Deviation estimation and classification fusion method based on underwater multi-base passive detection

Abstract

The invention provides a deviation estimation and classification fusion method based on underwater multi-base passive detection, and the method comprises the steps: firstly obtaining a plurality of groups of passive sonar sensor positioning data containing random angle estimation deviation, building a likelihood function, and obtaining the estimation value of the deviation of each group of passive sonar sensor through a maximum likelihood estimation method; secondly, classifying the estimated sensor groups according with different normal distribution deviations by adopting an expectation maximization algorithm according to the obtained estimated values of the deviations of each group of passive sonar sensors, and reversely deducing different types of distribution parameters through iteration; and finally, considering the influence of the deviation on the credibility of passive detection, carrying out variable step weight fusion on observation data of each sensor group in the class, and then carrying out variable step weight fusion on fusion data of each class so as to obtain an estimated value of a target position. According to the invention, the target position estimation precision of the multi-base sonar network is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater passive target detection, and in particular to a deviation estimation and classification fusion method based on underwater multi-base passive detection. Background Art

[0002] Since the birth of sonar detection technology, the field has experienced significant development. Depending on whether the sensor emits a signal, positioning technology can be divided into active sonar and passive sonar. Passive sonar positioning technology does not actively emit a signal, but uses the noise radiated by the target for positioning. Due to its good concealment, it has become a hot topic of research today. Current research focuses include overcoming the challenges brought by environmental noise, improving detection capabilities in complex underwater environments, and combining passive sonar systems with other sensor technologies to achieve comprehensive underwater surveillance.

[0003] In order to improve the accuracy and reliability of passive sonar in target detection and positioning, a distributed fusion system based on underwater passive sonar has been proposed. The system integrates the information collected by multiple geographically dispersed sensors or platforms, combines the data of various underwater passive sonar systems deployed in a wide area, and fuses them to create a comprehensive and consistent target perception scenario. The distributed fusion system involves the collaboration between multiple sensor nodes or platforms. Each node or platform is equipped with a passive sonar system, and the information collected from different sensor nodes is fused through a fusion algorithm. The distributed characteristics of the system can prevent the impact of single sensor failure or performance limitations on the overall system. The redundancy of its data collection ensures that the entire system can remain functional and accurate even if some sensor nodes have problems. And by cross-comparing and verifying the signals of different sensor nodes, the fusion of data from multiple sources can help filter out false alarms or false detections, thereby distinguishing actual underwater targets from environmental noise or abnormal situations.

[0004] The focus of distributed fusion technology is to optimize the fusion algorithm, solve the communication and synchronization problems between distributed sensors, and explore the data fusion from various sensor platforms. However, due to the influence of various factors such as underwater detection process, signal processing technology and hydrological environment, passive sonar sensors usually have large angle estimation deviations. The appearance of angle deviation affects the positioning accuracy of the target to a certain extent. At present, there have been related studies on eliminating the influence of angle estimation deviation through data fusion of multiple passive sonars, but the focus of the research is mainly on the technical error of the sensor itself and the estimation of linear angle deviation, and the classification fusion process basically needs to ensure the stability of the sensor operation, that is, no singular value or missed state. However, due to the influence of various factors such as underwater hydrological environment and the time-varying characteristics of the underwater environment, the sensor deviation does not change linearly. At the same time, the performance of the sensor is difficult to guarantee. The appearance of singular values ​​and missed reports are problems often faced by underwater sensors and need to be solved urgently. Therefore, the problem of target fusion and positioning in passive sonar networks cannot be ignored. The impact of sensor deviation on positioning and the problem of data classification fusion of multiple passive sonars should be considered. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the present invention proposes a deviation estimation and classification fusion method based on underwater multi-base passive detection, which adopts random deviation estimation and combines with the corresponding adaptive information classification fusion method to finally obtain a more accurate target position estimation value.

[0006] The technical solution of the present invention is:

[0007] A method for bias estimation and classification fusion based on underwater multi-base passive detection includes the following steps:

[0008] Step 1: Obtain multiple groups of passive sonar sensor positioning data containing random angle estimation deviations, establish a likelihood function, and obtain the estimated value of each group of passive sonar sensor deviations through the maximum likelihood estimation method;

[0009] Step 2: Based on the estimated values ​​of the bias of each group of passive sonar sensors obtained in step 1, the maximum expectation algorithm is used to classify the estimated sensor groups that meet different normal distribution biases, and the distribution parameters of different categories are deduced through iteration;

[0010] Step 3: Considering the impact of deviation on the credibility of passive detection, the observation data of each sensor group in the class are firstly fused with variable step size weighted fusion, and then the fused data of each classification are fused with variable step size weighted fusion to obtain the estimated value of the target position.

[0011] Furthermore, in step 1, there are q groups of passive sonar sensors in the passive sonar sensor monitoring area, each group consists of 2 passive sonar sensors; the deviation estimation value vector of a group of sensors (i, j) is obtained by maximum likelihood estimation:

[0012]

[0013] in is the deviation estimation vector of sensor group (i, j) within time length K; is the bias vector for angle estimation;

[0014]

[0015] is the pseudo measurement vector obtained by the passive sonar sensor group (i, j) at time k;

[0016]

[0017] is the position estimation of the target by the sensor group (i, j), and represents the average value of the target positioning estimation results of q groups of passive sonar sensors; ω(k) is the pseudo-measurement value The correlated additive white Gaussian noise has a mean of zero.

[0018] Furthermore, the position estimation of the target by the sensor group (i, j) at time k is:

[0019]

[0020]

[0021] where x i and i are the initial deployment positions of the passive sonar sensor i on the X-axis and Y-axis of the set coordinate system, respectively. j and j are the initial deployment positions of the passive sonar sensor j on the X-axis and Y-axis of the set coordinate system, and Z i (k) and Z j (k) are the target angle information received by passive sonar sensors i and j respectively.

[0022] Furthermore, the specific process of step 2 is:

[0023] Step 2.1: Taking the estimated values ​​of a certain set of sensor groups (i, j) as the benchmark, assume the initial covariance matrix of the two classifications The means of the two normal distributions are where e = [ee] T, e represents the mean difference between sensors after they are divided into two categories;

[0024] The initial expressions of the two types of normal distribution are obtained as follows:

[0025]

[0026] Step 2.2: For q groups of sensors, evaluate the category to which the estimated value of the pseudo measurement vector of each group of sensors belongs:

[0027] For the Qth group of sensors (m,n), Q = 1, 2, ..., q, the estimated value of its pseudo measurement vector is Pick

[0028]

[0029] According to Bayes' theorem, belongs to p1(z b (k)|B1 0 ) is:

[0030]

[0031] belongs to p2(z b (k)|B2 0 ) is:

[0032]

[0033] Compare and Size, divide the sensor group Q into a category with a larger probability value, and complete the classification of the sensor group Q;

[0034] Step 2.3: According to the categories of the q groups of sensors obtained in step 2.2, recalibrate the Gaussian distribution of the two categories:

[0035]

[0036] Update the prior probabilities of both classes simultaneously:

[0037]

[0038] Step 2.4: Return to step 2.2 and perform iterative calculation until the set iteration termination condition is reached.

[0039] Furthermore, in step 2.1,

[0040] Among them, R ω (k) represents the covariance matrix of ω(k), is the covariance matrix of sensor group (i, j), R xy (k) represents the average value of the covariance matrix and the sum of the q group matrices:

[0041]

[0042] represents the covariance of the Qth group of sensors.

[0043] Furthermore, in step 2, assuming that q groups of sensors are divided into c categories, Q = 1, 2, ..., q, C = 1, 2, ..., c, then after the wth iteration:

[0044]

[0045]

[0046] Among them, B C w+1 Represents the mean of the normal distribution for the Cth category after w iterations; R C w+1 (k) represents the covariance matrix for the C-th category after w iterations; P(B C w+1 ) represents the prior probability of the Cth category classification after w iterations; p C (z b (k)|B C w+1 ) represents the updated equation of the normal distribution of the estimated bias for the C-th category classification after w iterations.

[0047] Furthermore, in step 3, the process of variable step-size weighted fusion of the observation data of each sensor group in the class is:

[0048] For the sensor group that is classified into the same category as the sensor group (i, j), it is set as category C. The number of sensor groups in this category is q1, and the fusion estimation result of this category is:

[0049]

[0050] in represents the x-position estimation of the target by the q1 sensor groups that are classified into the same category as the sensor group (i, j), represents the y-position estimation of the target by the q1 sensor groups that are classified into the same category as the sensor group (i, j); The weight matrix representing the sensor groups that are classified into the same category as sensor group (i, j):

[0051]

[0052] is the confidence weight of the Q1th group of sensors in the sensor group that is classified into the same category as the sensor group (i, j).

[0053] Furthermore, in step 3, the result of variable step-size weighted fusion of the fusion data of various sensor groups is:

[0054]

[0055] The weight of each class

[0056]

[0057] Beneficial Effects

[0058] Compared with the prior art, the beneficial effect of the present invention is that, unlike the traditional multi-base passive sonar network, the present invention takes into account the random deviation of the sensor and the classification and fusion problem after the deviation estimation. The random deviation is estimated by using the maximum likelihood estimation, and the target position is estimated by combining the deviation-based EM classification algorithm and the variable step-size adaptive fusion algorithm, thereby improving the target position estimation accuracy of the multi-base sonar network.

[0059] The advantages of the present invention are: (1) taking into account the possible random errors of underwater sensors, it is more in line with the actual underwater situation; (2) a sensor deviation estimation method based on the maximum likelihood function is designed, which can accurately estimate the random sensor deviation; (3) using the EM classification algorithm and the variable step-size weighted fusion algorithm, combined with the sensor deviation, the sensor results with different deviations are classified, and the estimation results of different categories are subjected to variable step-size weighted fusion, so as to obtain the estimation result of the target position; (4) in the case of missing sensor data or the presence of singular values, the algorithm can still guarantee the accuracy of the target position estimation; (5) the algorithm can still guarantee the accuracy of the target position estimation when processing large angle deviations; (6) under the same number of iterations, the convergence speed of the algorithm is faster and the target positioning accuracy is higher.

[0060] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0062] Figure 1 Flowchart of underwater multi-base passive detection bias estimation and classification fusion BECFL algorithm;

[0063] Figure 2Considering the relationship between the number of iterations of different algorithms and the distance estimation error of the sensor random deviation;

[0064] Figure 3 Schematic diagram of using the BECFL algorithm to classify passive detection results into two categories;

[0065] Figure 4 Schematic diagram of using the BECFL algorithm to classify passive detection results into four categories;

[0066] Figure 5 The relationship between the angle deviation and distance estimation error of different algorithms;

[0067] Figure 6 The distance estimation error of the BECFL algorithm when the sensor has extreme estimation values. DETAILED DESCRIPTION

[0068] Embodiments of the present invention are described in detail below. The embodiments are exemplary and intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0069] like Figure 1 As shown, the deviation estimation and classification fusion method based on underwater multi-base passive detection in this embodiment includes the following steps:

[0070] Step 1: Obtain multiple groups of passive sonar sensor positioning data containing random angle estimation biases, establish a likelihood function, and obtain the estimated value of each group of passive sonar sensor biases through the maximum likelihood estimation method.

[0071] Assuming a pair of passive sonar sensors i and j, the target position at time k is estimated by geometric intersection method, ignoring the measurement noise, and the following is obtained:

[0072]

[0073] in, and are the Cartesian estimates of the target X and Y axes, respectively, i and i represents the initial X and Y axis deployment position of the passive sonar sensor i, x j and j Similarly, Z i (k) and Z j (k) are the target angle information received by passive sonar sensors i and j respectively.

[0074] Assume that there are q groups of passive sonar sensors in the passive multi-sonar surveillance area, each group consists of 2 passive sonar sensors, that is, the number of sensors is 2q. In a certain period of time T, assuming that the time is independent of each other, T is divided into n independent time periods, that is, it can be written as T = nK, where n ≥ 2. Then, for each divided time period K, when k = 1, 2, ..., K. Among them, according to the position estimation of a certain group of passive sonar sensors (i, j) to the target, define The nonlinear pseudo-measurement vector is:

[0075]

[0076] in, Pseudo measurements obtained by the passive sonar sensor group. It represents the position estimation of the target by the passive sonar sensor group (i, j). and represents the average value of the target positioning estimation results of q groups of passive sonar sensors. ω(k) is the pseudo-measurement value The covariance matrix of the additive Gaussian white noise with a correlated mean of zero is defined as:

[0077]

[0078] Among them, R ω (k) represents the covariance matrix of ω(k), is the covariance matrix of sensor group (i, j), R xy (k) represents the average value of the covariance matrix and the sum of the q group matrices, represents the covariance of the Qth group of sensors. Then, the likelihood function of the sensor's bias parameter can be given as:

[0079]

[0080] Among them, the bias vector of angle estimation is It can be expressed as Nonlinear function of the bias vector It can be expressed as:

[0081]

[0082] Then, from formula (7), when k = 1, 2, ..., K, the likelihood function can be written as:

[0083]

[0084] in, Then, when the likelihood function is maximum, the bias estimate vector for sensor (i, j) is can be written as:

[0085]

[0086] According to formula (9), the deviation estimation value vector of a set of sensors (i, j) can be obtained by maximum likelihood estimation. It is expressed as the estimated value of the random deviation within a certain time length K. Similarly, the deviation vectors of the remaining sensor groups within a certain time length K can be calculated to obtain the deviation values ​​of all global sensor groups within time K. Similarly, for a system with a time length of T, after n iterations, the deviation value of each time length K can be obtained, which can be regarded as the deviation within this certain time period.

[0087] Step 2: Use the Expectation Maximum (EM) algorithm to classify the estimated sensor groups that conform to different normal distribution deviations, and infer the distribution parameters of different categories through iteration.

[0088] For a system with q groups of sensors in the monitoring area, within a certain period of time T, the time is independent of each other, and T is divided into n independent time periods of length K. In each period of time K, the above formula 6 assumes that the nonlinear pseudo-measurement vector for the sensor group (i, j) is Bias vector for angle estimation It conforms to the only normal distribution.

[0089] Afterwards, taking the estimated values ​​of sensor group (i, j) as the benchmark, we assume that the initial covariance matrix of the two classifications is The bias of the angle estimate obtained by the first calculation is used as the known value of the iteration. Then, within a certain period of time K, assuming that the means of the two types of normal distributions are and where e = [ee] T , e represents the mean difference between the sensors after they are divided into two categories, that is, the significance of the difference between the two categories of sensors after classification, which needs to be set to an empirical value that takes into account the average deviation. Here, since the classification is based on the normal distribution, the mean of the normal distribution affects the classification result, and the mean difference is the mean difference of the normal function between the two desired classifications. The mean is the mean of the normal function that the two classifications meet.

[0090] Then, the initial expressions of the two new normal distributions can be written as:

[0091]

[0092] Assume that the prior probabilities of the two classes are equal:

[0093] P(B1 0 )=P(B2 0 )=0.5 (12)

[0094] Where P(B1 0 ) and P(B2 0 ) respectively indicate the occurrence of p1(z b (k)|B1 0 ) and p2(z b (k)|B2 0 ), that is, the prior probabilities of the two categories.

[0095] For the EM algorithm, we first need to make an expected estimate. Let’s first evaluate which category the estimated pseudo measurement vector of each sensor group belongs to. Assume that for a group of sensors (m, n), the estimated pseudo measurement vector is It is the Qth group of sensors Q=1,2,...,q, then

[0096]

[0097] That is, according to Bayes' theorem, belongs to p1(z b (k)|B1 0 ) can be expressed as:

[0098]

[0099] Where Q = 1, 2, ..., q represents the Qth group of sensors. According to formula (14), we can find belongs to p2(z b (k)|B2 0 ) is:

[0100]

[0101] So far, in this iteration, we have learned that for a set of sensors Q, the estimated value of the pseudo measurement vector is The probability of belonging to two categories, comparing and The sensor group Q is divided into a class with a larger probability value, and the classification of the sensor group Q is completed. For the remaining sensor groups, this method is also needed to be used for calculation and classification, and a total of q operations are performed.

[0102] Now that we have obtained the probability that each set of estimated values ​​in the q sets of estimates belongs to each class, we can recalibrate the Gaussian distribution of the two categories, that is, calculate the mean and variance of the two categories after calibration:

[0103]

[0104]

[0105] Update the prior probabilities of both classes simultaneously:

[0106]

[0107] Then, based on the above derivation, the iterative process is generalized. Assuming that q groups of sensors need to be divided into c categories, Q = 1, 2, ..., q, C = 1, 2, ..., c, then after the wth iteration:

[0108]

[0109] Among them, B C w+1 Represents the mean of the normal distribution for the Cth category after w iterations. R C w+1 (k) represents the covariance matrix for the Cth category after w iterations. C w+1 ) represents the prior probability of the Cth category classification after w iterations. C (z b (k)|B C w+1 ) represents the updated equation of the normal distribution of the estimated bias for the C-th classification after w iterations. After obtaining these data, continue to use the EM method to classify them, and then repeat the iterations to classify each group of sensors with similar biases into one category.

[0110] Step 3: Considering the impact of deviation on the credibility of passive detection, the observation data of each sensor group in the class are firstly fused with variable step size weighted fusion, and then the fused data of each classification are fused with variable step size weighted fusion to obtain the estimated value of the target position.

[0111] After one iteration, for sensors classified into c categories, C = 1, 2, ..., c, we have

[0112]

[0113] Among them, q C Indicates the number of sensor groups classified into category C, and has Q C =1,2,...,q C .

[0114] A group of sensor groups is selected for the positioning algorithm description, and the sensor groups in the remaining categories also obtain the target positioning information according to this method. Take the number of sensor groups selected as q1 that are classified into the same category as the sensor group (i, j) as an example, and Q1 = 1, 2, ..., q1. Weighted data fusion is performed on this type of data, and the bias of different sensor groups is considered as the basis for weight calculation. Obviously, the estimation result of the sensor group with too large bias will have a lower credibility. On the contrary, the estimation credibility of the sensor group with smaller deviation will be higher, so the weight should be inversely proportional to the degree of sensor deviation. The confidence weight calculation formula is as follows:

[0115]

[0116] in, It is the confidence weight of the Q1th sensor group in the sensor group that is classified into the same category as the sensor group (i, j), () T Indicates the transposition of the matrix, () -1 Represents the inverse function of the matrix. Calculate the weights of the sensor groups that are in the same category as the sensor group (i, j) to obtain the weight matrix:

[0117] λ=[λ1λ2...λ q1 ] (28)

[0118] Then the estimated result is:

[0119]

[0120] in, It represents the x-position estimation of the target by the q1 sensor groups that are classified into the same category as the sensor group (i, j), where the estimation method is shown in formula (1). represents the confidence-weighted fusion estimation result of the sensor of this category. Similarly, the y-position estimation result of the target is obtained, where the estimation method is shown in formula (2):

[0121]

[0122] What you get and It is the estimated result of the target position for the sensor group that is classified into the same category as the sensor group (i, j). Extending it to all groups, formulas (27), (28), (29) and (30) can be rewritten as:

[0123]

[0124] Then for category C, the final first weighted target estimate is By extending this method to the iterative process, it can be seen that in the wth iteration, the mean of the Gaussian distribution that each classification meets is used as the basis for the weight of each class in the secondary fusion at the fusion center. Specifically, it is based on the classification result of q groups of sensors being divided into c classes in step 2, Q = 1, 2, ..., q, C = 1, 2, ..., c, and the weight of each class is:

[0125]

[0126] After obtaining the weight of each class, a second weighting is performed at the fusion center to obtain the final estimation result:

[0127]

[0128] in, It is the passive sonar distributed fusion result obtained by weighting twice after classification by the maximum deviation expectation algorithm. Since this algorithm uses likelihood estimation, EM classification and variable step-size fusion weighting to deal with the distributed multi-sonar fusion positioning problem under the condition of passive sonar bias, this algorithm is called the BECFL algorithm.

[0129] In order to illustrate the estimation of the target position deviation by the proposed BECFL algorithm, the root mean square error (RMSE) of the uncorrected target position deviation, the RMSE of the deviation corrected by the likelihood estimation, and the RMSE of the estimated deviation after fusion weighting by the BECFL algorithm are compared. Figure 2 As shown in the figure, the RMSE of the passive sonar with angular deviation is compared. No deviation correction is performed, only the likelihood estimation is used for deviation correction and then uniform weighting is performed, and the RMSE of the BECFL algorithm is used. It can be seen that the likelihood estimation of the passive sonar deviation has an effect on improving the accuracy of target position estimation. However, if the deviation correction is used and then uniform weighting is performed, the improvement in positioning accuracy is not as obvious as that of the EM classification and variable step fusion weighting algorithm. This proves the effectiveness of the proposed deviation model and BECFL algorithm, and proves the ability of the algorithm to estimate deviation and position fusion.

[0130] To illustrate the impact of the BECFL algorithm on target classification, the azimuth intersection method is used, and the sonar partial compliance standard deviation is σ θ =0.05rad and σ θ =0.02rad. Most of the estimated intersection points are near the target, but due to angle deviation, there are also intersection points far from the target. When the BECLF algorithm is used to divide the passive sonar positioning into two categories, the following is obtained: Figure 3As shown in the figure, due to the EM classification method used by the BECLF algorithm, points that meet the same normal distribution criteria are classified. To illustrate the relationship between the number of classifications and the sonar type, when the BECLF algorithm is used to divide the target into four categories, the classification obtained is as follows Figure 4 As shown in Figure 2, we can still classify the nodes that are farther away and the nodes that are closer with smaller errors. It can be seen that the classification method can effectively classify different estimation errors.

[0131] To illustrate the change in the standard deviation of the algorithm under the condition of angle error and its impact on the target position estimation accuracy, it is assumed that the standard deviation of the angle error of the two types of passive sonars is σ θ =0.02rad. Through 2000 Monte Carlo simulations, and the number of iterations w = 30. The results are as follows Figure 5 As shown in the figure, when the standard deviation of the angle error is small, the deviation has little effect on passive detection. Therefore, whether the angle deviation is estimated at this time has little effect on the target position estimation. However, when the standard deviation of the angle deviation increases, it can be seen that the target positioning performance using the error likelihood estimation and BECLF algorithm is better. Therefore, it is shown that the likelihood estimation of the error plays a key role in improving passive detection when the deviation is large. At the same time, when the BECLF algorithm with classification and variable step-size weighted fusion is used, the positioning accuracy is further improved compared to only performing deviation estimation.

[0132] To illustrate the impact of extreme angle deviation of passive sonar on positioning results using this patented algorithm. Figure 6 It can be seen that when there is one extreme sonar, the RMSE of its positioning is about 700m. When there are two extreme sonars, the RMSE of its positioning is about 900m. When there are three extreme sonars, the RMSE of its positioning is about 1200m. Using the BECFL algorithm to process, the fusion weight of sonar data with large deviations can be reduced. At the same time, this sonar will be distinguished from other normal sonars in the EM classification algorithm to reduce the impact of erroneous estimation on the fusion center. Figure 6 It can be seen that the final algorithm can maintain convergence under different extreme sonar conditions. Among them, when there is one extreme sonar, the RMSE of positioning is reduced and maintained at about 50m. When there are two and three extreme sonars, the RMSE of positioning is also reduced to about 80m and 130m. For the error itself, the improvement in positioning accuracy is acceptable. It verifies that when this algorithm has extreme points, the final error can still be reduced by data fusion weight allocation.

[0133] In summary, the present invention provides a bias estimation and classification fusion algorithm based on underwater multi-base passive detection, which takes into account the random estimation bias that may be generated by sonar and its influence on target fusion positioning. First, the dynamic random process bias modeling method of passive sensors is discussed. Using multiple groups of sensor observation data results, a likelihood function is established, and the estimated value of sensor bias is obtained by the maximum likelihood estimation method. Afterwards, the EM algorithm is used to classify the sensors with estimated deviations that conform to different normal distributions, and the distribution parameters of different categories are inferred by iteration. Considering the influence of bias on the credibility of passive detection, the observation data is weightedly fused with variable step length to obtain the estimated value of the target position, so that it can handle false alarms and misreports in a distributed fusion architecture. Through the positioning error of target fusion positioning, it is shown that the present invention can still maintain good positioning performance when the sensor is biased. At the same time, the classification effect simulation of the algorithm shows that when the appropriate number of classifications is selected, the accuracy of the algorithm can be guaranteed. In order to prove that the proposed method can work even when the sensor has false alarms and missed detections, the simulation results in the scenario of missing data or singular values ​​of some nodes are also shown, indicating that the proposed BECFL has excellent performance. In summary, the advantages of the present invention are: (1) taking into account the possible random errors of underwater sensors, it is more in line with the actual underwater situation; (2) designing a sensor deviation estimation method based on the maximum likelihood function, which can estimate the random sensor deviation; (3) using the EM classification algorithm and the variable step-size weighted fusion algorithm, combined with the sensor deviation, the sensor results with different deviations are classified, and the estimation results of different categories are subjected to variable step-size weighted fusion to obtain the target position estimation result; (4) in the scenario where individual sensor data is missing or singular values ​​occur, the algorithm can still guarantee the accuracy of the target position estimation; (5) the algorithm can guarantee the accuracy of the target position when the angle deviation is large; (6) under the same number of iterations, the convergence speed of the algorithm is faster and the target positioning accuracy is better.

[0134] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and intent of the present invention.

Claims

1. A deviation estimation and classification fusion method based on underwater multi-base passive detection, Features: The following steps are involved: Step 1: Obtain multiple groups of passive sonar sensor positioning data containing random angle estimation deviations, establish a likelihood function, and obtain the estimated value of each group of passive sonar sensor deviations through the maximum likelihood estimation method; Step 2: Based on the estimated values ​​of the bias of each group of passive sonar sensors obtained in step 1, the maximum expectation algorithm is used to classify the estimated sensor groups that meet different normal distribution biases, and the distribution parameters of different categories are deduced through iteration; Step 3: Considering the impact of deviation on the credibility of passive detection, the observation data of each sensor group in the class are firstly fused with variable step size weighted fusion, and then the fused data of each classification are fused with variable step size weighted fusion to obtain the estimated value of the target position.

2. According to the method for error estimation and classification fusion based on underwater multi-base passive detection according to claim 1, Features: In step 1, there are q groups of passive sonar sensors in the passive sonar sensor monitoring area, each group consists of 2 passive sonar sensors; the deviation estimation value vector of a group of sensors (i, j) is obtained by maximum likelihood estimation: in is the deviation estimation vector of sensor group (i, j) within time length K; is the bias vector for angle estimation; is the pseudo measurement vector obtained by the passive sonar sensor group (i, j) at time k; is the position estimation of the target by the sensor group (i, j), and represents the average value of the target positioning estimation results of q groups of passive sonar sensors; ω(k) is the pseudo-measurement value The correlated additive white Gaussian noise has a mean of zero.

3. According to the method for error estimation and classification fusion based on underwater multi-base passive detection according to claim 2, Features: The position estimate of the target by the sensor group (i, j) at time k is: where x i and i are the initial deployment positions of the passive sonar sensor i on the X-axis and Y-axis of the set coordinate system, respectively. j and j are the initial deployment positions of the passive sonar sensor j on the X-axis and Y-axis of the set coordinate system, and Z i (k) and Z j (k) are the target angle information received by passive sonar sensors i and j respectively.

4. According to the method for error estimation and classification fusion based on underwater multi-base passive detection according to claim 1, Features: The specific process of step 2 is: Step 2.1: Taking the estimated values ​​of a certain set of sensor groups (i, j) as the benchmark, assume the initial covariance matrix of the two classifications The means of the two normal distributions are and where e = [ee] T , e represents the mean difference between sensors after they are divided into two categories; The initial expressions of the two types of normal distribution are obtained as follows: Step 2.2: For q groups of sensors, evaluate the category to which the estimated value of the pseudo measurement vector of each group of sensors belongs: For the Qth group of sensors (m,n), Q = 1, 2, ..., q, the estimated value of its pseudo measurement vector is Pick According to Bayes' theorem, Belong to p 1 (z b (k)|B 1 0 ) is: Belong to p 2 (z b (k)|B 2 0 ) is: Compare and Size, divide the sensor group Q into a category with a larger probability value, and complete the classification of the sensor group Q; Step 2.3: According to the categories of the q groups of sensors obtained in step 2.2, recalibrate the Gaussian distribution of the two categories: Update the prior probabilities of both classes simultaneously: Step 2.4: Return to step 2.2 and perform iterative calculation until the set iteration termination condition is reached.

5. According to the method for error estimation and classification fusion based on underwater multi-base passive detection according to claim 4, Features: In step 2.1, Among them, R ω (k) represents the covariance matrix of ω(k), is the covariance matrix of sensor group (i, j), R xy (k) represents the average value of the covariance matrix and the sum of the q group matrices: represents the covariance of the Qth group of sensors.

6. According to claim 4, a method for error estimation and classification fusion based on underwater multi-base passive detection, Features: In step 2, assuming that q groups of sensors are divided into c categories, Q = 1, 2, ..., q, C = 1, 2, ..., c, then after the wth iteration: Among them, B C w+1 Represents the mean of the normal distribution for the Cth category after w iterations; R C w+1 (k) represents the covariance matrix for the C-th category after w iterations; P(B C w+1 ) represents the prior probability of the Cth category classification after w iterations; p C (z b (k)|B C w+1 ) represents the updated equation of the normal distribution of the estimated bias for the C-th category classification after w iterations.

7. According to the method for error estimation and classification fusion based on underwater multi-base passive detection according to claim 1, Features: In step 3, the process of variable step-size weighted fusion of the observation data of each sensor group in the class is: For the sensor group that is classified into the same category as the sensor group (i, j), it is set as category C, and the number of sensor groups in this category is q 1 , the fusion estimation result of this class is: in represents the q that is classified into the same category as the sensor group (i, j) 1 The x-position estimate of the target by the sensor group, represents the q that is classified into the same category as the sensor group (i, j) 1 The y position estimation of the target by the sensor groups; The weight matrix representing the sensor groups that are classified into the same category as sensor group (i, j): is the Qth sensor in the same class as sensor group (i, j). 1 Confidence weights for group sensors.

8. According to the method for error estimation and classification fusion based on underwater multi-base passive detection according to claim 7, Features: In step 3, the result of variable step-size weighted fusion of the fusion data of various sensor groups is: The weight of each class

Images