Real-time underwater acoustic positioning method based on IDE-BP neural network

Abstract

The present application relates to a kind of real-time underwater acoustic positioning method based on IDE-BP neural network, comprising the following steps: by constructing the acoustic positioning model based on non-convex RSS, considering path loss and absorption loss, obtain training sample set;BP neural network is constructed, the initial network parameter of the BP neural network is obtained based on differential evolution algorithm, the training sample set is used to train the BP neural network, and the BP neural network trained is obtained;Underwater acoustic data is obtained, and underwater acoustic positioning is carried out using the BP neural network trained.The present application solves the problem of multi-target underwater acoustic positioning based on RSS by using IDE algorithm combined with BP neural network compared with prior art, with the advantages of high precision, fast.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of positioning, and particularly to a real-time underwater acoustic positioning method based on an IDE-BP neural network. BACKGROUND

[0002] Wireless sensor networks (WSNs) are widely used in military and civilian applications, such as traffic control and environmental monitoring. Multi-target positioning in WSNs is essential because the collected data is usually only meaningful when geographically referenced. However, in dynamic underwater wireless sensor networks (UWSNs), accurate multi-target positioning is difficult. Traditional positioning techniques based on radio frequency, GPS, and optical signals are not suitable for underwater environments. In the present application, surface buoy nodes and underwater anchor nodes with known positions are randomly deployed in a UWSN, and acoustic signals are used for communication between nodes. Shore-based information from these nodes is collected and transmitted to a base station center. Using this information, the position of unknown target nodes can be determined. However, the combination of path loss and absorption loss makes it difficult for the base station center to achieve real-time and accurate positioning.

[0003] There have been studies on positioning in UWSNs. These studies are mainly divided into ranging-based or non-ranging-based methods. Ranging-based methods use signal measurements such as received signal strength (RSS), received signal strength difference (RSSD), time of arrival (TOA), and angle of arrival (AOA) to estimate target positions. Conversely, non-ranging-based methods mainly use parameters such as hop count and sensor density and do not consider signal measurements, so the estimation accuracy is usually not as good as ranging-based methods. Therefore, ranging-based methods are usually used for positioning applications. In ranging-based methods, RSS and RSSD-based techniques are popular due to their simplicity and low cost, but most applications are for terrestrial sensor network positioning. Some scholars have proposed a matrix decomposition-based method that provides good positioning accuracy. A method based on least squares relative error is also considered and uses semi-definite relaxation to solve non-convex problems. A robust fault-tolerant positioning scheme is proposed for the case of faulty nodes in a non-Gaussian environment. Differential evolution with opposition and redirection (DEOR) is used to handle homogeneous and non-homogeneous environments.

[0004] In contrast, the use of RSS and RSSD measurements for localization in UWSNs has received much less research attention because it is more challenging due to the presence of path loss and acoustic signal losses such as geometric loss and absorption loss. These losses can significantly degrade the localization accuracy and make many RSS-based schemes ineffective. Therefore, researchers have explored various methods to mitigate their effects. They developed a robust, non-cooperative localization algorithm (RNLA) to improve accuracy using bisection. A min-max strategy was considered to limit the attenuation due to absorption in shallow water. A two-stage robust approach was proposed to jointly estimate the position and the transmit power. The approach combined the interior point method and majorization-minimization (IPM-MM) to search for the optimal solution using an initial IPM solution. However, a poor initial solution can degrade the localization performance. Therefore, a weighted least squares technique was also proposed to transform the original problem into a generalized trust region subproblem under known and unknown transmit power conditions. However, the approximations used can affect the localization performance.

[0005] The underwater acoustic channel is very dynamic, which makes the data link unreliable. Therefore, it is difficult to develop a robust RSS-based localization technique for underwater environments. The data rates of most methods are very low, for example, 10 kHz to 100 kHz. Higher rates require higher transmission frequencies, but this increases absorption loss. Path loss and absorption loss can adversely affect performance, and mitigating the effects of these losses on real-time UWSN localization performance is a challenging task. SUMMARY

[0006] The purpose of the present application is to overcome the defects of the prior art, and provide a real-time underwater acoustic positioning method based on IDE-BP neural network, which can provide robust positioning in dynamic underwater wireless sensor networks under the condition that path loss and absorption loss exist, and improve the convergence speed of the neural network.

[0007] The purpose of the present application can be achieved by the following technical solutions:

[0008] The present application provides a real-time underwater acoustic positioning method based on IDE-BP neural network, comprising the following steps:

[0009] Acoustic data is obtained, and a trained BP neural network is used for underwater multi-target positioning,

[0010] The trained BP neural network is obtained by the following steps:

[0011] acquire a training sample set by constructing an acoustic positioning model based on a non-convex RSS, considering path loss and absorption loss;

[0012] construct a BP neural network, acquire initial network parameters of the BP neural network based on a differential evolution algorithm, train the BP neural network using the training sample set, and acquire the trained BP neural network.

[0013] As a preferred technical solution, the acoustic positioning model comprises a plurality of anchor nodes and a plurality of unknown target nodes, and the positions and quantities of the anchor nodes match the positions and quantities of underwater acoustic sensors in a real scene.

[0014] As a preferred technical solution, the acquisition process of the training sample set comprises the following steps:

[0015] acquire, based on distances between each anchor node and each unknown target node in the acoustic positioning model, received signal strength values between each anchor node and each unknown target node after considering path loss, absorption bias, and measurement noise;

[0016] acquire target position estimation information based on the received signal strength values;

[0017] acquire the training sample set based on the received signal strength values and the target position estimation information.

[0018] As a preferred technical solution, the acquisition of the target position estimation information comprises the following steps:

[0019] acquire a probability density function of an observation vector of each unknown target node based on the received signal strength values between each anchor node and each unknown target node, and obtain a maximum likelihood estimation of the position of the unknown target node by maximizing the probability density function.

[0020] As a preferred technical solution, the received signal strength values are acquired by the following formula:

[0021]

[0022] wherein P i j is a signal strength value received at the anchor node i for the target node j, d0 is a reference distance, P0 is a transmission power when the distance is d0, γ is a path loss exponent, n i is measurement noise, represents absorption bias, is a distance between the target node j and the anchor node i.

[0023] Preferably, the target position estimation information is:

[0024]

[0025] wherein, is the maximum likelihood estimation of u j i j is the signal strength value received at the anchor node i for the target node j, P0 is the transmission power at the reference distance d0, and d0 is the reference distance, represents the absorption bias, and γ is the path loss exponent, is the distance between the target node j and the anchor node i, is the variance of the measurement noise.

[0026] Preferably, the process of obtaining the initial network parameters of the BP neural network based on the differential evolution algorithm comprises the following steps:

[0027] determining the solution space of the network parameters, generating L forward vectors and L backward vectors based on the backpropagation learning, and selecting the L candidates with the lowest fitness as the initial population;

[0028] generating a mutation population from the current candidate population based on an adaptive mutation strategy, obtaining a cross population with new individuals by mating the candidate population and the mutation population, adjusting the individuals in the cross population that are outside the solution space to be within the solution space, selecting the individuals with the minimum fitness from the adjusted cross population to form a new candidate population, repeating the step multiple times until a preset iteration number is reached or the difference between the candidate populations before and after iteration is less than a preset value, and obtaining the initial network parameters of the BP neural network.

[0029] Preferably, the initial network parameters include the weights and biases of the BP neural network.

[0030] Preferably, the BP neural network comprises an input layer, a hidden layer, and an output layer connected in sequence.

[0031] Preferably, the underwater acoustic data is received signal strength difference data.

[0032] Compared with the prior art, the present application has the following advantages:

[0033] ​(1) To provide robust positioning in dynamic underwater wireless sensor networks in the presence of both path loss and absorption loss: The method uses the IDE algorithm combined with the BP neural network to solve the multi-target underwater acoustic positioning problem based on RSS. By constructing an acoustic positioning model based on non-convex RSS, considering path loss and absorption loss, a training sample set is obtained. The initial network parameters of the BP neural network are obtained based on the differential evolution algorithm. The training sample set is used to train the BP neural network, and the trained BP neural network is obtained. Finally, the trained BP neural network model is used for positioning. By constructing an acoustic positioning model considering path loss and absorption loss, robust positioning is provided in dynamic underwater wireless sensor networks in the presence of both path loss and absorption loss.

[0034] (2) Fast convergence speed and good training effect of neural network: Based on the opposition-based learning used to obtain a robust population to search for the global optimal value, the adaptive mutation strategy is used to improve the global search and convergence. The improved differential evolution (IDE) algorithm is used to obtain the initial network parameters. The opposition-based learning (OL) theory is used to obtain a robust population to search for the global optimal value, and the adaptive mutation strategy is used to improve the global search and convergence. The convergence speed of the neural network is improved, the training effect is improved, and the problem of poor performance and slow convergence speed caused by the fact that the BP network parameters are usually obtained based on experience is overcome.

[0035] (3) High positioning accuracy and fast calculation speed: In terms of real-time UWSN positioning accuracy, the proposed method RRTL is compared with unconventional methods (i.e. DEOR) and conventional methods (i.e. WLS, RNLA, RSS-SDP and IPM-MM). The results show that the RRTL algorithm provides better positioning accuracy than other methods. In addition, it is comparable to WLS in terms of calculation time and much better than other methods, which is important for real-time implementation. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 Flowchart of the real-time underwater acoustic positioning method based on IDE-BP neural network in Example 1;

[0037] Figure 2 Structure diagram of an underwater wireless sensor network (UWSN) positioning system;

[0038] Figure 3 Absorption coefficient α f Relationship diagram with frequency f;

[0039] Figure 4 This is a schematic diagram of a BP network structure;

[0040] Figure 5 This refers to a method of random node deployment in acoustic localization models.

[0041] Figure 6 The RMSE and CRLB of the existing method are compared with the absorption coefficient α. f A diagram illustrating the relationship between the two.

[0042] Figure 7 For time γ, α f A schematic diagram showing the relationship between the changes in sl and N. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0044] Example 1

[0045] like Figure 1 This embodiment provides a real-time underwater acoustic localization method based on an IDE-BP neural network, which provides robust localization in a UWSN even when both path loss and absorption loss exist. The method includes the following steps:

[0046] Step S1: Construct an acoustic localization problem based on non-convex RSS: Obtain a training sample set by constructing an acoustic localization model based on non-convex RSS that considers path loss and absorption loss.

[0047] Step S2: Construct a BP neural network. Obtain the initial network parameters of the BP neural network based on the differential evolution algorithm, and train the BP neural network using the training sample set to obtain the trained BP neural network.

[0048] An improved differential evolution (IDE) algorithm is employed to obtain optimized parameters. A robust population is obtained based on opposition-based learning (OL) theory to search for the global optimum, and an adaptive mutation strategy is used to improve global search and convergence. An error backpropagation (BP) neural network is trained using the optimized parameters, and the trained network is used to predict the locations of multiple underwater acoustic targets in real time.

[0049] Step S3, obtaining underwater acoustic data, using the trained BP neural network to perform underwater acoustic positioning.

[0050] Considering the positioning accuracy and computation time, the proposed method RRTL is evaluated and compared with non-traditional methods (i.e., DEOR) and traditional methods (i.e., WLS, RNLA, RSS-SDP, and IPM-MM).

[0051] The S1 is specifically:

[0052] Consider a 3-dimensional (3-D) UWSN positioning system as shown in Figure 2 The system has N known anchor nodes and M unknown target nodes. Let a i = [a i1 , a i2 , a i3 ] T , i = 1, 2, …, N, N > 4, denote the anchor node coordinates, and U = [u1, u2, …, u M ] T denote the target node positions u j = [u j1 , u j2 , u j3 ] T , j = 1, 2, …, M, where (·) T denotes the transpose. The distance between target node j and anchor node i is

[0053]

[0054] where ||·|| denotes the l2 norm. Based on the log-normal shadowing, the RSS received at anchor node i for target node j can be modeled as:

[0055]

[0056] where d0 is the reference distance, P0 is the transmit power at distance d0, γ is the path loss exponent, which ranges from [1, 6], n i is the measurement noise with mean 0 and variance α f is the absorption coefficient, which can be obtained by the Thorp formula

[0057]

[0058] where f is the signal frequency. The relationship between α f and f is shown in Figure 3 . α f = 0 represents a terrestrial environment.

[0059] Let denotes the absorption bias, the RSS value between the ith anchor node and the jth target node can be expressed as:

[0060]

[0061] Observation vector The probability density function (PDF) of

[0062]

[0063] Maximizing this PDF gives the maximum likelihood estimate of u j as

[0064]

[0065] The S2 is specifically: RSS measurement value P i j and target position u j are combined to obtain a data set, which can be expressed as

[0066] T = {P i j , u j}, (7)

[0067] where the dimension of P i j depends on the number of anchor nodes N, and the BP network structure is as shown in Figure 4 It includes an input layer with M(N+3) neurons, a hidden layer, and an output layer with 3M neurons. The network parameters are the weights and biases of these layers. Let the number of neurons in the hidden layer be b, then the output of the cth neuron in the hidden layer, c = 1, 2,..., b, can be expressed as:

[0068]

[0069] where f(·) is the activation function, w ct and θ c are the weights and biases, o t is the output of the previous layer, and t = 1, 2,..., b.

[0070] For the jth target node, the network output is

[0071]

[0072]

[0073]

[0074] The training error used to adjust the weights and biases in (8) and (9) can be expressed as follows:

[0075]

[0076] Among them, M t For the training set size, For the j-th target location (u j1 ,u j2 ,u j3 The estimated value of ). The trained network can be represented as l(·), and then the trained network can be used with the new RSSD value. Estimate the positions of M new targets

[0077]

[0078] in,

[0079] 2) IDE Algorithm

[0080] Backpropagation (BP) network parameters are typically obtained empirically, which can result in parameters that are either very small or very large. This can lead to vanishing or exploding gradients, resulting in poor performance and slow convergence. Therefore, an IDE algorithm is proposed to obtain network parameters. It iteratively generates the population using initialization, mutation, crossover, boundary handling, and selection. Furthermore, online search (OL) combined with an adaptive strategy is used to improve global search and convergence performance.

[0081] a) Initialization: Let the solution space of the BP network parameters (weights and biases) be S = {(s1, s2, ..., s...} m ):q k ≤s k ≤r k L vectors, k = 1, 2, ..., m. It can be generated by (12)

[0082]

[0083] Where, q k r k and are the lower bound and upper bound respectively, k = 1, 2, ..., m, where m is the dimension, μ l It is a random number and μ l ~U(0,1). Then, based on OL, the inverse vector. It can be used to improve search ability and convergence.

[0084]

[0085] In a set including 2L individuals In the initial stage, the L individuals with the lowest fitness are selected as the initial population, denoted as

[0086] b) Mutation: In the gth generation, the current population P(g) is mutated to generate a mutated population P'(g) using the self-adaptive mutation strategy Considering the underwater acoustic environment, this avoids the disadvantages of using a single operator, such as slow convergence and convergence to local optima. The self-adaptive control parameter η = |χ mean / χ best is compared with a threshold value δ to determine the strategy used, where χ mean is the average of all individual objective function values, and χ best is the objective function value of the best individual. Each set of strategies contains two mutation operations, given by

[0087]

[0088]

[0089] and

[0090]

[0091]

[0092] where α ∈ [0, 1] is a scaling factor, κ best represents the best individual in P'(g), and o1, o2, o3, o4, o5 ∈ {1, 2,..., L}, o1 ≠ o2 ≠ o3 ≠ o4 ≠ o5. When η > δ, (14a) and (14b) are used to generate the mutated population, otherwise (15a) and (15b) are used. The selection from (14a) and (14b) or (15a) and (15b) is based on the parameter β ∈ [0, 1].

[0093] c) Crossover: The mating of the original population and the mutated population is completed with a crossover probability ψ c ∈ [0, 1] to generate a new population P(g + 1) with individuals

[0094]

[0095] where c l ~ U(0, 1).

[0096] d) Boundary handling: Individuals that exceed the boundaries of the search space S can lead to incorrect solutions to the problem. To avoid this, individuals outside the search space are adjusted using (17)

[0097] ​

[0098] in, Follows a uniform distribution on [0,1] and

[0099]

[0100]

[0101] Where k = 1, 2, ..., m, and and These are the lower and upper bounds for the g-th generation individual, respectively.

[0102] e) Selection: In the (g+1)th generation, from individuals and Choose the individual with the smallest fitness value from the combination:

[0103]

[0104] Where l = 1, ..., L.

[0105] Iteration steps b) through e) continue until the maximum generating generation G is reached, satisfying either η < ξ. The best individual in the current generation g′ is the final selection.

[0106]

[0107] Specifically, S3 refers to the traditional algorithms compared in this invention, which mainly include WLS, RNLA, DEOR, RSS-SDP, and IPM-MM, as shown in Table 1.

[0108] Table 1 compares the algorithms with those of this invention.

[0109] Algorithm Description WLS Weighted Least Squares RNLA Robust Non-Cooperative Localization Algorithm DEOR Differential Evolution with Opposition-Based and Re-direction RSS-SDP Semi-Definite Programming IPM-MM Interior Point Method and Optimization-Minimization Algorithm RRTL Method of the Invention

[0110] The proposed algorithm can better mitigate the impact of absorption bias and provide accurate real-time positioning. Among traditional methods, RSS-SDP offers high accuracy, but its complex matrix operations lead to low real-time performance. WLS is a simple method with high real-time performance, but it is significantly affected by absorption bias and has lower accuracy. RNLA and IPM-MM methods are based on the LS approximation, therefore they cannot offer a good trade-off between accuracy and real-time performance. DEOR employs an evolutionary criterion to achieve a better trade-off, but its accuracy is lower than RRTL.

[0111] The specific implementation process of this invention is as follows:

[0112] Consider a 3D UWSN with a positioning area of ​​sl×sl×sl, where anchor points and unknown target node positions are randomly distributed, such as... Figure 5 As shown. The number of targets in the training set and the test set are M respectively.t and M g , whose RSS value is obtained using (4). The performance of the RRTL method is compared with WLS, RNLA, DEOR, RSS-SDP, and IPM-MM methods, as well as CRLB. The network employs a self-organizing routing protocol and does not require pre-established infrastructure or central nodes. It operates at a frequency of 6 kHz and a data rate of 1 kbps. All simulations are performed on a laptop with a 2.8 GHz Intel Core i5 processor and 12 GB of memory using MATLAB 2019a. The performance metric is the root mean square error (RMSE), defined as

[0113]

[0114] where, is the estimated value of the target u i .

[0115] Figure 6 The relationship between the RMSE of target estimation and a f , γ = 4, σ 2 = 3 dB, sl = 80 m, and N = 12 for the six methods is given. Remarkably, this coefficient range is greater than the expected range in real UWSNs. In theory, according to (4), an increase in a f will result in an increase in the absorption attenuation. This is reflected in the deterioration of the RMSE as a f increases, particularly for WLS, RNLA, RSS-SDP, and IPM-MM. Conversely, an increase in a f from 0.02 dB / m to 0.18 dB / m results in a slight decrease in the RMSE for RRTL and DEOR, with RRTL being closest to the CRLB. Thus, the proposed method is more resistant to absorption effects than other algorithms. For example: when a f = 0.1 dB / m, the RRTL performance is improved by 55.6%, 57.6%, 62.8%, 80.4%, and 89.2% compared to DEOR, IPM-MM, RSS-SDP, WLS, and RNLA, respectively.

[0116] Computational time is a key factor in real-time localization. Figure 7 (a) in Fig. 5 shows the computational time as a function of a f = 0.05 dB / m, sl = 80 m, and N = 12, Figure 7 (b) in Fig. 5 shows the computational time as a function of γ, with σ 2 = 3 dB, a f = 0.05 dB / m, sl = 80 m, and N = 12, Figure 7(c) in FIG. 6 shows the computation time as a function of a f with y = 4, s 2 = 3 dB, sl = 80 m, and N = 12, Figure 7 (d) in FIG. 6 shows the computation time as a function of sl with s 2 = 3 dB, y = 4, a f = 0.05 dB / m, and N = 12, Figure 7 (e) in FIG. 6 shows the computation time as a function of N with s 2 = 3 dB, y = 4, a f = 0.05 dB / m, and sl = 80 m. The results show that the computation time of WLS is the least, followed by RRTL. In addition, the computation time of RSS-SDP to search for a global solution is longer, so it is not a practical method. The performance of DEOR is good in terms of both computation time and positioning error, but it is worse than RRTL. Therefore, considering both positioning accuracy and real-time performance, the proposed RRTL method is the best choice, and its computation time is about 3.5 x 10 -4 s, which is slightly higher than that of WLS.

[0117] The method uses reverse learning to obtain a robust population and uses an adaptive mutation strategy to improve global search and convergence. The optimized network parameters are obtained using IDE to train the BP neural network, and the trained network is used to obtain multi-target prediction information. The IDE algorithm combined with the BP neural network solves the multi-target underwater acoustic positioning problem based on RSS. Reverse learning is used to obtain a robust population to search for a global optimal value, and an adaptive mutation strategy is used to improve global search and convergence. In terms of real-time UWSN positioning accuracy, the proposed method RRTL is compared with unconventional methods (i.e., DEOR) and conventional methods (i.e., WLS, RNLA, RSS-SDP, and IPM-MM). The results show that the RRTL algorithm provides better positioning accuracy than other methods. In addition, it is comparable to WLS in terms of computation time and much better than other methods, which is important for real-time implementation.

[0118] Example 2

[0119] The embodiment provides an electronic device, comprising: one or more processors and a memory, the memory having stored therein one or more programs, the one or more programs comprising instructions for performing the real-time underwater acoustic positioning method based on the IDE-BP neural network as described in embodiment 1.

[0120] Example 3

[0121] The embodiment provides a computer readable storage medium, comprising one or more programs for execution by one or more processors of an electronic device, the one or more programs comprising instructions for performing the real-time underwater acoustic positioning method based on the IDE-BP neural network as described in embodiment 1.

[0122] The above merely provides a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements shall be encompassed within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A real-time underwater acoustic localization method based on IDE-BP neural network, characterized in that, Includes the following steps: Acquire underwater acoustic data and use a trained BP neural network for underwater multi-target localization. The acquisition of the trained BP neural network includes the following steps: A training sample set was obtained by constructing an acoustic localization model based on non-convex RSS that considers path loss and absorption loss. A backpropagation (BP) neural network is constructed. Initial network parameters are obtained using the differential evolution algorithm. The BP neural network is then trained using the training sample set to obtain the trained BP neural network. The acoustic localization model is modeled as follows: in, for Maximum likelihood estimation, For the target node j Anchor node i The received signal strength value, The distance is The transmission power at that time This is a reference distance. This indicates absorption bias. This is the path loss index. For the target node j With anchor node i The distance between them To measure the variance of noise, It measures noise. N For the number of anchor nodes, The process of underwater multi-target localization using a trained BP neural network includes: initialization: Let the solution space of the BP network parameters be Generate by the following formula L vectors : in, , These are the lower limit and the upper limit, respectively. , m It is a dimension. It is a random number. Constructing the reverse vector : Including 2 L A collection of individuals Among them, choose the one with the lowest fitness. L The initial population consists of 10 candidates, denoted as 1. , Mutations: In the g-th generation, an adaptive mutation strategy is used to select from the current population. Generate a mutant population Adaptive control parameters With threshold Comparisons are made to determine the strategy to be adopted, where It is the average of the objective function values ​​of all individuals. It is the objective function value of the best individual, and each policy group includes the first mutation operation: Second mutation operation: in, It is a scaling factor. express The best individual, and , ,when The first mutation operation is used to generate a mutated population; otherwise, the second mutation operation is used. cross: The mating of the original population and the variant population is based on the crossover probability. Completed, resulting in a new population with individual members. : in, , Boundary handling: Adjustments are made for individuals outside the search space: in, Follows a uniform distribution on [0,1] and in, ,and and They are the first g The lower and upper bounds of an individual. choose: In the g+1 generation, from individuals and Choose the individual with the smallest fitness value from the combination. Iterate until the preset condition is met, then proceed to the next algebra. The best individual among them is the final selection.

2. The real-time underwater acoustic localization method based on the IDE-BP neural network according to claim 1, characterized in that, The acoustic localization model includes multiple anchor nodes and multiple unknown target nodes, and the position and number of the anchor nodes match the position and number of underwater acoustic sensors in the real scene.

3. The real-time underwater acoustic localization method based on the IDE-BP neural network according to claim 1, characterized in that, The process of obtaining the training sample set includes the following steps: Based on the distance between each anchor node and each unknown target node in the acoustic localization model, the received signal strength values ​​between each anchor node and each unknown target node are obtained after considering path loss, absorption bias and measurement noise. Based on the received signal strength value, target location estimation information is obtained; The training sample set is obtained based on the received signal strength value and the target location estimation information.

4. The real-time underwater acoustic localization method based on the IDE-BP neural network according to claim 3, characterized in that, The acquisition of the target location estimation information includes the following steps: Based on the received signal strength values ​​between each anchor node and each unknown target node, the probability density function of the observation vector of each unknown target node is obtained, and the maximum likelihood estimate of the position of the unknown target node is obtained by maximizing the probability density function.

5. The real-time underwater acoustic localization method based on the IDE-BP neural network according to claim 1, characterized in that, The initial network parameters include the weights and biases of the BP neural network.

6. The real-time underwater acoustic localization method based on the IDE-BP neural network according to claim 1, characterized in that, The BP neural network includes an input layer, a hidden layer, and an output layer connected in sequence.

7. The real-time underwater acoustic localization method based on the IDE-BP neural network according to claim 1, characterized in that, The underwater acoustic data mentioned above is the received signal strength difference data.

Images