Sonar detection target track prediction method
By combining gray theory and support vector machine algorithms to monitor and process the spatial data of sonar detection targets in real time, the problem of insufficient real-time and accuracy of sonar detection target track prediction in the existing technology is solved, and accurate track prediction of sonar detection targets is achieved.
Patent Information
- Application Number
- CN202411615507.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-13
- Publication Date
- 2025-02-25
AI Technical Summary
In the prior art, when sonar detects targets, it is difficult to establish a mathematical track model in real time and accurately, and the algorithm is difficult to apply to complex sea environments, resulting in insufficient real-time and accuracy of target track prediction.
A sonar detection target track prediction method is adopted, and through steps such as data preparation, data preprocessing, construction of gray prediction models, support vector machine residual correction models, gray residual prediction models and track prediction, combined with gray theory and support vector machine algorithm, spatial data is monitored and processed in real time to generate track prediction values.
Real-time and accurate track prediction of sonar detection targets is achieved, real-time and accuracy of predictions are improved, and can be applied to complex sea environments.
Smart Images
Figure CN119513828A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of track prediction, in particular to a method for predicting the track of a sonar-detected target. Background Art
[0002] With the rapid development of modern science and technology and the increasingly complex international situation, it is urgent to upgrade modern equipment based on maritime security. As the most effective means of underwater detection, sonar can predict the track of hydroacoustic targets, assist in the management of maritime targets, and play an important role in improving the maritime defense combat capability. However, when sonar detects targets, various factors may cause the target to be lost, such as limited array detection capability, target interference information, target motion posture changes, interference from marine environmental factors, etc. The marine environment is extremely complex, and the echo of sonar is complex and diverse, including not only useful targets such as ships, islands, and lighthouses, but also useless information such as waves, rain, snow, and system noise, which makes the tracking stability easily affected, and it is easy to have false tracks and difficulty in starting targets, which seriously restricts the track tracking performance. In addition, due to the randomness and diversity of interference, the motion state of the ship often changes frequently, and there is no regularity in the state information, making it difficult to accurately extract the trend of feature data changes, and some algorithms cannot meet the time series requirements and real-time requirements, making it extremely difficult to establish a real-time and accurate track mathematical model. More importantly, some algorithms are difficult to apply to actual projects.
[0003] Sonar target detection is the premise of sonar target tracking. Only after clutter is removed in the detection stage can the subsequent tracking process be performed. Therefore, how to further extract target features after target detection is an urgent problem to be solved in sonar target track prediction. After extracting target features, the nonlinearity and uncertainty characteristics of target information are enhanced based on the influence of multiple factors. Traditional target track prediction methods mainly include target track prediction based on serpentine maneuvers and track prediction using state transfer equations.
[0004] At present, there are mainly two methods for track prediction: prediction method based on particle motion model and prediction method based on hybrid estimation. The disadvantages of these two methods are lack of real-time performance and accuracy, and the calculation is too complicated and inconvenient to use. In recent years, on modern battlefields, acoustic wave data in complex acoustic wave environments have gradually become massive, redundant, rapid and diversified. Traditional methods of target track prediction and identification can no longer meet the needs. Summary of the invention
[0005] The technical problem to be solved by the present invention is to provide a sonar detection target track prediction method in view of the shortcomings of the existing technology, which can mine effective information from a large amount of data, and perform real-time monitoring, processing and aggregation of various spatial data, and then provide corresponding effective information to commanders for decision-making.
[0006] The technical problem to be solved by the present invention is achieved by the following technical solution. The present invention is a sonar detection target track prediction method, the method comprising the following steps:
[0007] (1) Data preparation:
[0008] Prepare single sonar detection target point track data;
[0009] (2) Data preprocessing:
[0010] Preprocess the target point data, dig out the inherent change rules of clean data sets and noisy data sets, and extract the main influencing factors;
[0011] (3) Select the starting position of the track:
[0012] Selecting an appropriate track starting position provides conditions for the implementation of the gray track residual correction algorithm;
[0013] (4) Constructing grey prediction model:
[0014] Build a grey prediction model and generate a residual sequence;
[0015] (5) Support vector machine residual correction model:
[0016] A residual correction model is built. Aiming at the nonlinear and uncertain characteristics of sonar detection targets, a support vector regression algorithm with good numerical convergence and stability is used to correct the target track residual.
[0017] (6) Grey residual prediction model:
[0018] The grey model is integrated with the residual prediction value to obtain an improved grey residual prediction model and generate a track prediction value;
[0019] (7) Track prediction:
[0020] The grey residual prediction model is applied to the sonar target prediction system to implement the sonar detection target track prediction.
[0021] The technical problem to be solved by the present invention can also be further achieved by the following technical solutions. For the sonar detection target track prediction method described above, in step (1), it is necessary to remove outliers and perform data normalization on the prepared data, wherein outliers refer to abnormal values in the measured data, and data normalization is to normalize all data sizes between [-1, 1]. The specific formula is:
[0022]
[0023] t normis the normalized value, t is the value before normalization, and t min is the minimum value before normalization, t max is the maximum value before normalization.
[0024] The technical problem to be solved by the present invention can also be further achieved by the following technical solution. For the sonar detection target track prediction method described above, in step (2), the data preprocessing method is the grey correlation analysis method, specifically:
[0025] (2.1) The influencing parameters with different dimensions are dimensionless. The dimensionless calculation formula is:
[0026]
[0027] Where n is the number of sample groups selected. is the initial sequence of x=(x(1),x(2),...x(n));
[0028] (2.2) Calculate the correlation coefficient:
[0029]
[0030] Where ζ(k) is the sequence x i With x 0 The correlation coefficient at time k, ρ∈[0,1] is the resolution coefficient, and ρ=0.5;
[0031] (2.3) First get n correlation coefficients, and then average the correlation coefficients:
[0032]
[0033] Among them, r i Represents the sequence x i With x 0 degree of relevance;
[0034] (2.4) Based on the correlation of each factor, several groups of comparison factors with correlations greater than 0.6 with longitude and latitude coordinates are obtained.
[0035] The technical problem to be solved by the present invention can also be further achieved by the following technical scheme. For the above-mentioned sonar detection target track prediction method, the output variables longitude and latitude coordinates are selected as the main factors, and the input variables are the influencing factors, including batch number, time, azimuth, distance, depth, heading, speed, characteristic spectrum, frequency, and label, wherein the number of sample groups n=800;
[0036] Then, based on the correlation of each factor, four groups of factors with the largest correlation between longitude and latitude coordinates were sorted out, namely time, direction, distance, and characteristic spectrum.
[0037] The technical problem to be solved by the present invention can be further achieved by the following technical solution. For the sonar detection target track prediction method described above, in step (3), the target prediction process includes a track starting module and a track maintaining module, and the track management is integrated into the track starting module and the track maintaining module;
[0038] The track initiation module is used to select the track initiation algorithm based on 3 / 3 logic when the track initiation reaction time is less than the system index requirement, that is, within three detection cycles, the track is successfully initiated only when the number of detections that meet the threshold condition in the time window reaches 3;
[0039] The track maintenance module adaptively selects the target prediction algorithm based on the target echo and target track information to ensure reliable track matching and correct measurement, and adjusts the size of the relevant threshold according to the track quality during the prediction process.
[0040] The technical problem to be solved by the present invention can also be further achieved by the following technical solution. For the above-mentioned sonar detection target track prediction method, the method of step (4) is:
[0041] First, the original sequence x(0) is accumulated to obtain the accumulated sequence x(1), and the GM(1,1) prediction model is established to obtain the residual sequence e(0) from time 1 to n, and the predicted value ^x(0)(n+1) at time n+1;
[0042] Then use the residual sequence to correct the prediction result: let e(0)(i) be the residual sequence, i = 1, 2, ..., n. If the prediction order is S, use e(0)(i-1), e(0)(i-2), ..., e(0)(iS) as the training samples of the support vector regression machine, and take e(0)(i) as the expected value;
[0043] Then, through training, the longitude residual, latitude residual and characteristic spectrum are used as input, and the residual longitude residual and latitude residual at time n+1 are obtained through the support vector regression machine. Finally, the predicted value at time n+1 is added to the residual value to obtain the corrected predicted value ^x(0)(n+1)+e(0).
[0044] The technical problem to be solved by the present invention can also be further achieved by the following technical solution. For the sonar detection target track prediction method described above, in step (5), the method of building a residual correction model is:
[0045] The grey model is corrected with machine learning method, and the support vector regression algorithm with good numerical convergence and stability is used to correct the target track residual.
[0046] The technical problem to be solved by the present invention can also be further achieved by the following technical solutions. For the sonar detection target track prediction method described above, the kernel function finally selected by the support vector machine regression algorithm for predicting the track is the most commonly used rdf kernel function. After the insensitive loss function and the kernel function are used in combination, the formula is converted to:
[0047]
[0048] In the formula, <ω,Φ(x i )> represents the vector product of the input weight and the kernel function, k(x,x i )=Φ(x)·Φ(x i ), constant C is the penalty factor, C>0, which is used to adjust the ratio of empirical risk and confidence range.
[0049] The technical problem to be solved by the present invention can also be further achieved through the following technical scheme. For the sonar detection target track prediction method described above, in step (6), a track prediction residual model is established based on the multi-dimensional feature information of the sonar point track. The model input includes n time series residual sequence values of the four-dimensional features of time, azimuth, distance, and characteristic spectrum. The output of the model is the longitude coordinate prediction residual value and the latitude coordinate prediction residual value.
[0050] The technical problem to be solved by the present invention can be further achieved by the following technical solution: for the sonar detection target track prediction method described above, the sonar target prediction system in step (7) is used to encapsulate the target prediction processing function, the nautical chart function, the elevation azimuth display function, the track information display function, the update point track display function and the data recording function into plug-ins respectively, and each plug-in is basically independent of each other;
[0051] After starting to run, the signal processing module is first called to obtain the point track data, and then the target prediction processing plug-in starts the prediction processing. After completion, the target track related information is sent out through the signal and slot mechanism. After the slot functions in the chart plug-in, elevation azimuth display plug-in, track information display plug-in, track update point track display plug-in and data recording plug-in receive the signal, they will call the function to implement the functions of each module.
[0052] Compared with the prior art, the present invention focuses on the sonar target track residual prediction model, and focuses on studying the model input parameters as the processed target parameters. According to the timing characteristics of the target, a time-series-based track residual prediction model is established. The model can be applied to the situation trend analysis of the integrated sonar display and control console to assist the combat decision-making of the overall combat system. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 A method flow chart of the present invention;
[0054] Figure 2 It is a flow chart of an improved grey model prediction embodiment of the present invention;
[0055] Figure 3 This is a flow chart of an embodiment of the support vector machine residual prediction model of the present invention. DETAILED DESCRIPTION
[0056] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0057] Reference Figure 1-3 , a sonar detection target track prediction method, which uses a support vector machine algorithm with good nonlinear solution ability to build a model, combines gray prediction with machine learning methods, uses gray theory to reduce the uncertainty of the original data and the interference of nonlinearity on the prediction model, and measures the prediction trend of its own curve, and then predicts the gray residual value through the machine learning model, and finally synthesizes the gray prediction value and the residual correction value to obtain the final result. This method can be applied to the sonar integrated display and control system, providing data support for target analysis for the overall combat system, thereby assisting the overall combat system to achieve scientific combat decision-making in actual combat.
[0058] like Figure 1 As shown, the sonar detection target track prediction method proposed by the present invention comprises the following steps:
[0059] Step 1, data preparation:
[0060] Prepare single sonar detection target point data, specifically:
[0061] First remove outliers:
[0062] Outliers refer to abnormal values in the measured data. Before the experiment, this part of the data needs to be removed, which can not only improve the clutter recognition rate but also reduce the experimental cost;
[0063] Then normalize the data:
[0064] The sonar point data used in the experiment contains a total of ten-dimensional feature information: batch number, time, direction, distance, depth, heading, speed, characteristic spectrum, frequency, and label. Because the dimensions of each feature information are different, and some feature data values are too large to cause trouble in calculation, the data needs to be normalized before the experiment, and all data sizes are normalized between [-1,1]. The specific formula is:
[0065]
[0066] t norm is the normalized value, t is the value before normalization, and t min is the minimum value before normalization, t max is the maximum value before normalization.
[0067] Step 2, data preprocessing:
[0068] Preprocess the target point data, dig out the inherent change rules of the clean data set and the noisy data set, and extract the main influencing factors, including:
[0069] The grey correlation analysis method mainly determines the correlation degree between factors through the fitting coefficient between the influencing variables and the reference variables;
[0070] 1. In practical problems, it is necessary to non-dimensionalize the influencing parameters with different dimensions. The dimensionless calculation formula is
[0071]
[0072] Where n is the number of sample groups selected. It is the initial sequence of x=(x(1),x(2),...x(n)).
[0073] 2. Calculate the correlation coefficient:
[0074]
[0075] Where ζ(k) is the sequence x i With x 0 The correlation coefficient at time k, ρ∈[0,1] is the resolution coefficient, and ρ=0.5 is taken.
[0076] 3. From the above formula, we can get n correlation coefficients. Since n pieces of information are not easy to measure, we need to average the correlation coefficients.
[0077]
[0078] Among them, r i Represents the sequence x i With x 0degree of correlation.
[0079] 4. Based on the correlation of each factor, several groups of comparison factors with correlation greater than 0.6 with longitude and latitude coordinates are obtained.
[0080] According to the above grey correlation analysis principle and implementation steps, the output variables longitude and latitude coordinates are selected as the main factors; the remaining 10 input variables are influencing factors: batch number, time, direction, distance, depth, heading, speed, characteristic spectrum, frequency, and label. Among them, the number of sample groups n = 800.
[0081] According to the correlation of each factor, four groups of factors with the largest correlation between longitude and latitude coordinates were sorted out, namely time, direction, distance, and characteristic spectrum.
[0082] Step 3, select the starting position of the track:
[0083] Selecting an appropriate track starting position provides conditions for the implementation of the gray track residual correction algorithm, including:
[0084] Target prediction processing mainly includes track initiation module and track maintenance module, and track management is integrated into track initiation and track maintenance modules;
[0085] Implementation of track start module
[0086] The track initiation algorithm based on 3 / 3 logic is selected under the requirement that the track initiation reaction time is less than the system index, that is, within three detection cycles, the track is successfully initiated only when the number of detections that meet the threshold condition in the time window reaches 3. This criterion can achieve rapid track initiation and ensure a high correct track initiation rate.
[0087] Implementation of track maintenance module
[0088] The key to track maintenance is to match the reliable track with the correct measurement. To ensure the effective association between the reliable track and the correct measurement, the target prediction algorithm can be adaptively selected based on the target echo, target track and other information, and the size of the relevant threshold can be adjusted according to the track quality during the prediction process.
[0089] Step 4: Construct a grey prediction model:
[0090] refer to Figure 2 , build a grey prediction model and generate a residual sequence, including:
[0091] The grey model is a relational calculation method for studying the trend of fuzzy and uncertain information. First, the original sequence x(0) is accumulated to obtain the accumulated sequence x(1), and the GM(1,1) prediction model is established to obtain the residual sequence e(0) from time 1 to n, and the predicted value ^x(0)(n+1) at time n+1.
[0092] The prediction results are corrected using the residual sequence. Let e(0)(i) be the residual sequence, i = 1, 2, ..., n. If the prediction order is S, then use e(0)(i-1), e(0)(i-2), ..., e(0)(iS) as the training samples of the support vector regression machine, and use e(0)(i) as the expected value. Through training, using the longitude residual, latitude residual and characteristic spectrum as input, the residual longitude residual and latitude residual at time n+1 are obtained through the support vector regression machine. The predicted value at time n+1 is added to the residual value to obtain the corrected predicted value ^x(0)(n+1)+e(0).
[0093] Step 5, support vector machine residual correction model:
[0094] A residual correction model is built, and the nonlinear and uncertain characteristics of sonar detection targets are considered. A support vector regression algorithm with good numerical convergence and stability is used to correct the target track residual.
[0095] refer to Figure 3 , the residual correction model is constructed, including:
[0096] The target point track has the characteristics of the gray model research object, but the gray prediction method cannot meet the expected requirements for the processing ability of the target click data with strong nonlinearity. Therefore, the gray model is corrected by machine learning method. Considering the nonlinearity and uncertainty characteristics of sonar detection targets, the support vector regression algorithm with good numerical convergence and stability is selected to correct the target track residual.
[0097] Step 6, grey residual prediction model:
[0098] The grey model is integrated with the residual prediction value to obtain an improved grey residual prediction model and generate a track prediction value;
[0099] Step 7, track prediction:
[0100] The grey residual prediction model is applied to the sonar target prediction system to implement the sonar detection target track prediction.
[0101] Among them, a track prediction residual model is established based on the multi-dimensional feature information of sonar tracks. The model input includes n time series residual sequence values of the four-dimensional features of time, azimuth, distance, and characteristic spectrum. The output of the model is the longitude coordinate prediction residual value and the latitude coordinate prediction residual value.
[0102] For sample data, traditional regression models are usually based on the allowable error between the model's input regression function value and the sample value being close to zero, that is, there exists a linear regression function f(x) that can estimate all sample data with ε accuracy.
[0103] If the function f(x) cannot estimate all (x) under the insensitive loss function ε accuracy, i ,y i ), the slack variable ξ can be introduced to calculate the model loss. The support vector machine algorithm has a deviation tolerance, namely the threshold. Let the sample point be outside the dotted area where most of the sample sets are located. Such a point can be regarded as an outlier or a support vector sample value. According to the principle of data statistics, if the distance of the sample point deviating from the dividing line exceeds 3 times the average value of the standard deviation of the sample set, such a point is regarded as an outlier. Excluding outliers, under the definition of the insensitive loss function ε, f(x i ) cannot estimate all sample points, so the slack variable ξ is introduced i ,ξ i * ≥0. Since sonar equipment is easily affected by the environment and its own slight differences, and the acquisition of acoustic wave signals affects the difference between the collected data and the actual value, this phenomenon further verifies that the use of support vector machine regression algorithm to build a prediction track model is effective.
[0104] The support vector machine regression algorithm uses kernel functions to transform nonlinear problems into linear problems in high-dimensional space. Currently, the commonly used kernel functions in the support vector machine regression algorithm are shown in the table:
[0105] Table 1 Kernel function of support vector machine regression algorithm
[0106]
[0107] After multiple comparisons, the kernel function finally selected by the support vector machine regression algorithm for predicting flight paths is the most commonly used rdf kernel function. After using the insensitive loss function and the kernel function, the formula is transformed into:
[0108]
[0109] In the formula, <ω,Φ(x i )> represents the vector product of the input weight and the kernel function, k(x,x i )=Φ(x)·Φ(x i ), the constant C (C>0) is a penalty factor that can adjust the ratio of empirical risk to confidence range.
[0110] The sonar target prediction system is used to encapsulate the target prediction processing function, chart function, elevation azimuth display function, track information display function, track update display function and data recording function into plug-ins, and each plug-in is basically independent of each other. After the sonar target prediction system starts running, the signal processing module is first called to obtain the track data, and then the target prediction processing plug-in starts the prediction processing. After completion, the target track related information is sent out through the signal and slot mechanism. After the slot functions in the chart plug-in, elevation azimuth display plug-in, track information display plug-in, track update track display plug-in and data recording plug-in receive the signal, they will call the function to implement the functions of each module.
[0111] The specific feasibility analysis of the present invention is as follows:
[0112] 1. Big data recorded in real time provides an objective basis
[0113] When the array detects and measures the target, it adopts the method of predicting and recording at the same time. The recording cycle records a set of data every 1ms, and the data can be filtered according to the needs. Therefore, the target motion information (time, direction, distance, characteristic spectrum) recorded after a period of prediction provides a large amount of high-speed, diverse and accurate first-hand information for track prediction. Through these data, the running trajectory characteristics and motion characteristics of the target can be displayed in real time. Big data does not require human investigation and collection. It only needs to continuously accumulate the target track data, use certain algorithms to extract the recorded information, establish a data model, and process the information from aspects such as time, distance, and spatial state, so as to obtain the target's track information in the lost stage, without time lag.
[0114] 2. The target’s motion characteristics provide a realistic possibility
[0115] The prediction method based on historical track refers to using the estimated historical motion information of the target within a period of time to predict the position of the target at the next moment or the next period of time. When the sonar loses lock, the target loses most of its characteristic information. At this time, its most reliable information is the previous motion information of the target. Because the target movement has inertia, its movement speed and acceleration generally do not change much in a short period of time. This algorithm assumes that the target movement follows certain rules. The historical motion information within a period of time can be used to infer the position of the target at a certain moment, thereby making full use of past and current information to maintain the continuity of the track and correctly and dynamically guide the sonar to make real-time adjustments and reconnaissance. When the target is out of the obscured area or the echo signal can be distinguished from noise in the time domain / frequency domain, the problem of data discontinuity and inability to intercept the target caused by re-capturing the target is avoided.
[0116] 3. Curve fitting method reveals the target movement trend
[0117] In order to predict the track when the sonar loses lock based on the existing hydroacoustic measurement data, it is necessary to use a suitable function to approximate (or fit) the known measurement data. The characteristics of this approximation are: moderate accuracy is required; a certain range of errors in the data is allowed; and the changing characteristics or trends at a certain moment or a certain period of time in the future can be predicted. The basic method of approximating discrete data is curve fitting. The purpose of data fitting is to minimize the error function, that is, to solve the convex quadratic programming problem.
[0118] The support vector machine method is a general learning algorithm proposed by Vapnik and Cortes in the 1990s based on the principle of structural risk minimization. This method has shown excellent characteristics and many successful applications. Its basic idea can be summarized as follows:
[0119] (1) Based on a limited number of sample points, a simulation function with high accuracy and low complexity is constructed, while seeking a compromise between the two;
[0120] (2) Since it solves a convex quadratic programming problem, the theoretical solution is the global optimal solution;
[0121] (3) The actual sample points are mapped to other feature spaces through nonlinear transformation, and approximate linear functions are constructed in this space to implement the nonlinear model in the actual problem, so that this method can handle the sample dimension problem well.
[0122] Support vector regression is a machine learning method based on support vector machine to deal with regression problems. It retains all the advantages of support vector machine and is mainly used in situations where the number of samples is limited or small.
[0123] Taking into account the constraints of existing conditions of sonar detection targets and the inherent uncertainty and timing characteristics of the detection targets, a machine learning method of support vector regression with good numerical convergence and stability is selected. Appropriate finite samples can be selected to train known historical measurement data and establish a target track prediction model that conforms to the target's motion trend and characteristics.
Claims
1. A sonar detection target track prediction method, characterized in that: The method comprises the following steps: (1) Data preparation: Prepare single sonar detection target point track data; (2) Data preprocessing: Preprocess the target point data, dig out the inherent change rules of clean data sets and noisy data sets, and extract the main influencing factors; (3) Select the starting position of the track: Selecting an appropriate track starting position provides conditions for the implementation of the gray track residual correction algorithm; (4) Constructing grey prediction model: Build a grey prediction model and generate a residual sequence; (5) Support vector machine residual correction model: A residual correction model is built. Aiming at the nonlinear and uncertain characteristics of sonar detection targets, a support vector regression algorithm with good numerical convergence and stability is used to correct the target track residual. (6) Grey residual prediction model: The grey model is integrated with the residual prediction value to obtain an improved grey residual prediction model and generate a track prediction value; (7) Track prediction: The grey residual prediction model is applied to the sonar target prediction system to implement the sonar detection target track prediction.
2. The sonar detection target track prediction method according to claim 1 is characterized in that: In step (1), the prepared data needs to be subjected to outlier removal and data normalization processing, where outliers refer to abnormal values in the measured data, and data normalization processing is to normalize all data sizes between [ω1, 1]. The specific formula is: t norm is the normalized value, t is the value before normalization, and t min is the minimum value before normalization, t max is the maximum value before normalization.
3. The sonar detection target track prediction method according to claim 1 is characterized in that: In step (2), the data preprocessing method is the grey correlation analysis method, specifically: (2.1) The influencing parameters with different dimensions are dimensionless. The dimensionless calculation formula is: Where n is the number of sample groups selected. is the initial sequence of x=(x(1), x(2), ...x(n)); (2.2) Calculate the correlation coefficient: Where ζ(k) is the sequence x i The correlation coefficient with x0 at time k, ρ∈[0,1] is the resolution coefficient, and ρ=0.5; (2.3) First get n correlation coefficients, and then average the correlation coefficients: Among them, r i Represents the sequence x i Correlation with x0; (2.4) Based on the correlation of each factor, several groups of comparison factors with correlations greater than 0.6 with longitude and latitude coordinates are obtained.
4. The sonar detection target track prediction method according to claim 3 is characterized in that: The output variables longitude and latitude coordinates are selected as the main factors, and the input variables are the influencing factors, including batch number, time, direction, distance, depth, heading, speed, characteristic spectrum, frequency, and label. The number of sample groups n = 800; Then, based on the correlation of each factor, four groups of factors with the largest correlation between longitude and latitude coordinates were sorted out, namely time, direction, distance, and characteristic spectrum.
5. The sonar detection target track prediction method according to claim 1 is characterized in that: In step (3), the target prediction process includes a track initiation module and a track maintenance module, and track management is integrated into the track initiation and track maintenance modules; The track initiation module is used to select the track initiation algorithm based on 3 / 3 logic when the track initiation reaction time is less than the system index requirement, that is, within three detection cycles, the track is successfully initiated only when the number of detections that meet the threshold condition in the time window reaches 3; The track maintenance module adaptively selects the target prediction algorithm based on the target echo and target track information to ensure reliable track matching and correct measurement, and adjusts the size of the relevant threshold according to the track quality during the prediction process.
6. The sonar detection target track prediction method according to claim 1 is characterized in that: The method of step (4) is: First, the original sequence x(0) is accumulated to obtain the accumulated sequence x(1), and the GM(1,1) prediction model is established to obtain the residual sequence e(0) from time 1 to n, and the predicted value ^x(0)(n+1) at time n+1; Then use the residual sequence to correct the prediction result: let e(0)(i) be the residual sequence, i = 1, 2, ..., n. If the prediction order is S, use e(0)(i-1), e(0)(i-2), ..., e(0)(iS) as the training samples of the support vector regression machine, and take e(0)(i) as the expected value; Then, through training, the longitude residual, latitude residual and characteristic spectrum are used as input, and the residual longitude residual and latitude residual at time n+1 are obtained through the support vector regression machine. Finally, the predicted value at time n+1 is added to the residual value to obtain the corrected predicted value ^x(0)(n+1)+e(0).
7. The sonar detection target track prediction method according to claim 1 is characterized in that: In step (5), the method of building the residual correction model is: The grey model is corrected with machine learning method, and the support vector regression algorithm with good numerical convergence and stability is used to correct the target track residual.
8. The sonar detection target track prediction method according to claim 7 is characterized in that: The kernel function finally selected by the support vector machine regression algorithm for predicting flight paths is the most commonly used rdf kernel function. After using the insensitive loss function and the kernel function, the formula is transformed into: In the formula, <ω, Φ(x i )> represents the vector product of the input weight and the kernel function, k(x, x i )=Φ(x)·Φ(x i ), constant C is the penalty factor, C>0, which is used to adjust the ratio of empirical risk and confidence interval.
9. The sonar detection target track prediction method according to claim 1 is characterized in that: In step (6), a track prediction residual model is established based on the multi-dimensional feature information of the sonar track. The model input includes n time series residual sequence values of the four-dimensional features of time, azimuth, distance, and characteristic spectrum. The output of the model is the longitude coordinate prediction residual value and the latitude coordinate prediction residual value.
10. The sonar detection target track prediction method according to claim 1, characterized in that: The sonar target prediction system in step (7) is used to encapsulate the target prediction processing function, the nautical chart function, the elevation and azimuth display function, the track information display function, the update point track display function and the data recording function into plug-ins respectively, and each plug-in is basically independent of each other; After starting to run, the signal processing module is first called to obtain the point track data, and then the target prediction processing plug-in starts the prediction processing. After completion, the target track related information is sent out through the signal and slot mechanism. After the slot functions in the chart plug-in, elevation azimuth display plug-in, track information display plug-in, track update point track display plug-in and data recording plug-in receive the signal, they will call the function to implement the functions of each module.