An edge side visual model dynamic switching method for substation patrol tasks

Abstract

The application discloses a kind of edge side visual model dynamic switching methods for substation inspection task, first in the process of substation inspection task execution, periodically collect the operating state parameters of edge intelligent analysis module, and the scene state parameters corresponding to current inspection image, then operating state parameters and scene state parameters are converted into operating state vector and scene state vector, to calculate the comprehensive matching score of each preset candidate visual model in the edge intelligent analysis module of substation field, and select the highest score target visual model;Based on target visual model, judge whether substation executes visual model switching at current time, if executing visual model switching, then load target visual model and output switching result;Otherwise, keep current visual model running.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent substation inspection, edge computing and machine learning applications. More specifically, it relates to a dynamic switching method for edge-side visual models for substation inspection tasks. Background Technology

[0002] With the development of intelligent substation inspection systems, more and more inspection tasks are shifting from traditional manual inspections to automated inspection modes that are collaboratively completed by fixed cameras, pan-tilt-zoom (PTZ) devices, inspection robots, drones, and station-end monitoring terminals. In these systems, a large amount of visible light images, infrared images, and short video data are continuously generated. If all of this data is uploaded to a central platform for unified processing, it will not only lead to high bandwidth consumption, high latency, and task interruptions due to unstable links, but also result in delayed alarm responses, making it difficult to meet the engineering requirements of real-time identification and immediate alarms for substation inspection tasks.

[0003] To address these issues, an increasing number of visual recognition tasks are being performed on edge devices closer to the data source. Edge intelligent analysis modules have thus become key units for tasks such as image preprocessing, target detection, status recognition, defect classification, and alarm linkage. Edge intelligent analysis modules are typically deployed at the camera side, station cabinet side, or inspection robot side. Their advantages include proximity to the field, reduced image transmission volume, and the ability to maintain partial autonomy when network access is limited. However, these modules generally have limited computing and storage resources, and significant power and heat dissipation constraints, resulting in fundamentally different operating conditions compared to central cloud servers.

[0004] On the other hand, current visual recognition models typically suffer from large scale, deep structure, and high computational cost. While large or high-precision models can achieve high recognition accuracy offline, when running on edge intelligent analysis modules, they often encounter problems such as excessive inference latency, high cache usage, significant power consumption increases, and rapid temperature rise. In severe cases, this can lead to frame drops, congestion, thermal throttling, or even task interruption on edge devices. If high-precision models are deployed permanently and permanently at the edge, although the static accuracy is high, better overall performance may not be achieved under resource constraints and complex operating conditions.

[0005] In contrast, while lightweight models offer lower computational complexity, faster execution speed, and lower power consumption, making them more suitable for long-term stable operation on edge devices, they may experience accuracy degradation in scenarios with complex backgrounds, severe occlusion, low illumination, and densely packed multiple targets. This is particularly true in substation inspection scenarios, where image conditions exhibit significant time-varying characteristics: for example, day-night cycles cause changes in illumination, rainy weather introduces blurring and reduced contrast, complex textured backgrounds near equipment increase false detection rates, and the movement of the inspection robot introduces jitter and dynamic blur. This means that a scenario suitable for running a lightweight model at one time may require switching to a stronger balanced model or even a high-precision model at another.

[0006] Most existing research focuses on lightweight design of single models, model compression, pruning quantization, knowledge distillation, or general edge inference optimization. It pays more attention to the compression and acceleration effects of the model structure itself, while rarely considering the real-time resource status of edge intelligent analysis modules and the complexity of inspection scenarios to perform online comparison, selection, and switching of multiple candidate visual models. While a few dynamic inference methods consider early exit mechanisms or model-level switching, they do not adequately address issues such as power consumption constraints, temperature rise constraints, cache migration costs, link bandwidth, and model switching jitter suppression in substation inspection scenarios, making it difficult to directly develop a systematic approach applicable to engineering sites.

[0007] Therefore, there is an urgent need to propose a visual model switching method for edge intelligent analysis modules for substation inspection defect identification. This method would enable the edge intelligent analysis module to no longer run a single visual model, but to dynamically switch between multiple candidate models based on its own operating status, the complexity of the inspection images, and the cost of model operation. This would ensure the feasibility of the inspection task and the stable operation of the system, while taking into account the recognition accuracy, inference latency, and resource consumption as much as possible. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a dynamic switching method for edge-side visual models for substation inspection tasks. This method addresses the technical problems of high resource consumption, poor model adaptability, difficulty in balancing recognition accuracy and real-time performance in complex inspection scenarios, and operational instability caused by frequent model switching in the prior art, which are all due to the fixed deployment of a single visual model on the edge side.

[0009] To achieve the above-mentioned objectives, the present invention provides a method for dynamic switching of edge-side visual models for substation inspection tasks, characterized by comprising the following steps:

[0010] (1) A set of candidate visual models is pre-configured in the edge intelligent analysis module at the substation site. The set of candidate visual models includes at least two visual models with different complexities.

[0011] (2) During the substation inspection task, the operating status parameters of the edge intelligent analysis module are periodically collected. The operating status parameters include processor utilization, available memory ratio, task queue pressure, power consumption ratio, and cache utilization. Then, an operating status vector is constructed based on the operating status parameters:

[0012] ;

[0013] in, express The running state vector at any given time; express Processor utilization at any given moment; express The percentage of available memory at any given time; express Constant task queue pressure; express The percentage of power consumption at any given moment; express Cache usage at any given moment;

[0014] (3) Synchronously collect scene state parameters corresponding to the current patrol image. Scene state parameters include image brightness, image sharpness, occlusion degree, target density, and weather conditions; then construct a scene state vector based on the scene state parameters:

[0015] ;

[0016] in, express The scene state vector at any given time; express The brightness of the patrol image at any given time; express The clarity of the patrol images at any given time; express The degree of occlusion at any given time; express The density of targets at any given moment; express The weather conditions at any given time;

[0017] (4) Calculate the resource availability score at the current moment based on the running state vector. ;

[0018] ;

[0019] in, These are the weighting coefficients;

[0020] (5) Calculate the scene complexity score at the current moment based on the scene state vector. ;

[0021] ;

[0022] in, Indicates the standard brightness reference value; These are the weighting coefficients;

[0023] (6) Calculate the model requirement score for each candidate visual model;

[0024] ;

[0025] in, Indicates the first Candidate visual models Model requirement score; express The computational load; express Storage usage; express The average inference latency; express Operating power consumption; These are the weighting coefficients;

[0026] (7) Calculate the overall matching score of each candidate visual model;

[0027] ;

[0028] in, express exist A comprehensive matching score for each moment; Indicates the current running visual model Switch to the Candidate visual models The switching cost; These are the weighting coefficients;

[0029] (8) Select the candidate visual model with the highest comprehensive matching score from the candidate visual model set as the target visual model;

[0030] ;

[0031] in, Represents the set of candidate visual models; This represents the target visual model selected at time t;

[0032] (9) Determine the current substation status. Whether to perform visual model switching at any given time;

[0033] Selected target visual model With the currently running visual model If there is a discrepancy, further determine whether the difference in the overall matching scores between the two exceeds a preset switching threshold. And the currently running visual model Whether the minimum stay time requirement has been met is determined by the following criteria:

[0034] ;

[0035] in, express The overall matching score; express The overall matching score; Indicates the preset switching threshold; Indicates the time of the last visual model switch; Indicates the minimum stay time;

[0036] When both the score difference condition and the minimum dwell time condition are met, the visual model is switched and the process proceeds to step (10); otherwise, the current visual model continues to run and the process ends.

[0037] (10) Load the target visual model and output the switching result;

[0038] The currently running visual model Switch to target vision model and the target visual model Loaded into the edge intelligent analysis module, the switching results are output after the model switching is completed. The switching results include the target model identifier, the current switching status, the resource availability score, the scene complexity score, the comprehensive matching score, and the switching basis, which can be called by subsequent substation inspection image processing tasks.

[0039] The objective of this invention is achieved as follows:

[0040] This invention provides a dynamic switching method for edge-side visual models in substation inspection tasks. First, during the substation inspection process, the operating status parameters of the edge intelligent analysis module and the scene status parameters corresponding to the current inspection image are periodically collected. Then, the operating status parameters and scene status parameters are converted into operating status vectors and scene status vectors, respectively. Based on these vectors, a comprehensive matching score is calculated for each preset candidate visual model in the substation's edge intelligent analysis module, and the target visual model with the highest score is selected. Based on the target visual model, it is determined whether the substation should perform a visual model switch at the current moment. If a visual model switch is performed, the target visual model is loaded and the switch result is output; otherwise, the current visual model remains running.

[0041] Meanwhile, the edge-side visual model dynamic switching method for substation inspection tasks proposed in this invention also has the following beneficial effects:

[0042] (1) This invention simultaneously collects the processor utilization, available storage space, task load, power consumption status, link bandwidth and temperature rise status of the edge intelligent analysis module, and combines the scene features such as target density, brightness fluctuation, blur degree, texture complexity and recognition uncertainty in the inspection image to establish resource availability score and scene complexity score, thereby realizing the joint perception of equipment operating status and inspection scene changes, making the model switching decision more in line with the actual working conditions, and improving the environmental adaptability of edge defect identification task.

[0043] (2) By introducing model demand scoring, inference delay prediction model, power consumption prediction model, comprehensive utility function and switching cost function, this invention realizes comprehensive evaluation and dynamic selection among multiple candidate visual models. It can not only select a more suitable model according to the current resource conditions, but also achieve a good balance between recognition accuracy, response delay and operating energy consumption, thereby improving the scientificity and practicality of model scheduling.

[0044] (3) By setting up an upgrade correction mechanism under time delay constraints, power consumption constraints, cache constraints, switching thresholds, minimum dwell time and confidence constraints, this invention can effectively suppress system oscillations caused by frequent model switching. At the same time, by combining cache reuse and model loading strategies, it reduces loading latency and migration overhead during model switching, thereby improving the stability and reliability of the edge intelligent analysis module during continuous operation.

[0045] (4) Compared with the traditional fixed deployment of a single high-precision model or a single lightweight model, the present invention can adaptively switch between candidate models of different capability levels according to the operating status and scenario complexity. It is significantly better than the fixed high-precision model in terms of average inference latency and average power consumption, while maintaining a recognition effect close to that of the high-precision model. It is also significantly better than the recognition ability of the fixed lightweight model in complex scenarios. Therefore, it is more suitable for the inspection edge ring of substations with limited resources and significant fluctuations in operating conditions. Attached Figure Description

[0046] Figure 1 This is a flowchart of a method for dynamic switching of edge-side visual models for substation inspection tasks according to the present invention;

[0047] Figure 2 This is a diagram illustrating the resource scoring, scenario complexity, and model selection results at different times. Detailed Implementation

[0048] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0049] Example

[0050] In this embodiment, for ease of explanation, a visible light inspection segment of the high-voltage side disconnecting switch and adjacent bushing area of ​​the main transformer of a 110 kV substation is used as an example. The image acquisition device outputs a 1920×1080 image, and the edge intelligent analysis module extracts one frame every 5 seconds for model switching judgment. The edge intelligent analysis module adopts an 8-core ARM processor, a 6 TOPS-level neural network acceleration unit and 8 GB of running memory. The rated available power consumption in the station cabinet is calculated as 30 W.

[0051] like Figure 1 As shown, the present invention provides a method for dynamic switching of edge-side visual models for substation inspection tasks, comprising the following steps:

[0052] (1) A set of candidate visual models is pre-configured in the edge intelligent analysis module at the substation site. The set of candidate visual models includes lightweight visual models, balanced visual models and high-precision visual models.

[0053] In this embodiment, the candidate visual model set includes at least two visual models with different complexities. The lightweight visual model is used for inspection scenarios where edge device resources are scarce or real-time requirements are high. The balanced visual model is used for inspection scenarios where the edge device resource status and the complexity of the inspection scenario are at an intermediate level. The high-precision visual model is used for inspection scenarios where edge device resources are sufficient and the complexity of the inspection scenario is high. The computational cost, storage usage, average inference latency, and power consumption of the lightweight visual model are all lower than those of the balanced and high-precision visual models. The model accuracy of the high-precision visual model is higher than that of the balanced and lightweight visual models. The balanced visual model is used to strike a balance between model accuracy and resource consumption. Each candidate visual model has its model identifier, computational cost, storage usage, average inference latency, and power consumption pre-recorded.

[0054] In this embodiment, the lightweight visual model is denoted as M. L Let the balanced visual model be denoted as M. B Let the high-precision visual model be denoted as M. H All three models were trained or distilled from the same batch of substation equipment defect samples and pre-stored in the local storage area of ​​the edge intelligent analysis module. The candidate models' model identifier, input size, computational cost, storage usage, average inference latency, operating power consumption, and normalized model capability value A are also considered. iAs shown in Table 1, where A i It is calculated by combining the defect detection rate of historical verification samples at the station, the false detection rate in complex backgrounds, and the small target recognition rate.

[0055] Table 1. Pre-configured parameters for candidate visual models;

[0056] <![CDATA[M L ]]> Lightweight visual model 512×512 4.6 24 38 6.8 0.70 <![CDATA[M B ]]> Balanced visual model 640×640 10.8 58 72 10.5 0.86 <![CDATA[M H ]]> High-precision visual model 768×768 24.5 126 145 17.8 0.94

[0057] (2) During the substation inspection task, the operating status parameters of the edge intelligent analysis module are periodically collected. The operating status parameters include processor utilization, available memory ratio, task queue pressure, power consumption ratio, and cache utilization. Then, an operating status vector is constructed based on the operating status parameters:

[0058] ;

[0059] in, express The running state vector at any given time; express Processor utilization at any given moment; express The percentage of available memory at any given time; express Constant task queue pressure; express The percentage of power consumption at any given moment; express The cache occupancy rate at any given time; the running state vector can be used to determine whether the edge intelligent analysis module is currently in a state of sufficient resources, moderate resources, or resource shortage.

[0060] In this embodiment, the current judgment time is... 09:42:15, the currently running visual model is M L The previous model switching time The timeframe is 09:35:00. At 09:42:15, the edge intelligence analysis module reads the average processor utilization, available memory, image frame queue length, total power consumption, and model cache usage over the past 5 seconds, and normalizes them to 0-1. The normalized running state vector is:

[0061] ;

[0062] in, This indicates that the average processor utilization rate over the past 5 seconds is 43%. This indicates that approximately 4.64 GB of the 8 GB memory is usable. This indicates that the current task queue pressure is 32%. This indicates that the actual measured power consumption is approximately 15.3 W, which is 51% of the rated usable power consumption of 30 W. This indicates that the model cache has occupied approximately 920 MB, which is 46% of the reserved 2.0 GB cache space.

[0063] (3) Synchronously collect scene state parameters corresponding to the current patrol image. Scene state parameters include image brightness, image sharpness, occlusion degree, target density, and weather conditions; then construct a scene state vector based on the scene state parameters:

[0064] ;

[0065] in, express The scene state vector at any given time; express The brightness of the patrol image at any given time; express The clarity of the patrol images at any given time; express The degree of occlusion at any given time; express The density of targets at any given moment; express The weather conditions at any given time; the scene state vector can reflect the impact of lighting, clarity, occlusion, number of targets, and weather changes on the visual model's performance during substation inspections.

[0066] In this embodiment, the edge intelligent analysis module performs image preprocessing on the inspection image extracted at 09:42:15. First, it extracts the areas containing the disconnect switch, bushing, lead clamp, and equipment nameplate. Then, it calculates the average brightness, Laplacian sharpness, target occlusion ratio, target density, and weather impact level. The normalized scene state vector is:

[0067] ;

[0068] Table 2 shows the sampled values ​​of specific operating status parameters and scenario status parameters;

[0069] Table 2 Sample values ​​of operating status parameters and scenario status parameters

[0070] Running status <![CDATA[Processor occupancy rate u t > 0.43 The average CPU / NPU scheduling utilization rate over the past 5 seconds is approximately 43%. Running status <![CDATA[Available memory ratio m t > 0.58 Approximately 4.64 GB of the 8 GB memory is available. Running status <![CDATA[Task queue pressure q t > 0.32 The length of the image frame queue to be processed is approximately 32% of the preset upper limit. Running status <![CDATA[Power consumption ratio p t > 0.51 Actual power consumption: 15.3 W / Rated power consumption: 30 W Running status <![CDATA[Cache occupancy rate c t > 0.46 The model cache currently in use is approximately 920 MB / 2.0 GB. Scene status <![CDATA[Image brightness l t > 0.32 Normalized to grayscale mean, below standard brightness. Scene status <![CDATA[Image sharpness g t > 0.57 The normalized Laplace variance is 0.57. Scene status <![CDATA[Occlusion degree o t > 0.34 Approximately 34% of the bushing is obscured by the frame and lead wires. Scene status <![CDATA[Target density d t > 0.52 Multiple targets, such as disconnect switches, bushings, and lead clamps, need to be identified in a single frame. Scene status <![CDATA[Weather condition w t > 0.40 Light fog following a light rain, with an impact level of 0.40.

[0071] (4) Calculate the resource availability score at the current moment based on the running state vector. The resource availability score is used to represent the current edge intelligence analysis module's ability to support the operation of the visual model, and is calculated according to the following formula:

[0072] ;

[0073] in, These are the weighting coefficients;

[0074] In this embodiment, The larger the value, the more abundant the available resources of the current edge intelligent analysis module; The smaller the value, the more strained the resources of the current edge intelligence analysis module are. The resource availability score decreases when processor utilization increases, available memory percentage decreases, task queue pressure increases, power consumption percentage increases, or cache utilization increases.

[0075] Since higher processor utilization, task queue pressure, power consumption percentage, and cache utilization indicate more strained resources, a certain threshold is used when calculating resource availability scores. , , and A higher percentage of available memory indicates more abundant resources, therefore, we can directly adopt... Weighting coefficients We take values ​​of 0.25, 0.25, 0.20, 0.15, and 0.15 respectively, with the sum of the five weights being 1. Substituting the above sampled values, we get:

[0076] ;

[0077] Therefore, the resource availability score of the edge intelligent analysis module at the current moment is 0.578, which is above average. This indicates that the module can support models that are stronger than lightweight models, but it is not suitable to directly load high-precision models for a long time.

[0078] (5) Calculate the scene complexity score at the current moment based on the scene state vector. The scene complexity score is used to represent the degree of requirement for visual model accuracy and computational power in the current inspected image or video frame, and is calculated according to the following formula:

[0079] ;

[0080] in, Indicates the standard brightness reference value; These are the weighting coefficients; The larger the value, the more complex the current inspection scene, making it more suitable to call a higher-precision visual model; The smaller the value, the simpler the current inspection scene, and the visual model with lower computational overhead can be prioritized. The scene complexity score increases when image clarity decreases, occlusion increases, target density increases, image brightness deviates from the standard brightness reference value, or weather conditions worsen.

[0081] In this embodiment, the standard brightness reference value The value is set to 0.50; the greater the deviation of brightness from the standard value, the lower the clarity, the more severe the occlusion, the denser the targets, and the more obvious the weather effects, the higher the scene complexity. Weighting coefficient. We take weights of 0.20, 0.25, 0.20, 0.25, and 0.10 respectively, with the sum of the five weights being 1; substituting the above scenario parameters, we obtain... Therefore, the current scene complexity score is 0.418. This value indicates that the current scene is not extremely complex, but low illumination, slight rain and fog blurring, partial occlusion, and relatively dense targets will reduce the recognition stability of the lightweight vision model, and further evaluation is needed to determine whether to switch to a balanced vision model.

[0082] (6) Calculate the model requirement score for each candidate visual model;

[0083] The requirement score for candidate visual models is calculated based on computational cost, storage usage, average inference latency, and power consumption. The specific formula for calculating the model requirement score is as follows:

[0084] ;

[0085] in, Indicates the first Candidate visual models Model requirement score; express The computational load; express Storage usage; express The average inference latency; express Operating power consumption; These are the weighting coefficients; The larger the value, the higher the resource requirements of the edge intelligent analysis module for the candidate visual model; The smaller the value, the easier it is for the candidate visual model to run under resource-constrained conditions.

[0086] In this embodiment, the maximum computational cost of 24.5 GFLOPs, the maximum storage usage of 126 MB, the maximum average latency of 145 ms, and the maximum power consumption of 17.8 W for the three candidate models in Table 1 are first used as normalization benchmarks. Then, the model requirement score is calculated based on the computational cost, storage usage, average inference latency, and power consumption. Weighting coefficients Take values ​​of 0.35, 0.25, 0.25, and 0.15 respectively; for example, , , The model requirement scores were 0.482, 0.236, and 1.000, indicating that... Minimal resource requirements It has the highest requirements for resources. It is at an intermediate level and is suitable as a candidate target model when resource availability is moderate and scenario complexity increases.

[0087] (7) Calculate the overall matching score of each candidate visual model;

[0088] The comprehensive matching score is used to determine whether a candidate visual model is suitable for the current edge device state and the current patrol scene state. The higher the comprehensive matching score, the more suitable the candidate visual model is to be used in the current state. It is calculated according to the following formula:

[0089] ;

[0090] in, express exist A comprehensive matching score for each moment; Indicates the current running visual model Switch to the Candidate visual models The switching cost; These are the weighting coefficients;

[0091] In this embodiment, , , The overall matching scores were 0.764, 0.688, and 0.557, respectively. It is suitable to be invoked in the current state.

[0092] (8) Select the candidate visual model with the highest comprehensive matching score from the candidate visual model set as the target visual model;

[0093] ;

[0094] in, Represents the set of candidate visual models; This represents the target visual model selected at time t;

[0095] In this embodiment, the target visual model that best suits the current state is automatically determined among the lightweight visual model, the balanced visual model, and the high-precision visual model.

[0096] In this embodiment, as Figure 2 As shown, the correspondence between resource availability score, scenario complexity score and model selection results at various typical time points is presented.

[0097] (9) Determine the current substation status. Whether to perform visual model switching at any given time;

[0098] Selected target visual model With the currently running visual model If there is a discrepancy, further determine whether the difference in the overall matching scores between the two exceeds a preset switching threshold. And the currently running visual model Whether the minimum stay time requirement has been met is determined by the following criteria:

[0099] ;

[0100] in, express The overall matching score; express The overall matching score; Indicates the preset switching threshold; Indicates the time of the last visual model switch; Indicates the minimum stay time;

[0101] When both the score difference condition and the minimum dwell time condition are met, the visual model is switched and the process proceeds to step (10); otherwise, the current visual model continues to run and the process ends.

[0102] In this embodiment, a preset switching threshold is used. Take 0.05, minimum dwell time Take 180 seconds. Since the current running model is... The target visual model selected in step (8) is The two are inconsistent, therefore, it is necessary to further determine whether the comprehensive matching score difference and the minimum stay time simultaneously meet the requirements. As calculated above, Compared to The score improvement exceeded the preset switching threshold, and The minimum dwell time has been exceeded, thus meeting the model switching conditions. If only the highest score is met but the score difference does not exceed 0.05, or if less than 180 seconds have passed since the last model switch, the current model will continue running to avoid frequent model switching caused by short-term brightness fluctuations in the inspection screen.

[0103] (10) Load the target visual model and output the switching result;

[0104] The currently running visual model Switch to target vision model and the target visual model Loaded into the edge intelligent analysis module, the switching results are output after the model switching is completed. The switching results include the target model identifier, the current switching status, the resource availability score, the scene complexity score, the comprehensive matching score, and the switching basis, which can be called by subsequent substation inspection image processing tasks.

[0105] In this embodiment, the edge intelligent analysis module performs the following: arrive The switching process is completed, and the output switching results are shown in Table 3.

[0106] Table 3. Example of model switching result output;

[0107] Target model identifier <![CDATA[M B ]]> Current switching state <![CDATA[Already switched by M L Switched to M B > Resource availability score 0.578 Scene complexity score 0.418 Target Model Comprehensive Matching Score 0.764 Current model comprehensive matching score 0.688 Rating difference 0.076 Switch judgment result It satisfies 0.076 > 0.05 and 435s > 180s. Output processing method <![CDATA[Call M starting from the next sampling period B Identify the inspection images]]>

[0108] After the above switch is completed, subsequent inspection images will be from The system will perform identification; if resource availability significantly decreases in subsequent sampling times, for example, if processor utilization rises above 75% and power consumption exceeds 80%, the comprehensive matching score will be recalculated. The system can revert to its initial state after meeting the threshold and minimum dwell time conditions. If strong nighttime glare, severe occlusion, or dense clusters of multiple targets further increase scene complexity, while resource availability remains high, the system can further assess whether to switch to [a different mode / approach]. Through the above process, the edge intelligent analysis module can dynamically switch models based on changes in resource status and scene during actual inspections.

[0109] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for dynamic switching of edge-side visual models for substation inspection tasks, characterized in that, Includes the following steps: (1) A set of candidate visual models is pre-configured in the edge intelligent analysis module at the substation site. The set of candidate visual models includes at least two visual models with different complexities. (2) During the substation inspection task, the operating status parameters of the edge intelligent analysis module are periodically collected. The operating status parameters include processor utilization, available memory ratio, task queue pressure, power consumption ratio, and cache utilization. Then, an operating status vector is constructed based on the operating status parameters: ； in, express The running state vector at any given time; express Processor utilization at any given moment; express The percentage of available memory at any given time; express Constant task queue pressure; express The percentage of power consumption at any given moment; express Cache usage at any given moment; (3) Synchronously collect scene state parameters corresponding to the current patrol image. Scene state parameters include image brightness, image sharpness, occlusion degree, target density, and weather conditions; then construct a scene state vector based on the scene state parameters: ； in, express The scene state vector at any given time; express The brightness of the patrol image at any given time; express The clarity of the patrol images at any given time; express The degree of occlusion at any given time; express The density of targets at any given moment; express The weather conditions at any given time; (4) Calculate the resource availability score at the current moment based on the running state vector. ; ； in, These are the weighting coefficients; (5) Calculate the scene complexity score at the current moment based on the scene state vector. ; ； in, Indicates the standard brightness reference value; These are the weighting coefficients; (6) Calculate the model requirement score for each candidate visual model; ； in, Indicates the first Candidate visual models Model requirement score; express The computational load; express Storage usage; express The average inference latency; express Operating power consumption; These are the weighting coefficients; (7) Calculate the overall matching score of each candidate visual model; ； in, express exist A comprehensive matching score for each moment; Indicates the current running visual model Switch to the Candidate visual models The switching cost; These are the weighting coefficients; (8) Select the candidate visual model with the highest comprehensive matching score from the candidate visual model set as the target visual model; ； in, Represents the set of candidate visual models; This represents the target visual model selected at time t; (9) Determine the current substation status. Whether to perform visual model switching at any given time; Selected target visual model With the currently running visual model If there is a discrepancy, further determine whether the difference in the overall matching scores between the two exceeds a preset switching threshold. And the currently running visual model Whether the minimum stay time requirement has been met is determined by the following criteria: ； in, express The overall matching score; express The overall matching score; Indicates the preset switching threshold; Indicates the time of the last visual model switch; Indicates the minimum stay time; When both the score difference condition and the minimum dwell time condition are met, the visual model is switched and the process proceeds to step (10); otherwise, the current visual model continues to run and the process ends. (10) Load the target visual model and output the switching result; The currently running visual model Switch to target vision model and the target visual model Loaded into the edge intelligent analysis module, the switching results are output after the model switching is completed. The switching results include the target model identifier, the current switching status, the resource availability score, the scene complexity score, the comprehensive matching score, and the switching basis, which can be called by subsequent substation inspection image processing tasks.

2. The method for dynamic switching of edge-side visual models for substation inspection tasks according to claim 1, characterized in that, The candidate visual model set includes lightweight visual models, balanced visual models, and high-precision visual models; lightweight visual models are used for inspection scenarios where edge device resources are scarce or real-time requirements are high; balanced visual models are used for inspection scenarios where the edge device resource status and the complexity of the inspection scenario are at an intermediate level. High-precision visual models are used for inspection scenarios where edge devices have sufficient resources and the inspection scenarios are complex; each candidate visual model is pre-recorded with model identifier, computational cost, storage usage, average inference latency and power consumption.

Images