Multi-agent reinforcement learning-based drainage basin rain and sewage cooperative regulation and control system and method

Abstract

The invention discloses a drainage basin rain and sewage cooperative regulation and control system and method based on multi-agent reinforcement learning. The method comprises the following steps: constructing a distributed agent network topology structure of a drainage basin rain and sewage system; defining an agent observation space, an action space and a multi-target reward function fusing feedforward and feedback information; constructing an intelligent agent cooperative communication and information fusion mechanism based on a low-delay communication network and a graph attention network; adopting a centralized training and distributed execution framework to train a multi-agent reinforcement learning model; and performing real-time coordinated regulation and control on the rain and sewage diversion facilities in the drainage basin based on the trained strategy model, and ensuring that the response time of a key control loop is not more than 200 milliseconds. According to the method, the problems that single-point decision-making in existing rain and sewage diversion control lacks collaboration and is difficult to adapt to dynamic complex working conditions are solved, collaborative optimization regulation and control of a plurality of diversion facilities in a drainage basin range are realized, and the operation efficiency and toughness of a rain and sewage system are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of smart water and advanced control, and particularly relates to a basin rain and sewage collaborative regulation system and method based on multi-agent reinforcement learning. BACKGROUND

[0002] In recent years, with the acceleration of urbanization, combined overflow pollution and urban waterlogging problems have become increasingly prominent, and building an efficient rain and sewage separation system has become the key to urban water environment management. Traditional rain and sewage separation control mainly relies on rule-based strategies based on fixed thresholds, such as performing simple opening and closing operations on the gate of the separation well according to the preset rainfall intensity or time parameters. Although this method is simple and stable, it is essentially a static and open-loop control logic that cannot adapt to uneven rainfall spatial distribution, dynamic fluctuations in water quality concentration, and complex and variable hydraulic states of the pipe network, resulting in insufficient control accuracy and often causing problems such as direct discharge of high-concentration rainwater into receiving water bodies or a large amount of rainwater suddenly flowing into downstream sewage treatment plants, causing load shocks.

[0003] To improve the intelligent level of control, single-point intelligent separation devices have been introduced into online monitoring and machine learning algorithms in the prior art. These solutions use algorithms such as support vector machines (SVM) to classify water quality, or use reinforcement learning algorithms such as Q-learning to develop control strategies for individual separation wells, achieving an upgrade from static rules to dynamic decision-making. However, this "single-point intelligence" solution has inherent defects: the drainage system in a basin is a complex network composed of multiple separation wells, storage tanks, and sewage treatment plants. Independent decision-making at a single point lacks system-level coordination, which can easily lead to conflicts between facilities. For example, the downstream separation well may cause a sharp increase in upstream pipe network pressure, or a large number of separation wells may simultaneously discharge to the end, far exceeding the processing capacity of the sewage treatment plant, forming a "fallacy of composition" and reducing the overall efficiency and safety of the system.

[0004] Therefore, the present application proposes a basin rain and sewage collaborative regulation system and method based on multi-agent reinforcement learning. The present application deeply integrates the core business rule of "direct discharge of heavy rain and treatment of light rain" into the multi-agent collaborative decision-making framework, aiming to fundamentally solve the pain points of isolated single-point decision-making and lack of global coordination in the prior art, and to achieve collaborative optimization and efficient and precise regulation of multiple rain and sewage separation facilities within a basin. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application aims to provide a basin rain and sewage collaborative regulation system and method based on multi-agent reinforcement learning, which solves the technical defects of isolated single-point decision-making, lack of system coordination, and difficulty in adapting to dynamic and complex working conditions in existing rain and sewage separation control.

[0006] The technical solution of the present application is realized in the following manner:

[0007] The present application provides a watershed rain and pollution collaborative regulation method based on multi-agent reinforcement learning, comprising the following steps:

[0008] S1: Constructing a distributed agent network topology structure of the watershed rain and pollution system;

[0009] S2: Defining the individualized observation space, action space and multi-objective reward function of the fusion feedforward and feedback information of each agent;

[0010] S3: Based on low-delay communication network and graph attention network, constructing a collaborative communication and information fusion mechanism between agents;

[0011] S4: Using a centralized training and distributed execution framework to train a multi-agent reinforcement learning model that integrates feedforward prediction and feedback regulation;

[0012] S5: Based on the trained strategy model, performing real-time collaborative regulation of the watershed rain and pollution separation facilities, and ensuring that the response time of the key control loop is not greater than 200 milliseconds.

[0013] Preferably, in step S1, the construction of the distributed agent network topology structure further comprises the following parameterized settings:

[0014] The node abstraction parameters include: the total number of agents N is set to 5-50 according to the size of the watershed; a three-level control architecture is adopted: upstream nodes (L=1), midstream nodes (L=2), and downstream nodes (L=3); the network topology parameters include: the hydraulic connection strength H ij The pipe water conveying capacity is calculated based on the Manning formula; the number of neighbor nodes of a single node N i is limited to no more than 8; the communication distance between nodes is controlled within 2 km; the space-time scale parameters include: the control time step Δt is set to 1-5 minutes adjustable; the topology structure is updated once every 24 hours;

[0015] The weight parameters include: the node importance weight W i is calculated based on the flow proportion and pollution load; the edge weight W ij takes into account the hydraulic connection strength and geographical distance.

[0016] Preferably, in step S2, the parameterized settings of the hydraulic calculation include:

[0017] Hydraulic calculation parameters: Manning roughness coefficient n is selected according to pipe material: 0.013-0.015 for concrete pipe and 0.009-0.011 for plastic pipe; pipe slope S is selected in the range of 0.1%-5% according to actual terrain; hydraulic radius R is calculated based on pipe fullness, and the default fullness is 0.75; split ratio parameter: rainwater and sewage split ratio λ i Dynamic range is 0-1, and the initial value is set to 0.7; total flow Q total According to the design flow, the value range is 50-500 L / s; adjacency matrix parameter: hydraulic dependence strength W ij Normalization processing is performed to ensure that ∑W ij = 1; pipe conveying capacity Q pipe Real-time calculation and update based on the Manning formula; dynamic adjustment parameter:

[0018] The calculation time step is synchronized with the control system, and the default is 2 minutes; when the rainfall intensity exceeds 30 mm / h, the calculation period is automatically shortened to 1 minute.

[0019] Preferably, in step S2, the parameterization of the reward function is set as follows: the conversion coefficient in the pollution reduction reward is selected in the range of 0.8-1.2, the settling decay coefficient is set to 0.05-0.2 h-1, and the time window is set to 2-6 hours; the safety liquid level in the flood control reward is set to 70%-80% of the design liquid level, and the maximum allowed liquid level is 85%-95% of the pipe top elevation; the action frequency weight in the economic reward is selected in the range of 0.1-0.3, and the action amplitude weight is selected in the range of 0.05-0.15; in the weight distribution of each reward, pollution reduction accounts for 0.4-0.6, flood control accounts for 0.3-0.5, economy accounts for 0.1-0.2, and coordination accounts for 0.1-0.2, and the flood control weight is automatically increased under strong rainfall conditions, and the economic weight is dynamically adjusted according to the operation cost.

[0020] Preferably, in step S3, the parameterization of the information fusion mechanism is preferably set as follows: the graph attention network adopts 2-4 attention heads, and the hidden layer dimension is set to 64-256; the hydraulic weight W ij Based on the hydraulic dependence strength in claim 2, and using a leakage coefficient of 0.1-0.3 for LeakyReLU activation; the dimension of the feature transformation matrix W matches the observation space dimension and is kept in the range of 32-128; the time step of information fusion is synchronized with the control system, the communication delay is controlled within 100-500 milliseconds to ensure real-time requirements; the normalization of attention weight adopts softmax with temperature parameter, and the temperature coefficient is selected in the range of 0.8-1.2; the learning rate of network training is set to 0.0001-0.001, and the batch size is 32-128.

[0021] Preferably, in step S4, the parameterized setting of the multi-agent reinforcement learning model is as follows: the centralized critic network adopts 3-5 hidden layers, the number of nodes in each layer is 128-512, and the activation function is selected from ReLU or Tanh; the distributed executor network adopts 2-3 hidden layers, the number of nodes in each layer is 64-256; the discount factor γ is 0.9-0.99, the target network update frequency is 100-500 training steps; the learning rate is set to 0.0001-0.001, and the Adam optimizer is used; the experience replay buffer capacity is 10000-50000 samples, the batch size is 32-128; the initial value of the exploration rate is 0.9-1.0, and it is linearly decayed to 0.05-0.1; the weight coefficient of the flow balance index ΔQ is 0.1-0.3, and the network training period is 1000-5000 rounds.

[0022] The beneficial effects of the present application are:

[0023] Through multi-agent collaborative decision-making, the dynamic regulation of "heavy rain directly into the river and light rain into the plant for treatment" is intelligently realized, which not only fully plays the flood control benefit of direct discharge of heavy rain, but also ensures that high-concentration polluted water bodies are effectively treated, and significantly improves the operation efficiency and environmental safety of the rain and sewage system of the river basin. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 A flow chart of the technical scheme of the present application is shown.

[0025] Figure 2 A technical scheme diagram of the present application is shown. DETAILED DESCRIPTION

[0026] The technical scheme in the embodiments of the present application will be described clearly and completely below, and obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0027] As shown in Figure 2 , the present application is a kind of based on multi-agent reinforcement learning's flow basin rain and sewage collaborative regulation system and method, including the following steps: the present embodiment is aimed at a certain city suburb small and medium-sized river basin (river basin area 60km 2 , total length of pipe network 85km, 12 key rain and sewage separation facilities), downstream link 1 day processing capacity 100,000 m 3The core goal is to achieve a response time of the key control loop of ≤200 ms, an increase of ≥30% in pollution reduction rate, a decrease of ≥70% in the number of times the pipe network liquid level exceeds the limit, and dynamic coordination of "direct discharge during heavy rain and sewage treatment during light rain" through formula calculation.

[0028] S1, Distributed intelligent agent network topology construction

[0029] Upstream nodes (L = 1): v1 (diversion well), v2 (diversion well), v3 (diversion well), v4 (storage tank); Midstream nodes (L = 2): v5 (diversion well), v6 (diversion well), v7 (storage tank), v8 (storage tank); Downstream nodes (L = 3): v9 (pump station), v 10 (pump station), v 11 (wastewater plant inlet gate), v 12 (end interception gate). Directed graph G = (V, E): based on SWMM hydraulic model to identify flow direction, such as v1 (upstream) → v5 (midstream), there is a directed edge e 15 ∈E; Neighbor node set N i : The neighbors of v5 are v1, v2, v6, v7, v 11 (|N i | = 5 ≤ 8).

[0030] Node weight calculation

[0031] Node importance weight W i Based on the proportion of flow Q ratio,i and the proportion of pollution load P ratio,i Weighted calculation:

[0032] W i = 0.6 × Q ratio,i + 0.4 × P ratio,i

[0033] Example: Q 11 of v ratio,i (wastewater plant inlet gate) = 28% (total flow of the watershed Q total,all = 800 L / s, its excess flow is 224 L / s), P ratio,i = 32% (total COD load 120 kg / h, its received load is 38.4 kg / h), substituting:

[0034] W 11 = 0.6 × 0.28 + 0.4 × 0.32 = 0.296

[0035] S2, Define the individualized observation space, action space, and multi-objective reward function of each intelligent agent's fusion feedforward and feedback information

[0036] Pipeline transport capacity Q pipe :

[0037]

[0038] Parameter values ​​(main pipes from v1 to v5): A (cross-sectional area of ​​water flow): diameter D of circular pipe = 1.2m, fullness α = 0.75. n (Manning coefficient): for concrete pipes, n = 0.014; R (hydraulic radius): (P is the wetted perimeter), P = α × πD = 0.75 × π × 1.2 = 2.827 m, therefore S (pipe slope): 0.3% = 0.003; Substitute into the calculation:

[0039]

[0040] Rainwater and sewage separation ratio λ i :

[0041]

[0042] v1 Light rain scene, r t =3mm / h:

[0043] Q sewer,i =140L / s (flowing into the sewage pipe), Q rain,i = 60L / s (flowing into the rainwater pipe), Q total,i =200L / s, substituting, we get:

[0044]

[0045] Suspended sedimentation velocity v s (Stokes correction formula):

[0046]

[0047] Value: g = 9.8 m / s 2 , ρ p =1100kg / m 3 (Density of suspended matter), ρ w =1000kg / m 3 (Water density), d p =50μm=5×10 -5 m, μ=0.001Pa·s (hydrodynamic viscosity at 20℃), C c =1.1

[0048] Substitute the Cunningham correction factor into the calculation:

[0049]

[0050] Observation space o i :

[0051]

[0052] Time step Δt = 1 min, so:

[0053]

[0054] (upstream neighbor flow and): upstream neighbors v1, v2 of v5 have q1 = 200 L / s, q2 = 180 L / s, so:

[0055]

[0056] Reward function formulaic computation

[0057] Define multi-objective reward R i = R pollution + R flood + R cost + R business Each sub-item is computed as follows:

[0058] Pollution reduction reward R pollution

[0059]

[0060] v5, T = 4 h, η = 1.0, k s = 0.1 h -1 :

[0061] |τ(h)|Q sewer,i (τ)(m 3 / h)|C i (τ)(kg / m 3 )|exp(-0.1τ)| term value (kg)

[0062] |0|504(140 L / s x 3.6)|0.15(150 mg / L)|1.0|75.6|

[0063] |1|486|0.14|0.905|61.3|

[0064] |2|468|0.13|0.819|49.8|

[0065] |3|450|0.12|0.741|39.8|

[0066] |4|432|0.11|0.670|31.9|

[0067] - Summing up gives: Rpollution = 75.6 + 61.3 + 49.8 + 39.8 + 31.9 = 258.4

[0068] Flood control reward R flood

[0069]

[0070] v3, l safe = 2.6 m, l max = 3.15 m, l i = 2.8 m:

[0071]

[0072] Economic reward R cost

[0073] R cost = -(ω1·I(Δa i ≠ 0) + ω2·|Δa i |·Δt)

[0074] v2, ω1 = 0.2, ω2 = 0.1, Δt = 1 min, Δa i = λ i,t - λ i,t-1 = 0.7 - 0.68 = 0.02): I(Δa i ≠ 0) = 1, by substitution:

[0075] R cost = -(0.2 x 1 + 0.1 x 0.02 x 1) = -0.202

[0076] Business rule reward R business

[0077] Heavy rain scenario (r t = 35 mm / h): R business = +5.0;

[0078] Light rain scenario (r t = 3 mm / h): R business = +3.0.

[0079] Total reward instance (v5 light rain scenario): R5 = 258.4 - 0.132 - 0.202 + 3.0 = 261.066

[0080] S3, based on low latency communication network and graph attention network, construct the collaborative communication and information fusion mechanism between agents:

[0081]

[0082] Calculate v5 against vAttention coefficient a of 11 5,11 Hydraulic weight w 5,11 Based on claim 3 adjacency matrix formula calculation, w 5,11 = 0.35 (v5 and v 11 Strongest hydraulic connection); Feature transformation: W is a 64-dimensional feature matrix, h5 (v5 feature vector), h 11 (v 11 Feature vector) is transformed by W to Wh5 = [x1,..., x 64 ], Wh 11 = [y1,..., y 64 ], and spliced to [Wh5||Wh 11 ](128-dimensional); LeakyReLU activation: a T [Wh5||Wh 11 ] = z, LeakyReLU(z) = z(z > 0), exp(z) = e z = 120; Denominator calculation: Final a 5,11 :

[0083]

[0084] S4, using a centralized training and distributed execution framework, training a multi-agent reinforcement learning model that integrates feedforward prediction and feedback regulation

[0085] Flow balance index AQ:

[0086]

[0087] N = 12, taking 3 key node data:

[0088]

[0089] - Sum of 12 nodes ∑ = 0.0052, substitute:

[0090]

[0091] Centralized critic loss function:

[0092]

[0093] Gamma = 0.95 (discount factor), theta - (Target network parameters, updated every 300 steps), (12-agent total reward);

[0094] Q total (o, a; theta) = 3200,

[0095] Substituting the timing difference error:

[0096] TD_error = 3200 - (3120 + 0.95 * 3100) = 3200 - (3120 + 2945) = -2865 loss:

[0097] L(θ) = (-2865) 2 = 8,208,225 (loss after batch averaging is about 1.2 * 10 6 )

[0098] Actor policy gradient

[0099]

[0100] v5: (action pair log probability gradient), (Q value gradient with respect to action), substituting:

[0101] (update φ5 along the gradient)

[0102] S5, based on the trained policy model, perform real-time collaborative regulation of rain and sewage shunting facilities within the watershed, and ensure that the response time of the key control loop is not greater than 200 milliseconds,

[0103] Key control loop response time T total = T 采集 + T 计算 + T 执行 :

[0104] T 采集 : sensor data collection (0.5 seconds / time) -> 100ms;

[0105] T 计算 : actuator network inference (lightweight model) -> 40ms;

[0106] T 执行 : shunting facility action response -> 10ms;

[0107] Total response time: T total = 100 + 40 + 10 = 150ms ≤ 200ms (satisfies the requirement) Implementation effects are shown in the following table:

[0108] Table 1. Implementation effect comparison

[0109]

[0110]

[0111] Compared with the traditional rain and pollution control scheme, the present application takes multi-agent collaborative decision-making as the core and realizes significant breakthroughs in four key dimensions: the pollution reduction rate is increased from total reduction to 97%, the pollution direct discharge is greatly reduced to improve the water environment; the pipe network liquid level overrun rate is reduced from the total number of overrun times to 5% of the total number of monitoring times, effectively reducing the risk of waterlogging; the sewage treatment plant load fluctuation is narrowed from the design to 10, ensuring the stability of the treatment process; the energy consumption per unit of water quantity is reduced from the energy consumption per unit of water quantity to 0.6 kWh / m 3 , significantly saving operation and maintenance costs. It completely solves the pain points of traditional control single point isolation, static and extensive, and realizes the multi-objective collaborative improvement of "environment-safety-stability-economy".

[0112] The above-mentioned embodiments only express the implementation of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the protection scope of the present application patent should be subject to the appended claims.

Claims

1. A watershed stormwater and wastewater coordinated regulation system and method based on multi-agent reinforcement learning, characterized in that, Includes the following steps: S1: Construct a distributed intelligent agent network topology for the watershed stormwater and sewage system; S2: Define the personalized observation space, action space, and multi-objective reward function for each agent by fusing feedforward and feedback information; S3: Construct a collaborative communication and information fusion mechanism among agents based on low-latency communication networks and graph attention networks; S4: A centralized training and distributed execution framework is used to train a multi-agent reinforcement learning model that integrates feedforward prediction and feedback regulation. S5: Based on the trained strategy model, perform real-time coordinated control of rainwater and sewage separation facilities within the watershed, and ensure that the response time of key control loops is no more than 200 milliseconds.

2. The watershed stormwater and sewage coordinated regulation system and method based on multi-agent reinforcement learning according to claim 1, characterized in that, In step S1, all key stormwater and sewage separation facilities within the watershed are identified and abstracted into a set of intelligent agent nodes. Where N is the total number of agents; Based on the watershed network topology and hydrological connectivity, a directed graph is constructed among the agents. Where the edge set ε represents the hydraulic connection between nodes, if node v j Located at node v i Downstream, there exists a directed edge e. ij ∈ε; For each agent node v i Construct its neighbor node set This serves as the perception range for its local collaborative decision-making.

3. The watershed stormwater and wastewater coordinated regulation system and method based on multi-agent reinforcement learning according to claim 1, characterized in that, In step S2, when constructing the network topology, hydraulic calculations are required to quantify the hydraulic connections between nodes: for pipe connections, the Manning formula is used to calculate the flow capacity Q under full pipe flow. pipe : Where A is the cross-sectional area of ​​the pipe, n is the Manning roughness coefficient, R is the hydraulic radius, and S is the pipe slope; the calculation node v i Rainwater and sewage separation ratio λ i : Among them, Q sewer,i Q is the flow rate into the sewage pipe. rain,i Q is the flow rate to rainwater pipes or natural water bodies. total,i Total flow; Based on the split ratio and pipeline transport capacity, an adjacency matrix W is constructed to describe the hydraulic dependence strength between nodes, and its elements w ij Defined as:

4. The watershed stormwater and sewage coordinated regulation system and method based on multi-agent reinforcement learning according to claim 1, characterized in that, In step S2, the intelligent agent v i observation space o i Defined as: Where, q i ,c i ,l i They are nodes v i Real-time inflow rate, influent water concentration, and node liquid level; Δl i h is the rate of change of liquid level. i For historical action sequences; r t The real-time rainfall intensity of the region; r t+1:t+H For the rainfall intensity sequence over the next H time steps; v s ∑q represents the settling velocity of suspended solids. j This is the sum of the traffic of the upstream neighboring nodes; Among them, the settling velocity of suspended matter v s Calculated using the modified Stokes' law formula: Where g is the acceleration due to gravity, ρ p ρ is the particle density. w Let d be the density of water. p Let C be the equivalent diameter of the particle, μ be the dynamic viscosity of water, and C be the viscosity of the water. c This is the Cunningham correction factor.

5. The watershed stormwater and wastewater coordinated regulation system and method based on multi-agent reinforcement learning according to claim 3, characterized in that, In step S2, the reward function R i Introducing dynamic expectations based on water quality and sedimentation: Pollution Reduction Incentive R pollution It is related to the total amount of pollutants trapped, and its calculation takes into account the pollutant deposition effect: Where η is the conversion coefficient, C i (τ) represents the pollutant concentration at time τ, k s The settling attenuation coefficient characterizes the rate of pollutant settling loss in the pipeline network; Flood control reward R flood The degree of liquid level exceeding the limit at the node is negatively correlated, and an exponential penalty term is used: Among them, l safe For a safe liquid level, l max Maximum permissible liquid level; Economic reward R cost It is negatively correlated with the gate's operating frequency and energy consumption. in, Let Δa be the indicator function. i Let Δt be the change in motion, ω1 and ω2 be the time step, and ω1 and ω2 be the weights. The actual business rule of "heavy rain flows directly into the river, and light rain discharges more sludge into the sewage treatment plant" is further integrated into the multi-agent reinforcement learning framework. Through business rule reward items, mode switching mechanisms and business feasibility constraints, it is ensured that the agents meet both actual business needs and environmental safety requirements when learning collaborative strategies.

6. The watershed stormwater and sewage coordinated regulation system and method based on multi-agent reinforcement learning according to claim 1, characterized in that, The information fusion mechanism in step S3 has a graph attention coefficient α. ij The calculation incorporates hydraulic weights: Among them, w ij The hydraulic dependence strength defined in claim 2 is used as a prior bias for attention calculation, making the model pay more attention to neighboring nodes with close hydraulic connections.

7. The watershed stormwater and sewage coordinated regulation system and method based on multi-agent reinforcement learning according to claim 5, characterized in that, In step S4, a centralized critic network Q is constructed. total (o,a;θ), whose input is the joint observation of all agents o=(o1,...,o N ) and the combined action a=(a1,...,a N The output is a global state-action value function, used to evaluate the long-term benefits of joint actions; each agent v i Having a distributed actuator network π i (o i ;φ i ), based on local observation o i Output action a i ; During the training phase, the central critic learns from the collected empirical tuples (o,a,R,o') and updates the network parameters θ by minimizing the temporal difference error: Where, θ - Here, γ represents the target network parameters, and γ is the discount factor. For global instant rewards; π is the actuator network for each agent. i Update using the policy gradient provided by the central commentator to maximize the global expected return. The centralized critic network Q total The input includes system-level traffic balancing metrics: Define the instantaneous flow imbalance ΔQ of the system: Among them, Q in,i and Q out,i They are nodes v i Instantaneous inflow and outflow rates, Q cap,i Design its traffic capacity; Inputting ΔQ as a global state feature into the centralized commentator network guides the model to learn a collaborative strategy for flow balancing.

8. A watershed stormwater and sewage coordinated regulation system for performing the method of any one of claims 1-6, characterized in that, include: The distributed intelligent control unit is deployed at each rainwater and sewage separation facility site. It has built-in lightweight actuator network and local sensors and is responsible for calculating the flow rate, settling velocity and separation ratio of the Manning formula in real time. Edge computing gateways are responsible for data exchange, communication management, and degradation mode switching between intelligent agents within their respective regions. The cloud-based training and collaboration platform is responsible for collecting operational data across the entire domain, conducting centralized model training and evaluation, and distributing updated model parameters to each edge node.

Images