A method and system for intelligent layout and interference prediction analysis of strong and weak power lines

Abstract

The application relates to the technical fields of electrical engineering and artificial intelligence, and discloses a strong and weak electric circuit intelligent layout and interference prediction and analysis method and system, wherein the strong and weak electric circuit intelligent layout and interference prediction and analysis method comprises the following steps: constructing a state representation and an interference propagation model of an electronic system layout; performing uncertainty quantification based on Bayesian inference and a particle filtering method; applying an improved multi-objective optimization algorithm to solve a Pareto optimal layout scheme set of the strong and weak electric circuit; realizing robust reinforcement learning strategy optimization based on a distributed deep Q network; and performing decision support through a preference learning model; the application solves the problems of multi-objective balance difficulty, weak anti-interference capability and imperfect decision support in traditional wiring technology, improves the line layout design quality and efficiency, and reduces the electromagnetic interference risk.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of electrical engineering and artificial intelligence, in particular to an intelligent layout and interference prediction analysis method and system for strong and weak power lines. BACKGROUND

[0002] With the increasing complexity of modern electrical systems, strong power lines and weak power lines need to coexist in limited space, and the problem of electromagnetic interference between them is increasingly prominent. However, the existing technology has the following main shortcomings: the traditional single-objective optimization method is difficult to balance multiple conflicting objectives such as interference suppression, space utilization and cost control; most methods do not consider environmental uncertainties such as load changes and material property deviations, resulting in poor performance of the theoretically optimal layout in practical applications; the existing methods are inefficient in calculation, making it difficult to process large-scale layout problems within a reasonable time; there is a lack of effective mechanism to systematically integrate expert experience into the optimization process, and the human-machine collaboration capability is limited.

[0003] Therefore, there is a need for an intelligent layout and interference prediction analysis method for strong and weak power lines that can simultaneously consider multi-objective balance, environmental uncertainty, computational efficiency and expert knowledge integration to improve the quality, efficiency and robustness of line layout. SUMMARY

[0004] The application provides an intelligent layout and interference prediction analysis method and system for strong and weak power lines, which solves the technical problem of being difficult to simultaneously consider multi-objective balance, environmental uncertainty, computational efficiency and expert knowledge integration in related technologies.

[0005] The application provides an intelligent layout and interference prediction analysis method for strong and weak power lines, which includes:

[0006] Constructing a state representation and interference propagation model of the electronic system layout includes: establishing an electromagnetic interference propagation model; forming an interference influence matrix;

[0007] Using the constructed interference propagation model and interference influence matrix, based on Bayesian inference and particle filtering method, the uncertainty is quantified, including: collecting interference data under different working conditions; obtaining the probability distribution of the main uncertainty factors;

[0008] Based on the state representation and interference propagation model and the probability distribution of the uncertainty factors, an improved multi-objective optimization algorithm is applied to solve the Pareto optimal layout scheme set of the strong and weak power lines, including: setting multiple optimization objectives; a non-dominated solution set should be generated;

[0009] Robust reinforcement learning strategy optimization is implemented based on distributed deep Q network using the non-dominated solution set as training data, including: defining the environment state space, action space and reward function; optimizing the layout decision strategy; evaluating and optimizing the trained strategy for robustness.

[0010] In combination with the non-dominated solution set and the optimized layout decision strategy, decision support is performed through a preference learning model, including: constructing a preference learning model; evaluating and ranking all non-dominated solutions; and recommending an optimal layout scheme.

[0011] In a preferred embodiment, the state representation and interference propagation model for constructing the electronic system layout specifically includes:

[0012] Discretize the layout space into a three-dimensional grid structure of G={g i，j，k}, where G represents the grid set of the entire layout space, i, j, k represent the X-axis, Y-axis, and Z-axis coordinates of the space, respectively, and g i，j，k represents a grid unit at a specific location in the three-dimensional space.

[0013] For the potential interference I(e a , e b ) between any two circuit elements e a and e b , a calculation model is established based on electromagnetic field theory:

[0014]

[0015] where I(e a , e b ) represents the electromagnetic interference intensity from element e a to element e b , P a represents the interference source power, S b represents the sensitivity of the disturbed element, d ab represents the distance between the two elements, f(θ ab ) represents the directional factor, g(ω a , ω b ) represents the frequency-dependent function, and α is the correction coefficient, and α is the comprehensive correction coefficient.

[0016] The calculation results are organized into an interference influence matrix M n×n , where n is the total number of circuit elements in the system.

[0017] In a preferred embodiment, the uncertainty quantification is performed based on Bayesian inference and particle filtering, specifically including:

[0018] Install multiple electromagnetic interference monitoring sensors {s1, s2,..., s ms1, s2, s m respectively represent the 1st, 2nd, mth electromagnetic interference monitoring sensors deployed in the system, and m represents the total number of sensors;

[0019] establish a state transition model p(x t |x t-1 ) and a measurement model p(z t |x t ), wherein x t represents the interference state at time t, x t-1 represents the interference state at time t-1, and z t represents the corresponding measurement value;

[0020] estimate the posterior probability of the interference distribution using a particle filtering algorithm:

[0021]

[0022] wherein p(x t |z 1：t ) represents the posterior probability distribution of the interference state x 1：t at time t under the condition of the observation sequence z t , represents a particle set and the corresponding weight, δ is a Dirac delta function, represents the i th particle, and N represents the total number of particles, represents the summation of all N particles;

[0023] obtain the probability distribution of the main uncertainty factor through the sampling and importance resampling steps.

[0024] In a preferred embodiment, the application improves the multi-objective optimization algorithm, specifically including:

[0025] set the optimization objective function set:

[0026] F(x)={f1(x),f2(x),...,f k (x)};

[0027] wherein F(x) represents the objective function set in the multi-objective optimization problem, x represents the decision variable vector of the layout scheme, f1(x) represents the interference suppression quantity, f2(x) represents the space utilization rate, f k (x) represents the cost control, and k represents the total number of optimization objectives;

[0028] apply the improved NSGA-III algorithm for multi-objective optimization, including:

[0029] use the adaptive crossover probability p cand mutation probability p m :

[0030]

[0031] where p c represents the crossover probability in the current iteration, p c0 represents the initial value of the crossover probability, p m represents the mutation probability in the current iteration, p m0 represents the initial value of the mutation probability, G is the current generation number, β1 and β2 represent the amplitude coefficients of the crossover probability and the mutation probability respectively, and γ1 and γ2 represent the decay rates of the crossover probability and the mutation probability respectively.

[0032] A reference point-based selection mechanism is adopted to ensure the diversity of the solution set in the target space.

[0033] A non-dominated solution set P * is generated, which represents the optimal compromise between multiple objectives.

[0034] In a preferred embodiment, the distributed deep Q network-based implementation realizes robust reinforcement learning policy optimization, specifically including:

[0035] Define the state space S, the action space A, and the reward function R:

[0036] The state space includes the current layout state and the uncertainty estimate;

[0037] The action space is defined as the possible line placement positions and directions;

[0038] The reward function takes into account interference suppression, space utilization, and cost control;

[0039] Train the distributed deep Q network:

[0040] Use a double Q learning strategy to reduce Q value estimation bias;

[0041] Apply an experience replay mechanism D to store and convert samples (s t , a t , r t , s t+1 ), where s t represents the state at time t, a t represents the action performed at time t, r t represents the immediate reward obtained after performing action a t , and s t+1 represents the new state transferred after performing action a t ;

[0042] Minimize the loss function:

[0043] L(0) = E (s,a,r,s')～D [(r + max a′ Q(s', a'; 0 - ) - Q(s, a; 0)) 2 ];

[0044] where L(0) denotes the loss function of network training, E (s，a，r，s′)～D denotes the expectation sampled from the experience replay buffer D, r denotes the immediate reward, g denotes the discount factor, max a′ Q(s', a'; 0 - ) denotes the maximum Q value of the next state s' calculated using the target network parameter 0 - , Q(s, a; 0) denotes the Q value of the state-action pair (s, a) calculated using the current network parameter 0, and 0 and 0 - denote the current network and target network parameters, respectively.

[0045] Through Monte Carlo simulation and scenario analysis, the robustness of the strategy in an uncertain environment is evaluated and optimized.

[0046] In a preferred embodiment, the decision support by the preference learning model specifically includes:

[0047] Collecting an expert evaluation dataset D p = {(x i , x j , y ij )}, where D p denotes the expert evaluation dataset, x i and x j denote the i-th and j-th layout schemes, respectively, and y ij denotes the preference degree of the expert for the schemes x i and x j ;

[0048] Constructing an enhanced twin neural network for the preference learning model, and the loss function thereof is defined as:

[0049]

[0050] where L p (0) denotes the loss function of the preference learning model, denotes the summation of all data samples in the expert evaluation dataset D p , f θ denotes the preference learning model, y ij denotes the preference degree of the expert for the schemes x i and x j , and l is a regularization parameter, denotes the regularization term.

[0051] score(x) = f * (x) + a · diversity(x) for all layout schemes in the non-dominated solution set P

[0052] score(x) = f θ (x) + a · diversity(x) for all layout schemes in the non-dominated solution set P

[0053] where score(x) denotes the comprehensive score of layout scheme x, f θ (x) denotes the utility score of layout scheme x by the preference model, diversity(x) is the diversity measure, and a is the balancing parameter.

[0054] The optimal layout scheme is recommended based on the comprehensive score.

[0055] In a preferred embodiment, an adaptive weighting mechanism is further included to consider the importance of different objectives, where the objective weights are dynamically adjusted according to environmental conditions and user demands:

[0056]

[0057] where W i (t) denotes the weight of the i-th objective at time t, b i (t) is a parameter that changes with environmental conditions, denotes the exponential transformation of the importance parameter, denotes the sum of the exponential transformations of all k objectives, and k denotes the total number of optimization objectives.

[0058] In a preferred embodiment, the Bayesian inference and particle filtering method for uncertainty quantification further includes an uncertainty propagation analysis method for evaluating the propagation characteristics of uncertainty in the system:

[0059] An uncertainty propagation graph is established:

[0060] G u = (V, E, W);

[0061] where G u denotes the uncertainty propagation graph, V denotes the system components, E denotes the associations between components, and W denotes the uncertainty propagation strength.

[0062] The Fokker-Planck equation is applied to describe the evolution of uncertainty in the system:

[0063]

[0064] where, denotes the rate of change of the probability density function with respect to time, p(x, t) denotes the probability density of the system state x at time t, A(x) and B(x) denote the drift and diffusion coefficients, respectively, denotes the deterministic drift term, denotes the stochastic diffusion term.

[0065] In a preferred embodiment, a system for implementing a smart layout and interference prediction analysis method for power and communication lines comprises:

[0066] a state representation and interference modeling module for constructing a state representation and interference propagation model of the electronic system layout;

[0067] an uncertainty quantification module for performing uncertainty quantification based on Bayesian inference and particle filtering methods;

[0068] a multi-objective optimization module for applying an improved multi-objective optimization algorithm to solve a Pareto optimal layout solution set for the power and communication lines;

[0069] a reinforcement learning strategy module for implementing a robust reinforcement learning strategy optimization based on a distributed deep Q-network;

[0070] a preference learning and decision support module for providing decision support through a preference learning model;

[0071] a human-machine interaction interface for visualizing the layout solutions and receiving user feedback.

[0072] In a preferred embodiment, a computer-readable medium having a computer program recorded thereon, the computer program being executed by a processor to implement a smart layout and interference prediction analysis method for power and communication lines.

[0073] The present application has the following advantages:

[0074] Multi-objective balancing capability: the present application can simultaneously consider multiple conflicting objectives such as interference suppression, space utilization, and cost control, forming a Pareto front representing different trade-off relationships, and providing decision-makers with a variety of compromise solution options. Experimental verification shows that the layout solution quality is improved by about 35% compared to traditional methods in complex engineering environments.

[0075] Anti-interference robustness: through Bayesian inference and uncertainty quantification techniques, the present application can effectively handle the uncertainty of environmental perception and the randomness of noise interference, predict and analyze various interference conditions, and generate robust layout solutions. Actual measurements show that the reliability of the layout solution is improved by about 50% and the actual environmental adaptability is enhanced by about 40% under complex working conditions.

[0076] Optimization efficiency improvement: By combining deep reinforcement learning with improved genetic algorithm, the invention breaks through the computational bottleneck of traditional optimization methods, significantly improving the efficiency of layout optimization. In large-scale layout problems, the optimization speed is improved by about 80%, and the convergence speed under complex constraint conditions is improved by about 125%.

[0077] Adaptive capacity enhancement: Through partial observable Markov decision process modeling and distributed deep Q network learning, the invention can perceive environmental changes and automatically adjust the layout strategy, effectively dealing with dynamic factors such as load changes and material aging. In practical applications, the system's adaptability to new scenarios is improved by about 40%, and the running stability is improved by about 30%.

[0078] Decision support improvement: Through preference learning and interactive visualization technology, the invention provides intuitive decision support tools for designers, allowing expert knowledge to be effectively integrated into the system and quickly identifying the most suitable layout scheme from a large number of Pareto optimal solutions. Practice has shown that this technology reduces the scheme evaluation time by about 70% and the optimization iteration number by about 40%. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 is a flowchart of the intelligent layout and interference prediction analysis method for strong and weak electric lines of the invention;

[0080] Figure 2 is a detailed flowchart of the state representation and interference propagation model for constructing the layout of the electronic system of the invention;

[0081] Figure 3 is a detailed flowchart of the uncertainty quantification based on Bayesian inference and particle filtering method of the invention;

[0082] Figure 4 is a detailed flowchart of the application of the improved multi-objective optimization algorithm to solve the Pareto optimal layout scheme set of the strong and weak electric lines of the invention;

[0083] Figure 5 is a detailed flowchart of the robust reinforcement learning strategy optimization based on distributed deep Q network of the invention;

[0084] Figure 6 is a detailed flowchart of the decision support through the preference learning model of the invention. DETAILED DESCRIPTION

[0085] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that discussions of these implementations are merely provided to enable those skilled in the art to better understand so as to best use the subject matter described herein, and variations of elements can be made dependant on the needs and preferences of a particular implementation without departing from the scope of the present description. Various processes or components can be omitted, substituted, or added as desired. Additionally, features described in relation to some examples can be combined in other examples.

[0086] A method for intelligent layout and interference prediction of strong and weak power lines is disclosed in at least one embodiment of the present application, as shown in Figures 1 to 6 comprising the following steps:

[0087] Step 1, constructing a state representation and interference propagation model of the electronic system layout, including: establishing an electromagnetic interference propagation model; forming an interference influence matrix;

[0088] comprising the following sub-steps:

[0089] Step 1.1, line system state representation;

[0090] Discretize the line layout space into a three-dimensional grid:

[0091] G grid ={g i，j,k |1≤i≤I max ,1≤j≤J max ,1≤k≤K max};

[0092] where G grid represents the grid set of the entire layout space, i, j, k represent the X-axis, Y-axis, Z-axis coordinates of the space, respectively, g i，j，k represents the grid unit at a specific position in the three-dimensional space, I max , J max , K max represent the maximum number of grids in the X-axis, Y-axis, and Z-axis directions, respectively.

[0093] Each grid point g i，j，k can carry the following information:

[0094] Whether it is occupied: Boolean value O i，j，k ∈{0,1};

[0095] Occupancy type: T i，j，k ∈{empty, strong power line, weak power line, obstacle};

[0096] Line properties: P i，j，k ={current size, frequency, signal type}(as applicable);

[0097] The overall state of the system is represented as:

[0098] S svstem = {G grid , C connect};

[0099] where S system represents the overall state of the system, G grid represents the grid set of the entire layout space, and C connect represents the set of line connection constraints, defined as the pairs of end points that must be connected:

[0100] C connect = {(g start , g end ) l | l = 1, 2,..., L total};

[0101] where C connect represents the set of line connection constraints, g start and g end represent the start and end grid points of the lth line, respectively, and l represents the number of the line, and L total is the total number of lines to be routed.

[0102] Step 1.2, Electromagnetic interference model creation;

[0103] Based on the physical electromagnetic field theory, an electromagnetic interference propagation model is created for each line.

[0104] For the electromagnetic field intensity around any grid point g i，j，k , the calculation is as follows:

[0105] Electric field intensity calculation:

[0106]

[0107] Magnetic field intensity calculation:

[0108]

[0109] where E represents the electric field intensity vector at grid point g i，j，k , L total represents the total number of lines in the system, C l represents the path of the lth line, represents the position vector of grid point g i，j，k , represents the position vector of the microelement dl' on line l, dl' represents a micro vector element on the line, and p(l') represents the charge density at position l' on line l, represents the position vector of grid point gi，j，k The magnetic field strength vector of the line l, I current，l represents the current in the line l, ε0is the vacuum permittivity, μ0is the vacuum permeability, and x represents the vector cross product operation.

[0110] Step 1.3, multi-objective optimization function definition;

[0111] Definition of multi-objective optimization function of line layout:

[0112] F obj (x layout ) = (f1(x layout ), f2(x layout ), f3(x layout ));

[0113] where F obj (x layout ) represents the multi-objective optimization function of the line layout, x layout represents a layout scheme, including the path information of all lines, f1, f2, and f3 represent three different dimensions of optimization sub-objective functions, respectively, which are electromagnetic interference suppression, space utilization, and cost control. The definitions of each sub-objective function are as follows:

[0114] Interference suppression objective:

[0115]

[0116] Space utilization objective:

[0117]

[0118] Cost control objective:

[0119]

[0120] where f1(x layout ) represents the interference suppression objective, f2(x layout ) represents the space utilization objective, and f3(x layout ) represents the cost control objective, represents the summation of all grid points in the entire three-dimensional grid space, α i，j，k is the weight coefficient of the grid point g i，j，k , reflecting the sensitivity of the point to interference; c balance is the electromagnetic field weight balance coefficient; length(l) represents the length of the line l; and cost(l) represents the unit length cost of the line l, represents the summation of all lines in the system, L total represents the total number of lines that need to be routed in the system, and cost(l) represents the unit length cost of the line l.

[0121] Through the above sub-steps, the state representation and interference modeling of the line layout problem are completed, providing a mathematical basis for subsequent optimization and learning algorithms.

[0122] Step 2, using the constructed interference propagation model and interference influence matrix, based on Bayesian inference and particle filtering method for uncertainty quantification, including: collecting interference data under different working conditions; obtaining the probability distribution of main uncertainty factors;

[0123] Including the following sub-steps:

[0124] Step 2.1, identification and modeling of environmental uncertainty factors;

[0125] Identify and model the main uncertainty factors in the line layout environment:

[0126] Measurement uncertainty: sensor measurement noise is modeled as a Gaussian distribution:

[0127]

[0128] where p(z|s real ) represents the conditional probability density of the measurement value z under the condition that the true state is s real , represents a multivariate Gaussian distribution with mean s real and covariance matrix∑ z .

[0129] Environmental parameter uncertainty: parameter uncertainty caused by factors such as material properties and environmental temperature is modeled as a prior probability distribution p(θ env ), where θ env represents the set of environmental parameters, such as θ env ={ε r , μ r , σ conduct ,...}, where ε r , μ r , σ conduct represent the relative permittivity, relative permeability, and conductivity, respectively.

[0130] Load change uncertainty: line load change is modeled as a conditional probability p(I current，t |I current，t-1 , ω), where I current，t represents the load current at time t, I current，t-1 represents the load current value at time t-1, and ω represents external factors affecting load change.

[0131] Step 2.2, Bayesian inference model implementation;

[0132] Based on the above uncertainty model, a Bayesian inference framework is implemented to estimate the posterior distribution of the environment true state:

[0133] State transition model:

[0134] p(s t |s t-1 ,a t-1 )=∫p(s t |s t-1 ,a t-1 ,θ env )p(θ env )dθ env ;

[0135] Observation model:

[0136] p(o t |s t )=∫p(o t |s t ,θ env )p(θ eny )dθ env ;

[0137] Posterior distribution calculation: apply Bayes' rule to calculate the state posterior distribution:

[0138] p(s t |o 1：t )∝p(o t |s t )∫p(s t |s t-1 ,a t-1 )p(s t-1 |o 1:t-1 )ds t-1 ;

[0139] where p(s t |s t-1 ,a t-1 ) represents the conditional probability of the current state, s t represents the system state at time t, s t-1 represents the system state at time t-1, a t-1 represents the layout adjustment action performed at time t-1, θ env represents the set of environment parameters, d represents the differential calculation, p(o t |s t ) represents the conditional probability of specific observation data, o t represents the observation data at time t, p(s t |o 1：t ) represents the state posterior probability distribution at time t under the condition of observing all observation data from time 1 to time t, and p(st-1 |o 1：t-1 ) represents the state posterior distribution at the previous time, and a represents the proportionality to.

[0140] Since the above integral is usually difficult to solve analytically in high-dimensional state space, a particle filter algorithm is used for approximate calculation:

[0141] M samples particles are sampled from the proposal distribution ;

[0142] The particle weight is calculated:

[0143]

[0144] wherein, represents the weight of the i-th particle at time t, represents the weight of the particle at the previous time, represents the probability of observing o t under the condition that the state is ; represents the proposal distribution;

[0145] The normalized weight is:

[0146]

[0147] wherein, represents the normalized particle weight, represents the weight of the i-th particle at time t, represents the sum of all particle weights as a normalization factor;

[0148] Resampling is performed according to the effective particle number, and a particle set representing the posterior distribution is obtained

[0149] Step 2.3, Monte Carlo simulation scenario generation and robustness evaluation;

[0150] Based on the implemented Bayesian model, multiple groups of simulated scenarios are generated and the robustness of the layout scheme is evaluated:

[0151] M samples environmental states are sampled from the posterior distribution p(s t |o 1：t );

[0152] For each sampled state, the objective function value of the given layout scheme x layout is calculated

[0153] The robustness evaluation index is constructed:

[0154] Expected performance:

[0155] Performance variance:

[0156]

[0157] Worst-case performance:

[0158]

[0159] Robustness index:

[0160]

[0161] where, represents the average performance of the layout scheme under all sampled environments, represents the weight factor of the average calculation, represents the sum of the performance of all sampled states, M samples represents the number of samples, represents the environment state obtained by the jth sampling, x layout represents the circuit layout scheme to be evaluated, represents the performance score of the layout scheme x under the jth sampled environment state layout , represents the fluctuation degree of the performance of the layout scheme under different environments, F worst (x layout ) represents the worst performance of the layout scheme in all sampled environments, represents the minimum value in all sampled states, R robust (x layout ) represents the comprehensive robustness score of the layout scheme, λ1, λ2, λ3 respectively represent the weight coefficients of the average performance, performance variance and worst-case performance.

[0162] Through the above sub-steps, the quantification of environmental uncertainty and the implementation of Bayesian inference model are completed, which provides a theoretical basis for robust optimization.

[0163] Step 3, based on the state representation and disturbance propagation model and the probability distribution of uncertainty factors, an improved multi-objective optimization algorithm is applied to solve the Pareto optimal layout scheme set of strong and weak power lines, including: setting multiple optimization objectives; a non-dominated solution set should be generated;

[0164] Including the following sub-steps:

[0165] Step 3.1, layout encoding and initial population generation;

[0166] Encode the layout scheme of the line into a chromosome structure suitable for genetic algorithm processing:

[0167] Path encoding: for each line l, its path is composed of a series of consecutive grid points, represented by relative direction encoding, where each step of movement is represented by a vector (dx, dy, dz), dx, dy, dz ∈ {-1, 0, 1}, dx, dy, dz represent the movement in X-axis, Y-axis, Z-axis direction respectively.

[0168] In actual implementation, the upper limit of the encoding length of each line is set to 200, and the line exceeding 200 grid points is encoded in segments;

[0169] Complete layout scheme x layout is represented as a set of all line paths: where x layout represents a complete line layout scheme, represents the complete path of the 1st, 2nd, L total th line respectively, L total represents the total number of lines in the system; for a typical line layout problem, L total is generally 10 to 50 lines;

[0170] Initial population generation: generate N pop initial layout schemes by combining heuristic rules and random disturbance. The heuristic rules include:

[0171] Shortest path priority: use A* algorithm to generate the approximate shortest path of each line, heuristic function:

[0172] h heuristic (n) = α heuristic ·Manhattan(n, goal) + β heuristic ·EMI(n);

[0173] where h heuristic (n) represents the estimated cost from node n to the target node, α heuristic represents the weight coefficient of the Manhattan distance term, Manhattan(n, goal) represents the Manhattan distance from node n to the target node, β heuristic represents the weight coefficient of the electromagnetic interference term, EMI(n) is the electromagnetic interference intensity at node n;

[0174] Obstacle avoidance: ensure that the path avoids the preset obstacles, which is achieved by setting the cost of the obstacle nodes to infinity in the A* algorithm;

[0175] Strong and weak current separation principle: the strong and weak current lines in the initial layout maintain a preset minimum distance, which is set to 3 grid units in the specific implementation;

[0176] Random perturbation: add some random variation to the above generated path by randomly selecting 20% of the points on the path and replacing them with the adjacent feasible points;

[0177] In practical applications, the population size N pop is set to 100.

[0178] Step 3.2, non-dominated sorting and crowding distance calculation;

[0179] Multi-objective evaluation and non-dominated sorting of layout schemes in the population:

[0180] For each layout scheme x i , calculate its function value on each objective:

[0181] F(x i ) = (f1(x i ), f2(x i ),..., f k (x i ));

[0182] Where F(x i ) represents the multi-objective function value vector of layout scheme x i , f1(x i ), f2(x i ), f k (x i ) respectively represent the function value of layout scheme x i on the 1st, 2nd, kth objective, and k represents the total number of optimization objectives;

[0183] In the layout problem of strong and weak power lines, k = 3, corresponding to the interference suppression, space utilization and cost control objectives respectively;

[0184] Based on the Pareto dominance relationship, the population is divided into different levels of non-dominated frontiers {F1, F2,...};

[0185] If for all j ∈ {1, 2,..., k}, f j (x m ) ≥ f j (x n ), and at least one j makes f j (x m ) > f j (x n ), then scheme x m dominates scheme x n , denoted as x m > x n ;

[0186] The first front F1 consists of solutions in the population that are not dominated by any other solution;

[0187] For r > 1, the front F r After removing all solutions in the fronts F1,..., F r-1 , the remaining solutions consist of solutions that are not dominated by any other solution;

[0188] The sorting implementation has a complexity of O(MN 2 ), where M is the number of objectives and N is the population size;

[0189] Crowding distance calculation based on reference points:

[0190] The objective space is regularized and a uniform distribution of reference points set W = {w1, w2,..., w H} is generated, where W represents the reference points set, w1, w2, W H represent the 1st, 2nd, Hth reference point respectively, and H represents the total number of reference points;

[0191] For the case of 3 objectives, 91 reference points are generated using a two-level grid method;

[0192] For each solution x i , the perpendicular distance to each reference point is calculated and assigned to the nearest reference point;

[0193] The local crowding distance of each reference point is calculated as the number of solutions in that region;

[0194] The KD-tree data structure is used in the implementation to accelerate the search for the nearest reference point, reducing the time complexity from O(NH) to O(NlogH), where N is the number of solutions and H is the number of reference points.

[0195] Step 3.3, selection, crossover, and mutation operations;

[0196] Based on non-dominated sorting and crowding distance, genetic operations are performed to generate a new population:

[0197] Selection operation: Tournament selection is used, randomly selecting T individuals from the population each time, and selecting the best one based on the following rules:

[0198] If the individuals belong to different non-dominated fronts, select the individual with a lower (better) front level;

[0199] If the individuals belong to the same front, select the individual with a smaller crowding distance in the reference point;

[0200] Crossover operation: To ensure that the path after crossover still connects the starting point and the ending point, a piece-based crossover operation is used:

[0201] For two parent layout solutions xa and x b , randomly select a crossing point for each line;

[0202] exchange the path segments before and after the selected crossing point, and add connection segments as needed to ensure path continuity;

[0203] check whether the path after the crossing meets the constraint conditions, and adjust it through a local repair algorithm if it does not meet the constraint conditions;

[0204] Mutation operation: according to the characteristics of the line layout, the following mutation operations are used:

[0205] Path node movement: randomly select a node on the path and move it to an adjacent position under the premise of meeting the constraints;

[0206] Local path re-routing: randomly select a segment on the path and use the A* algorithm to recalculate the route of the segment;

[0207] Path smoothing: reduce unnecessary inflection points to make the path smoother;

[0208] Step 3.4, Pareto front surface generation and update;

[0209] Iteratively perform the above genetic operations to constantly update the Pareto front surface:

[0210] Merge the parent population and the child population to form a hybrid population with a size of 2N pop ;

[0211] Non-dominated sorting is performed on the hybrid population to obtain the front set {F1, F2,...};

[0212] Add the fronts F1, F2,... to the new population in turn until the size of the new population is close to but not more than N pop ;

[0213] For the last front F1 that cannot be completely included, based on the reference point association and congestion degree, select part of the solution to be included in the new population, so that the size of the new population is exactly N pop ;

[0214] Repeat the above steps until the termination condition is met (such as reaching the maximum number of iterations, or the Pareto front surface is stable and unchanged);

[0215] Through the above sub-steps, the execution of the multi-objective Pareto optimization algorithm is completed, and a set of non-dominated layout schemes representing different trade-off relationships is obtained.

[0216] Step 4, using the non-dominated solution set as training data, implementing robust reinforcement learning policy optimization based on distributed deep Q-network, including: defining the environment state space, action space and reward function; optimizing the layout decision-making strategy; evaluating the robustness of the trained strategy and optimizing the adjustment;

[0217] Including the following sub-steps:

[0218] Step 4.1, modeling a partially observable Markov decision process;

[0219] Formalize the circuit layout problem as a partially observable Markov decision process (POMDP):

[0220] State space S: complete environment state, including circuit layout, electromagnetic field distribution, material properties, etc., which cannot be directly and completely observed;

[0221] Observation space O: observable data such as electromagnetic field intensity and temperature obtained through sensors;

[0222] Action space A: layout adjustment actions, including adding new lines, moving lines, adjusting shielding measures, etc.;

[0223] Transition function T(s'|s,a): the probability of state transition from s to s' after performing action a;

[0224] Observation function Z(o|s',a): the probability of observing o after performing action a in state s';

[0225] Reward function R(s,a): reward obtained by performing action a in state s, considering multiple objectives:

[0226] R(s,a)=w1·R interference (s,a)+w2·R space (s,a)+w3·R cost (s,a);

[0227] Where R(s,a) represents the overall reward value obtained after performing action a in state s, R interference (s,a) represents the interference suppression reward component, R space (s,a) represents the space utilization reward component, R cost (s,a) represents the cost control reward component, and w1, w2, w3 represent the weight coefficients of the interference suppression, space utilization and cost control reward components, respectively.

[0228] Discount factor γ∈[0,1]: controls the importance of future rewards;

[0229] The policy optimization goal is to find the optimal policy π * that maximizes the expected cumulative discounted reward:

[0230]

[0231] where π * denotes the optimal policy, π(b t ) denotes the action selected by policy π in belief state b t , and E denotes the mathematical expectation, denotes the cumulative discounted reward obtained by performing policy π from the initial state, γ t denotes the discount factor of the reward at the t-th step, b t is the belief state at time t, i.e., the probability distribution of the real environment state, i.e., b t = p(s t |o 1∶t , a 0∶t-1 ), where p(s t |o 1∶t , a 0∶t-1 ) represents the probability estimate of the current real state s 1：t under the condition of observing the historical observation sequence o 0：t-1 and performing the historical action sequence a t .

[0232] Step 4.2, distributed deep Q network implementation and training;

[0233] Based on the partially observable characteristics, a distributed deep Q network (D-DQN) structure is implemented:

[0234] Belief state encoder: encode the observation history o 1：t and the action history a 0：t-1 into a belief state representation b t ;

[0235] Adopt a recurrent neural network structure: h t = f RNN (h t-1 , [o t , a t-1 ]), where h t denotes the hidden vector at the current time t, h t-1 denotes the hidden state vector at the previous time t-1, o t denotes the observation data at the current time t, a t-1 denotes the action performed at the previous time t-1, and f RNN denotes the state update function of the recurrent neural network;

[0236] In specific implementation, a double-layer LSTM network is used, each layer containing 256 hidden units, and the input is observation data o t(contains the current electromagnetic field intensity, temperature and other environmental indicators, vector, dimension 16) and the previous time action a t-1 Concatenation of (one-hot encoding representation, dimension of action space size), the output of the LSTM passes through a 256x128 fully connected layer to generate the belief state representation b t ;

[0237] Output belief state representation: b t = f b (h t ), where b t represents the belief state representation vector at time t, f b represents the mapping function from the LSTM hidden state to the belief state, h t represents the hidden state of the LSTM network;

[0238] Distributed Q function: not only predicts the expected Q value, but also predicts its distribution to represent uncertainty;

[0239] Model the Q value as a set of different quantiles: Where, The first, second, N quantiles of the Q value distribution under the belief state b t are executed a, and N represents the total number of quantiles used;

[0240] In specific implementation, N = 51 quantiles are selected, uniformly distributed in the interval [0, 1];

[0241] The network structure adopts a three-layer fully connected network with layer sizes of [128, 256, 51x|A|], where |A| is the action space size, and the last layer outputs the value of each action at each quantile;

[0242] Where τ i ∈ [0, 1] is the preset quantile level, and N is the number of quantiles;

[0243] Double network architecture:

[0244] Main network Q(b, a; θ): for action selection;

[0245] Target network Q(b, a; θ - ): for target value calculation, parameter θ - is copied from the main network regularly;

[0246] In actual implementation, after every 10000 steps of interaction, the main network parameters are completely copied to the target network;

[0247] Distributed Q network training:

[0248] Sample a batch (o t , a t , r t , o t+1 ) from the experience replay buffer with a batch size of 64, where o t denotes the observation at time t, a t denotes the action performed at time t, r t denotes the reward obtained after performing the action, o t+1 denotes the observation at the next time step;

[0249] Compute the belief state: b t = f b (f RNN (h t-1 , [o t , a t ])), where b t denotes the belief state, f RNN denotes the state update function of the recurrent neural network, and f b denotes the mapping function from the RNN hidden state to the belief state;

[0250] For each quantile τ i , compute the target value:

[0251]

[0252] where Q denotes the target Q-value for the i-th quantile, r t denotes the immediate reward, γ denotes the discount factor, and argmax a denotes the action with the highest value under the main network evaluation, and θ denotes the parameter set of the main network, and θ - denotes the parameter set of the target network;

[0253] Update the network using the quantile regression loss:

[0254]

[0255] where L(θ) denotes the loss function, Q denotes the number of quantiles, B denotes the batch size, denotes the quantile regression loss function, denotes the i-th quantile target value for the j-th sample in the batch, denotes the i-th quantile Q-value estimate of the main network for the j-th sample;

[0256] Use the Adam optimizer with a learning rate of 0.0005, and train the process for 1 million steps of interaction in the routing task.

[0257] Step 4.3, progressive policy exploration and experience replay optimization;

[0258] Implement progressive policy exploration mechanism and optimize experience replay policy:

[0259] Progressive policy exploration: adopt parameterized noise policy, which can provide more structured exploration than simple ∈-greedy policy;

[0260] Add noise to network parameters: θ noisy = θ + σ ⊙ ∈, where θ noisy represents the original network parameters, σ is a learnable noise scale parameter, ⊙ represents an element-level multiplication operator, and ∈ is a random noise sampled from a standard normal distribution; the noise scale gradually decreases with the training process, realizing the smooth transition from exploration to utilization

[0261] Priority experience replay: set priority according to the sample's temporal difference error, improve learning efficiency;

[0262] Sample priority calculation: p i = |δ i | + ∈ small ; where p i represents the priority of the i-th sample, |δ i | represents the absolute value of the temporal difference error of the i-th sample, ∈ small represents a small positive number;

[0263] Sampling probability calculation: where P(i) represents the probability of the i-th sample being sampled, represents the i-th sample priority raised to the power of \, represents the sum of all sample priority raised to the power of \, and α controls the degree of priority usage;

[0264] Importance weight calculation: where w i represents the importance weight of the i-th sample, N is the size of the replay buffer, P(i) represents the probability of the i-th sample being sampled, and β controls the degree of importance sampling;

[0265] Multiple time scale experience replay: maintain replay buffers of different time scales to capture short-term and long-term dependencies;

[0266] Short-term buffer: store recent interaction experience, relatively small capacity, high update frequency;

[0267] Long-term buffer: store historical representative experience, larger capacity, low update frequency;

[0268] Interactive training samples from two buffers in proportion;

[0269] Step 4.4, policy evaluation and adaptive deployment;

[0270] Evaluate the performance of the learning policy under different environmental conditions and implement an adaptive deployment mechanism:

[0271] Robustness evaluation: evaluate policy performance under multiple simulated environmental conditions;

[0272] For the set of environmental state samples generated in Step 2 Evaluate the performance of the policy π on each sample;

[0273] Calculate robustness indicators: expected return E[R(π)], performance variance Var[R(π)], and worst-case performance

[0274] Policy adaptive adjustment: dynamically adjust policy parameters according to environmental changes;

[0275] Maintain a set of policies {π1, π2,..., π K}, each policy is suitable for different environmental conditions, where π1, π2,..., π K represent the 1st, 2nd, Kth policy respectively, K represents the total number of policies in the policy library;

[0276] Based on the current observation o t and historical information, estimate the current environmental conditions and select the most suitable policy;

[0277] Real-time monitoring of policy performance, triggering policy switching when environmental conditions change beyond a threshold;

[0278] Through the above sub-steps, the robust reinforcement learning policy optimization is completed, making the layout scheme adapt to the uncertainty and changes of the environment.

[0279] Step 5, combine the non-dominated solution set and the optimized layout decision strategy, and make decisions through a preference learning model, including: building a preference learning model; evaluate and sort all non-dominated solutions; recommend the optimal layout scheme

[0280] Including the following sub-steps:

[0281] Step 5.1, expert preference data collection and representation;

[0282] Collect and represent expert preference data for subsequent preference learning:

[0283] Preference query system implementation:

[0284] Select a pair of layout schemes (x i , x j ) from the Pareto frontier and present them to the expert;

[0285] Record expert preference feedback y i，j ∈{1, 0, -1}, representing preferences x respectively. i No obvious preference, preference x j ;

[0286] By employing proactive learning strategies, query pairs with the most information are selected, reducing the number of expert feedback sessions required.

[0287] Preference data representation:

[0288] Create a preference dataset D = {(x i x j y i，j )}, where D represents the preference dataset, (x i x j ) is a pair of layout schemes, y i，j These are the expert's preference tags;

[0289] Each layout scheme X is represented by its performance vector F(x) = (f1(x), f2(x), ..., f...) on each objective. k (x)), where F(x) serves as the input to the utility function, used to learn the expert's implicit preferences, f1(x), f2(x), f k (x) represents the performance score of layout scheme x on the 1st, 2nd and kth different objectives, respectively, where k represents the total number of optimization objectives.

[0290] Step 5.2, learning the implicit utility function;

[0291] Based on collected preference data, learn the utility function implicit in expert decision-making:

[0292] Utility function parameterization:

[0293] The expert's implicit utility function is parameterized as follows:

[0294] u(x; ω)=g ω (F(x));

[0295] Where u(x; ω) represents the parameterized utility function, g ω It is a neural network model controlled by parameter ω. The input is the performance vector of the scheme on each objective, and F(x) is used as the input of the utility function to learn the implicit preferences of experts.

[0296] In the specific implementation, g ω A four-layer fully connected neural network structure is adopted, with the following specific configuration:

[0297] Input layer: The dimension is the number of targets (3 in this embodiment, corresponding to interference suppression, space utilization and cost control);

[0298] First hidden layer: 32 neurons, ReLU activation function;

[0299] Second hidden layer: 16 neurons, ReLU activation function;

[0300] Third hidden layer: 8 neurons, ReLU activation function;

[0301] Output layer: 1 neuron, linear activation function;

[0302] To improve the expression ability of the model, a batch normalization layer and a Dropout layer (dropout rate of 0.2) are added after each hidden layer;

[0303] Bayesian preference learning:

[0304] Define preference probability model: when experts compare schemes x i and x j , the probability of preferring x i is:

[0305] P(y i，j =1|ω) = σ(u(x i ;ω)-u(x j ;ω));

[0306] where P(y i，j =1|ω) represents the probability that the expert prefers scheme x i to x j given the parameters ω, u(x i ;ω) and u(x j ;ω) represent the utility values of schemes x i and x j respectively, and σ is the sigmoid function;

[0307] Based on the preference dataset D, the posterior distribution of the utility function parameters is calculated:

[0308]

[0309] where p(ω) represents the posterior probability distribution of the parameters ω given the preference dataset D, represents the product of the likelihood functions of all preference data, and p(ω) represents the prior distribution of the parameters ω, which is a multivariate Gaussian distribution in implementation where λ=0.01 is the regularization parameter, and I represents the identity matrix;

[0310] Variational inference is applied to approximate the posterior distribution:

[0311] The parameterized distribution q φ (ω) is used to approximate the true posterior distribution p(ω|D);

[0312] In implementations, q φ (ω) adopts a diagonal Gaussian distribution where φ = {μ φ , σ φ} are variational parameters, μ φ represents the mean vector of the variational distribution, and σ φ represents the standard deviation vector of the variational distribution.

[0313] Minimizing the KL divergence KL(q φ (ω) || p(ω|D));

[0314] is equivalent to maximizing the evidence lower bound (ELBO):

[0315]

[0316] where ELBO(φ) represents the evidence lower bound, represents the expected value of the log-likelihood under the variational distribution, and KL(q φ (ω) || p(ω)) represents the KL divergence between the variational distribution and the prior distribution.

[0317] The unbiased estimation of the gradient is achieved through the reparameterization trick, and the Adam optimizer is used for optimization of the variational parameters, with a learning rate of 0.001.

[0318] In the layout scene, usually 50 to 100 pairs of expert preference data are collected for training, and convergence is achieved after 10,000 iterations.

[0319] Step 5.3, Pareto scheme evaluation and ranking;

[0320] Using the learned utility function to evaluate and rank the layout schemes on the Pareto frontier:

[0321] Scheme utility evaluation:

[0322] For each scheme x on the Pareto frontier, calculate its expected utility:

[0323]

[0324] where, represents the expected utility of scheme x, represents the expected value of the utility function u(x; ω) under the variational distribution q φ (ω), and ω represents the parameters of the utility function, q φ represents the approximation of the parameter posterior distribution obtained through variational inference.

[0325] Calculate the utility uncertainty represents the confidence in this evaluation, where σu (x) represents the standard deviation of the utility evaluation of scheme x, represents the variance of the utility function value under the variational distribution;

[0326] Scheme ranking and selection:

[0327] Ranking based on upper confidence bound criterion: where UCB(x) represents the upper confidence bound value of scheme x, represents the expected utility of scheme x, σ u (x) represents the uncertainty of the utility of the scheme, β is the exploration-exploitation balance parameter; recommend the K schemes with the highest utility upper confidence bounds to the decision maker.

[0328] Step 5.4, interactive visual decision interface implementation;

[0329] Implement an interactive visualization interface to support the decision maker to explore and select the final layout scheme:

[0330] Multi-dimensional target space visualization:

[0331] Use parallel coordinate charts to display the Pareto front in the multi-objective space;

[0332] Use dimension reduction techniques (such as t-SNE, UMAP) to map high-dimensional target space to 2D / 3D space for visualization;

[0333] Scheme comparison and screening tool:

[0334] Provide a detailed comparison view of the scheme to show the specific performance of different schemes on each target;

[0335] Support dynamic setting of target weights and real-time updating of scheme ranking;

[0336] Provide a scheme filtering function that allows setting threshold conditions for each target;

[0337] Interactive exploration of layout schemes:

[0338] 3D visualization interface displays the specific line layout of the selected layout scheme;

[0339] Support interactive operations such as zooming, rotating, and adjusting transparency;

[0340] Superimposed display of electromagnetic field strength heat map to intuitively display interference distribution;

[0341] Decision feedback and learning update system:

[0342] Record the decision maker's interactive behavior and final choice;

[0343] Include new decision data into the preference dataset and continuously update the preference model;

[0344] The system accumulates experience during use and improves the quality of recommendations;

[0345] Through the above sub-steps, the preference learning and decision support system is implemented, helping decision-makers to efficiently select the layout scheme that best meets the actual needs from the numerous Pareto optimal solutions.

[0346] The intelligent layout and interference prediction analysis method for strong and weak electric lines provided by the embodiment greatly improves the quality and efficiency of line layout design, reduces the risk of electromagnetic interference, enhances the adaptability of the system to environmental changes, and has significant technical value and application prospects.

[0347] Real application examples of the embodiment:

[0348] This example is applied to the electrical control cabinet design scenario of a large-scale automated production line. A large number of strong and weak electric lines need to be arranged in the control cabinet, including three-phase 380V power lines, single-phase 220V control power lines, 24V DC control lines, and various signal transmission lines (analog 4-20mA signal lines, digital switch signal lines, high-speed communication buses, etc.). The control cabinet size is 2000mm×800mm×600mm, the internal space is compact, and there are multiple structural obstacles (such as fixed supports, cooling systems, etc.).

[0349] The specific line requirements are shown in Table 1:

[0350] Table 1: Line requirement list in the control cabinet;

[0351]

[0352] The line layout faces the following challenges:

[0353] Limited space: the available space in the cabinet is limited and needs to be used efficiently;

[0354] Complex interference: strong electric lines will generate strong magnetic fields, which will interfere with weak electric signal lines, especially analog signal lines and communication buses;

[0355] Load changes: the load of each device on the production line will change during operation, causing dynamic changes in electromagnetic interference characteristics;

[0356] Environmental uncertainty: the temperature distribution in the cabinet is uneven, and the electromagnetic shielding effect at different positions also differs;

[0357] Cost control: reduce the amount of cable and shielding material used while ensuring performance.

[0358] Implementation process example:

[0359] For the above application scenarios, we implemented the intelligent layout and interference prediction analysis of strong and weak power lines according to the steps of the example:

[0360] Step 1 instance: state representation and interference modeling;

[0361] Discretize the control cabinet space into a 100x40x30 three-dimensional grid, with each cell size of 20mmx20mmx20mm. Based on electromagnetic field theory, an interference propagation model between lines is established. Table 2 shows the electromagnetic interference reference values of some key positions:

[0362] Table 2: Electromagnetic interference reference values of key positions in the control cabinet;

[0363] Position description Coordinate (x, y, z) Electric field intensity (V / m) Magnetic field intensity (μT) Interference sensitivity coefficient α PLC analog input area (30,25,15) 0.5 0.8 10.0 Communication bus junction area (45,15,20) 0.4 0.6 15.0 Strong current line concentration area (20,10,10) 25.0 35.0 0.1 Power transformer periphery (10,5,5) 30.0 50.0 0.05 Equipment cooling area (60,30,25) 1.2 1.5 0.5

[0364] Step 2 instance: uncertainty quantification and Bayesian inference;

[0365] By installing 12 interference monitoring sensors in the control cabinet, interference data under different load conditions are collected. The particle filter algorithm is applied to estimate the posterior probability of the interference distribution based on historical observation data. Table 3 shows the main uncertainty factors identified by the system and their modeling parameters:

[0366] Table 3: Environmental uncertainty modeling parameters;

[0367]

[0368] Based on the above uncertainty model, 1000 groups of simulated scenarios are generated to evaluate the performance of each layout scheme under different conditions.

[0369] Step 3 instance: multi-objective Pareto optimization algorithm execution;

[0370] For the control cabinet layout problem, an improved NSGA-III algorithm is implemented to simultaneously optimize interference suppression, space utilization, and cost control. The initial population size is set to 100, and through 1000 generations of iteration optimization, 28 non-dominated solutions are finally obtained on the Pareto frontier. Table 4 shows the performance of 5 typical Pareto optimal solutions:

[0371] Table 4: Performance comparison of some Pareto optimal layout schemes;

[0372]

[0373] Step 4 instance: robust reinforcement learning strategy optimization;

[0374] A layout strategy optimization system based on distributed deep Q network (D-DQN) is implemented, and a robust strategy is obtained by interacting with the layout environment for 1 million times. In the control cabinet layout, the input observation dimension is set to 16, including the current position electromagnetic field strength, nearby obstacle distribution, and wired circuit conditions. The action space is defined as 26 discrete actions, including 6 direction movements, various line placements, and special handling actions. Table 5 shows the performance of the strategy under different load conditions:

[0375] Table 5: Performance of reinforcement learning strategy under different conditions;

[0376] Working condition description Average reward value Interference control effect Path average length Adaptive adjustment times Light load (30%) 845.6 Excellent 325.7 3 Medium load (60%) 782.3 Good 331.4 5 Heavy load (90%) 720.8 Medium 338.2 8 Fluctuating load (30%-90%) 755.4 Good 335.6 12 Sudden interference working condition 698.2 Medium 342.1 15

[0377] Step 5 example: preference learning and decision support;

[0378] 85 pairs of layout scheme preference data were collected from 10 experienced electrical design engineers to build a preference learning model. After model training, 28 non-dominated solutions on the Pareto frontier were evaluated and ranked. Table 6 shows the evaluation results of some schemes by the model:

[0379] Table 6: Scheme evaluation results based on preference learning;

[0380]

[0381] Based on the above evaluation results, the final selected scheme P2 is the line layout scheme for the control cabinet, which achieves the best balance between interference suppression and cost control, and best meets the implicit preferences of design experts.

[0382] Technical effect verification:

[0383] The layout scheme P2 of the present embodiment is compared with the traditional manual layout scheme and the single-objective optimization scheme for testing, focusing on verifying the two core technical effects of "multi-objective balance ability" and "anti-interference robustness". The test is conducted in an actual control cabinet, using professional electromagnetic interference test equipment and industrial control system state monitoring system to collect data.

[0384] Multi-objective balance ability verification:

[0385] The present method is compared with the traditional single-objective optimization method. The traditional method optimizes for single objective of minimizing interference, shortest path, and minimum cost, respectively. Table 7 shows the performance comparison of each layout scheme on three objectives:

[0386] Table 7: Multi-objective performance comparison of different layout schemes;

[0387]

[0388] From Table 7, it can be seen that the method can achieve better overall balance on multiple targets compared to single-target optimization and traditional manual layout. Although it is not as good as the single-target method optimized for a single dimension, it has the best overall performance, with an overall score improvement of about 35%, while the implementation difficulty is moderate.

[0389] Anti-interference robustness verification:

[0390] To verify the robustness of different layout schemes in the face of environmental uncertainty and interference changes, the performance of each scheme under 5 different load conditions was tested, and key signal quality indicators were recorded. Table 8 shows the error rate of high sensitivity signals under different working conditions:

[0391] Table 8: Comparison of signal quality under different working conditions (signal error rate %);

[0392] Load working condition The method (P2) Traditional manual layout Single-objective optimization method Improvement range (%) No load (10%) 0.05 0.08 0.06 37.5 / 16.7 Light load (30%) 0.08 0.15 0.12 46.7 / 33.3 Medium load (60%) 0.12 0.28 0.22 57.1 / 45.5 Heavy load (90%) 0.18 0.45 0.35 60.0 / 48.6 Load fluctuation (20%-80%) 0.15 0.52 0.41 71.2 / 63.4 Sudden interference condition 0.22 0.68 0.54 67.6 / 59.3 Average improvement - - - 56.7 / 44.5

[0393] In addition, the adaptability of the system under load mutation was also tested. Figure 1 The real-time changes in communication bus signal quality when the load was suddenly changed from 30% to 90% were recorded (in table form):

[0394] Table 9: Signal quality recovery ability under load mutation (signal-to-noise ratio SNR, dB);

[0395] Time point (seconds) The method (P2) Traditional manual layout Single-objective optimization method t=0 (before mutation) 65.8 63.2 64.5 t=0.5 (after mutation) 48.6 32.5 38.2 t＝1.0 52.3 34.6 39.5 t＝2.0 58.7 38.2 42.1 t＝5.0 61.4 42.8 48.3 t＝10.0 63.5 45.6 52.7 Recovery rate (%) 96.5 72.2 81.7

[0396] From the data in Table 8 and Table 9, it can be seen that the method exhibits excellent anti-interference robustness in the face of environmental uncertainty and interference changes, improving the anti-interference ability by about 56.7% and 44.5% compared to traditional manual layout and single-target optimization methods, respectively. Especially under load fluctuations and sudden interference conditions, the improvement effect is more significant, fully verifying the anti-interference robustness improvement achieved by the method through Bayesian inference and uncertainty modeling.

[0397] This application example proves that the strong and weak electrical line intelligent layout and interference prediction analysis method proposed by the embodiment can be effectively applied in actual industrial control systems, significantly improving the layout quality and enhancing the reliability of the system under complex working conditions, and has obvious technical and economic value.

[0398] The above describes embodiments of the present application, but the embodiments are not limited to the specific embodiments described above, which are only illustrative and not limiting. Those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection scope of the embodiments.

Claims

1. A method for intelligent layout and interference prediction analysis of strong and weak current lines, characterized in that, The method comprises the following steps: Constructing a state representation and an interference propagation model of the electronic system layout, comprising: establishing an electromagnetic interference propagation model; forming an interference influence matrix; Using the constructed interference propagation model and interference influence matrix, quantifying uncertainty based on Bayesian inference and particle filtering methods, comprising: collecting interference data under different working conditions; obtaining the probability distribution of main uncertainty factors; The main uncertainty factors include measurement uncertainty, environmental parameter uncertainty, and load variation uncertainty; Based on the state representation and the interference propagation model and the probability distribution of the uncertainty factors, applying an improved multi-objective optimization algorithm to solve the Pareto optimal layout scheme set of the strong and weak power lines, comprising: setting multiple optimization objectives; generating a non-dominated solution set; Applying the improved multi-objective optimization algorithm, specifically including: Setting the optimization objective function set: ； wherein, denotes a set of objective functions in a multi-objective optimization problem, denotes a decision variable vector of a layout scheme, denotes an interference suppression metric, denotes a space utilization, denotes a cost control, denotes a total number of optimization objectives; Applying the improved NSGA-III algorithm for multi-objective optimization, including: Using adaptive crossover probability and mutation probability : ； wherein, denotes the crossover probability in the current iteration, denotes an initial value for the crossover probability, denotes the mutation probability in the current iteration, denotes an initial value for the mutation probability, is the current generation number, , denote the amplitude coefficients for the crossover probability and the mutation probability, respectively, , denote the decay rates for the crossover probability and the mutation probability, respectively. Using a reference point-based selection mechanism to ensure the diversity of the solution set in the target space; Generating a set of non-dominated solutions representing an optimal trade-off between a plurality of objectives; Using the non-dominated solution set as training data, implementing robust reinforcement learning strategy optimization based on distributed deep Q network, including: defining the environment state space, action space, and reward function; optimizing the layout decision strategy; evaluating and optimizing the robustness of the trained strategy; Combining the non-dominated solution set and the optimized layout decision strategy, supporting decision-making through a preference learning model, including: constructing a preference learning model; evaluating and ranking all non-dominated solutions; recommending the optimal layout scheme.

2. The method according to claim 1, wherein, The method of constructing a state representation and an interference propagation model of the electronic system layout specifically includes: discretize the layout space into a three-dimensional grid structure of wherein, represents a set of grids that represents the entire layout space, , , represents the X-axis, Y-axis, Z-axis coordinates of the space, respectively, represents a grid cell at a particular location in the three-dimensional space; For potential interference between any two circuit elements and a computational model is established based on electromagnetic field theory: , ； wherein, Ei represents the electromagnetic interference strength from an element to an element Ei represents the electromagnetic interference strength from an element Ei represents the electromagnetic interference strength from an element Ei represents the electromagnetic interference strength from an element Ei represents the electromagnetic interference strength from an element Ei represents the electromagnetic interference strength from an element Ei represents the electromagnetic interference strength from an element Ei represents the electromagnetic interference strength from an element Ei represents the electromagnetic interference strength from an element The results of the calculations are organized as an interference influence matrix wherein is the total number of circuit elements in the system.

3. The method of claim 1, wherein, The method of quantifying uncertainty based on Bayesian inference and particle filtering methods specifically includes: Installing multiple electromagnetic interference monitoring sensors Interference data is collected under different load conditions, wherein, , , respectively represent the first , , electromagnetic interference monitoring sensors deployed in the system, represent the total number of sensors; Establishing a state transition model and a measurement model wherein denotes the interference state at time instant denotes the interference state at time instant denotes the corresponding measurement value; Using a particle filtering algorithm to estimate the posterior probability of the interference distribution: ； in, Indicating in the observation sequence conditions, Constant Disturbance State The posterior probability distribution, This represents the set of particles and their corresponding weights. For the Diracdelta function, Indicates the first One particle, Represents the total number of particles. Indicates all Summing is performed on each particle; Through sampling and importance resampling steps, the probability distribution of the main uncertainty factors is obtained.

4. The method of claim 1, wherein, The method of implementing robust reinforcement learning strategy optimization based on distributed deep Q network specifically includes: Define state space , action space and reward function : The state space includes the current layout state and uncertainty estimation; The action space is defined as the possible line placement positions and directions; The reward function considers interference suppression, space utilization, and cost control; Training the distributed deep Q network: Using a double Q learning strategy to reduce Q value estimation bias; Application experience playback mechanism Store converted samples ,in, Indicates time state, Indicates time The action performed Indicates the execution of an action The instant reward obtained afterward Indicates the execution of an action The new state that was subsequently transferred to; Minimizing the loss function: ； wherein, represents a loss function for network training, represents an expectation sampled from an experience replay buffer represents an immediate reward, represents a discount factor, represents a maximum value of a next state computed using target network parameters represents a value of a state-action pair computed using current network parameters and represent current network and target network parameters, respectively;​​​​ Through Monte Carlo simulation and scenario analysis, the robustness of the strategy in an uncertain environment is evaluated, and adjustments are made.

5. The method of claim 1, wherein, The method of supporting decision-making through a preference learning model specifically includes: collection of expert evaluation data wherein, denotes a collection of expert evaluation data, , denote the first , layout scheme, denotes the preference degree of the expert for the scheme and . Constructing an enhanced twin neural network for the preference learning model, with its loss function defined as: ； wherein, denotes a loss function of the preference learning model, denotes a sum over all data samples of the expert evaluation dataset denotes a sum over all data samples of the expert evaluation dataset denotes the preference learning model, denotes the preference of the expert for the solution and denotes the preference of the expert for the solution is a regularization parameter, denotes a regularization term; Scoring all layout schemes in the non-dominated solution set using the trained preference model : ； in, Indicate layout scheme Overall score The preference model represents the layout scheme Utility rating For the measurement of diversity, For balance parameters; Based on the comprehensive score, the optimal layout scheme is recommended.

6. The method of claim 1, wherein, It also includes an adaptive weighting mechanism that considers the importance of different objectives, where the objective weights are dynamically adjusted based on environmental conditions and user needs: ； wherein, denotes the weight of the th objective at the time instant , is a parameter that varies with the environmental conditions, denotes an exponential transformation of the importance parameter, denotes the sum of the exponential transformations of all objectives, denotes the total number of optimization objectives.

7. The method of claim 1, wherein, The method of quantifying uncertainty based on Bayesian inference and particle filtering methods also includes an uncertainty propagation analysis method for evaluating the propagation characteristics of uncertainty in the system: Establishing an uncertainty propagation graph: ； wherein, represents an uncertainty propagation graph, represents a system component, represents an association between components, represents an uncertainty propagation strength; Applying the Fokker-Planck equation to describe the evolution of uncertainty in the system: ； where denotes the rate of change of the probability density function with respect to time, denotes the system state at time with probability density and denote the drift and diffusion coefficients, respectively, denotes the deterministic drift term, denotes the stochastic diffusion term.

8. A system for realizing the method for intelligent layout and interference prediction analysis of strong and weak power lines according to any one of claims 1-7, characterized in that, ​ a state representation and interference modeling module for constructing a state representation and interference propagation model of the electronic system layout; an uncertainty quantification module for uncertainty quantification based on Bayesian inference and particle filtering methods; a multi-objective optimization module for applying an improved multi-objective optimization algorithm to solve a Pareto optimal layout solution set of the power and signal lines; a reinforcement learning strategy module for implementing a robust reinforcement learning strategy optimization based on a distributed deep Q-network; a preference learning and decision support module for decision support through a preference learning model; a human-machine interaction interface for visualizing the layout solution and receiving user feedback.

9. A computer readable medium having a computer program recorded thereon, characterized in that, The computer program is executed by a processor to implement the method of any one of claims 1-7.

Images