A sensor network deployment method based on scene analysis for operator selection

Abstract

The application provides a sensor network deployment method based on scene analysis for operator selection, based on a scene analysis technology for describing characteristics of a solution space where a population is located, states of multiple deployment schemes in the solution space are extracted, including Euclidean distances between all individuals in the population P and an optimal individual, differences in coverage rates, Euclidean distances between the top 10% individuals in the population P, and position differences of all individuals in the population P, and then a mutation operator is adaptively selected according to the extracted states to perform mutation on the population, a strategy mapping from solution space state information to a suitable mutation operator is realized, a user obtains a deployment scheme with higher coverage rate in limited time, the solving efficiency of the deployment optimization problem is greatly improved, and the sensor network deployment optimization problem is effectively solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of operator selection, data analysis and sensor network deployment, and particularly relates to a sensor network deployment method based on scene analysis for operator selection. BACKGROUND

[0002] Engineering optimization problems exist widely in various fields, and the correlation between their decision variables makes the optimization objective function exhibit complex characteristics of non-convexity and non-linearity. For example, in the field of mechanical design, the design of a gear reducer needs to consider various complex factors such as the number of teeth, modulus, diameter of shaft, etc. to achieve the goal of maximizing torque conversion efficiency. In the field of sensor network deployment, it is often necessary to select several points in the area of interest of the user for deploying sensors, and each sensor works cooperatively to maximize the coverage rate of the area and avoid monitoring blind spots. It is particularly important to solve the deployment scheme through optimization algorithms to make full use of the detection capability and achieve the optimal coverage effect.

[0003] For the complex objective function in the actual engineering optimization problem of sensor network deployment, evolutionary computation methods have emerged. These methods draw on the biological evolution mechanism in nature, such as the principle of survival of the fittest and survival of the fittest, and can give feasible solutions or even optimal solutions within a reasonable time and computing resources. Common evolutionary computation methods include genetic algorithm, particle swarm optimization algorithm, differential evolution algorithm, etc. Among them, the differential evolution algorithm is widely used in various engineering optimization problems due to its simplicity, strong robustness, etc.

[0004] However, the performance of the differential evolution algorithm is highly dependent on the complexity of the solution space and the mutation operator used, and the advantages of the mutation operators are different. For example, when dealing with simple unimodal functions, using the DE / rand / 2 mutation operator with strong exploration ability will make it difficult for the optimization process to converge. For objective functions with multiple local optima, using the DE / best / 2 mutation operator with strong development ability is easy to make the algorithm fall into local optimum and cannot find the global optimal solution. By analyzing the scene of the solution space of the optimization population and adaptively selecting the appropriate mutation operator, the solving efficiency of the algorithm can be significantly improved to make the user get a higher coverage rate solution within a limited time, thereby better solving the complex sensor network deployment problem. SUMMARY

[0005] To solve the above problems, the present application provides a sensor network deployment method based on scene analysis for operator selection, which analyzes the scene of the solution space of the optimization population and adaptively selects the appropriate operator to make the user get a higher coverage rate deployment scheme within a limited time.

[0006] A sensor network deployment method based on scene analysis for operator selection, comprising the following steps:

[0007] S1: randomly setting initial values for N different individuals contained in the population P in the differential evolution algorithm, wherein each individual contains the position coordinates of all sensors to be deployed, and one individual corresponds to one sensor deployment scheme;

[0008] S2: obtaining the population state S of the current population P at the current time t by using scene analysis technology t ={s t1 , s t2 , s t3 ), wherein s t1 is a first substate for measuring the Euclidean distance between all individuals in the current population P and the optimal individual and the coverage difference, s t2 is a second substate for measuring the Euclidean distance between the top 10% of individuals in the current population P, and s t3 is a third substate for measuring the position difference of all individuals in the current population P;

[0009] S3: determining whether the current iteration number reaches the upper limit T, if yes, the sensor position coordinates contained in the optimal individual in the current population are the final sensor deployment scheme, if no, entering step S4;

[0010] S4: generating a random number between 0 and 1, and determining whether the random number is less than the set exploration probability ε, if yes, randomly selecting a mutation operator op as the action A t of the current time t, if no, inputting the population state S t of the current time t into the current Q network, and selecting the mutation operator op with the maximum Q function value as the action A t of the current time t;

[0011] S5: updating the current population P by using the mutation operator op;

[0012] S6: obtaining the reward R t corresponding to the action A t according to the fitness function values of all individuals in the updated and unupdated current population P, and obtaining the population state S t+1 of the next time of the updated current population P by using scene analysis technology;

[0013] S7: storing the quadruple (S t , A t , R t , S t+1 ) into the experience replay pool B, and then sampling multiple quadruples (S k , Ak , R k , S k+1 ) as the training set, k = 1, ..., Z, where Z is the number of four-tuple samples in the training set;

[0014] S8: The state S of each quadruple sample in the training set k 、S k+1 and action A k As targets The input of the network and Q network, for each four-tuple sample, its corresponding target Network output Function value maximum and reward value R k The sum of the values ​​is taken as the corresponding target value y k ; According to the target value y corresponding to each quadruple sample k The difference between the Q function value output by the Q network constructs the mean square error loss function L;

[0015] S9: Backpropagate the Q network through the mean square error loss function L to update the parameters of the Q network;

[0016] S10: Determine whether the current moment t is the target A multiple of the network's update frequency C. If yes, the target The network is updated to the Q network obtained in step S9, and then step S3 is re-executed; if not, the target The network remains unchanged, and step S3 is directly re-executed.

[0017] Furthermore, the first substate s t1 The method to obtain is:

[0018]

[0019] Among them, f i is the fitness function value of the i-th individual in the population P, i = 1, ..., N, is the average value of the fitness function of all individuals in the population P, and the fitness function value is the detection coverage of the sensor deployment scheme corresponding to the individual, d i is the Euclidean distance between the i-th individual in population P and the optimal individual with the largest fitness function value, is the average value of the Euclidean distances corresponding to all individuals in population P, σ F is the standard deviation of the fitness function values ​​of all individuals in population P, σ d is the standard deviation of the Euclidean distances corresponding to all individuals in population P.

[0020] Furthermore, the second substate s t2 The method to obtain is:

[0021]

[0022] wherein q is the number of the top 10% of individuals in the current population P with the largest fitness function values, and q = 0.1N, is the individual ranked i in the q individuals, is the individual ranked j in the q individuals, is the individual ranked i in the fitness value and the individual ranked j in the fitness value between the individual ranked i in the fitness value

[0023] Further, the third sub-state s t3 is obtained by the following method:

[0024]

[0025] wherein D is the dimension of each individual, and the number of sensors is D / 2, the ith dimension vector x i = {P 1，i , P 2，i , …, P N，i}, P 1，i ~ P N，i are the ith dimension of the 1st individual P1 in the current population P to the ith dimension of the Nth individual P N , iqr(x i ) is the interquartile distance of the ith dimension vector x i , up i and low i are the set upper limit value and the set lower limit value corresponding to the ith dimension, respectively.

[0026] Further, the method for updating the current population P by the mutation operator op in step S5 is as follows:

[0027] S51: mutate all individuals in the current population P by the mutation operator op to obtain N mutation intermediates;

[0028] S52: determine whether the position coordinates contained in each mutation intermediate exceed the set upper limit value or are lower than the set lower limit value, and for the mutation intermediate with the determination result of yes, directly set the position coordinates exceeding the set upper limit value to the upper limit value and the position coordinates lower than the set lower limit value to the lower limit value, and the mutation intermediate with the determination result of no remains unchanged;

[0029] S53: cross the N individuals in the current population P with the mutation intermediates obtained in step S52 by using the binomial crossover to obtain N crossover intermediates;

[0030] S54: selecting N individuals with better fitness values from the N individuals of the current population P and the N crossover intermediates to form a next generation population based on a greedy criterion, and completing the update of the current population.

[0031] Further, the mutation operation of all individuals of the current population P in step S51 using the mutation operator op is specifically:

[0032] When the mutation operator op is DE / rand / 2, the mutation operator op mutates the i-th individual of the current population P into the i-th mutation intermediate M i in the following method:

[0033]

[0034] wherein r i1 to r i5 are randomly selected positive integers in [1, N] when mutating the i-th individual, and r i1 ≠ r i2 ≠ r i3 ≠ r i4 ≠ r i5 , to are the corresponding individuals of the randomly selected integers r i1 to r i5 in the current population P, and W ∈ (0, 1] is a scaling factor;

[0035] When the mutation operator op is DE / current-to-rand / 1, the mutation operator op mutates the i-th individual P i of the current population P into the i-th mutation intermediate M i in the following method:

[0036]

[0037] wherein r i1 to r i3 are randomly selected positive integers in [1, N] when mutating the i-th individual, and r i1 ≠ r i2 ≠ r i3 , to are the corresponding individuals of the randomly selected integers r i1 to r i3 in the current population P;

[0038] When the mutation operator op is DE / best / 2, the mutation operator op mutates the i-th individual into the i-th mutation intermediate M i in the following method:

[0039]

[0040] Among them, P best is the individual with the largest fitness function value in the current population P, r i1 to r i4 is a positive integer randomly selected in [1, N] when mutating the i-th individual, and r i1 ≠r i2 ≠r i3 ≠r i4 , to is a random integer r in the current population P i1 to r i4 The corresponding individual.

[0041] Furthermore, any cross intermediate U in step S53 i The method to obtain is:

[0042]

[0043] Where j = 1, 2, ..., D, D is the dimension of each individual, and the number of sensors is D / 2, U i，j is the i-th crossover intermediate U i The jth dimension, M i，j is the jth dimension of the ith variant intermediate, P i，j is the jth dimension of the i-th individual, rand i To obtain the cross intermediate U i When , a random number is randomly selected between (0, 1], and CR∈(0, 1] is the set crossover probability; To obtain the cross intermediate U i is a randomly selected positive integer in the range [1, D].

[0044] Furthermore, in step S54, N individuals are selected from the N individuals of the current population P and the N crossover intermediates U to form the next generation population based on the greedy criterion as follows:

[0045]

[0046] in, is the i-th individual in the next generation population, f(U i ) is the i-th crossover intermediate U i The fitness function value, f(P i ) is the fitness function value of the i-th individual in the current population P.

[0047] Furthermore, in step S6, the reward R t The calculation method is:

[0048] R t =(min Ft-1 - min F t ) / min F t-1

[0049] Wherein, min F t-1 is the minimum value of the fitness function value of all individuals in the current population P before updating, min F t is the minimum value of the fitness function value of all individuals in the current population P after updating.

[0050] Further, the mean square error loss function L in step S8 is specifically:

[0051]

[0052] Wherein, Q(S k , A k ) is the Q function value obtained by the Q network when taking state S k and action A k as input; is the maximum value of the function value that the target network can output, and A is the action that makes the maximum value of the function value that the target network can output. Beneficial effects:

[0053] Advantages:

[0054] 1. The application provides a sensor network deployment method based on scene analysis for operator selection, based on scene analysis technology for describing the characteristics of the solution space where the population is located, the states of multiple deployment schemes in the solution space are extracted, including the Euclidean distance between all individuals in the population P and the optimal individual and the coverage difference, the Euclidean distance between the top 10% of individuals in the population P, and the position difference of all individuals in the population P, and then the extracted states are used to adaptively select a mutation operator to mutate the population, realizing the strategy mapping from the solution space state information to the appropriate mutation operator selection, so that the user can obtain a deployment scheme with higher coverage in a limited time, greatly improving the solution efficiency of the deployment optimization problem, and effectively solving the sensor network deployment optimization problem.

[0055] 2. The application provides a sensor network deployment method based on scene analysis for operator selection, compared with the traditional sensor network deployment based on fixed mutation operators, the optimization result standard deviation of the sensor network deployment method based on adaptive operator selection is smaller, and the performance is more stable, so that the sensitivity to noise in the actual sensor network deployment optimization problem is lower.

[0056] ​3. The application provides a sensor network deployment method based on scene analysis for operator selection, which first obtains a population state based on scene analysis, and then combines the population state with a DQN network to select a mutation operator, so that the convergence speed of the optimization problem of the sensor network deployment is faster, and the optimization effect of the sensor network deployment is more obvious. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 A flowchart of a sensor network deployment method based on scene analysis for operator selection provided by the application is shown in the figure.

[0058] Figure 2 A performance comparison diagram of a sensor network deployment method based on adaptive operator selection provided by the application and a sensor network deployment method based on three fixed mutation operators under a set experiment. DETAILED DESCRIPTION

[0059] In order to enable personnel in the technical field to better understand the scheme of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in combination with the drawings in the embodiments of the present application.

[0060] The application provides a sensor network deployment method based on scene analysis for operator selection, which adaptively selects a suitable operator from a set according to the solution space state obtained by the scene analysis technology, and further improves the efficiency and quality of the algorithm for solving complex deployment problems. First, some mathematical symbols used in the application are defined as follows: a population P with a size of N in the differential evolution algorithm, wherein each individual corresponds to a sensor network deployment scheme; the dimension D, that is, the two-dimensional coordinates of D / 2 sensors need to be determined; the fitness function value F of all individuals, that is, the area coverage rate of all deployment schemes; the upper limit T of the number of iterations of the algorithm; a set of mutation operators O, a selected operator op; a training network Q and a target network Exploration probability epsilon, state S at time t t , action A t (i.e., the selected operator op) and reward R t , experience replay pool B, target network update frequency C, batch sampling size Z, and learned operator selection method pi.

[0061] A sensor network deployment optimization problem in a two-dimensional area is taken as an implementation example. The optimization objective of the problem is to maximize the area coverage rate. The average coverage probability of all monitoring points is calculated to evaluate the pros and cons of the deployment scheme, and the objective function is as follows:

[0062]

[0063] Among them, n p Represents the number of monitoring points. The sensor network is a monitoring point (x j ,y j ), j = 1, 2, ..., n p ;S i is the i-th sensor, i = 1, 2, ..., n p .

[0064]

[0065] Among them, C x，y (S i ) is a single sensor S i The coverage probability of the monitoring point (x, y), d is the coverage probability of the sensor S i The Euclidean distance between the monitoring point and the sensor; r is the detection radius of the sensor, r e Characterizes the uncertainty of detection, here we take r e =0.1·r; α1, α2, β1, β2 are the performance parameters of the sensor, λ1 = r e -r+d,λ2=r e +rd. Formula (14) is the joint coverage probability of the sensor network for a single monitoring point (x, y). If a monitoring point is covered by multiple sensors, its coverage probability will increase accordingly.

[0066] Specifically, such as Figure 1 As shown, a sensor network deployment method for operator selection based on scene analysis includes the following steps:

[0067] S1: Randomly set initial values ​​for the N different individuals in the population P in the differential evolution algorithm, where each individual contains the location coordinates of all sensors to be deployed, and one individual corresponds to one sensor deployment plan;

[0068] It should be noted that when initializing the population P in the differential evolution algorithm, the upper and lower bounds of the sensor network deployment location are first defined. The position coordinates of each sensor initially randomly generated need to fall between the set upper and lower bounds.

[0069] S2: Use the image analysis technology to obtain the population state S of the current population P at the current time t t ={s t1 , s t2 , s t3}, where s t1 is the first substate that measures the Euclidean distance and coverage difference between all individuals in the current population P and the optimal individual, s t2The second substate of the Euclidean distance between the best 10% individuals in the current population P is measured, s t3 The third substate is used to measure the position differences of all individuals in the current population P;

[0070] It should be noted that landscape analysis is a technique that describes the characteristics of the solution space where the population is located, and characterizes the characteristics of the initial deployment plan in the solution space.

[0071] Specifically, the first substate s t1 The method to obtain is:

[0072]

[0073] Among them, f i is the fitness function value of the i-th individual in the population P, i = 1, ..., N, is the average value of the fitness function of all individuals in the population P, and the fitness function value is the detection coverage of the sensor deployment scheme corresponding to the individual. The calculation method of the detection coverage can refer to the above C x，y (S i ) calculation process, d i is the Euclidean distance between the i-th individual in population P and the optimal individual with the largest fitness function value, is the average value of the Euclidean distances corresponding to all individuals in population P, σ F is the standard deviation of the fitness function values ​​of all individuals in population P, σ d is the standard deviation of the Euclidean distances corresponding to all individuals in the population P. It can be seen that in the sensor network deployment problem, the first sub-state s t1 The relationship between the distance to the optimal deployment solution and the area coverage is provided for each solution.

[0074] The second substate s t2 The method to obtain is:

[0075]

[0076] Where q is the number of the top 10% individuals with the largest fitness function value in the current population P, and q = 0.1N, is the individual with fitness ranking i among q individuals, is the individual with the fitness value ranking j among q individuals, The individual with fitness value ranking i and the individual with fitness ranking j The Euclidean distance between them. It can be seen that the second substate s t2The average pairwise distance between the q best individuals in the population P is defined, which represents whether the positions of the excellent deployment schemes are close in the sensor network deployment problem.

[0077] The third sub-state s t3 The acquisition method is:

[0078]

[0079] Wherein, D is the dimension of each individual, and the number of sensors is D / 2, the i-th dimension vector x i ={P 1，i , P 2，i , …, P N，i}, P 1，i ~P N，i is the i-th dimension of the 1st individual P1 in the current population P to the i-th dimension of the Nth individual P N , iqr(x i ) is the quartile distance of the i-th dimension vector x i , that is, the difference between the third quartile and the first quartile, up i and low i are the set upper limit value and the set lower limit value corresponding to the i-th dimension. It can be seen that the third sub-state s t3 The diversity of the population P is calculated, that is, the difference between each deployment scheme is measured; the higher the value of s t3 , the higher the upper limit of the current population to improve the area coverage rate.

[0080] S3: Determine whether the current iteration number reaches the upper limit T, if yes, the sensor position coordinates contained in the optimal individual in the current population are the final sensor deployment scheme, if no, go to step S4;

[0081] S4: Based on the current state S t , an action is selected using the ε-greedy strategy, specifically: a random number between 0 and 1 is generated, and it is determined whether the random number is less than the set exploration probability ε, if yes, a mutation operator op is randomly selected as the action A t of the current t time, if no, the population state S t of the current t time is input into the current Q network, and the mutation operator op that makes the Q function value maximum is taken as the action A t of the current t time, that is

[0082] It should be noted that the mutation operator set O includes 3 mutation operators, which are DE / rand / 2, DE / current-to-rand / 1 and DE / best / 2, and a randomly selected mutation operator op is one of DE / rand / 2, DE / current-to-rand / 1 and DE / best / 2.

[0083] S5: updating the current population P by using the mutation operator op, specifically:

[0084] S51: performing mutation operation on all individuals in the current population P by using the mutation operator op to obtain N mutation intermediates;

[0085] When the mutation operator op is DE / rand / 2, the mutation operator op mutates the i-th individual in the current population P into the i-th mutation intermediate M i The method is:

[0086]

[0087] wherein r i1 to r i5 are randomly selected positive integers in [1, N] when mutating the i-th individual, and r i1 ≠ r i2 ≠ r i3 ≠ r i4 ≠ r i5 , to are the individuals corresponding to the randomly selected integers r i1 to r i5 in the current population P, and W e (0, 1] is a scaling factor.

[0088] When the mutation operator op is DE / current-to-rand / 1, the mutation operator op mutates the i-th individual P i in the current population P into the i-th mutation intermediate M i The method is:

[0089]

[0090] wherein r i1 to r i3 are randomly selected positive integers in [1, N] when mutating the i-th individual, and r i1 ≠ r i2 ≠ r i3 , to are the individuals corresponding to the randomly selected integers r i1 to r i3corresponding individual, i.e., a sensor network deployment scheme;

[0091] When the mutation operator op is DE / best / 2, the mutation operator op mutates the ith individual into the ith mutation intermediate M i The method is:

[0092]

[0093] wherein P best is the individual with the largest fitness function value in the current population P, r i1 to r i4 are randomly selected positive integers in [1, N] when the ith individual is mutated, and r i1 ≠ r i2 ≠ r i3 ≠ r i4 , to are the individuals corresponding to the random integers r i to r i，j in the current population P, i.e., sensor network deployment schemes.

[0094] It should be noted that when different mutation operators are used to mutate the same individual i in the current population P, although the randomly selected positive integers r i1 to r i5 in [1, N] appear, r i1 to r i5 may have different values in the mutation process of different mutation operators.

[0095] S52: Determine whether the position coordinates contained in each mutation intermediate exceed the set upper limit value or are lower than the set lower limit value. For the mutation intermediate with a determination result of yes, directly set the position coordinates exceeding the set upper limit value to the upper limit value, and directly set the position coordinates lower than the set lower limit value to the lower limit value. Meanwhile, the mutation intermediate with a determination result of no remains unchanged;

[0096] S53: Cross the N individuals of the current population P with the mutation intermediates obtained in step S52 in a binomial cross manner to obtain N cross intermediates;

[0097] The acquisition method of any one cross intermediate U i is:

[0098]

[0099] wherein j = 1, 2, …, D, D is the dimension of each individual, and the number of sensors is D / 2, U i，j is the jth dimension of the ith cross intermediate U i , and M i，jP is the jth dimension of the ith mutation intermediate i，j rand is the jth dimension of the ith individual i P is the jth dimension of the ith mutation intermediate i rand is a random number randomly selected between (0, 1], and CR is a set crossover probability; P is the jth dimension of the ith mutation intermediate i rand is a positive integer randomly selected in the range of [1, D].

[0100] S54: N individuals with better fitness values are selected from the N individuals of the current population P and the N crossover intermediates based on a greedy criterion to form a next generation population, and the update of the current population is completed, and the specific process is as follows:

[0101]

[0102] wherein, P is the ith individual of the next generation population, f(U i ) is the fitness function value of the ith crossover intermediate U i , that is, the detection coverage, and f(P i ) is the fitness function value of the ith individual of the current population P.

[0103] It can be seen that the update process of the population of the application specifically includes mutation, crossover and selection. According to the selected operator op, the corresponding difference strategy is used to perform mutation operation on all individuals, and it is checked whether the newly generated mutation intermediate M satisfies the boundary condition of the complex optimization sensor network deployment problem. Then, the binomial crossover is used to cross the population P at time t and the mutation intermediate M to obtain the crossover intermediate U. Finally, the next generation population P is selected based on the greedy criterion.

[0104] S6: Based on the fitness function improvement criterion, the action A t corresponding to the reward R t is obtained according to the fitness function values of all individuals in the updated and unupdated current population P, and the scene analysis technology is used to obtain the population state S t+1 of the updated current population P at the next time;

[0105] The calculation method of the reward R t is as follows:

[0106] R t = (min F t-1 -min F t ) / min F t-1

[0107] wherein, min F t-1is the minimum value of the fitness function of all individuals in the current population P before the update, minF t is the minimum value of the fitness function of all individuals in the updated current population P. t >0 means that the deployment solutions included in the updated population have improved regional coverage.

[0108] S7: The quadruple (S t , A t , R t , S t+1 ) is stored in the experience replay pool B, and then multiple quadruple groups (S k , A k , R k , S k+1 ) as the training set, k = 1, ..., z, where Z is the number of four-tuple samples in the training set;

[0109] S8: The state S of each quadruple sample in the training set k 、S k+1 and action A k As targets The input of the network and Q network, for each four-tuple sample, its corresponding target Network output Function value maximum and reward value R k The sum of the values ​​is taken as the corresponding target value y k ; According to the target value y corresponding to each quadruple sample k The difference between the Q function value output by the Q network constructs the mean square error loss function L;

[0110] Among them, the calculation formula of the square error loss function L is specifically as follows:

[0111]

[0112] Among them, Q(S k , A k ) is the Q network in state S k and action A k The Q function value obtained as input; Target The network can output The maximum value of the function, A is the target The network can output The action at which the function value is maximized.

[0113] S9: Back propagate the Q network through the mean square error loss function L to update the parameters of the Q network so that Q(S k, A k ) as close as possible to y k ;

[0114] S10: judging whether the current time t is a multiple of the update frequency C of the target network, if yes, updating the target network to the Q network obtained in step S9, and re-executing step S3; if no, the target network remains unchanged, and directly re-executing step S3.

[0115] To further illustrate the effectiveness of the provided method, the adaptive operator selection method provided by the present application and the differential evolution algorithm with three fixed mutation operators are simulated and verified under the coverage problem. The simulation parameters of the problem in the experiment are set as: the deployment space is a two-dimensional region of 40*40, r=5, r e =0.1, r=0.5, alpha1=1, alpha2=0, beta1=1, and beta2=0.5. The algorithm parameters are set as: N=50, q=0.1, N=5, D=40, T=100, U={DE / rand / 2, DE / best / 2, DE / current-to-rand / 1}, epsilon=0.1, C=5, Z=64, W=0.5, and CR=0.8. Each decision variable is a continuous value between 0 and 40, i.e., up i =0, and low i =40.

[0116] Considering 50 random experiments, Figure 2 the simulation results are shown, which are the performance comparison between the adaptive operator selection method (DE-DQN) of the present application and the differential evolution algorithm with three fixed mutation operators. The results show that the convergence speed of the adaptive operator selection method (DE-DQN) based on scene analysis is faster, and the optimization effect is more obvious, which can improve the regional coverage rate of the deployment scheme by 6.1% based on the existing method. In addition, Table 1 shows the average value and standard deviation of the fitness value of the adaptive operator selection method (DE-DQN) and the differential evolution algorithm with three fixed mutation operators under 50 random experiments. The optimization result standard deviation of the adaptive operator selection method is smaller, and the performance is more stable, so that its sensitivity to noise in the actual sensor network deployment optimization problem is lower.

[0117] Table 1

[0118]

[0119] ​​​Based on the above experiments, the adaptive operator selection based on scene analysis provided in the application effectively solves the sensor network deployment optimization problem, realizes the strategy mapping from the solution space state information to the suitable mutation operator selection, makes the user obtain the deployment scheme with higher coverage in limited time, and greatly improves the solution efficiency of the deployment optimization problem.

[0120] Of course, the present application can have other various embodiments, and those skilled in the art can certainly make various corresponding changes and modifications according to the present application without departing from the spirit and essence of the present application, but these corresponding changes and modifications should all belong to the protection scope of the claims attached to the present application.

Claims

1. A sensor network deployment method based on scene analysis for operator selection, characterized by, The method comprises the following steps: S1: randomly setting initial values for N different individuals included in a population P in a differential evolution algorithm, wherein each individual includes position coordinates of all sensors to be deployed, and one individual corresponds to one sensor deployment scheme; S2: obtaining a population state S of the current population P at the current time t by using a landscape analysis technology t = {s t1 ,s t2 ,s t3}, wherein s t1 is a first substate for measuring a Euclidean distance between all individuals in the current population P and an optimal individual and a coverage difference, s t2 is a second substate for measuring a Euclidean distance between the top 10% of individuals in the current population P, and s t3 is a third substate for measuring a position difference of all individuals in the current population P; S3: determining whether the current iteration number reaches an upper limit T, if yes, the sensor position coordinates included in the optimal individual in the current population are the final sensor deployment scheme, if no, entering step S4; S4: Generate a random number between 0 and 1, and determine whether the random number is less than the set exploration probability ε. If so, randomly select a mutation operator op as the action A at the current time t. t If not, the population state S at the current time t t Input the current Q network and use the mutation operator op that maximizes the Q function value as the action A at the current time t t ; S5: updating the current population P by using a mutation operator op; S6: obtaining an action A according to the fitness function values of all individuals in the current population P after and before the update t a corresponding reward R t , and obtaining the population state S of the current population P at the next time by using the scene analysis technology t+1 ; S7: store the quadruple (S t ,A t ,R t ,S t+1 ) into the experience replay pool B, and then sample a plurality of quadruples (S k ,A k ,R k ,S k+1 ) from the experience replay pool B in batches as a training set, k = 1, …, Z, wherein Z is the number of quadruple samples in the training set; S8: The state S of each quadruple sample in the training set k 、S k+1 and action A k As targets The input of the network and Q network, for each four-tuple sample, its corresponding target Network output Function value maximum and reward value R k The sum of the values ​​is taken as the corresponding target value y k ; According to the target value y corresponding to each quadruple sample k The difference between the Q function value output by the Q network constructs the mean square error loss function L; S9: performing back propagation on the Q network through a mean square error loss function L to update parameters of the Q network; S10: judge whether the current time t is the target a multiple of the update frequency C of the network, if yes, the target network is updated to the Q network obtained in step S9, and then step S3 is re-executed; if no, the target network is unchanged, and step S3 is directly re-executed.

2. The sensor network deployment method based on scene analysis for operator selection according to claim 1, wherein, The first sub-state s t1 The acquisition method is: where f i is the fitness function value of the i-th individual in the population P, i = 1,..., N, is the average of the fitness function values of all individuals in the population P, and the fitness function value is the detection coverage of the sensor deployment scheme corresponding to the individual, d i is the Euclidean distance between the i-th individual in the population P and the optimal individual with the largest fitness function value, is the average of the Euclidean distances corresponding to all individuals in the population P, σ F is the standard deviation of the fitness function values of all individuals in the population P, σ d is the standard deviation of the Euclidean distances corresponding to all individuals in the population P.

3. The sensor network deployment method based on scene analysis for operator selection according to claim 1, wherein, Second sub-state s t2 The acquisition method is: where q is the number of individuals in the current population P whose fitness function values are in the top 10%, and q = 0.1N, is the individual whose fitness value ranks i among the q individuals, is the individual whose fitness value ranks j among the q individuals, is the individual whose fitness value ranks i and the individual whose fitness value ranks j is the Euclidean distance between the two individuals.

4. The sensor network deployment method based on scene analysis for operator selection according to claim 1, wherein, Third sub-state s t3 The acquisition method is: where D is the dimension of each individual, and the number of sensors is D / 2, the ith dimension vector x i = {P 1,i 1, P 2,i 2, …, P N,i N}, P 1,i 1, P N,i 2, …, P N N} are the 1st~Nth individual of the ith dimension in the current population P, respectively, iqr(x i ) is the interquartile range of the ith dimension vector x i , up i and low i are the set upper bound and lower bound of the ith dimension, respectively.

5. The sensor network deployment method based on scene analysis for operator selection according to claim 1, wherein, The method for updating the current population P by using the mutation operator op in step S5 is: S51: performing mutation operation on all individuals of the current population P by using the mutation operator op to obtain N mutation intermediates; S52: determining whether the position coordinates included in each mutation intermediate exceed a set upper limit value or are lower than a set lower limit value, for the mutation intermediates with a yes determination result, directly setting the position coordinates exceeding the set upper limit value to the upper limit value and the position coordinates lower than the set lower limit value to the lower limit value, and the mutation intermediates with a no determination result remain unchanged; S53: performing crossing between the N individuals of the current population P and the mutation intermediates obtained in step S52 in a binomial crossing manner to obtain N crossing intermediates; S54: selecting N individuals with better fitness values from the N individuals of the current population P and the N crossing intermediates to form a next generation population based on a greedy criterion, and completing updating of the current population.

6. The sensor network deployment method based on scene analysis for operator selection according to claim 5, wherein, The specific method for performing mutation operation on all individuals of the current population P by using the mutation operator op in step S51 is: When the mutation operator op is DE / rand / 2, the mutation operator op mutates the i-th individual in the current population P into the i-th mutation intermediate M i The method is as follows: where r i1 r i5 is a random selected positive integer within [1, N] at the i-th iteration of the mutation of the i-th individual, and r i1 ≠ r i2 ≠ r i3 ≠ r i4 ≠ r i5 , r is a random integer within the current population P, r i1 r i5 is the corresponding individual, and W e (0, 1] is a scaling factor. When the mutation operator op is DE / current-to-rand / 1, the mutation operator op mutates the ith individual P i in the current population P i to the ith mutated intermediate M i in the following way: where r i1 r i3 is a random selected positive integer within [1, N] at the mutation of the i-th individual, and r i1 ≠ r i2 ≠ r i3 , r is a random integer within the current population P, and r i1 r i3 is the corresponding individual. When the mutation operator op is DE / best / 2, the mutation operator op mutates the ith individual into an ith mutated intermediate M i The method is as follows: where P best is the individual with the largest fitness function value in the current population P, r i1 to r i4 is a randomly selected positive integer in [1, N] when the i-th individual is mutated, and r i1 ≠ r i2 ≠ r i3 ≠ r i4 , to is the individual corresponding to the randomly selected integer r i1 to r i4 in the current population P.

7. The sensor network deployment method based on scene analysis for operator selection according to claim 5, wherein, Any one of the cross intermediates U in step S53 i The acquisition method is: where j = 1, 2, …, D, D is the dimension of each individual, and the number of sensors is D / 2, U i,j is the jth dimension of the ith crossover intermediate, M i is the jth dimension of the ith mutation intermediate, P i,j is the jth dimension of the ith individual, rand i,j is a random number randomly selected between (0, 1] when obtaining the crossover intermediate U i is a random number randomly selected between (0, 1] when obtaining the crossover intermediate U i , CR ∈ (0, 1] is a set crossover probability; is a positive integer randomly selected within the range of [1, D] when obtaining the crossover intermediate U i .

8. The sensor network deployment method based on scene analysis for operator selection according to claim 5, wherein, The specific method for selecting N individuals from the N individuals of the current population P and the N crossing intermediates to form a next generation population based on a greedy criterion in step S54 is: wherein, is the fitness function value of the ith individual of the next generation population, f(U i ) is the fitness function value of the ith crossover intermediate U i , and f(P i ) is the fitness function value of the ith individual of the current population P.

9. The sensor network deployment method based on scene analysis for operator selection according to claim 1, wherein, The reward R in step S6 t The calculation method is: R t = (min F t-1 - min F t ) / min F t-1 where min F t-1 is the minimum value of the fitness function values of all individuals in the current population P before the update, min F t is the minimum value of the fitness function values of all individuals in the current population P after the update.

10. The sensor network deployment method based on scene analysis for operator selection according to claim 1, wherein, The specific mean square error loss function L in step S8 is: where Q(S k , A k ) is a Q function value obtained when the Q network takes the state S k and the action A k as inputs; is a target network can output a maximum value of a function value, and A is an action that makes the target network can output a maximum value of a function value.

Images