Pumped storage power station cloud management side end collaborative video monitoring system and method

Abstract

The invention discloses a cloud management side end collaborative video monitoring system and method for a pumped storage power station. The method comprises the following steps of: acquiring a video stream in real time by a sensing terminal; identifying and analyzing the video stream by an edge computing node; transmitting the video stream by a pipeline network; and processing the video stream by a cloud platform. According to the technical scheme, technologies of artificial intelligence, cloud computing, edge computing, intelligent sensing and the like are combined, a cloud, management, edge and end collaborative method is adopted, multi-dimensional resources such as communication, computing, storage, energy and the like are integrated, and real-time processing and intelligent research and judgment of video monitoring data of the pumped storage power station are achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and particularly relates to a cloud-pipeline-edge collaborative video monitoring method and system for pumped storage power stations. Background Art

[0002] In the wave of global energy transformation, promoting the low-carbonization of the energy structure has become the core goal of achieving sustainable development, and the rise of smart grids marks that the power system is moving towards a new era of intelligence and high efficiency. In this transformation, energy storage systems, as the cornerstone of smart grids, are becoming increasingly important, stimulating the industry's unprecedented enthusiasm and exploration for energy storage technologies. Energy storage has the same important status as power generation, and once a failure occurs in an energy storage power station, it will also have a significant impact on the country's economy and security. Currently, the video monitoring of energy storage power stations mainly relies on manual labor, and it is very easy to have omissions. Summary of the Invention

[0003] In order to solve the problems existing in the video monitoring of energy storage power stations in the present technology, the present application provides a cloud-pipeline-edge collaborative video monitoring method and system for pumped storage power stations, which can enable the instant processing of video monitoring data at the production site, reduce data transmission delay, save network bandwidth, and has extremely important application value especially in the fields of intelligent operation and maintenance, intelligent inspection, intelligent video monitoring, etc. Making full use of cloud-pipeline-edge collaborative technology will play a positive role in promoting the successful transformation of the intelligent production safety of pumped storage power stations.

[0004] To achieve the above object, an intelligent body system for cloud-pipeline-edge collaborative video monitoring of a pumped storage power station according to an embodiment of the present invention includes: (1) a cloud platform, (2) a pipeline network, (3) an edge computing node equipped with an edge computing framework and edge-side application programs, and (4) a sensing terminal;

[0005] Among them, (1) the cloud platform is the core component, providing the capabilities of data storage, calculation, and aggregation, and is the brain of the intelligent body for video monitoring of the pumped storage power station, a management platform deployed in the cloud;

[0006] Among them, (2) the pipeline network is the connecting bridge, responsible for transmitting the data collected by the edge devices to the cloud platform, and is the limbs of the intelligent body for video monitoring of the pumped storage power station, various remote communication networks connecting the edge and the cloud platform, realizing data access and backhaul.

[0007] Among them, (3) the edge computing node equipped with an edge computing framework and edge-side application programs is the infrastructure, the source of data analysis and transmission, and is the skeleton of the intelligent body for video monitoring of the pumped storage power station, software and hardware resources deployed on-site with edge computing and cloud-edge interaction capabilities, realizing the recognition, analysis, caching, cleaning, and uploading of video data;

[0008] Among them, the (4) perception terminal is a video acquisition terminal, which realizes the perception of source data and is the sensory organ of the video monitoring intelligent agent of the pumped-storage power station. It is deployed in various terminals inside or near the acquisition object to realize the acquisition of on-site safety production video data.

[0009] At the same time, the embodiment of the present invention proposes an edge computing framework for the cloud-pipe-edge-end collaborative video monitoring intelligent agent system of a pumped-storage power station, including: (1) the edge computing basic platform of the video monitoring intelligent agent of the pumped-storage power station, (2) the edge computing enabling application of the video monitoring intelligent agent of the pumped-storage power station, and (3) the edge computing collaborative hub of the video monitoring intelligent agent of the pumped-storage power station;

[0010] Among them, the (1) edge computing basic platform of the video monitoring intelligent agent of the pumped-storage power station includes: the edge computing framework integrates edge capabilities such as network communication, analysis and calculation, data storage, resource and application management, and is the platform basis for the edge computing of the video monitoring intelligent agent of the pumped-storage power station. Based on container technology, it can be installed on various hardware at the sites of different pumped-storage power stations to provide edge platform services;

[0011] Among them, the (2) edge computing enabling application of the video monitoring intelligent agent of the pumped-storage power station includes: it can be oriented to the edge-side applications of the video monitoring intelligent agent of the pumped-storage power station. By opening its basic service capabilities such as communication, data, and models, the edge-side video monitoring applications of the video monitoring intelligent agent of the pumped-storage power station can focus on the on-site safety production of the pumped-storage power station itself, meeting the needs of rapid development and flexible deployment;

[0012] Among them, the (3) edge computing collaborative hub of the video monitoring intelligent agent of the pumped-storage power station includes: the edge computing framework uses standard communication protocols to achieve video monitoring data collaboration and the interactive collaboration of devices, containers, and applications for the cloud above, and provides communication hub services such as data reporting and command issuing for the devices on the edge side of the video monitoring intelligent agent of the pumped-storage power station below.

[0013] At the same time, the embodiment of the present invention proposes a cloud-edge collaborative hub for the cloud-pipe-edge-end collaborative video monitoring intelligent agent system of a pumped-storage power station, including: (1) the cloud data center of the video monitoring intelligent agent of the pumped-storage power station and (2) the edge computing platform of the video monitoring intelligent agent of the pumped-storage power station;

[0014] Among them, the (1) cloud data center of the video monitoring intelligent agent of the pumped-storage power station is used to deploy the AI platform to realize sample management and model management. Among them, sample management provides the capabilities of sample collection, sample annotation, sample review, and sample sharing, is responsible for the aggregation and release of sample data, and provides a data basis for the design and development of AI algorithms. Model management provides the capabilities of model training, model encapsulation, model release, and model evaluation, and is responsible for generating AI algorithms according to business scenarios and training, iterating, and optimizing the algorithms;

[0015] Among them, the (2) intelligent edge computing platform for video monitoring of pumped-storage power stations consists of three parts: a business center, a cloud-edge-end collaborative management component, and edge nodes. Among them, the business center includes business management functions such as scenario registration, scenario routing, data processing, permission management, and event management, and is responsible for receiving the algorithm calculation results of edge nodes, docking with downstream business systems as the center, and realizing the application of the analysis and calculation results of edge nodes. The cloud-edge-end collaborative management component is the core of edge node management and control, realizing functions such as edge node management, container application management, cloud-edge network management, node monitoring management, edge pooling and sharing, and parameter configuration management, and managing devices, applications, and algorithms. The cloud-edge-end collaborative management component includes an edge node management platform and edge node software. The edge node management platform can uniformly manage edge nodes in terms of resources, container applications, function applications, operation and maintenance monitoring, cloud-edge messages, etc., and has the collaborative management capabilities in aspects such as cloud-edge message collaboration, metering and billing, and security risks between the cloud and edge nodes; the cloud manages the terminal devices of the edge nodes connected through the edge node management platform and deploys AI algorithms to the edge nodes. The edge node software combines a scheduling task framework orchestration (K8S) and a lightweight container runtime environment (Docker) to provide a basic environment for the operation of AI algorithms. In addition, the cloud-edge-end collaborative management component adopts an image repository and image caching mechanism to realize the management and distribution of business model algorithm application images. The image cache is deployed on the edge nodes to reduce the network pressure during image distribution. Terminal devices such as cameras, robots, and drones collect data and hand it over to the edge nodes for calculation. The edge nodes receive the data collected by the terminal devices, run AI algorithms, and output edge computing results. On the one hand, it provides them to the business center for violation warnings and picture pushing, and on the other hand, it provides them to the AI middle platform for sample aggregation.

[0016] Meanwhile, the embodiment of the present invention proposes an application of the business process of the cloud-pipe-edge collaborative video monitoring intelligent agent system for a pumped-storage power station, including: (1) Analyze business requirements and develop algorithms. Business personnel analyze business pain points, discover business problems that need to be solved using artificial intelligence, and submit requirements to the algorithm development department; the algorithm development department sorts out business logic according to business requirements, develops corresponding AI algorithms, and uses the AI training model to complete the algorithm training and release for targeted scenarios; (2) Develop applications and publish images. After the application development is completed, it is made into a container image and uploaded to the image server. After the edge computing platform issues the application, the edge node can pull the application image from the image repository; (3) Configure business parameters. Business personnel complete the configuration of business parameters for corresponding scenarios on the platform, such as camera angles, employee clothing, identification area division, etc.; (4) Deploy applications. The algorithm image and business parameter configuration are scheduled and issued by the edge computing platform to become the application load on the edge device; (5) Configure the intelligent center. Configure scenario rules and scenario routing in the business center; (6) Incorporate edge nodes and access terminal devices. The edge computing platform incorporates the edge nodes into the management of the platform and binds the terminal devices to the edge nodes; (7) Algorithm application. After the algorithm and parameters are issued, the application on the edge device uses local computing resources for real-time AI computing, and the real-time computing results are uniformly sent to the business center. The computing results are further processed and forwarded by the business center; (8) Disposal and feedback of computing results. The processed computing results are sent by the business center to each downstream business system, and the business system conducts specific disposal and feedback.

[0017] Meanwhile, the embodiment of the present invention proposes a process for constructing an algorithm model of the cloud-pipe-edge collaborative video monitoring intelligent agent system for a pumped-storage power station:

[0018] Step 1, Algorithm model for processing latency

[0019] (1) The computational latency of the edge device (camera) executing the neural network task is expressed as follows:

[0020]

[0021] Where

[0022] T loc u,l (t) = Y u,l (t).ξ u,loc (t),

[0023] Here, T u,loc (t) represents the overall computational latency of the edge device u at a specific time point t, that is, all local latencies T loc u,lThe sum of (t). Among them, l represents different local computing tasks or levels, and ku is the number of local tasks or levels on the edge device u. T loc u,l (t) represents the computing latency of the l-th local task of the edge device u at time point t, which consists of Y u,l (t) and ξ u,loc (t). Y u,l (t) represents the computing requirement or computing complexity of the local task at time t, and ξ u,loc (t) represents the computing resource or performance metric of the edge device u at time t.

[0024] (2) The computing latency of the edge computing server for executing the neural network task is expressed as follows:

[0025]

[0026] Among them,

[0027] T mec u,l (t) = Y u,l (t) · ξ mec (t),

[0028] Here, T u,mec (t) represents the overall computing latency of the edge computing server u at a specific time point t, that is, the sum of all local latencies T mec u,l (t), where l represents different local computing tasks or levels, Lu is the total number of local tasks or levels on the edge computing server u, and k u is a specific level marker. T mec u,l (t) represents the computing latency of the l-th local task of the edge computing server u at time point t, which consists of Y u,l (t) and ξ mec (t). Y u,l (t) represents the computing requirement or computing complexity of this local task at time t. ξ mec (t) represents the computing resource or performance metric of the edge computing server at time t.

[0029] (3) The transmission latency of uploading the output data of the edge device (camera) to the edge computing server is:

[0030] The latency is:

[0031]

[0032] Among them,

[0033]

[0034] Here, T u,up (t) represents the transmission delay when the edge device u uploads data to the edge computing server at time point t. represents the amount of data output by the ku-th layer of the edge device u at time point t. R u,up (t) represents the upload rate from the edge device u to the edge computing server at time point t. Among them, B u (t) represents the available bandwidth of the edge device u at time point t. h u (t) represents the channel gain of the edge device u at time point t. P up y (t) represents the upload power of the edge device u at time point t. N 0 represents the noise power spectral density.

[0035] (4) The download delay of data from the edge computing server to the edge device (camera) is:

[0036]

[0037] Among them,

[0038]

[0039] Here, T u,do (t) represents the download delay from the edge computing server to the edge device u at time point t. represents the amount of data sent from the edge computing server to the edge device u at time point t. R u,do (t) represents the download rate from the edge device u to the edge computing server at time point t. Among them, B u (t) represents the available bandwidth of the edge device u at time point t. h u (t) represents the channel gain of the edge device u at time point t. P M (t) represents the download power of the edge device u at time point t. N 0 represents the noise power spectral density.

[0040] (5) The total delay of the edge device (camera) can be modeled as:

[0041]

[0042] Among them,

[0043]

[0044] Here, T u(t) represents the total latency of the edge device (camera), which is divided into 4 cases: (1) 0 indicates that in the ideal state, there is no latency in the edge device (camera). (2) T u,loc (t) represents the local computing latency of the edge device at time point t. (3) represents the total latency of the edge device in the first hybrid operation mode at time point t, including the edge computing latency T u,mec (t), the upload latency T u,up (t), and the download latency T u,do (t). (4) represents the total latency of the edge device in the second hybrid operation mode at time point t, including the local computing latency T u,loc (t), the edge computing latency T u,mec (t), the upload latency T u,up (t), and the download latency T u,do (t).

[0045] Step 2. Energy consumption model

[0046] (1) The computing energy consumption of the neural network task on the edge device (camera) can be modeled

[0047] as:

[0048] E u,loc (t) = P exe u (t) T u,loc (t),

[0049] Here, E u,loc (t) represents the computing energy consumption of the edge device at time point t. P exe u (t) represents the computing power consumption of the edge device at time point t. T u,loc (t) represents the computing latency of the edge device at time point t.

[0050] (2) The energy consumed for uploading the output data after the edge device (camera) executes to the edge computing server is

[0051] as follows:

[0052] E u,up (t) = P up u (t) min{T u,up (t), τ},

[0053] Here, E u,up (t) represents the energy consumed for uploading data from the edge device to the edge computing server at time point t. P up u(t) represents the power consumption of the edge device when uploading data at time point t. T u,up (t) represents the transmission delay of the edge device when uploading data to the edge computing server at time point t. τ represents a preset time threshold.

[0054] (3) The total energy consumption of the edge device (camera) is expressed as follows:

[0055]

[0056] Among them,

[0057] E u,M (t) = E u,loc (t) + E u,up (t)

[0058] Here, E u (t) represents the total energy consumption of the edge device (camera), which is divided into 4 cases:

[0059] (1) 0 represents the ideal state where there is no energy consumption in the edge device (camera). (2)

[0060] E u,loc (t) represents the local computing energy consumption of the edge device at time t. (3) E u,up (t) represents the energy consumed by the edge device when uploading data to the edge computing server at time point t. (4) E u,M (t) represents the total energy consumption of the edge device in the hybrid operation mode at time point t, including the local computing energy consumption E u,loc (t) and the energy consumed E u,up (t).

[0061] Step 3, Problem Modeling

[0062] (1) The total cost of the edge device (camera) consists of delay, energy cost, and penalty term, and its

[0063] is formalized as:

[0064] C u (t) = u·T u (t) + v·E u (t) + w·Q ′ u (t),

[0065] Here, C u (t) represents the total cost of the edge device (camera). T u (t) represents the total delay of the edge device at time point t, including local computing delay, data upload delay, and data download delay. v·E u(t) represents the total energy consumption of the edge device at time point t, including local computing energy consumption and data transmission energy consumption. Q ′ u (t) is a penalty term, representing the additional cost when certain performance metrics are not met (such as latency or energy consumption exceeding a preset threshold). u, v, and w are weight coefficients used to adjust the importance of latency, energy consumption, and the penalty term in the total cost.

[0066] (2) The task processing cost minimization problem can be expressed as:

[0067] P1:

[0068]

[0069] min

[0070] ξ mec (t) ≥ ξ mec ,

[0071]

[0072] Here, P1 represents the task processing cost minimization problem, aiming to minimize the long-term average total cost. C u (t) represents the total cost of edge device u at time point t, consisting of latency, energy cost, and penalty term. The goal is to find the optimal task allocation and resource allocation strategies to minimize this long-term average total cost. The constraints include: (1) This constraint ensures that each edge device u either executes the task locally at any time point t

[0073] (a u,L (t) = L u ,a u,M (t) = 0), or offloads the task to the edge computing server (a u,L (t) = 0,a u,M (t) = L u ). (2) This constraint ensures that the local computing resources of edge device u at time point t reach at least the minimum threshold ξ mim u,loc . (3) ξ mec (t) ≥ ξ min mec : This constraint ensures that the computing resources of the edge computing server at time point t reach at least the minimum threshold ξ min mec . (4) This constraint ensures that the upload rate of edge device u at time point t is between 0 and the maximum upload rate R maxu,up Between. (5) This constraint ensures that the download rate of the edge device u at time point t is between 0 and the maximum download rate R max u,do Between.

[0074] The above problem is a mixed integer programming problem with high computational complexity, and the time complexity and space complexity of using traditional optimization methods are too high. Therefore, a Markov decision process is used for dynamic modeling. Finally, a deep reinforcement learning algorithm is used to optimize and solve the Markov decision process.

[0075] Action:

[0076] a u (t)=[a u,L (t),a u,M (t)]

[0077] Here, the action a u (t) represents the operation taken by the edge device u at time point t, and it is a vector containing elements a u,L (t) and a u,M (t) elements. a u,L (t) represents whether the edge device u chooses to execute the task locally at time point t. If a u,L (t)=L u , the device executes the task locally; if a u,L (t)=0, the device does not execute the local task. a(u,M)(t) represents whether the edge device u chooses to offload the task to the edge computing server at time point t. If a u,L (t)=L u , the device offloads the task to the edge server; if a u,L (t)=0, the device does not offload the task.

[0078] State:

[0079] S u (t)=[D u (t),h u (t),Q u (t)]∈S

[0080] Here, the state S u (t) describes the situation of the edge device u at time point t, and it is a vector containing multiple elements, and each element represents a different attribute of the device. D u (t) represents the amount of data or tasks of the edge device u at time point t. This may include the amount of sensor data to be processed, the number of tasks to be executed, etc. h u(t) represents the channel state of the edge device u at time point t, which may include information such as the signal-to-noise ratio, signal strength, and channel capacity of the wireless channel. Q u (t) represents the queue length or buffer state of the edge device u at time point t, which may include the amount of data waiting to be processed, the number of tasks to be uploaded, etc.

[0081] Reward:

[0082] R u (t) = C u (t)

[0083] Here, the reward R u (t) can be defined as the total cost C u (t) of the edge device u at time point t.

[0084] Meanwhile, the embodiment of the present invention proposes an algorithm model for the cloud-pipe-edge-end collaborative video surveillance intelligent body system of a pumped storage power station: y = γ t +γmaxa'Qθ'(S t+1 ,a'). Where, θ represents the parameters in the evaluation network; θ' represents the parameters in the target network; S t represents the state of the system at time t; S t+1 represents the state of the system at time t+1; a t represents the decision-making action at time t; a' represents the decision-making action at time t+1; Qθ(S t ,a t ) represents the Q value obtained by the evaluation network taking the action a t under the state S t ; Qθ'(S t+1 ,a') represents the Q value obtained by the target network taking the action a' under the state S t+1 ; γ t represents the reward obtained by taking the action a t under the state S t ; y represents the reward decay ratio. The specific algorithm process design:

[0085] (1) Initialize the evaluation network, target network, and memory bank in deep reinforcement learning. The current system state is S t , t is initialized to 1, and the iteration count k is initialized to 1;

[0086] (2) When k is less than or equal to the given iteration count K, if k modulo m is 0, then update the current state S t to the best state so far; if k modulo m is not 0, then randomly select a probability p;

[0087] (3) If p is less than or equal to the greedy policy probability ξ, then select the decision-making action a output by the evaluation network t , otherwise randomly select an action;

[0088] (4) After taking the decision-making action a, obtain the reward γ and the next state S, and save them in the memory bank in the format (S t , a t , γ t , S t+1 );

[0089] (5) Combine the output of the target network and calculate the output of the evaluation network: y = γ t + γ max a' Qθ'(S t+1 , a');

[0090] (6) Use the minimized error (y - Qθ(S t+1 , a t )) 2 to evaluate, obtain the evaluation network parameter θ, and update it;

[0091] (7) Every S steps, assign the parameters of the evaluation network to the target network, and at the same time set k = k + 1 and return to step (2) for iteration;

[0092] (8) When k is greater than the given number of iterations K, the learning process ends and the optimal decision-making is obtained.

[0093] The beneficial effects of the above technical solutions of this application are as follows: The solution of the embodiment of this application can enable the video surveillance data to be processed immediately at the production site, reduce the data transmission delay, save network bandwidth, and has extremely important application value especially in the fields of intelligent operation and maintenance, intelligent inspection, intelligent video surveillance, etc. Making full use of the cloud-edge-end collaborative technology will play a positive role in promoting the successful transformation of the safety production intelligence of pumped storage power stations. The technical solutions of the embodiments of this application have at least one of the following advantages:

[0094] (1) Reduce latency: Since the video surveillance data of the pumped storage power station is processed immediately on the edge-side device, the data transmission time from the edge-side device to the cloud-side is reduced, thereby reducing latency and enhancing the ability of real-time application response in typical scenarios.

[0095] (2) Improve network efficiency: The video surveillance intelligent agent of the pumped storage power station based on cloud-edge-end collaboration can use edge computing to share part of the data processing tasks during the process of processing large-scale and high-concurrency video data, reduce the load on the cloud-side network, and improve network efficiency.

[0096] (3) Protect data security: The monitoring video data of the on-site safety production process of the pumped storage power station does not need to be transmitted to the cloud side, reducing the risk of data leakage and improving data security, which is of great significance for protecting the sensitive information of the pumped storage power station.

[0097] (4) Support diversity: The video monitoring intelligent agent of the pumped storage power station based on cloud-pipe-edge-end collaboration supports video monitoring devices and platforms of different models and forms, realizing seamless access to video acquisition devices such as cameras, robots, and drones, and effectively supporting diverse integration and expansion.

[0098] (5) Reduce costs: The video monitoring intelligent agent of the pumped storage power station based on cloud-pipe-edge-end collaboration can effectively reduce the dependence on the network on the pipe-end side by using the edge computing method, so it can reduce the network bandwidth cost and cloud computing cost on the pipe-end side.

[0099] (6) Enhance user experience: The video monitoring intelligent agent of the pumped storage power station based on cloud-pipe-edge-end collaboration can respond to the needs and operations of end-users faster on the edge-end side through edge computing, providing a smoother and more convenient real-time interaction experience for end-users.

[0100] (7) Promote technological innovation: The video monitoring intelligent agent of the pumped storage power station based on cloud-pipe-edge-end collaboration provides a powerful impetus for intelligent video monitoring. It not only improves the data processing efficiency but also reduces the consumption of network resources, enabling better integration of artificial intelligence technology and Internet of Things technology, and further promoting the development of the intelligent Internet of Things. Description of the Drawings

[0101] The following drawings are used to provide a further understanding of the present invention. The schematic examples and descriptions of the present invention are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0102] Figure 1 is the schematic structural diagram of the cloud-pipe-edge-end collaborative video monitoring intelligent agent system of the pumped storage power station of the present invention;

[0103] Figure 2 is the edge computing framework diagram of the cloud-pipe-edge-end collaborative video monitoring intelligent agent system of the pumped storage power station of the present invention;

[0104] Figure 3 is the cloud-edge collaborative hub diagram of the cloud-pipe-edge-end collaborative video monitoring intelligent agent system of the pumped storage power station of the present invention;

[0105] Figure 4 is the business process application diagram of the cloud-pipe-edge-end collaborative video monitoring intelligent agent system of the pumped storage power station of the present invention;

[0106] Figure 5It is a diagram for constructing the algorithm model of the cloud-pipeline-edge-end collaborative video monitoring intelligent agent system of the pumped storage power station of the present invention;

[0107] Figure 6 It is a diagram of the algorithm flow method of the cloud-pipeline-edge-end collaborative video monitoring intelligent agent system of the pumped storage power station of the present invention; Detailed implementation manners

[0108] It should be noted that, without conflict, the examples and features in this application can be combined with each other. Unless explicitly required, individual components and functions are optional, and the order of operations can be changed. Parts and features of some embodiments can be included in or replace parts and features of other embodiments. The scope of the embodiments of the present invention includes the entire scope of the claims and all available equivalents of the claims. The present invention will be described in detail below with reference to the drawings and in conjunction with examples.

[0109] According to an embodiment of the present invention, a cloud-pipeline-edge-end collaborative video monitoring intelligent agent system for a pumped storage power station is provided; as Figure 1 shown, it includes: (1) a cloud platform, (2) a pipeline network, (3) an edge computing node equipped with an edge computing framework and edge-side application programs, and (4) a sensing terminal;

[0110] Among them, (1) the cloud platform is the core component, providing the capabilities of data storage, computing, and aggregation. It is the brain of the video monitoring intelligent agent of the pumped storage power station, and is a management platform deployed in the cloud;

[0111] Among them, (2) the pipeline network is the connection bridge, responsible for transmitting the data collected by the edge devices to the cloud platform. It is the limbs of the video monitoring intelligent agent of the pumped storage power station, and is various remote communication networks that connect the edge and the cloud platform to realize the access and backhaul of data.

[0112] Among them, (3) the edge computing node equipped with an edge computing framework and edge-side application programs is the infrastructure, the source of data analysis and transmission. It is the skeleton of the video monitoring intelligent agent of the pumped storage power station, and is the software and hardware resources deployed on-site with edge computing and cloud-edge interaction capabilities to realize the recognition, analysis, caching, cleaning, and uploading of video data;

[0113] Among them, (4) the sensing terminal is a video acquisition terminal, realizing the perception of the source data. It is the sensory organ of the video monitoring intelligent agent of the pumped storage power station, and is various terminals deployed inside or near the acquisition object to realize the acquisition of on-site safety production video data.

[0114] According to an embodiment of the present invention, an edge computing framework for a cloud-pipeline-edge-end collaborative video monitoring intelligent agent of a pumped storage power station is provided; as Figure 2As shown in the figure, it includes: (1) the edge computing basic platform of the video monitoring intelligent agent for pumped-storage power stations, (2) the edge computing enabling applications of the video monitoring intelligent agent for pumped-storage power stations, and (3) the edge computing collaborative hub of the video monitoring intelligent agent for pumped-storage power stations;

[0115] Among them, (1) the edge computing basic platform of the video monitoring intelligent agent for pumped-storage power stations includes: the edge computing framework integrates edge capabilities such as network communication, analysis and calculation, data storage, resource and application management, and is the platform basis for the edge computing of the video monitoring intelligent agent for pumped-storage power stations. Based on container technology, it can be installed on various types of hardware at the sites of different pumped-storage power stations to provide edge platform services;

[0116] Among them, (2) the edge computing enabling applications of the video monitoring intelligent agent for pumped-storage power stations include: it can be oriented to the edge-side applications of the video monitoring intelligent agent for pumped-storage power stations. By opening its basic service capabilities such as communication, data, and models, the edge-side video monitoring applications of the video monitoring intelligent agent for pumped-storage power stations can focus on the on-site safety production of pumped-storage power stations themselves, meeting the needs of rapid development and flexible deployment;

[0117] Among them, (3) the edge computing collaborative hub of the video monitoring intelligent agent for pumped-storage power stations includes: the edge computing framework uses standard communication protocols to achieve video monitoring data collaboration and the interactive collaboration of devices, containers, and applications with the cloud above, and provides communication hub services such as data reporting and command issuing to the edge-side devices of the video monitoring intelligent agent for pumped-storage power stations below.

[0118] According to an embodiment of the present invention, a cloud-edge collaborative hub of the video monitoring intelligent agent for the cloud-pipe-edge-end collaboration of pumped-storage power stations is provided; as Figure 3 shown in the figure, it includes: (1) the cloud data center of the video monitoring intelligent agent for pumped-storage power stations and (2) the edge computing platform of the video monitoring intelligent agent for pumped-storage power stations;

[0119] Among them, (1) the cloud data center of the video monitoring intelligent agent for pumped-storage power stations is used to deploy an AI platform to realize sample management and model management. Among them, sample management provides the capabilities of sample collection, sample annotation, sample review, and sample sharing, is responsible for the aggregation and release of sample data, and provides a data basis for the design and development of AI algorithms. Model management provides the capabilities of model training, model encapsulation, model release, and model evaluation, and is responsible for generating AI algorithms according to business scenarios and training, iterating, and optimizing the algorithms;

[0120] Among them, the (2) intelligent edge computing platform for video monitoring of pumped-storage power stations consists of three parts: a business center, a cloud-edge-end collaborative management component, and edge nodes. Among them, the business center includes business management functions such as scenario registration, scenario routing, data processing, permission management, and event management, responsible for receiving the algorithm calculation results of edge nodes, docking with downstream business systems as the center, and realizing the application of the analysis and calculation results of edge nodes. The cloud-edge-end collaborative management component is the core of edge node management, realizing functions such as edge node management, container application management, cloud-edge network management, node monitoring management, edge pooling and sharing, and parameter configuration management, and managing devices, applications, and algorithms. The cloud-edge-end collaborative management component includes an edge node management platform and edge node software. The edge node management platform can uniformly manage edge nodes in terms of resources, container applications, function applications, operation and maintenance monitoring, cloud-edge messages, etc., and has the collaborative management capabilities in aspects such as cloud-edge message collaboration, metering and billing, and security risks between the cloud and edge nodes; the cloud manages the terminal devices of the edge nodes connected through the edge node management platform and deploys AI algorithms to the edge nodes. The edge node software combines a scheduling task framework orchestration (K8S) and a lightweight container runtime environment (Docker) to provide a basic environment for the operation of AI algorithms. In addition, the cloud-edge-end collaborative management component adopts an image repository and image caching mechanism to realize the management and distribution of business model algorithm application images. The image cache is deployed on the edge nodes to reduce the network pressure during image distribution. Terminal devices such as cameras, robots, and drones collect data and hand it over to the edge nodes for calculation. The edge nodes receive the data collected by the terminal devices, run AI algorithms, and output edge computing results. On the one hand, it provides them to the business center for violation warning and picture push, and on the other hand, it provides them to the AI middle platform for sample aggregation.

[0121] According to an embodiment of the present invention, there is provided a business process application of a cloud-edge-end collaborative video monitoring intelligent agent for a pumped-storage power station; as Figure 4As shown in the figure, it includes: (1) Analyze business requirements and develop algorithms. Business personnel analyze business pain points, discover business problems that need to be solved using artificial intelligence, and put forward requirements to the algorithm development department; the algorithm development department sorts out business logic according to business requirements, develops corresponding AI algorithms, and uses the AI training model to complete algorithm training and release for targeted scenarios; (2) Develop applications and publish images. After the application development is completed, it is made into a container image and uploaded to the image server. After the edge computing platform issues the application, the edge node can pull the application image from the image repository; (3) Configure business parameters. Business personnel complete the configuration of business parameters for corresponding scenarios on the platform, such as camera angle, employee clothing, identification area division, etc.; (4) Deploy applications. The algorithm image and business parameter configuration are scheduled and issued by the edge computing platform and become the application load on the edge device; (5) Configure the intelligent center. Configure scenario rules and scenario routing in the business center; (6) Incorporate edge nodes and access terminal devices. The edge computing platform incorporates edge nodes into the platform management and binds terminal devices to edge nodes; (7) Algorithm application. After the algorithm and parameters are issued, the application on the edge device uses local computing resources for real-time AI calculation, and the real-time calculation results are uniformly sent to the business center. The calculation results are further processed and forwarded by the business center; (8) Disposal and feedback of calculation results. The processed calculation results are sent by the business center to each downstream business system, and the business system conducts specific disposal and feedback.

[0122] According to an embodiment of the present invention, there is provided a construction of an intelligent body algorithm model for collaborative video surveillance of a pumped storage power station cloud tube edge end; as Figure 5 shown, it includes:

[0123] Step 1, an algorithm model for processing delay

[0124] (1) The calculation delay of the end-side device (camera) performing neural network tasks is expressed as follows:

[0125]

[0126] Among them,

[0127] T loc u,l (t) = Y u,l (t).ξ u,loc (t),

[0128] Here, T u,loc (t) represents the overall calculation delay of the end-side device u at a specific time point t, that is, the sum of all local delays T loc u,l (t). Among them, l represents different local computing tasks or levels, and ku is the number of local tasks or levels on the end-side device u. T locu,l (t) represents the computing latency of the l-th local task of the edge device u at time point t, which consists of Y u,l (t) and ξ u,loc (t). Y u,l (t) represents the computing requirement or computing complexity of the local task at time t, and ξ u,loc (t) represents the computing resource or performance metric of the edge device u at time t.

[0129] (2) The computing latency of the edge computing server executing the neural network task is expressed as follows:

[0130]

[0131] Among them,

[0132] T mec u,l (t) = Y u,l (t) · ξ mec (t),

[0133] Here, T u,mec (t) represents the overall computing latency of the edge computing server u at a specific time point t, that is, the sum of all local latencies T mec u,l (t), where l represents different local computing tasks or levels, Lu is the total number of local tasks or levels on the edge computing server u, and k u is a specific level marker. T mec u,l (t) represents the computing latency of the l-th local task of the edge computing server u at time point t, which consists of Y u,l (t) and ξ mec (t). Y u,l (t) represents the computing requirement or computing complexity of this local task at time t. ξ mec (t) represents the computing resource or performance metric of the edge computing server at time t.

[0134] (3) The transmission latency of uploading the output data of the edge device (camera) to the edge computing server is:

[0135] Latency is:

[0136]

[0137] Among them,

[0138]

[0139] Here, T u,up(t) represents the transmission delay when the edge device u uploads data to the edge computing server at time point t. represents the amount of data output by the ku-th layer of the edge device u at time point t. R u,up (t) represents the upload rate from the edge device u to the edge computing server at time point t. Among them, B u (t) represents the available bandwidth of the edge device u at time point t. h u (t) represents the channel gain of the edge device u at time point t. P up y (t) represents the upload power of the edge device u at time point t. N 0 represents the noise power spectral density.

[0140] (4) The download delay of data from the edge computing server to the edge device (camera) is:

[0141]

[0142] Among them,

[0143]

[0144] Here, T u,do (t) represents the download delay from the edge computing server to the edge device u at time point t. represents the amount of data sent from the edge computing server to the edge device u at time point t. R u,do (t) represents the download rate from the edge device u to the edge computing server at time point t. Among them, B u (t) represents the available bandwidth of the edge device u at time point t. h u (t) represents the channel gain of the edge device u at time point t. P M (t) represents the download power of the edge device u at time point t. N 0 represents the noise power spectral density.

[0145] (5) The total delay of the edge device (camera) can be modeled as:

[0146]

[0147] Among them,

[0148]

[0149] Here, T u (t) represents the total delay of the edge device (camera), which is divided into 4 cases: (1) 0 represents the ideal state where there is no delay for the edge device (camera). (2) T u,loc(t) represents the local computing delay of the edge device at time point t. (3) represents the total delay of the edge device in the first hybrid operation mode at time point t, including the edge computing delay T u,mec (t), upload delay T u,up (t) and download delay T u,do (t). (4) represents the total delay of the edge device in the second hybrid operation mode at time point t, including the local computing delay T u,loc (t), edge computing delay T u,mec (t), upload delay T u,up (t) and download delay T u,do (t).

[0150] Step 2, Energy consumption model

[0151] (1) The computing energy consumption of the neural network task on the edge device (camera) can be modeled

[0152] as:

[0153] E u,loc (t) = P exe u (t) T i,loc (t),

[0154] Here, E u,loc (t) represents the computing energy consumption of the edge device at time point t. P exe u (t) represents the computing power consumption of the edge device at time point t. T u,loc (t) represents the computing delay of the edge device at time point t.

[0155] (2) The energy consumed when uploading the output data after the edge device (camera) executes to the edge computing server is

[0156] as follows:

[0157] E u,up (t) = P up u (t) min{T u,up (t), τ},

[0158] Here, E u,up (t) represents the energy consumed when the edge device uploads data to the edge computing server at time point t. P up u (t) represents the power consumption when the edge device uploads data at time point t. T u,up(t) represents the transmission delay of the edge device uploading data to the edge computing server at time point t. τ represents a preset time threshold.

[0159] (3) The total energy consumption of the edge device (camera) is expressed as follows:

[0160]

[0161] Where,

[0162] E u,M (t) = E u,loc (t) + E u,up (t)

[0163] Here, E u (t) represents the total energy consumption of the edge device (camera), which is divided into 4 cases:

[0164] (1) 0 represents that in the ideal state, there is no energy consumption for the edge device (camera). (2)

[0165] E u,loc (t) represents the local computing energy consumption of the edge device at time t. (3) E u,up (t) represents the energy consumed by the edge device uploading data to the edge computing server at time point t. (4) E u,M (t) represents the total energy consumption of the edge device in the hybrid operation mode at time point t, including the local computing energy consumption E u,loc (t) and the energy consumed for uploading data E u,up (t).

[0166] Step 3, Problem Modeling

[0167] (1) The total cost of the edge device (camera) consists of delay, energy cost, and penalty term, and its

[0168] is formalized as:

[0169] C u (t) = u·T u (t) + v·E u (t) + w·Q ′ u (t),

[0170] Here, C u (t) represents the total cost of the edge device (camera). T u (t) represents the total delay of the edge device at time point t, including local computing delay, data upload delay, and data download delay. v·E u (t) represents the total energy consumption of the edge device at time point t, including local computing energy consumption and data transmission energy consumption. Q ′u (t) is a penalty term representing the additional cost when certain performance metrics are not met, such as the latency or energy consumption exceeding the preset thresholds. u, v, and w are weight coefficients used to adjust the importance of latency, energy consumption, and the penalty term in the total cost.

[0171] (2) The problem of minimizing the task processing cost can be expressed as:

[0172] P1:

[0173]

[0174] min

[0175] ξ mec (t) ≥ ξ mec ,

[0176]

[0177] Here, P1 represents the problem of minimizing the task processing cost, aiming to minimize the long - term average total cost. C u (t) represents the total cost of the edge device u at time point t, which consists of latency, energy cost, and the penalty term. The goal is to find the optimal task allocation and resource allocation strategies to minimize this long - term average total cost. The constraints include: (1) This constraint ensures that each edge device u either executes the task locally at any time point t

[0178] (a u,L (t) = L u , a u,M (t) = 0), or offloads the task to the edge computing server (a u,L (t) = 0, a u,M (t) = L u ). (2) This constraint ensures that the local computing resources of the edge device u at time point t reach at least the minimum threshold ξ min u,ioc . (3) ξ mec (t) ≥ ξ min mec : This constraint ensures that the computing resources of the edge computing server at time point t reach at least the minimum threshold ξ min mec . (4) This constraint ensures that the upload rate of the edge device u at time point t is between 0 and the maximum upload rate R max u,up . (5) This constraint ensures that the download rate of the edge device u at time point t is between 0 and the maximum download rate R. max u,do

[0179] The above problem is a mixed-integer programming problem with high computational complexity, and the time complexity and space complexity of traditional optimization methods are too high. Therefore, a Markov decision process is used for dynamic modeling. Finally, a deep reinforcement learning algorithm is used to optimize and solve the Markov decision process.

[0180] Action:

[0181] a u (t)=[a u,L (t),a u,M (t)]

[0182] Here, the action a u (t) represents the operation taken by the edge device u at time point t, which is a vector containing elements a u,L (t) and a u,M (t). a u,L (t) represents whether the edge device u chooses to execute the task locally at time point t. If a u,L (t)=L u , the device executes the task locally; if a u,L (t)=0, the device does not execute the local task. a(u,M)(t) represents whether the edge device u chooses to offload the task to the edge computing server at time point t. If a u,L (t)=L u , the device offloads the task to the edge server; if a u,L (t)=0, the device does not offload the task.

[0183] State:

[0184] S u (t)=[D u (t),h u (t),Q u (t)]∈S

[0185] Here, the state S u (t) describes the situation of the edge device u at time point t. It is a vector containing multiple elements, and each element represents a different attribute of the device. D u (t) represents the amount of data or tasks of the edge device u at time point t. This may include the amount of sensor data to be processed, the number of tasks to be executed, etc. h u (t) represents the channel state of the edge device u at time point t. This may include information such as the signal-to-noise ratio, signal strength, and channel capacity of the wireless channel. Q​u (t) represents the queue length or buffer status of the edge device u at time point t, which may include the amount of data waiting to be processed, the number of tasks to be uploaded, etc.

[0186] Reward:

[0187] R u (t) = C u (t)

[0188] Here, the reward R u (t) can be defined as the total cost C u (t) of the edge device u at time point t.

[0189] According to an embodiment of the present invention, a cloud-edge-end collaborative video surveillance agent algorithm model for a pumped-storage power station is provided: y = γ t +γmaxa'Qθ'(S t+1 , a'). Where, θ represents the parameters in the evaluation network; θ' represents the parameters in the target network; S t represents the state of the system at time t; S t+1 represents the state of the system at time t+1; a t represents the decision-making action at time t; a' represents the decision-making action at time t+1; Qθ(S t , a t ) represents the Q value obtained by the evaluation network taking the action a t under the state S t ; Qθ'(S t+1 , a') represents the Q value obtained by the target network taking the action a' under the state S t+1 ; γ t represents the reward obtained by taking the action a t under the state S t ; y represents the reward decay ratio. Specific algorithm process design; as Figure 6 shown, including:

[0190] (1) Initialize the evaluation network, target network, and memory bank in deep reinforcement learning. The current system state is S t , t is initialized to 1, and the iteration count k is initialized to 1;

[0191] (2) When k is less than or equal to the given iteration count K, if k modulo m is 0, then update the current state S t to the best state so far; if k modulo m is not 0, then randomly select a probability p;

[0192] (3) If p is less than or equal to the greedy policy probability ξ, then select the decision-making action a t, otherwise randomly select an action;

[0193] (4) After taking the decision-making action a, obtain the reward γ and the next state S, and save them in the memory bank in the format (S t , a t , γ t , S t+1 );

[0194] (5) Combine the output of the target network and calculate the output of the evaluation network: y = γ t + γ max a' Qθ'(S t+1 , a');

[0195] (6) Use the minimization error (y - Qθ(S t+1 , a t )) 2 to evaluate, obtain the evaluation network parameter θ, and update it;

[0196] (7) Every S steps, assign the parameters of the evaluation network to the target network, and at the same time set k = k + 1 and return to step (2) for iteration;

[0197] (8) When k is greater than the given number of iterations K, the learning process ends and the optimal decision is obtained.

[0198] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer programs. When the computer program is loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer program may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another. For example, the computer program may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more integrated available media. The available medium may be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a high-density digital video disc (DVD)), or a semiconductor medium (such as a solid-state disk (SSD)), etc.

[0199] Those of ordinary skill in the art can understand that the various numerical numbers such as the first and second involved in the present application are only for the convenience of description and are not used to limit the scope of the embodiments of the present application, nor do they represent the order of precedence.

[0200] At least one in the present application can also be described as one or more. The plurality can be two, three, four, or more, and the present application does not make any restrictions. In the embodiments of the present application, for a technical feature, the technical features in this technical feature are distinguished by "first", "second", "third", "A", "B", "C", and "D", etc. There is no order of precedence or size order among the technical features described by the "first", "second", "third", "A", "B", "C", and "D".

[0201] The corresponding relationships shown in each table in this application can be configured or predefined. The values of the information in each table are only examples and can be configured as other values, which are not limited in this application. When configuring the corresponding relationships between the configuration information and each parameter, it is not necessarily required to configure all the corresponding relationships shown in each table. For example, in the tables of this application, the corresponding relationships shown in some rows can also not be configured. Another example is that appropriate deformation adjustments can be made based on the above tables, such as splitting, merging, etc. The names of the parameters shown in the titles of the above tables can also be other names understandable by the communication device, and the values or representation methods of the parameters can also be other values or representation methods understandable by the communication device. When implementing the above tables, other data structures can also be used, such as arrays, queues, containers, stacks, linear lists, pointers, linked lists, trees, graphs, structures, classes, heaps, hash tables or hash maps, etc.

[0202] The predefined in this application can be understood as definition, pre - definition, storage, pre - storage, pre - negotiation, pre - configuration, solidification, or pre - firing.

[0203] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or by a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0204] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above - described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0205] As described above, this is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed by this application can easily think of changes or substitutions, which should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claimed rights.

Claims

1. A method for cloud-pipe-edge collaborative video monitoring of a pumped storage power station, characterized in that: include: The sensing terminal collects video streams in real time, the edge computing node identifies and analyzes the video streams, the pipeline network transmits the video streams, and the cloud platform processes the video streams; wherein, The step of real-time video stream collection by the perception terminal is performed by various terminals deployed inside or near the collection object, which is used to capture video data and realize the collection of on-site safety production video data. It is the sense organ of the video monitoring intelligent body of the pumped storage power station; The step of edge computing node identification and analysis of video streams is performed by software and hardware resources deployed on site with edge computing and cloud-edge interaction capabilities. It is used to extract continuous frames of video stream data through the edge computing framework and edge-side applications, and realize the identification, analysis, caching, cleaning, and uploading of video data. It is the skeleton of the video monitoring intelligent body of the pumped storage power station; The pipeline network video stream transmission step is performed by various remote communication networks connecting the edge computing nodes and the cloud platform. It is used to transmit the data collected by the edge computing nodes to the cloud platform to achieve data access and return. It is the limbs of the video monitoring intelligent body of the pumped storage power station; The steps of processing video streams on the cloud platform are executed by the management platform deployed on the cloud. It is used for data storage, calculation and aggregation of video streams and is the brain of the video surveillance intelligent body of the pumped-storage power station.

2. According to the pumped storage power station cloud-pipe-edge collaborative video monitoring method as claimed in claim 1, the step of edge computing node identifying and analyzing the video stream comprises: Edge computing video stream acquisition, sending, preprocessing sub-steps, edge computing target recognition enabling application sub-steps, edge computing collaborative hub sub-steps; among them, The edge computing video stream acquisition, transmission, and preprocessing sub-steps include: using a streaming media protocol to obtain real-time video stream data of a target area from a video source; sending the video stream data to a target edge device; preprocessing the video stream data at the target edge device to obtain a number of image frames; The edge computing target recognition enabling application sub-steps include: extracting continuous frames from the video stream data; adjusting the resolution of the frames to a uniform size; performing enhancement operations on the frames after the uniform size to obtain enhanced frames; converting the enhanced frames into the format required by the algorithm; The edge computing collaborative hub sub-steps include: using standard communication protocols to carry out video surveillance data collaboration and interactive collaboration of devices, containers, and applications to the cloud, and to provide communication hub services such as data reporting and command issuance to edge devices.

3. The method for cloud-pipe-edge coordinated video monitoring of a pumped storage power station according to claim 2 is characterized in that: In the edge computing collaborative hub sub-step, the cloud is the cloud data center and the side device is the edge computing platform; The cloud data center is used to deploy the AI ​​platform and implement sample management and model management. Sample management provides the capabilities of sample collection, sample labeling, sample review and sample sharing, and is responsible for the aggregation and release of sample data, providing a data basis for the design and development of AI algorithms. Model management provides model training, model packaging, model release and model evaluation capabilities, and is responsible for generating AI algorithms based on business scenarios, and training, iterating and optimizing the algorithms. The edge computing platform includes: a business hub, a cloud-edge collaborative management component and an edge node; the business hub is used to provide business management functions, including scene registration, scene routing, data processing, permission management, and event management. It is responsible for receiving the edge node algorithm calculation results, and as a hub to connect with the downstream business system to realize the application of edge node analysis and calculation results; the cloud-edge collaborative management component is the core of edge node management, and is used to realize the following functions: edge node management, container application management, cloud-edge network management, node monitoring management, edge pooling sharing, parameter configuration management, and management of devices, applications and algorithms; the cloud-edge collaborative management component includes an edge node management platform and edge node software; the edge node management platform can uniformly manage edge nodes in resources, container applications, function applications, operation and maintenance monitoring, and cloud-edge messages, and has cloud-edge message collaboration between the cloud and edge nodes. , metering and billing, security risks, etc.; the cloud manages the terminal devices connected to the edge nodes through the edge node management platform, and deploys AI algorithms to the edge nodes; the edge node software adopts a combination of scheduling task framework orchestration and lightweight container operating environment to provide a basic environment for the operation of AI algorithms; in addition, the cloud-edge collaborative management component adopts an image warehouse and image cache mechanism to realize the management and distribution of business model algorithm application images; the image cache is deployed on the edge node to reduce the network pressure when the image is distributed; data collection is realized through terminal devices and handed over to the edge node for calculation. The terminal devices include: cameras, robots, and drones; the edge node receives the data collected by the terminal device, runs the AI ​​algorithm, and outputs the edge computing results; on the one hand, it provides it to the business center for violation alarms and picture push, and on the other hand, it provides it to the AI ​​middle station for sample aggregation.

4. The method for cloud-pipe-edge coordinated video monitoring of a pumped storage power station according to claim 1, characterized in that: The method further comprises: a business process application step; The business process application steps include: Analyze business needs and develop algorithms; business personnel analyze business pain points, discover business problems that need to be solved using artificial intelligence, and submit requirements to the algorithm development department; the algorithm development department sorts out business logic based on business needs, develops corresponding AI algorithms, and uses AI training models to complete algorithm training and release for targeted scenarios; Develop applications and publish images: After the application is developed, it is made into a container image and uploaded to the image server; after the edge computing platform sends the application, the edge node can pull the application image from the image warehouse; Configure business parameters: Business personnel complete the business parameter configuration of the corresponding scenario on the platform, including: camera angle, employee clothing, and recognition area division; Deploy applications: Algorithm images and business parameter configurations are scheduled and issued by the edge computing platform and become application loads on edge devices; Configure the smart hub: Configure scenario rules and scenario routing in the business hub; Manage edge nodes and access terminal devices: The edge computing platform manages edge nodes and binds terminal devices to edge nodes. Algorithm application: After the algorithm and parameters are issued, the application on the edge device uses local computing resources to perform real-time AI calculations, and the real-time calculation results are uniformly sent to the business center; the calculation results are further processed and forwarded by the business center; Processing and feedback of calculation results: The processed calculation results are sent by the business center to various downstream business systems, and the business systems perform specific processing and feedback.

5. The method for cloud-pipe-edge coordinated video monitoring of a pumped storage power station according to claim 1, characterized in that: The method further comprises: an algorithm model construction step; wherein the algorithm model construction step comprises: Step 1: Algorithmic model for processing delays (1) The computational delay of the end-side device in executing the neural network task is expressed as follows: in, T loc u,l (t)=Y u,l (t).ξ u,loc (t), Here, T u,loc (t) represents the overall computational delay of the end-side device u at a specific time point t, that is, all local delays T loc u,l (t); where l represents different local computing tasks or levels, ku is the number of local tasks or levels on the end-side device u; T loc u,l (t) represents the computation delay of the lth local task of the end-side device u at time point t, represented by Y u,l (t) and ξ u,loc (t) consists of two parts; Y u,l (t) represents the computational requirements or computational complexity of the local task at time t, ξ u,ooc (t) represents the computing resources or performance indicators of the end-side device u at time t; (2) The computational delay of the edge computing server in executing the neural network task is expressed as follows: in, T mec u,l (t)=Y u,l (t)·ξ mec (t), Here, T u,mec (t) represents the overall computing delay of edge computing server u at a specific time point t, that is, all local delays T mec u,l (t), where l represents different local computing tasks or levels, Lu is the total number of local tasks or levels on edge computing server u, and k u is a specific level marker; T mec u,l (t) represents the computation delay of the lth local task of edge computing server u at time point t, represented by Y u,l (t) and ξ mec (t) consists of two parts; Y u,l (t) represents the computational requirements or computational complexity of the local task at time t; ξ mec (t) represents the computing resources or performance indicators of the edge computing server at time t; (3) The transmission delay of uploading the output data of the end-side device to the edge computing server is: in, Here, T u,up (t) represents the transmission delay of the end device u uploading data to the edge computing server at time point t; represents the amount of data output by the kuth layer of the end-side device u at time point t; R u,up (t) represents the upload rate from the end device u to the edge computing server at time point t; where B u (t) represents the available bandwidth of the end-side device u at time point t; h u (t) represents the channel gain of the end-side device u at time point t; P up y (t) represents the upload power of the end-side device u at time point t; N0 represents the noise power spectrum density; (4) The download delay of data from the edge computing server to the end device (camera) is: in, Here, T u,do (t) represents the download delay from the edge computing server to the end-side device u at time point t; R represents the amount of data sent from the edge computing server to the end device u at time point t; u,do (t) represents the download rate from the end device u to the edge computing server at time point t; where B u (t) represents the available bandwidth of the end-side device u at time point t; h u (t) represents the channel gain of the end-side device u at time point t; P M (t) represents the download power of the end-side device u at time point t; N0 represents the noise power spectrum density; (5) The total latency of the end device (camera) can be modeled as: in, Here, T u (t) represents the total delay of the end-side device (camera), which is divided into four cases: (1) 0 means that in an ideal state, there is no delay in the end-side device; (2) T u,loc (t) represents the local computation delay of the end-side device at time point t; (3) represents the total delay of the end-side device in the first hybrid operation mode at time point t, including the edge computing delay T u,mec (t), upload delay T u,up (t) and download delay T u,do (t); (4) represents the total delay of the end-side device in the second hybrid operation mode at time point t, including the local calculation delay T u,loc (t), edge computing delay T u,mec (t), upload delay T u,up (t) and download delay T u,do (t); Step 2: Energy Consumption Model (1) The computing energy consumption of neural network tasks on the end device can be modeled for: E u,loc (t)=P exe u (t)T u,loc (t), Here, E u,loc (t) represents the computing energy consumption of the end-side device at time point t; P exe u (t) represents the computing power consumption of the end-side device at time point t; T u,loc (t) represents the computation delay of the end-side device at time point t; (2) The energy consumed by uploading the output data of the end-side device (camera) to the edge computing server is: E u,up (t)=P up u (t)min{T u,up (t),τ}, Here, E u,up (t) represents the energy consumed by the end device to upload data to the edge computing server at time point t; P up u (t) represents the power consumption of the end-side device when uploading data at time point t; T u,up (t) represents the transmission delay of the end-side device uploading data to the edge computing server at time point t; τ represents a preset time threshold; (3) The total energy consumption of the end-side equipment is expressed as follows: in, E u,M (t)=E u,loc (t)+E u,up (t), Here, E u (t) represents the total energy consumption of the end-side device, which is divided into four cases: (1) 0 means that in an ideal state, there is no energy consumption on the end-side device; (2) E u,loc (t) represents the local computing energy consumption of the end-side device at time t; (3) E u,up (t) represents the energy consumed by the end device to upload data to the edge computing server at time point t; (4) E u,M (t) represents the total energy consumption of the end-side device in the mixed operation mode at time point t, including the local computing energy consumption E u,loc (t) and the time consumed by uploading data E u,up (t) energy; Step 3: Problem Modeling (1) The total cost of the end-side device consists of delay, energy cost, and penalty, which can be formalized as follows: C u (t)=u·T u (t)+v·E u (t)+w·Q ′ u (t), Here, C u (t) represents the total cost of the end-side equipment (camera); T u (t) represents the total delay of the end-side device at time point t, including local calculation delay, data upload delay and data download delay; v·E u (t) represents the total energy consumption of the end-side device at time point t, including local computing energy consumption and data transmission energy consumption; Q ′ u (t) is a penalty term, which indicates the additional cost when certain performance indicators are not met (such as delay or energy consumption exceeding a preset threshold); u, v, and w are weight coefficients used to adjust the importance of delay, energy consumption, and penalty terms in the total cost; (2) The task processing cost minimization problem can be expressed as: min x mec (t)≥ξ mec , Here, P1 represents the task processing cost minimization problem, which aims to minimize the long-term average total cost; C u (t) represents the total cost of the end device u at time point t, which consists of delay, energy cost and penalty term. The goal is to find the optimal task allocation and resource allocation strategy to minimize this long-term average total cost. The constraints include: (1) This constraint ensures that each end-side device u either executes tasks locally (a u,L (t) = L u ,a u,M (t)=0), or offload the task to the edge computing server (a u,L (t)=0,a u,M (t) = L u ); (2) This constraint ensures that the local computing resources of the end-side device u at time point t at least reach the minimum threshold ξ min u,loc ; (3)ξ mec (t)≥ξ min mec :This constraint ensures that the computing resources of the edge computing server at time point t at least reach the minimum threshold ξ min mec ; (4) This constraint ensures that the upload rate of the end device u at time point t is between 0 and the maximum upload rate R max u,up between; (5) This constraint ensures that the download rate of the end device u at time point t is between 0 and the maximum download rate R max u,do between; The above problem is a mixed integer programming problem with high computational complexity. The time complexity and space complexity of traditional optimization methods are too high. Therefore, the Markov decision process is used for dynamic modeling. Finally, the deep reinforcement learning algorithm is used to optimize the Markov decision process. That is: action: a u (t)=[a u,L (t),a u,M (t)], Here, action a u (t) represents the operation taken by the end device u at time point t, including element a u,L (t) and a u,M (t) vector of elements; a u,L (t) indicates whether the end-side device u chooses to execute the task locally at time point t; if a u,L (t) = L u , the device executes the task locally; if a u,L (t) = 0, the device does not perform local tasks; a(u, M)(t) indicates whether the end-side device u chooses to offload the task to the edge computing server at time point t; if a u,L (t) = L u , the device offloads the task to the edge server; if a u,L (t) = 0, the device does not offload tasks; state: S u (t)=[D u (t),h u (t),Q u (t)]∈S, Here, the state S u (t) describes the situation of the end-side device u at time point t. It is a vector containing multiple elements, each element represents a different attribute of the device; D u (t) represents the amount of data or tasks on the end device u at time point t; this may include the amount of sensor data to be processed, the number of tasks to be executed, etc.; h u (t) represents the channel status of the end device u at time point t, which may include information such as the signal-to-noise ratio, signal strength, and channel capacity of the wireless channel; Q u (t) represents the queue length or buffer status of the end-side device u at time point t, which may include the amount of data waiting to be processed, the number of tasks to be uploaded, etc.; award: R u (t)=C u (t), Here, the reward R u (t) can be defined as the total cost C of the end-side device u at time point t u (t).

6. The method for cloud-pipe-edge coordinated video monitoring of a pumped storage power station according to claim 1, characterized in that: In order to enhance the decision-making ability of deep learning and find the optimal path decision, the intelligent agent algorithm model is introduced: y = γ t +γmaxa'Qθ'(S t+1 ,a'); where θ represents the parameters in the evaluation network; θ' represents the parameters in the target network; S t Indicates the state of the system at time t; S t+1 Indicates the state of the system at time t+1; a t represents the decision action at time t; a' represents the decision action at time t+1; Qθ(S t , a t ) indicates that the evaluation network is in state S t Take action a t , the obtained Q value; Qθ'(S t+1 ,a') indicates that the target network is in state S t+1 The Q value obtained by taking action a'; γ t Indicates that in state S t Take action a t , the reward obtained; y represents the reward decay ratio; the method also includes an algorithm model design step; The algorithm model design steps include: (1) Initialize the evaluation network, target network, and memory bank in deep reinforcement learning; The current system status is S t , t is initialized to 1, and the number of iterations k is initialized to 1; (2) When k is less than or equal to the given number of iterations K, if the modulus of k over m is 0, then update the current state S t is the best state at present; if the modulus of k over m is not 0, a probability p is randomly selected; (3) If p is less than or equal to the greedy strategy probability ξ, then the decision action a output by the evaluation network is selected t , otherwise randomly select an action; (4) After taking decision action a, we get reward γ and next state S, and follow the format (S t 、a t , γ t , S t+1 ) is stored in the memory bank; (5) Combined with the output of the target network, calculate the output of the evaluation network: y = γ t +γmaxa'Qθ'(S t+1 ,a'); (6) Using the minimized error (y-Qθ(S t+1 ,a t )) 2 Evaluate, obtain the evaluation network parameter θ, and update it; (7) Every S steps, assign the parameters of the evaluation network to the target network, and set k = k + 1 to return to step (2) for iteration; (8) When k is greater than the given number of iterations K, the learning process ends and the optimal decision is obtained.

7. A system for executing the pumped storage power station cloud-pipe-edge collaborative video monitoring method according to any one of claims 1 to 6, characterized in that: The system comprises: Cloud, pipe, edge, and end; The cloud includes various business applications, as well as business middle platform, data middle platform, three-dimensional middle platform, and artificial intelligence middle platform; The pipe is a dedicated power grid; The edge is an edge computing node and an edge computing framework, and also includes at least one edge-side application; The terminals include but are not limited to: cameras, robots, and drones.

Images