Neural Network Parameter Optimization Method for Sonar Reflection Signal Judgment

Abstract

The present invention relates to a neural network parameter optimization method for sonar reflection signal judgment. The method requires obtaining a sonar reflection signal, inputting the signal into a trained signal judgment model, and outputting a judgment result. The method requires large-scale parameter optimization of the neural network signal judgment model, with the lowest proportion of non-zero weights and the smallest model prediction error as the objective function, and ultimately obtaining the optimal weight parameters of the neural network to achieve accurate signal judgment. The specific optimization process is: (1) randomly setting the initial population; (2) calculating the linear transfer matrix; (3) dividing the decision variables into two categories: convergence variables and diversity variables; (4) generating a first new population; (5) retaining individuals in the convergence optimization population; (6) generating a second new population; (7) screening the next generation population, using the next generation population as the new current population, and returning to (4). Compared with the existing technology, the present invention has the advantages of fast convergence and high judgment accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of sonar signal judgment, and in particular to a neural network parameter optimization method for sonar reflection signal judgment. Background Art

[0002] When exploring the ocean, a sonar detector usually emits a signal and then recovers the bounced signal. By analyzing the bounced signal, the nature of the detected object can be quickly determined, such as whether the object is rock or metal.

[0003] Prior art often employs neural networks for judgment. When training a neural network, two objective functions can be set: minimizing the proportion of non-zero weights and minimizing the neural network's error rate on a dataset. Furthermore, the dimensions of the decision variables are very high, meaning that training such a neural network is a multi-objective optimization problem. When faced with large-scale optimization problems, where the dimensions of the decision variables are 500 or greater, most multi-objective optimization algorithms fail to achieve the desired results or converge very slowly. On the one hand, the search space for feasible solutions grows exponentially with the number of decision variables, making the search for the optimal solution prohibitively time-consuming. On the other hand, high-dimensional spaces often contain numerous local optimal solutions, which severely impacts the accuracy and stability of the algorithm.

[0004] In summary, the existing multi-objective optimization method for training neural networks for reflection signal judgment, when faced with a large number of decision variables, has a high-dimensional target space that significantly slows down the solution process, converges slowly, and affects the final diversity, which in turn affects the accuracy of rebound signal judgment. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a neural network parameter optimization method for sonar reflection signal judgment. By analyzing the decision variables, the large-scale decision variables are divided into two categories: convergence variables and diversity variables.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A neural network parameter optimization method for sonar reflection signal judgment includes the following steps:

[0008] The sonar reflection signal is obtained and input into a trained signal judgment model to output the judgment result. The signal judgment model is built based on a neural network, and the neural network is subjected to large-scale parameter optimization to obtain the optimal weight parameters of the neural network to minimize the classification error of the neural network. The specific objective function is to minimize the proportion of non-zero weights and the prediction error of the neural network on the data set:

[0009]

[0010] min f2(x)=ErrorRate(x)

[0011] Where x is the decision vector, representing the neural network weight parameters, ||x||0 represents the number of non-zero weight parameters, D represents the dimension of the decision vector, ErrorRate(x) represents the prediction error of the neural network, the objective function f1(·) represents the proportion of non-zero weights, and the objective function f2(·) represents the error rate of the neural network on the dataset;

[0012] The specific optimization process is:

[0013] (1) The weight parameters of the neural network are used as the population, and the initial population is randomly set. The population size of the initial population is set to N, the target number is set to M, the dimension of the decision vector is set to D, and the initial population is used as the current population;

[0014] (2) Construct the decision variable matrix X and target value matrix Y from the current population, and calculate the linear transfer matrix T;

[0015] (3) Using the decision variable analysis method, the decision variables in the decision vector are divided into convergence variables D S and diversity variable D G Two categories;

[0016] (4) Convergence variable D S , a nonlinear inverse mapping model is used to map from the target space to the decision space to generate the first new population;

[0017] (5) Calculate the target value matrix corresponding to the first new population, compare the corresponding target value matrix with the target value matrix Y constructed in the current population, screen according to the comparison results, and retain the individuals that meet the conditions into the converged optimization population;

[0018] (6) For the diversity variable D G , a nonlinear inverse mapping model is used to map from the target space to the decision space to generate the second new population;

[0019] (7) Calculate the corresponding target value matrix for the second new population, mix the target value matrix with the converged optimized population, perform non-dominated sorting, and stratify it according to the sorting. According to the stratification results, the next generation population is selected and the next generation population is used as the new current population. Return to (4) until the number of function evaluations reaches the set value.

[0020] Furthermore, the nonlinear inverse mapping model is established based on the activation function of the neural network introduced on the basis of the linear transfer matrix T.

[0021] Furthermore, in (4), the process of generating the first new population is:

[0022] Generate a new decision value matrix X1′, which is: Among them, Y is the target value matrix of the current population, represents the vector product, T is the corresponding linear transfer matrix, R is the random number matrix between 0 and 1, f σ (·) is the activation function, Indicates that the target values ​​of the population are all reduced by a random multiple, Indicates mapping back to the target space, Indicates that a new decision value matrix is ​​obtained through a nonlinear activation function;

[0023] Assign the current population X to the new population X 1new ;

[0024] The convergence variable D of the new decision value matrix X1′ S Assign value to new population X 1new , the new population X after assignment 1new The first new population.

[0025] Furthermore, in (6), the process of generating the second new population is:

[0026] Generate a new decision value matrix X2′, which is: Among them, Y is the target value matrix of the current population, represents the vector product, T is the corresponding linear transfer matrix, R is the random number matrix between 0 and 1, f σ (·) is the activation function, Indicates that the target values ​​of the population are all reduced by a random multiple, Indicates mapping back to the target space, Indicates that a new decision value matrix is ​​obtained through a nonlinear activation function;

[0027] Assign the current population X to the new population X 2new ;

[0028] The diversity variable D of the new decision value matrix X2′ G Assign value to new population X 2new , the new population X after assignment 2new The second new population.

[0029] Furthermore, the activation function is a tanh(·) function.

[0030] Furthermore, the target value matrix Y is N×M, the dimension of X is N×D, and the dimension of T is M×D, where N is the population size of the initial population, M is the target size, and D is the decision vector dimension.

[0031] Furthermore, the dimension of the random number matrix R is the same as that of the target value matrix Y, which is a random number matrix between 0 and 1.

[0032] Furthermore, the expression of the linear transfer matrix T is:

[0033] T=Y -1 X

[0034] Among them, T is the linear transfer matrix, Y is the target value matrix of the current population, X is the decision variable matrix, and Y -1 represents the generalized inverse of Y.

[0035] Furthermore, the screening is performed according to the comparison results as follows:

[0036] Compare the target value vector corresponding to the kth individual of the first new population with the target value vector corresponding to the kth individual of the current population. If each value of the former is not greater than the latter, then retain the kth individual of the first new population and enter the convergent optimization population. Otherwise, retain the kth individual of the current population and enter the convergent optimization population.

[0037] Furthermore, based on the stratification results, the next generation population is selected as follows:

[0038] The individuals of each layer are added to the next generation population in the order of the number of layers from small to large, until for the k-1th layer, |L1∪L2∪…∪L k |>N, where L1 represents the number of the first layer and N is the population size of the initial population. At this point, individuals in the k+1th layer are stopped from being added to the next generation population. Individuals in the kth layer of the next generation population are selected based on fitness until the size of the next generation population reaches the population size of the initial population.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The present invention introduces the activation function in the neural network and establishes a nonlinear inverse mapping model from the target space to the decision space based on the linear transfer matrix; then, by generating a new target vector in the target space, it is mapped back to the decision space using the nonlinear inverse mapping model to generate a new individual, and through decision variable analysis, large-scale decision variables are divided into two categories: convergence variables and diversity variables, thereby improving the efficiency of evolutionary search and better balancing population diversity and convergence, ultimately enhancing the method's ability to solve large-scale high-dimensional multi-objective optimization problems, effectively avoiding blind large-scale searches, and improving the accuracy of judging rebound signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a flow chart of the present invention;

[0042] Figure 2 Schematic diagram of the generation of offspring individuals for the linear and nonlinear target space inverse mapping of the present invention;

[0043] Figure 3 Schematic diagram of the target space of the present invention mapped back to the decision space. DETAILED DESCRIPTION

[0044] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0045] The present invention proposes a neural network parameter optimization method for sonar reflection signal judgment, comprising the following steps:

[0046] The sonar reflection signal is obtained and input into a trained signal judgment model to output the judgment result. The signal judgment model is built based on a neural network and undergoes large-scale parameter optimization of the neural network to obtain the optimal weight parameters of the neural network to minimize the classification error of the neural network. The specific objective function is to minimize the proportion of non-zero weights and minimize the prediction error of the neural network on the data set:

[0047]

[0048] minf2(x)=ErrorRate(x) (1)

[0049] Where x is the decision vector, representing the neural network weight parameters, ||x||0 represents the number of non-zero weight parameters, D represents the total number of weight parameters (the dimension of the decision vector), and ErrorRate(x) represents the prediction error of the neural network. The objective function f1(·) represents the proportion of non-zero weights, and the objective function f2(·) represents the error rate of the neural network on the dataset.

[0050] In this embodiment, a three-layer feedforward neural network with 20 hidden neurons is used as the model. The neural network input dimension is 60 and the output dimension is 1. The name of the test dataset is:

[0051] Connectionist_bench_Sonar(http: / / archive.ics.uci.edu / ml / index.php);

[0052] The data in this sonar dataset has two categories: rock and metal. The decision variable dimension D is 1241, and the target dimension is 2, where -1≤x≤1.

[0053] The flowchart of the process of solving the above objective function is as follows Figure 1 As shown in Figure 2, the training process is as follows:

[0054] Step 1: Randomly generate a parent population. In some embodiments, the population size N is 100, the target size M is 2, and the decision vector dimension D is 1241.

[0055] Step 2: Calculate the linear transfer matrix T using the following formula:

[0056] T=Y -1 X

[0057] Among them, Y is the target value matrix of the current population, X is the decision variable matrix, and Y -1 Represents the generalized inverse of Y. The dimensions of matrix Y are N×M, the dimensions of X are N×D, and the dimensions of T are M×D, where M and D are the target and decision vector dimensions respectively, and N is the population size.

[0058] Once T is determined, the target vector can be mapped back to the decision space at any time, as Figure 3 shown.

[0059] Step 3: Use the LMEA decision variable analysis method to divide large-scale decision variables into convergence variables D S and diversity variable D G Two categories. D S With D G are all sequences of decision variables (assuming the decision variable dimension D = 10, they can be

[1357] , [2468910] respectively).

[0060] Step 4: Converge the relevant variables D one by one S Generate the first new population using the following formula:

[0061]

[0062] X new =X

[0063] X new (:,D S )=X′(:,D S ) (4)

[0064] The process of generating the first new population is:

[0065] Assign the current population X to the new population X 1new ;

[0066] The convergence variable D of the new decision value matrix X1′ S Assign value to new population X 1new , the new population X after assignment 1new The first new population.

[0067] Where Y is the target value matrix of the current population, X is the decision variable matrix of the current population, and T is the linear transfer matrix. The dimensions of matrix Y are N×M, the dimensions of X are N×D, and the dimensions of T are M×D. σ (·) is the activation function, which can be tanh(·). The dimension of R is the same as Y, which is a random number matrix between 0 and 1. Represents the vector product. X new represents the new population, D S represents the convergence-related decision variable.

[0068] The dimensions of matrix Y are N×M, the dimensions of X are N×D, the dimensions of T are M×D, and the dimensions of R are the same as Y, which is a matrix of random numbers between 0 and 1. Indicates that the target values ​​of the population are all reduced by a random multiple, Indicates mapping back to the target space, Indicates that a new decision value matrix is ​​obtained through a nonlinear activation function. new For new population.

[0069] Step 5: Through the new population X new Calculate the corresponding target value matrix Y new Compare the target value Y with the corresponding individual in the original population, and keep the individual with the smaller target value. new kth individual The target value vector and the kth individual x of the original population X k The target value vector y k For comparison, if Every value of is not greater than y k , then retain Enter the converged optimization population X c , otherwise keep x k Enter the converged optimization population Xc, 1≤k≤N. Obtain the converged optimization population X by one-to-one comparison c Calculate the corresponding target value matrix Y c , matrix X c Dimensions N×D, Y c The dimension is N×M.

[0070] Step 6: For each diversity-related variable D G Generate the second new population using formula (5):

[0071]

[0072] X new =X

[0073] X new(:,D G )=X′(:,D G ) (5)

[0074] The process of generating the second new population is:

[0075] Generate a new decision value matrix X2′, which is: Among them, Y is the target value matrix of the current population, represents the vector product, T is the corresponding linear transfer matrix, R is the random number matrix between 0 and 1, f σ (·) is the activation function, Indicates that the target values ​​of the population are all reduced by a random multiple, Indicates mapping back to the target space, Indicates that a new decision value matrix is ​​obtained through a nonlinear activation function;

[0076] Assign the current population X to the new population X 2new ;

[0077] The diversity variable D of the new decision value matrix X2′ G Assign value to new population X 2new , the new population X after assignment 2new The second new population.

[0078] Step 7: Pass the new population X new Calculate the corresponding target value matrix Y new , and the convergent optimization population X c After mixing, perform non-dominated sorting and stratify them according to the sorting. Individuals with smaller layers enter the next generation population in turn, and continue until the k-1th layer so that |L1∪L2∪…∪L k |>N, (L1 represents the number of the first layer, and so on), at this time, stop adding individuals in the k+1th layer to the next generation population, and use I for individuals in the kth layer ∈+ The indicator selection is continued until the next generation population size reaches N.

[0079] If the population size N is 100, the second new population will be mixed with the converged optimized population and non-dominated sorting will be performed, and the layers will be layered according to the sorting. If the first layer L1 has 10 individuals, the second layer L2 has 30 individuals, the third layer L3 has 50 individuals, and the fourth layer L4 has 20 individuals. Obviously, |L1+L2+L3|<100, |L1+L2+L3+L4|>100, so the individuals in L1, L2, and L3 will all enter the next generation, and 10 individuals in L4 need to be selected to enter the next generation. The selection method is I ∈+ Indicators, namely:

[0080] L kAny two individuals s1, s2 in L, that is, s1, s2∈L k , will calculate the Chebyshev distance matrix,

[0081] I(s1,s2)=max(f j (s1)-f j (s2), 1≤j≤M

[0082] Where M is the number of targets;

[0083] For L k For each individual s1 in , its fitness is calculated:

[0084]

[0085] s2∈L k \s1 means L k The rest of the individuals except s1, ∑ represents the summation symbol, κ is the coefficient, usually 0.05;

[0086] e (.) is an exponential function. The basic logic is that individuals with high fitness will enter the next generation.

[0087] Step 8: Repeat steps 4-7 until the number of function evaluations reaches the set value. After reaching the set value, the decision vector is determined, the neural network weight parameters are determined, and the signal judgment model under the weight parameters is obtained to determine the category of the reflection signal.

[0088] In steps 4 and 6, a new target vector is generated in the target space and mapped back to the decision space using a nonlinear inverse mapping model to generate a new individual, as shown in formula (3):

[0089]

[0090] Where: r=(r1, ..., r j ,…,r M ) is a vector of random numbers between 0 and 1, represents the vector product, i.e. Methods for generating offspring individuals through linear and nonlinear mapping are as follows Figure 2 shown.

[0091] This paper utilizes activation functions in neural networks to establish a nonlinear inverse mapping model from the objective space to the decision space in multi-objective optimization. It also designs a method for generating high-quality offspring populations using this nonlinear inverse mapping. This method improves the functional expressiveness of the mapping model while ensuring the rapid construction of a stable and usable mapping relationship, thereby more efficiently generating high-quality offspring.

[0092] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A neural network parameter optimization method for sonar reflection signal judgment, characterized in that: The following steps are involved: The sonar reflection signal is obtained and input into a trained signal judgment model to output the judgment result. The signal judgment model is built based on a neural network, and the neural network is subjected to large-scale parameter optimization to obtain the optimal weight parameters of the neural network to minimize the classification error of the neural network. The specific objective function is to minimize the proportion of non-zero weights and the prediction error of the neural network on the data set: in, is the decision vector, which represents the neural network weight parameters, represents the number of non-zero weight parameters, represents the dimension of the decision vector, Represents the prediction error of the neural network, the objective function Represents the proportion of non-zero weights, the objective function Represents the error rate of the neural network on the dataset; The specific optimization process is: (1) Take the weight parameters of the neural network as the population, randomly set the initial population, and set the population size of the initial population to N , the target number is M , the decision vector dimension is D , take the initial population as the current population; (2) Construct the decision variable matrix X and target value matrix Y from the current population, and calculate the linear transfer matrix T; (3) Using the decision variable analysis method, the decision variables in the decision vector are divided into convergence variables and diversity variables Two categories; (4) Convergence variables , a nonlinear inverse mapping model is used to map from the target space to the decision space to generate the first new population; (5) Calculate the corresponding target value matrix for the first new population, compare the corresponding target value matrix with the target value matrix Y constructed in the current population, screen according to the comparison results, and retain the individuals that meet the conditions into the converged optimization population; (6) Diversity variables , a nonlinear inverse mapping model is used to map from the target space to the decision space to generate the second new population; (7) Calculate the corresponding target value matrix for the second new population, mix the target value matrix with the converged optimized population, perform non-dominated sorting, and stratify it according to the sorting. According to the stratification results, the next generation population is selected and the next generation population is used as the new current population. Return to (4) until the number of function evaluations reaches the set value; In (4), the process of generating the first new population is: Generate a new decision value matrix , decision value matrix for: ,in, is the target value matrix of the current population, represents the vector product, T is the corresponding linear transfer matrix, is a random number matrix between 0 and 1, is the activation function, Indicates that the target values ​​of the population are all reduced by a random multiple, Indicates mapping back to the target space, Indicates that a new decision value matrix is ​​obtained through a nonlinear activation function; The current population Assign value to new population ; The new decision value matrix Convergence variables Assign value to new population , the new population after assignment It is the first new population; In (6), the process of generating the second new population is: Generate a new decision value matrix , decision value matrix for: ,in, is the target value matrix of the current population, represents the vector product, T is the corresponding linear transfer matrix, is a random number matrix between 0 and 1, is the activation function, Indicates that the target values ​​of the population are all reduced by a random multiple, Indicates mapping back to the target space, Indicates that a new decision value matrix is ​​obtained through a nonlinear activation function; The current population Assign value to new population ; The new decision value matrix Diversity variables Assign value to new population , the new population after assignment The second new population.

2. The neural network parameter optimization method for sonar reflection signal judgment according to claim 1, characterized in that: The nonlinear inverse mapping model is established based on the introduction of an activation function of a neural network on the basis of a linear transfer matrix T.

3. The neural network parameter optimization method for sonar reflection signal judgment according to claim 1, characterized in that: The activation function is tanh function.

4. The neural network parameter optimization method for sonar reflection signal judgment according to claim 1, characterized in that: The target value matrix Y is N M , Dimension N D , The dimension is M D , N is the population size of the initial population, M is the target number, D is the decision vector dimension.

5. The neural network parameter optimization method for sonar reflection signal judgment according to claim 4 is characterized in that: The dimension of the random number matrix R is the same as that of the target value matrix Y, and it is a random number matrix between 0 and 1.

6. The neural network parameter optimization method for sonar reflection signal judgment according to claim 5, characterized in that: The expression of the linear transfer matrix T is: Where T is the linear transfer matrix, is the target value matrix of the current population, is the decision variable matrix, represents the generalized inverse of Y.

7. The neural network parameter optimization method for sonar reflection signal judgment according to claim 1, characterized in that: Filter by comparison results: Compare the target value vector corresponding to the kth individual of the first new population with the target value vector corresponding to the kth individual of the current population. If each value of the former is not greater than the latter, then retain the kth individual of the first new population and enter the convergent optimization population. Otherwise, retain the kth individual of the current population and enter the convergent optimization population.

8. The neural network parameter optimization method for sonar reflection signal judgment according to claim 1, characterized in that: According to the results of stratification, the next generation population is selected as follows: Add the individuals of each layer to the next generation population in the order of the number of layers from small to large, until the layer, satisfy| L 1 | > N ,in, L 1 Indicates the number of the first layer, N is the population size of the initial population, at this time, stop adding individuals from the k+1th layer to the next generation population, and for the next generation population Individuals in the layer are selected based on fitness until the size of the next generation population reaches the population size of the initial population.

Images