A data field joint decision graph clustering method based on mesh division

Abstract

The application discloses a data field joint decision graph clustering method based on grid division, relates to the field of radar signal sorting, and comprises the following steps: firstly, the number of grid division in each dimension interval is determined, and the grid side length is calculated; then, the equivalent data points and equivalent quality of each grid are calculated, and the influence factor is set according to the grid side length; then, the data field is constructed according to the influence factor and the equivalent data points, and the potential energy value of each data point is calculated; then, the local density of the data object and the distance of the nearest high-density point are calculated by using the equivalent quality; then, the potential energy distance product of each data object is calculated, and the descending order is arranged; then, the descending rate of the potential energy distance product of each data object is calculated, the data point corresponding to the maximum value is selected as the division line, the number of clustering centers and each clustering center are obtained; finally, the data object clustering is carried out according to the clustering center; the application has the advantages of fast operation speed, high automation degree, less manual setting parameters, and the ability of quickly determining the clustering center and the number of clusters in the radar signal sorting process.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of radar signal sorting, in particular to a data field joint decision graph clustering method based on grid division. BACKGROUND

[0002] The statements in this section merely provide background information related to the present disclosure and can not constitute the prior art.

[0003] Electronic reconnaissance is an important part of electronic warfare, with the richness of radar system and the diversification of signal modulation patterns, the electromagnetic environment faced by electronic reconnaissance is becoming increasingly complex. How to effectively, accurately and quickly sort radar pulse signals in a complex and variable electromagnetic environment has become the focus and difficulty of today's radar signal sorting. As an important part of radar signal sorting, the clustering center and the estimation accuracy of the number of clusters directly affect the subsequent sorting results. However, with the rapid increase of radar pulses and the increasing demand for signal reconnaissance real-time, it is necessary to find a clustering method that is both fast and can determine the number of clustering centers autonomously.

[0004] The clustering algorithms suitable for radar signal sorting mainly include: parameter matching-based clustering algorithm, random search-based clustering algorithm, data field-based clustering algorithm, etc.; among them, the parameter matching-based clustering algorithm is simple to implement, but has high requirements for parameter estimation accuracy and fewer parameters that can be used for sorting, and the sorting error will be accumulated step by step; the random search-based clustering algorithm is very effective for small data sets, but it has no good scalability for large data sets; the data field-based clustering sorting algorithm has good sorting results in complex electromagnetic environments, but the determination of the influence factor needs multiple iterations, which consumes a lot of computing resources. SUMMARY

[0005] The present application aims to: in view of the problems of few parameters that can be used for sorting, large computing consumption and the need for manual intervention to select the number of clustering centers in the clustering algorithms suitable for radar signal sorting, a data field joint decision graph clustering method based on grid division is provided, which uses the data field theory idea to normalize all sorting parameters and unify them as field strength functions, thereby avoiding the high requirement for sorting parameter estimation accuracy and the limitation of the number of sorting parameters; at the same time, combined with the idea of grid division, the problem of consuming a lot of computing resources due to the need for a large number of iterations of the influence factor in the data field algorithm can be solved, and finally the number of clustering centers is automatically determined according to the maximum principle of potential energy distance product drop rate by joint decision graph, thereby realizing fast sorting without manual intervention; thereby solving the above problems.

[0006] The technical scheme of the present application is as follows:

[0007] A data field joint decision graph clustering method based on grid division, comprising:

[0008] Step S1: determining the number of grid divisions K according to the total number N of PDW data objects, and calculating the grid side length l;

[0009] Step S2: diluting the data objects according to the grid division, and calculating the equivalent data points and equivalent mass of each grid;

[0010] Step S3: setting the influence factor σ according to the grid side length;

[0011] Step S4: constructing the data field according to the influence factor σ and the equivalent data points, and calculating the potential energy value of each data point;

[0012] Step S5: calculating the local density of the data objects and the distance to the nearest high-density point by using the equivalent mass after the grid division;

[0013] Step S6: calculating the potential energy distance product of each data object, and arranging them in descending order;

[0014] Step S7: calculating the drop rate of the potential energy distance product of each data object, and selecting the data point corresponding to the maximum value as the dividing line to obtain the number of clustering centers and each clustering center;

[0015] Step S8: clustering the data objects according to the clustering centers.

[0016] Further, the step S1 comprises:

[0017] Step S11: normalizing the N PDW data objects by using the max-min rule;

[0018] Step S12: determining the number of grid divisions according to the total number of data and the PDW parameter dimension;

[0019] Step S13: calculating the grid side length l according to the normalized data;

[0020] Further, the normalization process comprises:

[0021]

[0022] In the formula:

[0023] PDW m (n) represents the nth data of the mth dimension parameter in the PDW;

[0024] PDW m (all) represents all data of the mth dimension parameter;

[0025] The step S12 comprises:

[0026]

[0027] wherein:

[0028] d represents the PDW parameter dimension;

[0029] The step S13 comprises:

[0030] l = 1 / K.

[0031] Further, the step S2 comprises:

[0032] For the data in the same grid, the mean value is calculated as the equivalent data point x in the grid; the total number of data objects divided into the same grid (m, n) is the equivalent mass p of the grid x→(m,n) The grid to which no corresponding data object is mapped is defined as having an equivalent mass of 0.

[0033] Further, the step S4 comprises:

[0034] Step S41: According to the principle that 99.74% of the data objects are contained in the 3σ interval of the Gaussian function, the field strength function of each data is optimized to reduce the calculation amount;

[0035] Step S42: According to the optimized field strength function, the potential energy value of each data point is calculated.

[0036] Further, the step S41 comprises:

[0037]

[0038] wherein:

[0039] f y (x) represents the field strength function of the data object y to the data object x;

[0040] d(x, y) represents the Euclidean distance between the data object x and the data object y;

[0041] The step S42 comprises:

[0042]

[0043] wherein:

[0044] F y represents the potential energy value of the data object y.

[0045] Further, the step S5 comprises:

[0046] Step S51: Gaussian blur is performed once on the equivalent mass;

[0047] Step S52: According to the scope of the two-dimensional Gaussian blur kernel, the local density of the data object is determined;

[0048] Step S53: determining the nearest density point distance of the data object according to the local density of the data object.

[0049] Further, the two-dimensional Gaussian blur kernel in step S52 is:

[0050]

[0051] The step S52 includes:

[0052] The Gaussian kernel scope S is a k×k matrix, and the local density of the data object x is:

[0053] ρ x = gauss(ρ x→(m,n) )

[0054] In the formula:

[0055] gauss(·) represents a Gaussian blur operation;

[0056] The step S53 includes:

[0057] The nearest density point distance δ x of the data object x is defined as:

[0058]

[0059] It is stipulated that if there is a data object with equivalent mass greater than the data object x within the Gaussian kernel scope S, the smallest Euclidean distance from the data object x in the data object set is found as the nearest density point distance, otherwise the largest Euclidean distance from the data object x in the entire data set is taken as the nearest density point distance.

[0060] Further, the step S6 includes:

[0061] Step S61: normalizing the potential energy value and the nearest density point distance of each data object and calculating the potential energy distance product;

[0062] Step S62: based on the calculation result, performing descending arrangement.

[0063] Further, the S61 includes:

[0064] τ i = F norm (i)·δ norm (i)

[0065] In the formula:

[0066] F norm (i) represents the normalized potential energy value of the data object;

[0067] delta norm (i) represents the nearest large density point distance of the normalized data object;

[0068] The step S7 comprises:

[0069] The descending rate of each data object potential energy distance product is defined as:

[0070]

[0071] In the formula:

[0072] tau prime i represents the descending rate of the potential energy distance product of the data object i.

[0073] Compared with the prior art, the beneficial effects of the present application are:

[0074] A data field joint decision graph clustering method based on grid division, based on the grid division idea, the dense pulse description word data is diluted, and the influence factor in the data field algorithm is quickly determined according to the length of the grid, so as to avoid the process of iteratively calculating the influence factor; at the same time, the local density matrix is constructed according to the grid division, and the density matrix is smoothed by Gaussian blur, so as to correct the local density quantization error introduced by the grid division; finally, the data field and the decision graph theory are combined for joint clustering, the potential value of the data object and the distance of the nearest large density point are used to realize the automatic determination of the clustering center and the number of clustering centers, and the potential risk of manual determination of the number of clustering centers is avoided; it has the characteristics of fast operation speed, high automation degree and less human setting parameters, and can quickly determine the clustering center and the number of clusters in the radar signal sorting process. BRIEF DESCRIPTION OF DRAWINGS

[0075] Figure 1 It is a flow chart of a data field joint decision graph clustering method based on grid division;

[0076] Figure 2 It is a two-dimensional distribution graph of radar signal pulse width and amplitude parameters in Example 2;

[0077] Figure 3 It is a three-dimensional schematic diagram of the potential value of the equivalent data after grid division in Example 2;

[0078] Figure 4a It is a three-dimensional schematic diagram of the equivalent quality of the equivalent data after grid division in Example 2;

[0079] Figure 4b It is a three-dimensional schematic diagram of the equivalent quality after Gaussian blur in Example 2;

[0080] Figure 5 It is a potential energy distance graph after descending order sorting in Example 2;

[0081] Figure 6 Potential energy distance rate of decline graph for Example 2;

[0082] Figure 7 Grid-based data field joint decision graph clustering results for Example 2. DETAILED DESCRIPTION

[0083] It should be noted that the relational terms herein, such as first and second, and the like, are used solely to distinguish one from another entity or action without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0084] The features and advantages of the present application will be further described in the following embodiments.

[0085] Example 1

[0086] Referring to Figure 1 A grid-based data field joint decision graph clustering method, specifically comprising the following steps:

[0087] Step S1: Determine the number of grid partitions K for each dimension interval according to the total number N of PDW data objects, and calculate the grid side length l; it should be noted that the PDW is a radar signal pulse description word;

[0088] Step S2: Dilute data objects according to grid partition, and calculate equivalent data points and equivalent quality of each grid;

[0089] Step S3: Set the influence factor σ according to the grid side length; preferably, the influence factor can be set to σ = 1.01 x l according to the empirical value;

[0090] Step S4: Construct a data field according to the influence factor σ and the equivalent data points, and calculate the potential energy value of each data point;

[0091] Step S5: Calculate the local density of data objects and the distance to the nearest high-density point using the equivalent quality after grid partition;

[0092] Step S6: Calculate the potential energy distance product of each data object and arrange it in descending order;

[0093] Step S7: Calculate the descending rate of each data object potential energy distance product, and select the data point corresponding to the maximum value as the demarcation line to obtain the number of clustering centers and each clustering center;

[0094] Step S8: Cluster the data objects according to the clustering centers.

[0095] In this embodiment, specifically, the step S1 comprises:

[0096] Step S11: Normalize the N PDW data objects by using the max-min rule;

[0097] Step S12: Determine the number of grid divisions according to the total number of data and the PDW parameter dimension;

[0098] Step S13: Calculate the grid side length l according to the normalized data;

[0099] In this embodiment, specifically, the normalization processing comprises:

[0100]

[0101] In the formula:

[0102] PDW m (n) represents the nth data of the mth dimension parameter in the PDW;

[0103] PDW m (all) represents all data of the mth dimension parameter;

[0104] The step S12 comprises:

[0105]

[0106] In the formula:

[0107] d represents the PDW parameter dimension;

[0108] The step S13 comprises:

[0109] l = 1 / K

[0110] That is, it is easy to know that the side length l is the reciprocal of K, that is, l = 1 / K.

[0111] In this embodiment, specifically, the step S2 comprises:

[0112] For the data in the same grid, the mean value is taken as the equivalent data point x in the grid; the total number of data objects divided into the same grid (m, n) is taken as the equivalent mass p of the grid x→(m,n) The equivalent mass of the grid to which no corresponding data object is mapped is defined as 0.

[0113] In the embodiment, specifically, the step S4 comprises:

[0114] Step S41: According to the principle that 99.74% of data objects are contained in the 3σ interval of the Gaussian function, the field strength function of each data is optimized to reduce the calculation amount.

[0115] Step S42: According to the optimized field strength function, the potential energy value of each data point is calculated.

[0116] In the embodiment, specifically, the step S41 comprises:

[0117]

[0118] In the formula:

[0119] f y (x) represents the field strength function of data object y to data object x;

[0120] d(x,y) represents the Euclidean distance between data object x and data object y;

[0121] The step S42 comprises:

[0122]

[0123] In the formula:

[0124] F y represents the potential energy value of data object y.

[0125] In the embodiment, specifically, the step S5 comprises:

[0126] Step S51: Considering the quantization error of grid division, the equivalent mass needs to be Gaussian blurred once;

[0127] Step S52: According to the scope of the two-dimensional Gaussian blur kernel, the local density of the data object is determined;

[0128] Step S53: According to the local density of the data object, the distance of the nearest high-density point of the data object is determined.

[0129] In the embodiment, specifically, the two-dimensional Gaussian blur kernel in the step S52 is:

[0130]

[0131] The step S52 comprises:

[0132] The Gaussian kernel scope S is a k×k matrix, and the local density of the data object x is:

[0133] ρ x= gauss(p x→(m,n) )

[0134] wherein:

[0135] gauss(·) represents a Gaussian blur operation;

[0136] The step S53 comprises:

[0137] The nearest large density point distance δ of the data object x is defined as: x The nearest large density point distance δ of the data object x is defined as:

[0138]

[0139] It is provided that if there is a data object with equivalent mass greater than the data object x within the Gaussian kernel scope S, the minimum Euclidean distance of the data object set with the data object x is found as the nearest large density point distance, otherwise the maximum Euclidean distance of the entire data set relative to the data object x is taken as the nearest large density point distance.

[0140] In this embodiment, specifically, the step S6 comprises:

[0141] Step S61: normalizing the potential energy value and the nearest large density point distance of each data object and calculating the potential energy distance product;

[0142] Step S62: performing descending arrangement based on the calculation result.

[0143] In this embodiment, specifically, the S61 comprises:

[0144] τ i = F norm (i) · δ norm (i)

[0145] wherein:

[0146] F norm (i) represents the potential energy value of the normalized data object;

[0147] δ norm (i) represents the nearest large density point distance of the normalized data object;

[0148] The step S7 comprises:

[0149] The drop rate of the potential energy distance product of each data object is defined as:

[0150]

[0151] wherein:

[0152] τ′ i represents the drop rate of the potential energy distance product of the data object i.

[0153] Example two

[0154] Example two is based on the simulation verification of the data field joint decision graph clustering method based on grid division proposed in example one, and the simulation platform is Matlab R2019b; the present application will be described in detail in combination with specific implementation schemes and drawings.

[0155] Step one: considering the possibility of missing angle signal, only the pulse width (PW) and pulse amplitude (PA) of the radar pulse signal are selected to construct the data object set, the parameter dimension is 2, and the total number of data objects is N = 1726; 4 different radar signal parameter pairs are set, the signal pulse width and pulse amplitude are set as [20us, 5GHz], [25us, 5.3GHz], [32us, 5.6GHz], [2us, 4.7GHz], and the number of radar samples is 724, 334, 334, and 334 respectively; the pulse width and pulse amplitude parameters are normalized by max-min rule, and the two-dimensional distribution diagram of the normalized radar signal pulse width and amplitude parameters is as shown in Figure 2 ; the number of grid division is and K is rounded up; the grid length is

[0156]

[0157] Step two: dilute the data objects according to the grid division, and calculate the equivalent data points and equivalent quality of each grid.

[0158] The mean value of the data in the same grid is taken as the equivalent data point x in the grid. The total number of data objects divided into the same grid (m, n) is taken as the equivalent quality p of the grid x→(m,n) The equivalent quality of the grid to which the corresponding data object is not mapped is defined as 0, and the three-dimensional schematic diagram of the equivalent quality is as shown in Figure 4a .

[0159] Step three: set the influence factor σ according to the grid length, and the influence factor is

[0160] Step four: construct the data field according to the calculated influence factor and equivalent data point, and calculate the potential value of each data point.

[0161] According to the principle that 99.74% of the data objects are contained in the 3σ interval of the Gaussian function, the field strength function of each data is optimized to reduce the calculation amount:

[0162]

[0163] The potential value of the data object y is:

[0164]

[0165] The potential energy distribution diagram of the entire data domain is as follows Figure 3 As shown.

[0166] Step 5: Calculate the local density ρ of the data object using the equivalent mass after mesh generation. x Distance δ to the nearest high-density point x .

[0167] Considering the quantization error in mesh generation, the equivalent quality needs to be subjected to Gaussian blurring. The two-dimensional Gaussian blur kernel is:

[0168]

[0169] If the scope S of the Gaussian kernel is a 5×5 matrix, then the local density of the data object x is:

[0170] ρ x =gauss(ρ x→(m,n) )

[0171] Specifically, the local density matrix result after Gaussian blurring is as follows: Figure 4b As shown.

[0172] The distance δ of the nearest high-density point of data object x x Defined as:

[0173]

[0174] The rule is that if there is a data object with an equivalent mass greater than data object x within the scope S of the Gaussian kernel, then the Euclidean distance with the smallest data object x within the set of data objects is taken as the nearest high-density point distance; otherwise, the Euclidean distance with the largest data object x in the entire set is taken as the nearest high-density point distance.

[0175] Step 6: Calculate the potential distance product of each data object and sort them in descending order.

[0176] First, the potential energy value and the distance to the nearest high-density point of each data object are normalized, and the potential energy distance product is calculated:

[0177] τ i =F norm (i)·δ norm (i)

[0178] The results after sorting the potential energy distance product in descending order are as follows: Figure 5 As shown.

[0179] Step seven: calculate the rate of decline of each data object potential energy distance product, and select the maximum value corresponding to the data point as the demarcation line, to obtain the number of cluster centers and each cluster center, and the potential energy distance product decline rate calculation result is shown in Figure 6

[0180] The maximum point of the potential energy distance product decline rate is selected as the data classification boundary, and the number of points before the maximum value corresponds to the center of each class cluster. The number of cluster centers is 4.

[0181] Step eight: perform K-means clustering on the data objects according to the cluster centers, and the clustering result is shown in Figure 7

[0182] The above-described embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the technical concept of the present application, a number of modifications and improvements can be made, which are within the protection scope of the present application.

[0183] This background section is provided to generally present the context of the application, the work of the current named inventors, the work described in this background section to the extent that it is described, and the aspects described in this section at the time of filing, neither expressly nor implicitly, are recognized as prior art of the present application.​​

Claims

1. A data field joint decision graph clustering method based on grid partitioning, characterized in that, include: Step S1: Based on the total number of PDW data objects N Determine the number of grid divisions for each dimension. K And calculate the grid side length. l ; Step S2: Dilute the data objects according to the grid, and calculate the equivalent data points and equivalent quality of each grid; Step S3: Set the influence factor based on the grid side length ; Step S4: Based on the impact factor Construct a data field using equivalent data points and calculate the potential energy value of each data point; Step S5: Calculate the local density of the data object and the distance to the nearest high-density point using the equivalent mass after mesh generation; Step S6: Calculate the potential distance product of each data object and sort them in descending order; Step S7: Calculate the rate of decrease of the potential distance product of each data object, and select the data point corresponding to the maximum value as the boundary line to obtain the number of cluster centers and each cluster center; Step S8: Cluster the data objects based on the cluster centers; Step S1 includes: Step S11: For N Each PDW data object is normalized using the max-min rule; Step S12: Determine the number of grid divisions based on the total number of data points and the PDW parameter dimensions; Step S13: Calculate the grid edge length based on the normalized data. l； The normalization process includes: In the formula: Indicates the first in PDW m The first dimension parameter n One data point; Indicates the first m All data for the dimension parameter; Step S12 includes: In the formula: d Indicates the PDW parameter dimension; Step S13 includes: Step S2 includes: The mean value of the data within the same grid is taken as the equivalent data point within that grid. x ; will be divided into the same grid The total number of data objects in the grid is used as the equivalent quality of the grid. The mesh to which no corresponding data object is mapped specifies that its equivalent quality is 0; Step S4 includes: Step S41: Based on the Gaussian function Based on the principle that 99.74% of the data objects are contained within the interval, the field strength function of each data point is optimized to reduce the amount of computation; Step S42: Calculate the potential energy value of each data point based on the optimized field strength function; Step S41 includes: In the formula: Represents data objects y For data objects x The electric field strength function; Represents data objects x and data objects y Euclidean distance; Step S42 includes: In the formula: F y Represents data objects y The potential energy value.

2. The data field joint decision graph clustering method based on grid partitioning according to claim 1, characterized in that, Step S5 includes: Step S51: Apply a Gaussian blur to the equivalent mass; Step S52: Determine the local density of the data object based on the scope of the two-dimensional Gaussian blur kernel; Step S53: Determine the distance to the nearest high-density point of the data object based on the local density of the data object.

3. The data field joint decision graph clustering method based on grid partitioning according to claim 2, characterized in that, The two-dimensional Gaussian blur kernel in step S52 is: Step S52 includes: Gaussian kernel scope S for The matrix, then the data object x The local density is: In the formula: Indicates Gaussian blur operation; Step S53 includes: Data Objects x Distance of the nearest high-density point Defined as: The rule is that if in the Gaussian kernel domain S Memory has an equivalent quality greater than that of data objects x If the data object is a collection of data objects, then find the matching data object within that collection. x The minimum Euclidean distance is used as the distance to the nearest high-density point; otherwise, the distance relative to the data object is used within the entire dataset. x The largest Euclidean distance is taken as the distance to the nearest high-density point.

4. The data field joint decision graph clustering method based on grid partitioning according to claim 1, characterized in that, Step S6 includes: Step S61: Normalize the potential energy value and the distance to the nearest high-density point for each data object and calculate the potential energy distance product; Step S62: Sort the results in descending order based on the calculation results.

5. The data field joint decision graph clustering method based on grid partitioning according to claim 4, characterized in that, S61 includes: In the formula: This represents the potential energy value of the data object after normalization. This represents the distance to the nearest high-density point of a normalized data object. Step S7 includes: The rate of decrease of the potential distance product of each data object is defined as: In the formula: Represents data objects i The rate of decrease of the potential energy distance product.

Images