing power occupancy values ​​less than the preset occupancy value as distribution nodes, and selects the distribution node with the smallest computing power occupancy value. Simultaneously, it adjusts the amount of image data distributed based on the positive correlation between the computing power occupancy value and the computing power demand value calculated from the preset occupancy value, avoiding waste of computing power resources and data... According to the transmission overload, it effectively compensates for the limited computing power of a single edge acquisition hub; the distributed nodes retrieve the corresponding image algorithm according to the algorithm code, process the image data, and after obtaining the processed data, return a completion signal to the current edge acquisition hub. The current edge acquisition hub responds to the completion signal and reads the processed data, ensuring the continuity and real-time performance of multimodal data processing; by obtaining the label calculation time required by the distributed nodes from acquiring image data to obtaining processed data, and comparing it with the corresponding label reference time, if the label calculation time is longer, a label computing power comparison value is calculated based on the two, and the computing power occupancy value is amplified accordingly to further optimize the computing power scheduling accuracy, ultimately realizing real-time movement of computing power and improving the computing power utilization efficiency of multimodal data comprehensive processing.

[0010] Optionally, the method further includes the following steps:

[0011] When distributing image data, the current edge acquisition hub extracts temporary data from the end of the image data sent to the distributed nodes;

[0012] The current edge acquisition hub uses temporary data to calculate temporary end results. After the image data distribution is completed, the location of the temporary data and the temporary end results are sent to the distribution nodes.

[0013] After the distributed nodes generate the processing data, they analyze the location of the temporary data and extract the temporary processing results corresponding to the location of the temporary data from the processing data.

[0014] Verify the consistency between the temporary end result and the temporary processing result. If the verification fails, return a scheduling computing power processing anomaly signal to the current edge acquisition hub.

[0015] By adopting the above technical solution, anomalies in the image data distribution or processing process can be detected in a timely manner, avoiding the impact of erroneous data processing on the accuracy of multimodal data comprehensive analysis; the return of the scheduling computing power to process the anomaly signal when an anomaly occurs helps to ensure the reliability of the data.

[0016] Optionally, the method further includes the following steps:

[0017] The time taken for the edge acquisition hub to process sensor data using a preset sensing algorithm and obtain sensing results is the sensing duration.

[0018] The time taken from when the edge acquisition hub sends the algorithm code to when it obtains the corresponding processing data is recorded as the distributed computation time.

[0019] If the sensing duration is longer than the distributed computation duration, the sensing data and the corresponding sensing code are sent to the distributed node. The sensing demand value is calculated based on the sensing duration and the distributed computation duration, and the amount of sensing data distributed is adjusted according to the positive correlation of the sensing demand value.

[0020] The distributed nodes retrieve the corresponding sensing algorithm based on the sensing code to process the sensing data and obtain the analysis data; after completion, they return an end signal to the current edge acquisition hub, which then reads the analysis data returned by the distributed nodes in response to the end signal.

[0021] By adopting the above technical solutions, the computing power allocation for sensor data processing can be optimized, avoiding uneven load on edge acquisition hubs caused by excessively long sensor processing times; and the computing power load on edge and distributed nodes can be balanced by rationally scheduling sensor data to distributed nodes for processing.

[0022] Optionally, the method further includes the following steps:

[0023] The current edge acquisition hub extracts the latest field data from the sensor data sent to the distributed nodes;

[0024] The current edge acquisition hub uses field data to calculate the final field results and extracts the field analysis results corresponding to the field data from the analysis data.

[0025] Verify the consistency between the final on-site results and the on-site analysis results. If the verification fails, issue an error message regarding the scheduling computing power processing.

[0026] By adopting the above technical solutions, anomaly alerts can promptly address data processing issues, reduce the waste of computing power caused by data errors, and help ensure the stability of real-time computing power movement. This provides additional data verification and anomaly warnings to improve the reliability of power inspection defect diagnosis and ensure the stable implementation of multimodal inspection technology.

[0027] Optionally, the step of obtaining computing power usage data of the edge acquisition hub may also include the following steps:

[0028] Call the algorithm that processes the acquired data in the interrupt function;

[0029] Call the detection algorithm that generates preset fixed data in the non-interruptible function;

[0030] Calculate the time interval for generating fixed data and the time interval for calling the detection algorithm;

[0031] The interval average and interval uniformity are calculated based on multiple generation time intervals.

[0032] If the average interval is within the preset average range and the uniform interval is within the preset uniform range, then the computing power usage data is calculated based on the generation time interval and the call time interval; otherwise, the computing power usage data is the maximum value.

[0033] By adopting the above technical solution, the accuracy and validity of the computing power usage data of the edge acquisition hub can be ensured. When the time conditions are not met, the computing power usage data is set to the maximum value, which can respond to computing power anomalies in a timely manner and reduce scheduling deviations caused by computing power anomalies.

[0034] Optionally, the step of calculating the computing power occupancy value based on the computing power usage data and the preset usage data further includes the following steps:

[0035] Computing power consumption = Generation time interval / Call time interval;

[0036] Adjust the average range based on a positive correlation with computing power occupancy, and adjust the uniform range based on a negative correlation with computing power occupancy.

[0037] By adopting the above technical solutions, adjusting the average range based on the positive correlation of computing power occupancy value helps improve the accuracy of computing power usage data, while adjusting the uniform range based on the negative correlation of computing power occupancy value helps improve compatibility with fluctuations in computing power usage data.

[0038] Optionally, the step of obtaining the computing power occupancy values ​​of other edge acquisition hubs may also include the following steps:

[0039] The current edge acquisition hub searches for other nearby edge acquisition hubs based on a preset first communication protocol;

[0040] If other edge acquisition hubs are found, a handshake is established with the other edge acquisition hubs based on the first communication protocol to establish the first link; otherwise, the computing power occupancy value of the other edge acquisition hubs is set to the minimum value.

[0041] Based on the first link, obtain the signal strength of other edge acquisition hubs; if the signal strength is greater than the preset reference strength value, obtain the computing power occupancy value of other edge acquisition hubs; otherwise, set the computing power occupancy value of other edge acquisition hubs to the minimum value.

[0042] By adopting the above technical solutions, it is beneficial to ensure that the selected distributed nodes can participate in the scheduling stably, improve the effectiveness of real-time computing power movement, and provide a stable computing power scheduling foundation for the efficient processing of multimodal data in power inspection.

[0043] Optionally, the following steps may also be included:

[0044] If the signal strength is greater than the preset reference strength value, the control distributed node will handshake with other edge acquisition hubs and establish a second link;

[0045] The marked edge acquisition hub returns a secondary link signal to the current edge acquisition hub, and the current edge acquisition hub responds to the secondary link signal by invoking a forwarding command;

[0046] The marked edge acquisition hub responds to the received forwarding instruction by forwarding the image data and the corresponding algorithm code to other edge acquisition hubs;

[0047] Forward the data generated by other edge acquisition hubs and corresponding to the forwarding instructions to the current edge acquisition hub.

[0048] By adopting the above technical solutions, the coverage of computing power scheduling is expanded, the stability of image data and algorithm code forwarding is ensured, and computing power scheduling is made more flexible.

[0049] Optionally, the following steps may also be included:

[0050] The secondary signal comparison value is calculated based on the signal strength and the reference strength value. The amount of image data corresponding to the forwarding command is adjusted and forwarded based on the positive correlation of the secondary signal comparison value.

[0051] By adopting the above technical solutions, the amount of data forwarded can be adapted to the signal, improving the adaptability of computing power scheduling, ensuring smooth multimodal data processing, and providing transmission guarantee for efficient analysis of multimodal data in power inspection.

[0052] Secondly, this application provides a power inspection image recognition computing power scheduling system based on edge computing, which adopts the following technical solution:

[0053] A power line inspection image recognition computing power scheduling system based on edge computing includes a processor, wherein the processor executes the steps of the power line inspection image recognition computing power scheduling method based on edge computing as described in any one of the above. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the steps of a power inspection image recognition computing power scheduling method based on edge computing. Detailed Implementation

[0055] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0056] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0057] This application discloses a power dispatching method for power line inspection image recognition based on edge computing, referring to... Figure 1 It includes the following steps:

[0058] The edge acquisition hub synchronously acquires image and sensor data from power equipment through integrated image acquisition devices and sensors. In this embodiment, the edge acquisition hub is a mobile terminal carried by inspection personnel, such as an intelligent inspection instrument equipped with an edge computing module; the image acquisition device is such as a high-definition camera or an infrared thermal imager; the sensors are such as temperature sensors, current transformers, and voltage sensors; wherein:

[0059] The image data includes visible light images for identifying defects in the equipment's appearance, and infrared images for detecting abnormal temperature distribution in the equipment.

[0060] The sensor data includes real-time electrical parameters (current, voltage), environmental parameters (humidity, vibration), and infrared temperature data of the equipment.

[0061] The edge acquisition hub performs initial caching of the acquired data in preparation for subsequent processing.

[0062] The edge acquisition hub invokes preset preprocessing algorithms to preprocess the acquired image and sensor data, such as image noise reduction, size normalization, sensor data filtering, and outlier removal. During preprocessing, it collects its own computing power usage data in real time, including but not limited to CPU utilization, memory utilization, and data processing time.

[0063] Based on computing power usage data and preset usage data, namely the computing power thresholds of the edge acquisition hub, such as a maximum CPU load of 80% and a maximum memory usage of 70%, the current computing power usage of the edge acquisition hub is calculated. The specific calculation formula is as follows: Computing power usage = Actual computing power usage data / Preset usage data;

[0064] For example, if the actual CPU utilization rate is 60%, and the CPU threshold in the usage data is set to 80%, then the corresponding dimension of computing power utilization is 60% / 80%=0.75. After combining the computing power usage data of multiple dimensions (CPU, memory, time consumption), the final computing power utilization value is obtained, which ranges from 0 to 1. The larger the value, the higher the computing power load.

[0065] If the calculated computing power occupancy value is greater than the preset occupancy value, such as 0.8, which means the computing power load exceeds 80%, the current edge acquisition hub determines that its own computing power is insufficient and needs to request computing power support from other edge acquisition hubs.

[0066] The specific scheduling process is as follows:

[0067] Search and obtain the computing power status of other edge acquisition hubs: The current edge acquisition hub searches for other edge acquisition hubs in the same inspection network nearby, such as mobile terminals carried by other inspection personnel, through preset short-range communication protocols, such as 5G LAN, Bluetooth Mesh, ZigBee, etc., and obtains their computing power usage value.

[0068] Filtering distributed nodes: From the searched edge acquisition hubs, filter out devices with computing power occupancy values ​​less than the set occupancy value and mark them as distributed nodes; if there are multiple distributed nodes, prioritize the device with the smallest computing power occupancy value, that is, the device with the least computing power, as the target scheduling node.

[0069] Dynamically adjust data distribution volume: Calculate the computing power requirement based on the difference between the current computing power occupancy value of the edge acquisition hub and the set occupancy value. For example: Computing power requirement value = (Computing power occupancy value - Set occupancy value) / Set occupancy value. Adjust the amount of image data to be distributed according to the positive correlation with the computing power requirement value; that is, the higher the computing power requirement, the larger the amount of data distributed, to fully utilize the idle computing power of the target node. For example, if the computing power requirement value is 0.2, 20% of the total image data can be distributed to the distribution node; if the computing power requirement value is 0.5, then 50% of the data will be distributed.

[0070] Data and Algorithm Code Distribution: The current edge acquisition hub will send the filtered image data and the corresponding algorithm code, such as the algorithm number ALG-001 for insulator defect identification and the algorithm number ALG-002 for infrared image temperature analysis, to the selected distribution node.

[0071] After receiving the image data and algorithm code, the distributed node retrieves the corresponding image processing algorithm from the local algorithm library according to the algorithm code, such as a deep learning-based defect detection model or an image feature extraction algorithm, to process the image data and obtain processed data containing equipment defect information, such as defect location coordinates, defect type, and confidence level.

[0072] After processing is complete, the distributed node returns a completion signal to the current edge acquisition hub, such as an acknowledgment message containing a summary of the processed data. Upon receiving the completion signal, the current edge acquisition hub reads the processed data returned by the distributed node and performs fusion analysis with the locally processed sensor data or image data.

[0073] To further improve the accuracy of computing power scheduling, this embodiment introduces a processing time feedback mechanism:

[0074] Obtain the actual processing time: The total time taken by the current edge acquisition hub to record the distributed nodes from receiving image data to returning processed data is denoted as the mark calculation time.

[0075] Matching reference duration: Based on the amount of image data distributed, such as 100 images or 200MB of data, the corresponding marker reference time is matched from the preset duration mapping table. This is the theoretical processing time of the data volume under standard computing power. For example, the reference time for 200MB of data is 10 seconds.

[0076] Correcting Computing Power Utilization: If the marking calculation time is longer than the marking reference time (e.g., actual time 15 seconds > reference time 10 seconds), it indicates that the actual computing power of the distributed nodes may be lower than the initial assessment value. In this case, calculate the marking computing power comparison value, for example: Marking computing power comparison value = Marking calculation time / Marking reference time. Based on this value, amplify and correct the computing power utilization value of the distributed nodes. For example, if the original computing power utilization value is 0.3 and the marking computing power comparison value is 1.5, then the corrected value is 0.3 × 1.5 = 0.45. The corrected computing power utilization value will serve as the basis for subsequent scheduling to avoid unreasonable computing power allocation due to initial assessment deviations.

[0077] By monitoring the computing power status of edge acquisition hubs in real time, data is dynamically distributed to idle nodes, avoiding overload of individual edge devices. Simultaneously, data volume is adjusted based on computing power demand, reducing resource waste and transmission pressure. Relying on direct communication between edge nodes eliminates the need for cloud transmission, reducing data latency and ensuring real-time analysis of device status during inspections, making it particularly suitable for field inspection scenarios with unstable networks. By using processing time feedback to correct computing power assessments, scheduling decisions are made more closely aligned with actual computing power status, improving the reliability of multimodal data processing.

[0078] In this embodiment of the application, to further ensure the reliability of the image data distribution and processing process, the power grid inspection image recognition computing power scheduling method based on edge computing also includes an image data verification step, as follows:

[0079] When the current edge acquisition hub distributes image data to the distributed nodes, a data consistency verification mechanism is simultaneously initiated. First, the current edge acquisition hub extracts a segment of data from the end of the data sequence in the image data to be sent to the distributed nodes as temporary data. The length of the temporary data can be dynamically set according to the total capacity of the image data. For example, if the distributed image data is a data packet containing 100 frames of visible light images (total size 500MB), the last 10 frames (corresponding to 50MB of data) can be selected as temporary data; if it is a single high-definition infrared image (size 20MB), the last 10% of the pixel data in the image pixel matrix (corresponding to 2MB of data) can be selected as temporary data, ensuring that the temporary data reflects the characteristic attributes of the original image.

[0080] The current edge acquisition hub uses verification algorithms consistent with the subsequent processing algorithms of the distributed nodes, such as hash value calculation algorithms and feature value extraction algorithms, to perform calculations on the extracted temporary data and obtain temporary end results. For example, the SHA-256 hash algorithm is used to encrypt the temporary data and generate a unique hash value as the temporary end result; or an image feature extraction algorithm (such as the SIFT algorithm) is used to extract a set of feature points from the temporary image data, and the vector representation of this set of feature points is used as the temporary end result.

[0081] After the image data is fully distributed to the distribution nodes, the current edge acquisition hub immediately sends the location information of the temporary data in the original image data and the aforementioned temporary end result to the distribution nodes. The location information, such as the data start byte index, end byte index, or image frame sequence number range, can be reused with the image data transmission protocol or a separate lightweight protocol can be set through a preset verification information transmission protocol.

[0082] After receiving image data and processing it using the corresponding image algorithm to generate processed data, the distributed nodes parse the location information of the temporary data received from the current edge acquisition hub. Based on this location information, they locate and extract the processing segment corresponding to the temporary data from their own generated processed data as the temporary processing result. For example, if the temporary data is the last 10 frames of the original image data, the distributed nodes extract the defect detection results for these 10 frames from the processed data as the temporary processing result, such as defect coordinates and type labels. If the temporary data is the last part of the image pixel matrix, the distributed nodes extract the temperature analysis results corresponding to that part of the pixels from the processed data, such as the infrared temperature value distribution, as the temporary processing result.

[0083] The distributed nodes verify the consistency between the temporary end result and the temporary processing result. If hash value verification is used, the hash values ​​of the two are compared to see if they are completely identical. If feature point set verification is used, the matching degree between the two is calculated using a feature matching algorithm. If the matching degree exceeds a preset threshold, such as 95%, they are considered consistent. If the verification result is inconsistent, it indicates that the image data may have been lost or tampered with during distribution, or that there is an anomaly in the processing algorithm of the distributed node, such as an algorithm call error or parameter configuration deviation. In this case, the distributed node immediately returns a scheduling computing power processing anomaly signal to the current edge acquisition hub. This anomaly signal may include specific verification items of inconsistency, such as the hash value mismatch and the number of feature point matching failures, to help the current edge acquisition hub quickly locate the problem.

[0084] When the current edge acquisition hub receives an abnormal signal from the scheduling computing power, it can trigger an abnormal handling mechanism, such as redistributing image data to the distributed node, switching to other distributed nodes for processing, or enabling local redundant computing power to reprocess the data. This effectively prevents erroneous processed data from entering the subsequent multimodal data comprehensive analysis process, ensuring the accuracy of power inspection defect diagnosis results and further improving the data reliability of the entire computing power scheduling system.

[0085] In this embodiment of the application, in order to achieve dynamic balance of computing power load between edge acquisition hubs and distributed nodes and further optimize multimodal data processing efficiency, the method also includes a computing power scheduling step for sensor data, as follows:

[0086] When processing sensor data, current edge acquisition hubs simultaneously record the time taken to process electrical parameters such as infrared temperature, current, and voltage using preset sensing algorithms, such as temperature anomaly detection algorithms and electrical parameter trend analysis algorithms, and obtain the sensing results. This time is defined as the sensing duration. For example, if the edge acquisition hub uses a sliding window algorithm to perform fluctuation analysis on 100 continuously acquired current data sets, and the total time from data input to generating a sensing result containing abnormal fluctuation markers is 8 seconds, then this 8 seconds is the sensing duration for this processing.

[0087] Meanwhile, the current edge acquisition hub records the total time taken from sending the algorithm code to the distributed node to successfully receiving the corresponding processed data returned by the distributed node. This time is defined as the distributed computation time, where the algorithm code is as described above for image algorithms. For example, if it takes 5 seconds from sending the image algorithm code ALG-001 to receiving the defect detection data processed by the algorithm, then this 5 seconds is the current distributed computation time.

[0088] The current edge acquisition hub compares the recorded sensing duration with the distributed computing duration. If the sensing duration is longer than the distributed computing duration, such as 8 seconds > 5 seconds in the example above, it is determined that the current edge acquisition hub has a relatively high load on sensor data processing and needs to schedule some sensor data to distributed nodes for processing to balance computing power.

[0089] In the specific scheduling process, the sensing demand value is calculated based on the sensing duration and distributed computing duration. The formula for calculating the sensing demand value can be set as: Sensing demand value = (Sensing duration - Distributed computing duration) / Sensing duration. For example, when the sensing duration is 8 seconds and the distributed computing duration is 5 seconds, the sensing demand value = (8-5) / 8 = 0.375. The larger this value is, the higher the computing power requirement of the current edge acquisition hub for processing sensing data. The amount of sensing data to be distributed is adjusted according to the positive correlation with the sensing demand value, that is, the larger the sensing demand value, the more sensing data is distributed to the distributed nodes. For example, if the sensing demand value is 0.375, 37.5% of the total sensing data can be distributed to the distributed nodes; if the sensing demand value is 0.6, then 60% of the sensing data is distributed to make full use of the idle computing power of the distributed nodes.

[0090] The current edge acquisition hub sends the filtered sensor data and its corresponding sensor code, such as algorithm number SEN-001 for temperature data normalization and algorithm number SEN-002 for current-voltage correlation analysis, to the distributed nodes. Upon receiving the sensor data and sensor code, the distributed nodes retrieve the matching sensor algorithm from their local algorithm library, process the sensor data, and obtain analytical data containing information such as parameter anomaly detection and trend prediction. For example, for infrared temperature data, the SEN-001 algorithm calculates the difference between the device's hotspot temperature and a threshold, generating analytical data on the temperature anomaly level.

[0091] After completing the sensor data processing, the distributed node returns a termination signal to the current edge acquisition hub, such as a notification message containing an analysis data integrity check code. Upon receiving the termination signal, the current edge acquisition hub reads the analysis data returned by the distributed node and merges it with other locally processed data, such as the processing results of undistributed image data, to form a complete basis for device status assessment.

[0092] Through the above steps, the computing power allocation of sensor data can be dynamically adjusted according to the time difference between sensor data and distributed processing, avoiding the load backlog caused by long-term processing of sensor data at the current edge acquisition hub. At the same time, the idle computing power of distributed nodes can be fully utilized to achieve a balanced distribution of computing power load among edge nodes, further improving the overall efficiency of multimodal data integrated processing and providing more reliable support for the real-time performance and accuracy of power inspection.

[0093] In this embodiment of the application, to further ensure the reliability of the entire sensor data scheduling and processing process and to avoid erroneous data analysis interfering with the power equipment defect diagnosis results, the power inspection image recognition computing power scheduling method based on edge computing also includes sensor data consistency verification and anomaly alerting steps. The specific method includes the following steps:

[0094] Current edge data acquisition hubs, when distributing sensor data to distributed nodes, such as equipment infrared temperature data, real-time operating current / voltage data, and ambient humidity data, simultaneously perform on-site data extraction. Using the sensor data acquisition timestamp as the core criterion, the hub selects one or more sets of data with the latest timestamps from the set of sensor data to be distributed as on-site data. For example, if the sensor data to be distributed consists of 60 sets of voltage data collected at 1-second intervals within the past minute, with timestamps ranging from 14:30:00 to 14:30:59, then the voltage data set with timestamp 14:30:59 is extracted as the on-site data. This voltage data includes the instantaneous voltage value, acquisition accuracy, sensor number, and other information. If the data to be distributed is infrared temperature sequence data, collected at 2-second intervals for 20 sets, covering the temperature of different areas of the equipment, then the last collected set containing the complete data frame of temperature values ​​for key parts of the equipment, such as insulator skirts and conductor joints, is extracted as the on-site data, ensuring that the on-site data accurately reflects the latest operating status parameters of the power equipment.

[0095] The current edge acquisition hub uses a verification algorithm with completely identical logic to that employed by the distributed nodes in processing sensor data. This algorithm calculates the extracted field data and generates the final field result. For example, if the distributed nodes use a temperature threshold comparison algorithm to process infrared temperature data, with a preset normal temperature threshold of -20℃ to 80℃ (exceeding this threshold is considered abnormal), the current edge acquisition hub also uses this algorithm to calculate the deviation between the field temperature value and the threshold. If the field temperature is 85℃, the deviation is +5℃, and this deviation value along with the abnormality label is used as the final field result. Similarly, if the distributed nodes use a current fluctuation coefficient algorithm to process current data, calculating the standard deviation between the current value and the last 10 historical current values, a standard deviation > 0.5A is marked as an abnormal fluctuation. The current edge acquisition hub calculates the standard deviation based on the same historical data sample, and uses this standard deviation value along with the fluctuation status label as the final field result, ensuring a consistent verification basis between the final field result and the distributed node's processing logic.

[0096] After the distributed nodes complete the sensor data processing and return the analysis data to the current edge acquisition hub, the current edge acquisition hub first parses the structured format of the analysis data, such as JSON format or custom binary format. Based on the unique identifier of the field data, such as the acquisition timestamp, sensor ID, and data sequence number, it locates and extracts the processing results corresponding to the field data from the analysis data, defining them as field analysis results. For example, if the field data is voltage data at 14:30:59 and the sensor ID is VOL-003, then the analysis data extracts the voltage stability score of sensor VOL-003 at 14:30:59 and the conclusion of whether there is an overvoltage risk, etc., as field analysis results. If the field data is infrared temperature data of equipment wire connectors and the data sequence number is TEMP-158, then the analysis data extracts the temperature level and defect risk probability information corresponding to TEMP-158 as field analysis results.

[0097] The current edge acquisition hub performs consistency verification between the final results from the field and the field analysis results: For numerical results, such as temperature deviation and current standard deviation, it checks whether the absolute error between the two is less than a preset error threshold, such as ±0.3℃ for temperature error and ±0.05A for current error; for categorical results, such as normal / abnormal labels and risk levels, it checks whether the classification conclusions are completely consistent; for composite results, such as those containing both numerical values ​​and labels, it verifies both the numerical error and the classification conclusion. If the verification result is inconsistent, it indicates that there may be issues such as data packet loss, transmission errors, or algorithm call errors (e.g., calling an algorithm that does not match the sensor code) or parameter configuration deviations (e.g., using an incorrect temperature threshold) during the sensor data distribution process. In this case, the current edge acquisition hub immediately triggers a scheduling computing power processing anomaly warning.

[0098] The specific implementation of anomaly alerts can be configured according to the inspection scenario requirements, including but not limited to: triggering audible and visual alarms at the current edge data acquisition hub, such as a continuous buzzer and flashing red indicator light on the mobile terminal carried by the inspection personnel; displaying an anomaly pop-up window on the terminal screen, clearly indicating the anomaly type, such as inconsistent sensor data processing results, and clearly identifying the abnormal data, such as sensor ID and timestamp; sending anomaly messages to the power inspection management platform, simultaneously uploading the on-site final results, on-site analysis results, and verification logs, facilitating remote monitoring and problem troubleshooting by backend management personnel. Simultaneously, the current edge data acquisition hub automatically records the anomaly occurrence time, anomaly process, and handling measures to a local log file, providing a basis for subsequent fault tracing.

[0099] Through the above steps, anomalies can be detected in a timely manner during the sensor data processing stage, preventing erroneous analytical data from entering the multimodal data fusion process. This effectively reduces the waste of computing power caused by data errors, such as avoiding repeated scheduling of computing power based on erroneous data and ensuring the stability of real-time scheduling of computing power. In turn, it provides key verification support and anomaly early warning guarantee for improving the accuracy of power inspection defect diagnosis and promoting the stable implementation of multimodal inspection technology.

[0100] To avoid bias in computing power assessment caused by sudden load fluctuations or hardware variations at the edge acquisition hub, and to ensure that computing power usage data accurately reflects the actual load status, this application introduces a dual verification mechanism involving both interrupt and non-interrupt functions in the step of acquiring computing power usage data from the edge acquisition hub. The specific method includes the following steps:

[0101] During the current operation of the edge acquisition hub, algorithms for processing acquired data are invoked via interrupt functions. These interrupt functions are the highest-priority task functions in the edge acquisition hub's operating system, capable of responding to external data acquisition events in real time. For example, when the edge acquisition hub's image acquisition module acquires a visible light image of power equipment, the infrared thermal imager acquires a temperature image, or the sensor module acquires current or voltage data, the interrupt function is immediately triggered, invoking preset preprocessing algorithms, such as image noise reduction algorithms and sensor data filtering algorithms, to process the newly acquired data in real time. This process directly consumes the core computing power of the edge acquisition hub, and its operating status directly reflects the actual computing load.

[0102] Meanwhile, within the non-interruptible function of the edge acquisition hub, a pre-defined detection algorithm is periodically invoked. This non-interruptible function is a regular task with lower priority than the interruptible function in the operating system, and it does not affect real-time data processing. This detection algorithm can generate fixed data with a fixed format and fixed data volume, such as generating an array containing 2000 random floating-point numbers (fixed data volume of 16KB) or a solid color image with a resolution of 800×600 (fixed data volume of 1.44MB). The generation logic and data structure of this fixed data are pre-defined within the algorithm, ensuring that the fixed data generated each time the detection algorithm is invoked is completely consistent, serving as a benchmark for computing power evaluation.

[0103] During the above process, the edge acquisition hub synchronously records two sets of key time parameters: First, it calculates the generation time interval of the detection algorithm for each time it generates fixed data; that is, the time difference between two consecutive successful generation of fixed data. For example, if the first generation completion time is 15:00:00, the second is 15:00:03, and the third is 15:00:06, then the generation time intervals are 3 seconds and 3 seconds respectively. Second, it calculates the call time interval of the detection algorithm; that is, the time interval between the edge acquisition hub sending the detection algorithm call instruction to the non-interruptible function. This interval is preset to a fixed value, such as 3 seconds, to ensure a stable call frequency.

[0104] The edge acquisition hub collects at least 5 generation time intervals and calculates the average and uniform interval values. The specific number of times can be adjusted to 8 or 10 times according to the computing power stability requirements. The average interval value is the arithmetic mean of all generation time intervals. For example, if the 5 generation time intervals are 3.1 seconds, 2.9 seconds, 3.0 seconds, 3.2 seconds, and 2.8 seconds, the average interval value is (3.1 + 2.9 + 3.0 + 3.2 + 2.8) / 5 = 3.0 seconds. The uniform interval value is calculated by measuring the fluctuation of the generation time intervals. The specific formula is: Uniform interval value = 1 - (Sum of absolute deviations of all generation time intervals from the average value / (Average interval value × Number of collections)). The smaller the deviation, the closer the uniform interval value is to 1, which means that the generation time intervals are more stable and the computing power status of the edge acquisition hub is more stable.

[0105] The edge acquisition hub compares the average interval with a preset average range, such as 2.7 seconds to 3.3 seconds, set according to the call interval and the edge acquisition hub's basic computing power. Simultaneously, it compares the uniform interval with a preset uniform range, such as 0.85 to 1.0, indicating that the generation interval fluctuation does not exceed 15%. If both the average and uniform interval values ​​are within the average and uniform ranges, it indicates that the edge acquisition hub's current computing power is stable and free from abnormal interference. In this case, computing power usage data is calculated based on the generation and call intervals. The formula is: Computing power usage data = Generation interval / Call interval. For example, if the average generation interval is 3.0 seconds and the call interval is 3 seconds, the computing power usage data = 3.0 / 3 = 1.0, indicating that the computing power load is at a normal level. If the average generation interval is 3.6 seconds, the computing power usage data = 3.6 / 3 = 1.2, indicating that the computing power load exceeds the normal level. If the average interval exceeds the average range, or the uniform interval exceeds the uniform range, it indicates that there is a fluctuation in computing power at the edge acquisition hub, such as a sudden high load task or hardware response delay. In this case, the computing power usage data should be set to the maximum value. For example, if the preset maximum value is 2.0, it means that the computing power is already saturated and computing power scheduling should be triggered first to avoid data processing delays due to evaluation deviations.

[0106] To further improve the adaptability of computing power assessment when calculating computing power occupancy based on computing power usage data, this application introduces a dynamic adjustment mechanism based on computing power occupancy. The specific method includes the following steps:

[0107] The core formula for calculating the computing power occupancy value is: Computing power occupancy value = generation time interval / call time interval. This formula quantifies the computing power load of the edge acquisition hub by measuring the ratio of the actual running time of the detection algorithm to the theoretical call interval. For example, when the generation time interval is 3.6 seconds and the call time interval is 3 seconds, the computing power occupancy value = 3.6 / 3 = 1.2, which means that the current computing power load exceeds the theoretical load by 20%.

[0108] Based on the calculated computing power occupancy value, the average range and uniform range in the aforementioned steps are dynamically adjusted:

[0109] Positive correlation adjustment of the average range: When the computing power occupancy value increases, it indicates an increase in computing power load, and the upper and lower limits of the average range are expanded accordingly; when the computing power occupancy value decreases, it indicates a decrease in computing power load, and the upper and lower limits of the average range are narrowed accordingly. For example, the initial average range is 2.7 seconds - 3.3 seconds, corresponding to a computing power occupancy value of 1.0. If the computing power occupancy value rises to 1.2, it indicates an increase in computing power load, and the average range is adjusted to 2.4 seconds - 3.6 seconds to accommodate reasonable fluctuations in the generation time interval under high load; if the computing power occupancy value drops to 0.8, it indicates a decrease in computing power load, and the average range is adjusted to 2.9 seconds - 3.1 seconds to improve the accuracy of computing power assessment under low load conditions.

[0110] Negative correlation adjustment of the uniform range: When the computing power occupancy increases, the lower limit of the uniform range is expanded simultaneously, thus reducing the requirement for the stability of the generation time interval; when the computing power occupancy decreases, the lower limit of the uniform range is reduced simultaneously, thus increasing the requirement for stability. For example, if the initial uniform range is 0.85-1.0, corresponding to a computing power occupancy of 1.0, and the computing power occupancy rises to 1.2, the uniform range is adjusted to 0.75-1.0, allowing for greater fluctuations in the generation time interval under high load; if the computing power occupancy drops to 0.8, the uniform range is adjusted to 0.90-1.0, ensuring the stability of the generation time interval under low load conditions and further improving the accuracy of computing power assessment.

[0111] Through this dynamic adjustment mechanism, the average range can be adapted to a reasonable fluctuation range as the computing load changes, avoiding misjudgment of computing power anomalies under high load; the uniform range can adjust the stability requirements as the computing load changes, improving compatibility with different load states, and ultimately ensuring the accuracy of computing power occupancy calculation.

[0112] To ensure the communication stability and computing power reliability of the target node (distributed node) during computing power scheduling, this application introduces a communication protocol verification and signal strength detection mechanism in the step of obtaining the computing power occupancy value of other edge acquisition hubs. The specific method includes the following steps:

[0113] The current edge acquisition hub is based on a preset first communication protocol, such as 5G LAN protocol, LoRaWAN protocol, Bluetooth Mesh protocol. It selects the appropriate protocol based on the communication distance and anti-interference requirements of the power inspection scenario, and sends broadcast search commands to the surrounding environment to search for other edge acquisition hubs in the same inspection network, such as mobile terminals carried by other inspection personnel or fixedly deployed edge nodes.

[0114] If other edge acquisition hubs are found, the current edge acquisition hub sends a handshake request to the target node based on the first communication protocol. After the target node returns a handshake response, the two establish the first link, a dedicated data transmission link, for transmitting computing power status and scheduling instructions. If no other edge acquisition hubs are found, such as if only the current edge acquisition hub exists in the inspection area, the computing power occupancy value of other edge acquisition hubs is directly set to the minimum value. For example, the preset minimum value is 0.0, which means there is no available idle computing power, thus avoiding subsequent invalid scheduling.

[0115] After the first link is established, the current edge acquisition hub obtains the signal strength of other edge acquisition hubs in real time through the link, such as the RSSI value and ReceivedSignalStrengthIndicator, and compares it with a preset reference strength value, such as -80dBm, which is set according to the communication quality requirements of the inspection scenario. A signal strength greater than this value indicates that the link is stable. If the signal strength is greater than the reference strength value, it indicates that the link communication quality is reliable. The current edge acquisition hub sends a computing power occupancy request command to the target node, and the target node returns its own real-time computing power occupancy value. If the signal strength is less than or equal to the reference strength value, it indicates that there is a risk of link interruption, and data transmission may be delayed or lost. In this case, the computing power occupancy value of the target node is set to the minimum value, and it is excluded from the selection range of distributed nodes.

[0116] This step allows for the prioritization of edge acquisition hubs with stable communication and available computing power as candidate nodes, avoiding scheduling failures due to unstable links, improving the effectiveness of real-time computing power scheduling, and providing a stable computing power scheduling foundation for the efficient processing of multimodal data from power inspection.

[0117] To further expand the coverage of computing power scheduling and address scenarios where a single edge acquisition hub has insufficient computing power and requires the coordination of multiple nodes to process data, this application adds a secondary link construction and data forwarding mechanism based on the established first link. The specific method includes the following steps:

[0118] After acquiring the signal strength of other edge acquisition hubs, if the current edge acquisition hub determines that the signal strength of one of the other edge acquisition hubs is greater than a preset reference strength value (e.g., the reference strength value is set to -80dBm, and the target node's signal strength detection value is -72dBm), then the current edge acquisition hub sends a link extension command to the selected distributed node. This command controls the distributed node to initiate communication interaction with the other edge acquisition hub whose signal strength meets the standard. The other edge acquisition hubs are hereinafter referred to as forwarding nodes. The distributed node sends a handshake request frame to the forwarding node based on the same preset communication protocol as the first link, such as the 5G LAN protocol or Bluetooth Mesh protocol. This request frame contains the distributed node's device identifier, current computing power status, and data forwarding permission information. Upon receiving the handshake request, the forwarding node verifies the legitimacy of the distributed node's identity using a preset device whitelist or encryption key. If the verification is successful, a handshake response frame is returned. Thus, the distributed node and the forwarding node establish a second link for data forwarding.

[0119] Once the distributed node completes the establishment of the second link, it immediately returns a secondary link signal to the current edge acquisition hub. This signal contains key parameters such as the link identifier of the second link, the device ID of the forwarding node, the link bandwidth, and the signal strength. This signal is used by the current edge acquisition hub to confirm the establishment status and communication capability of the secondary link. After receiving and verifying the secondary link signal, the current edge acquisition hub determines that it has the conditions for multi-node collaborative processing and then invokes a preset forwarding instruction. This instruction contains the identifier of the image data to be forwarded, such as the data block number, the start / end byte index, and the corresponding algorithm code, such as ALG-001 for insulator defect identification. It also includes the data forwarding priority and result feedback requirements, such as the feedback timeout and data integrity verification method.

[0120] After receiving the forwarding instruction, the distributed node retrieves the corresponding image data from its local cache based on the image data identifier specified in the instruction. For example, a data block might contain 20 frames of infrared images of power equipment, which are then encapsulated into a forwarding data packet carrying an algorithm code. The forwarding data packet is then sent to the forwarding node via the established second link at a preset transmission rate, which may be dynamically adjusted to 10 Mbps based on the link bandwidth. Upon receiving the forwarding data packet, the forwarding node performs integrity verification, such as using a CRC32 checksum to verify whether the data is lost or tampered with. If the verification passes, the node retrieves a matching image processing algorithm from its local algorithm library based on the algorithm code, such as the ALG-001 defect detection algorithm consistent with the distributed node. This algorithm then processes the image data in parallel, generating processed data containing defect location, defect type, and confidence level.

[0121] After completing data processing, the forwarding node sends the processed data back to the distributed node via the second link, according to the feedback requirements specified in the forwarding instruction. The distributed node summarizes and performs preliminary verification on the processed data returned by the forwarding node, such as checking whether the data format conforms to preset standards. Then, it sends the summarized processed data, including its own processing results and the processing results of the forwarding node, back to the current edge acquisition hub via the first link. After receiving the complete processed data, the current edge acquisition hub integrates the processing results of multiple nodes to form comprehensive assessment data of the power equipment status, realizing collaborative computing power scheduling and data processing among multiple edge nodes.

[0122] This step allows for the expansion of secondary computing nodes based on distributed nodes, breaking the limitations of single-node scheduling. At the same time, the secondary link ensures the stability of data forwarding, making computing power scheduling more flexible. It is especially suitable for large-scale power inspection, such as scenarios where multiple devices collaboratively process massive image data in long-distance transmission line inspections.

[0123] To avoid data forwarding overload or resource waste due to differences in the communication capabilities of secondary links, this application introduces a dynamic adjustment mechanism for the forwarding data volume based on the comparison value of secondary signals, ensuring that data transmission is adapted to the link capacity. The specific method includes the following steps:

[0124] After establishing a second link with the forwarding node, the distributed node collects the signal strength of the second link in real time, such as collecting the RSSI value every 100ms, and calculates the secondary signal comparison value by combining it with a preset reference strength value. The formula for calculating the secondary signal comparison value is set as follows: Secondary signal comparison value = (Actual signal strength - Reference strength value) / (Maximum signal strength threshold - Reference strength value), where the maximum signal strength threshold is the optimal signal strength that the edge acquisition hub communication module can achieve, such as -40dBm. This formula ensures that the value of the secondary signal comparison value is in the range of 0~1, and the larger the value, the better the communication quality of the second link. For example, if the reference strength value is -80dBm, the actual signal strength is -72dBm, and the maximum signal strength threshold is -40dBm, then the secondary signal comparison value = (-72 - (-80)) / (-40 - (-80)) = 8 / 40 = 0.2; if the actual signal strength is -50dBm, then the secondary signal comparison value = (-50 - (-80)) / 40 = 30 / 40 = 0.75.

[0125] Before invoking forwarding commands, the current edge acquisition hub obtains secondary signal comparison values ​​from the distributed nodes and adjusts the amount of image data to be forwarded based on a positive correlation with these values. Specifically, the larger the secondary signal comparison value, the more image data is allocated to the forwarding node; the smaller the secondary signal comparison value, the less data is allocated, to match the actual communication capacity of the link. For example, if the total amount of image data to be processed is 100 frames, and the secondary signal comparison value is 0.75 (indicating excellent communication quality), then 60 frames of image data are allocated to the forwarding node for processing; if the secondary signal comparison value is 0.2 (indicating average communication quality), then only 20 frames of image data are allocated to the forwarding node to avoid link congestion or transmission timeouts due to excessive data volume.

[0126] During data forwarding, distributed nodes monitor the signal strength changes of the second link in real time and recalculate the secondary signal comparison value every 500ms. If the comparison value fluctuation exceeds a preset threshold (e.g., fluctuation amplitude > 0.1), the signal change information is immediately fed back to the current edge acquisition hub. The current edge acquisition hub dynamically adjusts the subsequent image data forwarding volume based on the latest secondary signal comparison value. For example, if the secondary signal comparison value drops from 0.75 to 0.4, it indicates a decrease in link quality, so the subsequent forwarding data volume is adjusted from 60 frames to 40 frames; if the comparison value rises from 0.2 to 0.5, it indicates an improvement in link quality, so the subsequent forwarding data volume is adjusted from 20 frames to 35 frames.

[0127] This dynamic adjustment mechanism enables precise matching between image data forwarding volume and secondary link communication capacity. It avoids wasting computing resources when communication quality is good and prevents data transmission overload when communication quality is poor. It effectively ensures the smoothness of multimodal data processing, provides stable transmission guarantee for efficient analysis of multimodal data in power inspection, and further improves the adaptability and reliability of computing power scheduling.

[0128] This application also discloses a power grid inspection image recognition computing power scheduling system based on edge computing, including a processor, wherein the processor executes the steps of the power grid inspection image recognition computing power scheduling method based on edge computing as described in any of the above embodiments.

[0129] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A power dispatching method for power line inspection image recognition based on edge computing, characterized in that, include: The edge acquisition hub acquires and preprocesses image data; Obtain computing power usage data from the edge acquisition hub, and calculate the computing power occupancy value by combining the set usage data; If the computing power occupancy value is greater than the set occupancy value, then other edge acquisition hubs with computing power occupancy values ​​less than the set occupancy value and the smallest occupancy value are designated as distribution nodes; the image data and the corresponding algorithm code are sent to the distribution nodes; the computing power requirement value is calculated based on the computing power occupancy value and the set occupancy value, and the amount of image data distributed is adjusted according to the positive correlation of the computing power requirement value; The distributed nodes retrieve the image data from the image algorithm based on the algorithm code, process the image data, obtain the processed data, and return it. Obtain the labeling computation time required for distributed nodes to process image data, match the labeling reference time based on the amount of image data, and if the labeling computation time is greater than the labeling reference time, calculate the labeling computing power comparison value based on the labeling computation time and the labeling reference time, and amplify the computing power occupancy value based on the labeling computing power comparison value; When distributing image data, the current edge acquisition hub extracts temporary data from the end of the image data sent to the distributed nodes; The current edge acquisition hub uses temporary data to calculate temporary end results. After the image data distribution is completed, the location of the temporary data and the temporary end results are sent to the distribution nodes. After the distributed nodes generate the processing data, they analyze the location of the temporary data and extract the temporary processing results corresponding to the location of the temporary data from the processing data. Verify the consistency between the temporary end result and the temporary processing result. If the verification fails, return a scheduling computing power processing anomaly signal to the current edge acquisition hub.

2. The power scheduling method for power inspection image recognition based on edge computing according to claim 1, characterized in that, The method also includes the following steps: The time taken for the edge acquisition hub to process sensor data using a preset sensing algorithm and obtain sensing results is the sensing duration. The time taken from when the edge acquisition hub sends the algorithm code to when it obtains the corresponding processing data is recorded as the distributed computation time. If the sensing duration is longer than the distributed computation duration, the sensing data and the corresponding sensing code are sent to the distributed node. The sensing demand value is calculated based on the sensing duration and the distributed computation duration, and the amount of sensing data distributed is adjusted according to the positive correlation of the sensing demand value. The distributed nodes retrieve the corresponding sensing algorithm based on the sensing code to process the sensing data and obtain the analysis data; after completion, they return an end signal to the current edge acquisition hub, which then reads the analysis data returned by the distributed nodes in response to the end signal.

3. The power scheduling method for power inspection image recognition based on edge computing according to claim 2, characterized in that, The method also includes the following steps: The current edge acquisition hub extracts the latest field data from the sensor data sent to the distributed nodes; The current edge acquisition hub uses field data to calculate the final field results and extracts the field analysis results corresponding to the field data from the analysis data. Verify the consistency between the final on-site results and the on-site analysis results. If the verification fails, issue an error message regarding the scheduling computing power processing.

4. The power scheduling method for power inspection image recognition based on edge computing according to claim 1, characterized in that, The steps for obtaining computing power usage data from edge acquisition hubs also include the following: Call the algorithm that processes the acquired data in the interrupt function; Call the detection algorithm that generates preset fixed data in the non-interruptible function; Calculate the time interval for generating fixed data and the time interval for calling the detection algorithm; The interval average and interval uniformity are calculated based on multiple generation time intervals. If the average interval is within the preset average range and the uniform interval is within the preset uniform range, then the computing power usage data is calculated based on the generation time interval and the call time interval; otherwise, the computing power usage data is the maximum value.

5. The power scheduling method for power inspection image recognition based on edge computing according to claim 4, characterized in that, The process of calculating the computing power occupancy value based on computing power usage data and preset usage settings also includes the following steps: Computing power consumption = Generation time interval / Call time interval; Adjust the average range based on a positive correlation with computing power occupancy, and adjust the uniform range based on a negative correlation with computing power occupancy.

6. The power scheduling method for power inspection image recognition based on edge computing according to claim 1, characterized in that, The steps for obtaining the computing power occupancy values ​​of other edge acquisition hubs also include the following steps: The current edge acquisition hub searches for other nearby edge acquisition hubs based on a preset first communication protocol; If other edge acquisition hubs are found, a handshake is established with the other edge acquisition hubs based on the first communication protocol to establish the first link; otherwise, the computing power occupancy value of the other edge acquisition hubs is set to the minimum value. Based on the first link, obtain the signal strength of other edge acquisition hubs; If the signal strength is greater than the preset reference strength value, the computing power occupancy value of other edge acquisition hubs is obtained; otherwise, the computing power occupancy value of other edge acquisition hubs is set to the minimum value.

7. The power scheduling method for power inspection image recognition based on edge computing according to claim 6, characterized in that, It also includes the following steps: If the signal strength is greater than the preset reference strength value, the control distributed node will handshake with other edge acquisition hubs and establish a second link; The marked edge acquisition hub returns a secondary link signal to the current edge acquisition hub, and the current edge acquisition hub responds to the secondary link signal by invoking a forwarding command; The marked edge acquisition hub responds to the received forwarding instruction by forwarding the image data and the corresponding algorithm code to other edge acquisition hubs; Forward the data generated by other edge acquisition hubs and corresponding to the forwarding instructions to the current edge acquisition hub.

8. The power scheduling method for power inspection image recognition based on edge computing according to claim 7, characterized in that, It also includes the following steps: The secondary signal comparison value is calculated based on the signal strength and the reference strength value. The amount of image data corresponding to the forwarding command is adjusted and forwarded based on the positive correlation of the secondary signal comparison value.

9. A power grid inspection image recognition computing power scheduling system based on edge computing, characterized in that, The device includes a processor that performs the steps of the edge computing-based power inspection image recognition computing power scheduling method as described in any one of claims 1-8.

Images