A real terrain coverage optimization method for wireless sensor networks

By applying the DBO algorithm based on quasi-reverse learning and Q-learning in 3DWSN, combining the line-of-sight detection and conceptual perception model, the dung beetle optimization algorithm is optimized, which solves the limitations of WSN coverage optimization in real terrain and improves the coverage rate and the algorithm's convergence accuracy achieves more efficient 3D terrain coverage.

CN118474672BActive Publication Date: 2024-11-05FUJIAN QIDOU NETWORK TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410551828.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-07
Publication Date
2024-11-05
Estimated Expiration
2044-05-07

AI Technical Summary

Technical Problem

The existing 3DWSN coverage optimization method has limitations in real terrain, the sensor nodes are small in size, limited storage information, and ordinary optimization algorithms are insufficient in complex problems and cannot meet the coverage requirements of real terrain.

Method used

Using the DBO algorithm (QOLDBO) based on quasi-reverse learning and Q-learning, the DBO algorithm (QOLDBO) is used to establish a three-dimensional coverage perception model, combined with the visual line detection and conceptual perception model, and using quantum state reverse learning and Q-learning optimization dung beetle optimization algorithm, add Individual randomness and search range, variable helical local neighborhood search and dimensional adaptive Gaussian variant strategy are adopted to improve the algorithm's global search ability and convergence efficiency.

Benefits of technology

It improves the coverage optimization performance of WSN under real 3D terrain, improves the coverage rate and algorithm convergence accuracy, enhances the ability to solve complex optimization problems, and proves its practical application capabilities in the real world.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118474672B_ABST
    Figure CN118474672B_ABST
Patent Text Reader

Abstract

The present invention provides a method for optimizing real terrain coverage of a wireless sensor network, comprising the following steps: S1, establishing a three-dimensional coverage perception model; S2, optimizing dung beetles based on quantum state reverse learning and Q-learning; comprising the following sub-steps: S2.1 quasi-reverse learning based on quantum computing; S2.2 behavior mode selection based on Q-learning; S2.3 variable spiral local neighborhood search; S2.4 optimal solution dimension-by-dimension adaptive Gaussian mutation. The present invention proposes a QOLDBO algorithm based on quasi-reverse learning and Q-learning, which improves the problem of insufficient global optimization of the QOLDBO algorithm, changes the degree of influence of the current solution on the quasi-reverse solution under different iterations, and increases the randomness of individuals; expands the search range of dung beetles in the breeding and stealing stages, and adopts a variable spiral local neighborhood search method to solve the neighborhood optimization problem; in order to improve the search accuracy of QOLDBO, a dimension-by-dimension adaptive Gaussian mutation strategy for the optimal solution is proposed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention provides a method for optimizing real terrain coverage of a wireless sensor network, belonging to the technical field of simulation scenes. Background Art

[0002] Although the simulated scene is closer to reality than the two-dimensional plane, it is still unrealistic. There are many complex problems in the study of 3DWSN coverage optimization. For example, 3D scenes require more nodes and adjacent nodes, and a large amount of memory is needed to store information, but the sensor nodes are small and cannot store much information. Some current studies are conducted on three-dimensional surfaces, but in real deployment environments, this method still has certain limitations. The curves generated by the function may be too regular and smooth and cannot meet the coverage requirements of real terrain. Based on this, ordinary optimization algorithms often have shortcomings when solving more complex problems, and it is necessary to use algorithms with better performance to optimize problems. Summary of the invention

[0003] In view of the above technical problems, the present invention proposes a DBO algorithm (Quasi-oppositional learning dung beetle optimizer, QOLDBO) based on quasi-oppositional learning and Q-learning to solve the WSN coverage problem under three-dimensional real terrain.

[0004] A method for optimizing real terrain coverage of a wireless sensor network comprises the following steps:

[0005] S1. Establishing a 3D coverage perception model

[0006] The line-of-sight detection method is used to calculate the straight-line distance between two sensor nodes and whether there is a direct line of sight between them; let there be N sensor nodes, and the node set is represented as S = {s 1 ,s 2 ,s 3 ,…,s n The set of monitoring area nodes M = {m 1 , m 2 , m 3 ,…,m n},(x i ,y i , z i ) and (x j ,y j , z j ) respectively correspond to s in the set i 、m j The three-dimensional space coordinates of the sensor node and the monitoring area node, the Euclidean distance Z between the sensor node and the monitoring area node is:

[0007]

[0008] The line-of-sight detection is integrated with the conceptual perception model, and the coverage model perceived under the 3D real terrain is adopted. The specific expression of perception probability is as follows:

[0009]

[0010] The monitored area is divided into L·W grids of equal area, and the monitoring node m is located at the center of the grid. The combined perception probability of all monitoring points is calculated, and the cumulative sum is the coverage area, and the coverage rate C r It is expressed as follows:

[0011]

[0012] S2. Dung beetle optimization based on quantum state reverse learning and Q-learning

[0013] S2.1 Quasi-reverse learning based on quantum computing

[0014] (1) Quasi-reverse learning

[0015] Define X i =[x 1 、x 2 , …, x pop ]A solution in N-dimensional space, when x i ∈[lb i ,ub i ], the solution of quasi-reverse learning is defined as follows:

[0016]

[0017] (2) Update quantum state

[0018] By updating the quantum state process in quantum computing, the influence of the current solution on the quasi-reverse solution is continuously changed to obtain more quasi-reverse individuals with mass. In quantum space, the quantum bit is the smallest unit. The quantum bit has two basic states |0> and |1>. The state of the quantum bit in quantum computing can be superimposed to form a basic storage unit, which is described as follows:

[0019]

[0020] represents the superposition state of the quantum, while ω and υ represent the probability in this state. The quantum bit is also represented by sine and cosine matrices:

[0021]

[0022] The rotation gate RY(θ) in the quantum logic gate will be used to update the quantum state. The RY gate is generated by the Pauli-Y matrix as a generator, which is specifically expressed as follows:

[0023]

[0024]

[0025] Where 0<θ<π / 2, let is [Γ Φ] T ,Γ,Φ range from 0 to 1, which is regarded as the influencing factor of the current solution on the quasi-reverse solution. The revolving gate in quantum computing continuously updates the quantum state to change the influence of the current solution on the quasi-reverse solution, and allows the two influencing factors to play a role in the population alternately, increasing the randomness of the reverse solution, thereby selecting a better initial population. The specific expression is as follows:

[0026]

[0027] (3) Specific steps

[0028] The main steps of QOL optimization algorithm in dung beetle are as follows:

[0029] (a) Set T o as the solution for the current iteration;

[0030] (b) Calculate the influence factor of the current solution on the reverse solution through equation (8);

[0031] (c) Calculate T by formula (9) o The quasi-reverse solution T qo ;

[0032] (d) Evaluate the quasi-reverse solution T for each population qo The fitness value of

[0033] (e) Comparison and The fitness value of , select the better one as the final population.

[0034]

[0035] S2.2 Behavior mode selection based on Q-learning

[0036] Each dung beetle is an individual, s represents the current state of each dung beetle, and a represents the transition of each individual from one state to another. The dung beetle adaptively switches from one state to another according to the Q-Table. The values ​​contained in the Q-Table describe the performance of each individual in the iteration. The Q-Table is used to help the algorithm control each individual to select the appropriate mode.

[0037] The mathematical formula for the learning rate α is as follows:

[0038]

[0039] α ini and α fin Represent the initial value and final value of the learning rate respectively, t and T are the current number of iterations and the maximum number of iterations.

[0040] The value in the Q-Table of each individual in the first iteration is 0. After making a selection in the current state, if the fitness value becomes smaller, it will be rewarded, otherwise it will be punished, which helps to select the appropriate behavior mode at different stages. With the help of Q-learning, the interaction between each dung beetle and the operation of the two behavior modes is divided into four steps:

[0041] (1) Initialize the Q-Table to a 2×2 table;

[0042] (2) According to the value in the current state Q-Table, obtain the optimal behavior mode for the current iteration to be executed by the dung beetle in the rolling ball stage.

[0043]

[0044] (3) Execute the selected operation and calculate the fitness value, and give a reward. The specific reward calculation method is as follows:

[0045]

[0046] (4) Update Q-Table according to formula (12);

[0047] S2.3 Variable spiral local neighborhood search

[0048] The variable spiral search strategy is adopted to improve the spiral operator β. The specific mathematical expression is as follows:

[0049] β=e bl ·cos(2πb) (14)

[0050]

[0051] In formula (14), b is a random number uniformly distributed from 0 to 1, and l is controlled by the cosine function with the number of iterations, so that the position update range of the dung beetle in this stage changes from large to small. In the early stage, good individuals are searched in a larger local neighborhood, which enhances the global search ability of the algorithm, reduces invalid searches in the later stage of iteration, and improves the convergence efficiency of the algorithm. After the improvement, the position updates of the dung beetle in the breeding and stealing stages are as follows:

[0052]

[0053]

[0054] S2.4 Optimal solution dimension-by-dimension adaptive Gaussian mutation

[0055] After each iteration, adaptive Gaussian mutation is applied to the best individual in each dimension, and a greedy strategy is used to retain the better individuals to achieve a balanced exploration and development effect. The specific mutation method is as follows:

[0056]

[0057]

[0058] In formula (19), x b (j) represents the jth dimension of the optimal individual, γ is the mutation probability, and randn indicates that it obeys the Gaussian normal distribution.

[0059] The optimal solution will be adaptively adjusted based on the result of the mutation. Compared to x b (j) has a low fitness value, then use Replace and amplify the probability of mutation, increase the individual movement distance, if the mutation point Compared to x b (j) does not have a low fitness value, then it remains unchanged and the mutation probability is reduced. The mutation probability update formula is as follows:

[0060]

[0061] In formula (20), μ is an expansion parameter and μ>1, λ is a contraction parameter and 0<λ<1.

[0062] The present invention proposes a QOLDBO algorithm based on quasi-reverse learning and Q-learning, which improves the problem of insufficient global optimization of the QOLDBO algorithm. In QOLDBO, the OBL strategy is to integrate quasi-opposition learning into the quantum state update process, which changes the influence of the current solution on the quasi-reverse solution under different iterations and increases the randomness of individuals. In order to expand the search range of dung beetles in the reproduction and stealing stages, a variable spiral local neighborhood search method is proposed to solve the neighborhood optimization problem. In addition, in order to improve the search accuracy of QOLDBO, a dimension-by-dimension adaptive Gaussian mutation strategy of the optimal solution is proposed. Finally, six classic swarm intelligent optimization algorithms such as QOLDBO are tested at CEC2017. The results show that QOLDBO has high convergence accuracy. A comparative analysis is conducted on the robustness, solution accuracy, convergence speed, stability of the algorithm and the effectiveness of related strategies, which proves that QOLDBO has obvious advantages. QOLDBO is applied to actual optimization problems in the real world and WSN coverage optimization problems on 3D real terrain to further prove its practical application capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 The line of sight detection method LOS of the present invention;

[0064] Figure 2 It is a schematic diagram of the process of updating the Q table of the present invention;

[0065] Figure 3 It is a schematic diagram of dimension-by-dimension variation of the present invention;

[0066] Figure 4 It is the flow chart of QOLDBO algorithm of the present invention;

[0067] Figure 5 A 3D real terrain map of an embodiment. DETAILED DESCRIPTION

[0068] The specific technical solution of the present invention is explained in conjunction with embodiments.

[0069] The present invention comprises the following steps:

[0070] S1. Establishing a 3D coverage perception model

[0071] The common communication methods between sensor nodes usually depend on the network topology and communication protocol, and can be divided into direct communication, relay communication, multi-hop communication, base station communication and routing protocol, etc. Sensor nodes generally use receiving antennas and transmitting antennas to achieve effective data exchange and communication, thereby realizing real-time monitoring of the environment, information collection and data transmission. When there is an obstacle between two sensor nodes, it usually affects the communication quality of the antenna transmission between them. In order to determine whether the communication is blocked, the present invention will adopt the line-of-sight detection method (LOS), by calculating the straight-line distance between the two sensor nodes and whether there is a direct line of sight between them. The specific principle is shown as follows Figure 1 As shown in the figure, it can be seen that if there is a direct line of sight between two points, it means that communication is not blocked; on the contrary, if the line of sight is blocked by obstacles, communication may be affected.

[0072] The present invention assumes that there are N sensor nodes, and the node set is represented by S = {s 1 ,s 2 ,s 3 ,…,s n The set of monitoring area nodes M = {m 1 , m 2 , m 3 ,…,m n},(x i ,y i , z i ) and (x j ,y j , z j ) respectively correspond to s in the set i 、m j The three-dimensional space coordinates of the sensor node and the monitoring area node, the Euclidean distance Z between the sensor node and the monitoring area node is:

[0073]

[0074] This paper integrates line-of-sight detection with the conceptual perception model and proposes a coverage model that can be perceived under 3D real terrain. The specific perception probability expression is as follows:

[0075]

[0076] The monitored area is divided into L·W grids of equal area, and the monitoring node m is located at the center of the grid. The combined perception probability of all monitoring points is calculated, and the cumulative sum is the coverage area, and the coverage rate C r It can be expressed as follows:

[0077]

[0078] S2. Dung beetle optimization based on quantum state reverse learning and Q-learning

[0079] S2.1 Quasi-reverse learning based on quantum computing

[0080] Quasi-oppositional learning is a method based on traditional reverse learning to improve the quality of the current solution. Because the original reverse learning conversion position is too large, the search accuracy is not high, and the generated solution is easy to stagnate at the local optimum. In order to significantly improve the quality of the initial solution, inspired by the improvement of chaotic mapping in the sparrow algorithm through quantum computing, the concept of quantum computing is introduced into quasi-oppositional learning. By continuously updating the quantum state, the influence of the current solution on the reverse solution is changed, so that the population is initialized more randomly. This method not only takes into account the current solution and the quasi-oppositional solution, but also uses the characteristics of continuous transformation in quantum logic gates to expand the search space and prove that the quasi-oppositional solution is closer to the optimal solution.

[0081] (1) Quasi-reverse learning

[0082] Define X i =[x 1 、x 2 , …, x pop ]A solution in N-dimensional space, when x i ∈[lb i ,ub i ], the solution of quasi-reverse learning is defined as follows:

[0083]

[0084] (2) Update quantum state

[0085] The present invention will obtain more quasi-reverse individuals with mass by updating the quantum state process in quantum computing and constantly changing the degree of influence of the current solution on the quasi-reverse solution. In quantum space, the quantum bit is the smallest unit, and the quantum bit has two basic states |0> and |1>. The state of the quantum bit in quantum computing can be superimposed to form a basic storage unit, which is specifically described as follows:

[0086]

[0087] Represents the superposition state of quantum, while ω and υ represent the probability in this state. The quantum bit can also be represented by sine and cosine matrices:

[0088]

[0089] In addition, in quantum computing, gate operators are used to operate data, also known as quantum logic gates, which convert input states into output states and are reversible. Common quantum logic gates include Pauli gate series, Hadamard gates, and rotation gates. The present invention will use the rotation gate RY(θ) in the quantum logic gate to update the quantum state. The RY gate is generated by the Pauli-Y matrix as a generator, which is specifically expressed as follows:

[0090]

[0091]

[0092] Where 0<θ<π / 2, let is [Γ Φ] T ,Γ,Φ range from 0 to 1, which is regarded as the influencing factor of the current solution on the quasi-reverse solution. The revolving gate in quantum computing continuously updates the quantum state to change the influence of the current solution on the quasi-reverse solution, and allows the two influencing factors to play a role in the population alternately, increasing the randomness of the reverse solution, thereby selecting a better initial population. The specific expression is as follows:

[0093]

[0094] (3) Specific steps

[0095] The lack of diversity in the population makes DBO prone to falling into a local optimal state. With the help of QOL, this deficiency can be compensated by improving the algorithm's exploration ability. The main steps of QOL in the dung beetle optimization algorithm are as follows:

[0096] (a) Set T o as the solution for the current iteration;

[0097] (b) Calculate the influence factor of the current solution on the reverse solution through equation (8);

[0098] (c) Calculate T by formula (9) o The quasi-reverse solution T qo ;

[0099] (d) Evaluate the quasi-reverse solution T for each population qo The fitness value of

[0100] (e) Comparison and The fitness value of , select the better one as the final population.

[0101]

[0102] S2.2 Behavior mode selection based on Q-learning

[0103] In the present invention, Q-learning is combined with the rolling ball stage in the DBO algorithm to select a better behavior mode for it, avoiding the uncertainty caused by random selection and the waste of position update in this stage. Each dung beetle is regarded as an individual, s represents the current state of each dung beetle, and a represents the transition of each individual from one state to another. The dung beetle adaptively switches from one state to another according to the Q-Table. The numerical value contained in the Q-Table describes the performance of each individual in the iteration. The Q-Table is used to help the algorithm control each individual to select a suitable mode.

[0104] From the updated Q-value formula, we can see that one of the key parameters of Q-learning is the learning rate α, which is related to the extent to which the newly learned information replaces the previously held knowledge. When α is close to 1, Q-learning assigns higher priority to the newly acquired information after exploring all states. On the contrary, a smaller α value gives a higher priority to the existing information in the Q-Table. Therefore, α is usually set to a high value at the beginning of the search process and gradually decreases. By adaptively adjusting the learning rate parameter, the information is reasonably used at each stage, which improves the exploration ability of QLDBO. Its mathematical formula is as follows:

[0105]

[0106] α ini and α fin Represent the initial value and final value of the learning rate respectively, t and T are the current number of iterations and the maximum number of iterations.

[0107] The value in the Q-Table of each individual in the first iteration is 0. After making a selection in the current state, if the fitness value becomes smaller, it will be rewarded, otherwise it will be punished, which helps to select the appropriate behavior mode at different stages. With the help of Q-learning, the main interaction between each dung beetle and the operation of the two behavior modes can be divided into four steps:

[0108] (1) Initialize the Q-Table to a 2×2 table, such as Figure 2 shown.

[0109] (2) According to the value in the current state Q-Table, obtain the optimal behavior mode for the current iteration to be executed by the dung beetle in the rolling ball stage.

[0110]

[0111] (3) Execute the selected operation and calculate the fitness value, and give a reward. The specific reward calculation method is as follows:

[0112]

[0113] (4) Update the Q-Table according to equation ((12), such as Figure 2 shown.

[0114] S2.3 Variable spiral local neighborhood search

[0115] In the reproduction and stealing stages of the dung beetle optimization algorithm, dung beetles mainly perform update searches around the local optimal individual. The search method is too blind and single, and it loses the exploration of the nearby search space while approaching the local optimal individual. The population diversity is reduced and it is easy to fall into the local optimum. Therefore, inspired by the whale optimization algorithm, the variable spiral search strategy is introduced into the position update of the dung beetle in the reproduction and stealing stages.

[0116] In the traditional whale optimization algorithm (WOA), when launching a bubble net attack, the whale approaches the best whale individual in a spiral, but adopts a fixed spiral shape. In order to adaptively reduce the spiral at different stages as the iteration proceeds, a variable spiral search strategy is proposed to improve the spiral operator β. The specific mathematical expression is as follows:

[0117] β=e bl ·cos(2πb) (14)

[0118]

[0119] In formula (14), b is a random number uniformly distributed from 0 to 1, and l is controlled by the cosine function with the number of iterations, so that the position update range of the dung beetle in this stage changes from large to small. In the early stage, good individuals are searched in a larger local neighborhood, which enhances the global search ability of the algorithm, reduces invalid searches in the later stage of iteration, and improves the convergence efficiency of the algorithm. After the improvement, the position updates of the dung beetle in the breeding and stealing stages are as follows:

[0120]

[0121]

[0122] S2.4 Optimal solution dimension-by-dimension adaptive Gaussian mutation

[0123] One of the main disadvantages of DBO is that it is not good at exploring the solution space. Each iteration of optimization is updated around the optimal solution, which makes the position of the optimal individual very important. Traditional DBO does not make full use of the information contained in the optimal individual and lacks the development and utilization of the optimal individual. In order to enhance the ability of DBO algorithm to explore local optimality and solve the problem of premature convergence, Figure 3As shown, after each iteration, the present invention will apply dimension-by-dimension adaptive Gaussian mutation to the optimal individual in each dimension, and use a greedy strategy to retain better individuals to achieve the effect of balanced exploration and development. The specific mutation method is as follows:

[0124]

[0125]

[0126] In formula (19), x b (j) represents the jth dimension of the optimal individual, γ is the mutation probability, and randn indicates that it obeys the Gaussian normal distribution.

[0127] When for the optimal individual with a certain position, when the optimal individual x is b When all dimensions of mutate, it can be regarded as the optimal solution x b Move to other positions in the solution space, where the mutation probability γ is positively correlated with the moving distance. Therefore, in the early and late stages of the algorithm, the present invention will make adaptive adjustments based on the results of the optimal solution mutation. If the mutation point Compared to x b (j) has a lower fitness value, then use Replace and amplify the probability of mutation, increase the individual movement distance, if the mutation point Compared to x b (j) does not have a lower fitness value, then it remains unchanged and the mutation probability is reduced. The mutation probability update formula is as follows:

[0128]

[0129] In formula (20), μ is an expansion parameter and μ>1, λ is a contraction parameter and 0<λ<1.

[0130] Based on the above strategies for improving DBO, the present invention draws out the entire process framework of QOLDBO. Figure 4 As shown, it is not difficult to see that the complexity of the QQLDBO algorithm is mainly composed of the following three processes: initialization process, updating Q-Table to select the best behavior mode, Gaussian mutation of the optimal solution, and comparison of fitness values. The complexity of the initialization process is O(M×N), where M represents the number of iterations and N represents the number of populations. The complexity of selecting the best solution when updating Q-Table is O(M×(N×D+1)), where D represents the dimension. The complexity of performing Gaussian mutation of each dimension for the optimal solution and comparing the fitness values ​​is O(M×N×D). Therefore, the complexity of QQLDBO is O(M×(N×D+1)).

[0131] Simulation experiment and analysis of this embodiment:

[0132] 1. Algorithm performance test

[0133] In this invention, the CEC2017 test function set is used to evaluate the performance of the proposed QLDBO. The 29 functions in CEC2017 can be divided into the following three categories: unimodal functions (f1, f3), multimodal functions (f4-f10), hybrid functions (f11-f20) and combined functions (f21-f29). In addition, since the algorithm has fewer variants, the present invention compares the algorithm with other classic metaheuristic algorithms: DBO, SO

[68] ,GWO,SSA,PSO,HHO,WOA,FLA

[69] For comparison, the population size, dimension, and iteration times of the CEC2017 test function were set to 50, 30, and 5000. To ensure the fairness of the experimental results, the average result of running each test function 30 times can represent the final result.

[0134] (1) Solution accuracy

[0135] The standard test function is not very convincing because it was adopted too early. As the optimization problem becomes more and more complex, the present invention will use the newer CEC2017 test suite to verify the ability of the QOLDBO algorithm to solve the current complex optimization problems, compare and rank the mean of QOLDBO and other metaheuristic algorithms, and more intuitively see the solution accuracy and performance of each algorithm. It can be seen from Table 1 that in the first three single-modal functions, except that the algorithm proposed by f3 is not ranked first, the other effects are the best compared with other algorithms, and the original DBO algorithm has been improved by more than 8 orders of magnitude, which proves the effectiveness of the QOLDBO algorithm in solving single-modal function problems. In multi-modal functions, the QOLDBO algorithm performs well in most cases, but ranks last in f8, but the gap is not large. In addition, the algorithm performs outstandingly in mixed and combined functions, and only f17 and f18 are second only to the PSO algorithm, which shows the superiority of the QOLDBO algorithm in solving mixed and combined functions.

[0136] After comprehensive consideration, the QOLDBO algorithm ranked first overall in the CEC2017 test suite. This is due to the role of each strategy at different stages. Through the Q-learning-based mechanism, each behavior pattern selection is positive, which helps to improve the accuracy of the algorithm. In the global development, the foraging behavior of tuna schools is incorporated to help expand the global search space. After the iteration, the optimal solution is adaptively mutated to explore more possible searches, etc. The successful implementation of these strategies proves that the QOLDBO algorithm can achieve excellent results in solving the latest optimization problems.

[0137] Table 1 Comparison of solution accuracy of different algorithms on CEC2017 (30 dimensions)

[0138]

[0139]

[0140]

[0141] (2) Stability comparison

[0142] The present invention uses box plots to verify the stability of the QOLDBO algorithm by using the CEC2017 test suite and compares it with 6 other classic intelligent optimization algorithms. Each function was run independently 30 times, and the red tick mark on each box represents the median of the 30 run results, while the "+" symbol is used to draw outliers. For unimodal functions, except for QOLDBO and HHO, which rank first and second, other algorithms have many outliers. For f4, f5, f6, and f9 in multimodal functions, the QOLDBO algorithm has higher convergence accuracy and stability. On f8, GWO and PSO have better effects. For hybrid functions, QOLDBO performs better than other compared algorithms and has no outliers. However, on f16 and f20, the function results have a large span and lack stability, and the QOLDBO algorithm needs further improvement. Finally, for combined functions, it should be noted that the algorithm encounters more outliers on f22, which may be because the quasi-inverse solution obtained from quantum computing is not stable enough. However, it still performs well on other functions. It can be seen from the above experimental results that the QOLDBO algorithm has higher convergence accuracy and better algorithm stability.

[0143] (3) Comparison of effectiveness of relevant strategies

[0144] The present invention analyzes the effectiveness of the proposed strategy on the CEC2017 test suite. Each function runs independently 30 times, and the average value is used to analyze the performance of the QOLDBO algorithm. The best results are marked in bold. QOLDBO1 does not use the TPOL strategy in population initialization, and will adopt the traditional population random deployment setting. In QOLDBO2, the proposed Q-learning strategy is not used to select the behavior mode selection in the dung beetle rolling ball search phase according to the value in the Q-table, but a random selection is made according to the original DBO algorithm. QOLDBO3 does not use the proposed variable spiral local neighborhood search method, but continues to keep the random vector in the traditional algorithm as the weight. In QOLDBO4, the optimal solution of each iteration is no longer dimensionally adaptive Gaussian mutation, reducing more searches of multiple pairs of solution spaces. The experimental results are shown in Table 2.

[0145] As can be seen from Table 2, in the QOLDBO1 algorithm, if the population initialized by quasi-reverse learning based on quantum computing is removed, the algorithm performance decreases significantly, which shows that the proposed TPOL strategy plays an important role in the performance of the QOLDBO algorithm, while the variant QOLDBO2 is slightly better than QOLDBO in the convergence accuracy of functions f10 and f22. It is possible that reinforcement learning only performs learning feedback in the rolling ball stage, which contributes little to the improvement of the overall performance of the algorithm. In the variant QOLDBOO3, the algorithm performs poorly under most test functions, but the algorithm effect is similar or equal to that of QOLDBO in combination functions such as f24, indicating that the variable spiral local neighborhood search strategy is not very helpful for multi-combination optimization problems. Finally, in the variant QOLDBO4, the convergence accuracy is comparable, indicating that the method of using Gaussian mutation for the optimal solution has not achieved significant improvement. The above experimental results show that different strategies have played different roles in improving the performance of the algorithm, but if these four strategies are adopted at the same time, not only the solution accuracy is improved, but also the stability of the algorithm is improved, proving that the improved algorithm is authentic and effective.

[0146] Table 2 Comparison of solution accuracy of QOLDBO and other variants

[0147]

[0148]

[0149] (4) Variable Dimension Comparison

[0150] The present invention compares the test results of different dimensions on the CEC2017 test suite. By comparing the mean values ​​in the cases of 10, 30 and 50 dimensions, the best result of each test function is also indicated in bold, so as to more intuitively see which algorithm has stronger adaptability and more advantages. The data results are shown in Table 3. It can be seen from the table that the mean value of the QOLDBO algorithm increases with the increasing dimension. In the single-mode f3 and multi-mode f7 functions, PSO has more advantages than QOLDBO when the dimension is 10. In the external f10, f17 and f18 test functions, QOLDBO performs better in 10 and 50 dimensions than in 30 dimensions, indicating that the algorithm has strong adaptability in mixed functions. Finally, in the combination function, the 50-dimensional SSA in f19 and the 50-dimensional PSO in f22 are better than QOLDBO, indicating that the algorithm needs to be improved and optimized in the combination function. In summary, in different dimensions, the QOLDBO algorithm performs well in most cases, can adapt to different test conditions, and has strong performance.

[0151] Table 3 Comparison of results of different algorithms based on different dimensions

[0152]

[0153]

[0154] 2. QOLDBO coverage optimization in 3D real terrain

[0155] Through the display and analysis of experimental results, it can be seen that the QOLDBO algorithm has good performance, but the optimization algorithm must serve to solve real problems after all, so it is not enough to perform well in the test function. Many researchers will use the improved algorithm to test the performance of the algorithm by solving the network coverage optimization problem. For this reason, the present invention decided to conduct coverage optimization experiments on three-dimensional real terrain to fully verify the real effectiveness of the algorithm.

[0156] The present invention combines the QOLDBO algorithm with the WSN coverage problem on 3D real terrain. The WSN coverage method on 3D real terrain is actually somewhat similar to the previous researchers' research method on WSN coverage on curved surfaces. The 3D curved surface WSN coverage method is to project the three-dimensional terrain, and finally calculate the coverage effect on the two-dimensional plane. This method is still very limited in its application on 3D real terrain, because the height of the real terrain is different, and there are blocked signals that cause the coverage expectation to be too high. Therefore, the present invention will save the specific information of the real terrain in a matrix, so when the two-dimensional coordinate information of a node is known, the corresponding third-dimensional coordinate can be found.

[0157] In order to verify that the QOLDBO algorithm proposed in the previous chapter is capable of solving the WSN coverage optimization problem in 3D real terrain, this paper will combine DBO, MPA, RLMMPA

[70] The deployment results of the algorithms EGWO, SSA, and HHO are compared and analyzed. In addition, the present invention also uses GlobalMapper and Tuxin Earth to obtain the specific location information coordinates of a real terrain. The 3D view of the terrain is as follows Figure 5 shown.

[0158] Depend on Figure 5 It can be seen that the projected area of ​​the monitoring area of ​​the terrain is 120×40m 2 , the entire monitored area is divided into 4800 grids, each point has a corresponding elevation value. The present invention plans to deploy 60-100 sensor nodes in the monitoring area. The specific experimental simulation parameters are shown in Table 4.

[0159] Table 4 Experimental simulation parameters required to be set

[0160]

[0161] Table 5 Coverage results of different algorithms

[0162]

[0163] Table 5 shows the average coverage of 7 different algorithms after running independently for 20 times. It can be seen from the table that the QOLDBO algorithm has the best coverage under different numbers of nodes. When 100 nodes are deployed, the coverage rate reaches 92.36%. The DBO and HHO algorithms are not very effective and may need to be improved or the algorithm itself is not suitable for 3D real terrain coverage. When the number of nodes reaches 80 or more, the network coverage rate begins to slow down, indicating that when about 80 nodes are deployed, an optimal solution for all aspects will be reached, and the monitoring area can be covered with fewer nodes and less redundant area.

Claims

1. A method for optimizing real terrain coverage of wireless sensor networks, characterized in that: The following steps are involved: S1. Establish a three-dimensional coverage perception model; the specific method is: The line-of-sight detection method is used to determine whether the communication is blocked by calculating the straight-line distance between two sensor nodes and whether there is a direct line of sight between them. Let there be N sensor nodes, and the node set is represented as S = {s1, s2, s3, ..., s n }; The set of monitoring area nodes M = {m1, m2, m3, ..., m n },(x i ,y i , z i ) and (x j ,y j , z j ) respectively correspond to s in the set i 、m j The three-dimensional space coordinates of the sensor node and the monitoring area node are as follows: The line-of-sight detection is integrated with the conceptual perception model, and the coverage model perceived under the 3D real terrain is adopted. The specific expression of perception probability is as follows: The monitored area is divided into L·W grids of equal area, and the monitoring node m is located at the center of the grid. The combined perception probability of all monitoring points is calculated, and the cumulative sum is the coverage area, and the coverage rate C r It is expressed as follows: S2, dung beetle optimization based on quantum state reverse learning and Q-learning; It includes the following sub-steps: S2.1 Quasi-reverse learning based on quantum computing; (1) Quasi-reverse learning Define X i =[x1, x2, ..., x pop ]A solution in N-dimensional space, when x i ∈[lb i ,ub i ], the solution of quasi-reverse learning is defined as follows: (2) Update quantum state By updating the quantum state process in quantum computing, the influence of the current solution on the quasi-reverse solution is continuously changed to obtain more quasi-reverse individuals with mass. In quantum space, the quantum bit is the smallest unit. The quantum bit has two basic states |0> and |1>. The state of the quantum bit in quantum computing can be superimposed to form a basic storage unit, which is described as follows: represents the superposition state of the quantum, while ω and v represent the probability in this state. The quantum bit is also represented by sine and cosine matrices: The rotation gate RY(θ) in the quantum logic gate will be used to update the quantum state. The RY gate is generated by the Pauli-Y matrix as a generator, which is specifically expressed as follows: Where 0<θ<π / 2, let is [Γ Φ] T ,Γ,Φ range from 0 to 1, which is regarded as the influencing factor of the current solution on the quasi-reverse solution. The revolving gate in quantum computing continuously updates the quantum state to change the influence of the current solution on the quasi-reverse solution, and allows the two influencing factors to play a role in the population alternately, increasing the randomness of the reverse solution, thereby selecting a better initial population. The specific expression is as follows: (3) Specific steps The main steps of QOL optimization algorithm in dung beetle are as follows: (a) Set T o as the solution for the current iteration; (b) Calculate the influence factor of the current solution on the reverse solution through equation (8); (c) Calculate T by formula (9) o The quasi-reverse solution T qo ; (d) Evaluate the quasi-reverse solution T for each population qo The fitness value of (e) Comparison and The fitness value of , select the better one as the final population; S2.2 Behavior mode selection based on Q-learning; Each dung beetle is an individual, s represents the current state of each dung beetle, and a represents the transition of each individual from one state to another. The dung beetle adaptively switches from one state to another according to the Q-Table. The values ​​contained in the Q-Table describe the performance of each individual in the iteration. The Q-Table is used to help the algorithm control each individual to select the appropriate mode. The mathematical formula of the learning rate α is as follows: α ini and α fin Represent the initial value and final value of the learning rate, t and T are the current number of iterations and the maximum number of iterations; The value in the Q-Table of each individual in the first iteration is 0. After making a selection in the current state, if the fitness value becomes smaller, it will be rewarded, otherwise it will be punished, which helps to select the appropriate behavior mode at different stages. With the help of Q-learning, the interaction between each dung beetle and the operation of the two behavior modes is divided into four steps: (1) Initialize the Q-Table to a 2×2 table; (2) According to the value in the current state Q-Table, obtain the optimal behavior mode of the current iteration to be executed by the dung beetle in the rolling ball stage; (3) Execute the selected operation and calculate the fitness value, and give a reward. The specific reward calculation method is as follows: (4) Update Q-Table according to formula (12); S2.3 variable spiral local neighborhood search; The variable spiral search strategy is adopted to improve the spiral operator β. The specific mathematical expression is as follows: β=e bl ·cos(2πb) (14) In formula (14), b is a random number uniformly distributed from 0 to 1, and l is controlled by the cosine function with the number of iterations, so that the position update range of the dung beetle in this stage changes from large to small. In the early stage, good individuals are searched in a larger local neighborhood, which enhances the global search ability of the algorithm, reduces invalid searches in the later stage of iteration, and improves the convergence efficiency of the algorithm. After the improvement, the position updates of the dung beetle in the breeding and stealing stages are as follows: S2.4 Optimal solution dimension-by-dimension adaptive Gaussian mutation; After each iteration, adaptive Gaussian mutation is applied to the best individual in each dimension, and a greedy strategy is used to retain the better individuals to achieve a balanced exploration and development effect. The specific mutation method is as follows: In formula (19), x b (j) represents the jth dimension of the optimal individual, γ is the mutation probability, and randn indicates that it obeys Gaussian normal distribution; The optimal solution will be adaptively adjusted based on the result of the mutation. Compared to x b (j) has a low fitness value, then use Replace and amplify the probability of mutation, increase the individual movement distance, if the mutation point Compared to x b (j) does not have a low fitness value, then it remains unchanged and the mutation probability is reduced. The mutation probability update formula is as follows: In formula (20), μ is an expansion parameter and μ>1, λ is a contraction parameter and 0<λ<1.