A data-driven hierarchical recursive weighted sensor information fusion recognition method

Abstract

The application discloses a data-driven hierarchical recursive weighted sensor information fusion recognition method, which fuses multi-source sensor information by inputting different information of multiple sensors, processes high conflict information in the fusion process, and obtains the recognition confidence of the target after the fusion processing. The key of the whole application is to measure the conflict degree of sensor information, calculate the weight of each evidence, and perform hierarchical recursive weighted fusion. The application can effectively process the conflict among multiple evidences, more effectively measure the direct conflict degree of different evidence groups, realize efficient fusion of heterogeneous sensor information, and improve the fusion recognition confidence of the target. For target recognition results from multiple sensors, the fusion result uncertainty given by the application is lower, and the final result is more explicit. The application implementation process does not depend on the original data of the underlying sensor, the fusion process is efficient and rapid, and the complexity is low.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of information fusion, and relates to a data-driven hierarchical recursive weighted sensor information fusion identification method. BACKGROUND

[0002] Information fusion technology is a multi-sensor data fusion technology, which can fuse different information about the same target from multiple sensors to obtain a comprehensive and integrated evaluation of the target. The advantage of information fusion technology lies in its ability to comprehensively utilize various information to obtain a more objective and more essential understanding of things or targets. Information fusion technology is constantly developing, and it is widely used in many fields such as target identification, classification, decision analysis, etc., and has achieved good results. In view of the multi-sensor target type fusion identification demand in complex environment, the target type fusion identification research is carried out, aiming to realize the type identification based on multi-source heterogeneous sensor information fusion. SUMMARY

[0003] (I) Invention purpose

[0004] The purpose of the present application is: multiple sensors exist multiple identifications of the same target, generating multi-source information about the same target; secondly, due to the differences in detection performance, field of view range and working mechanism of each sensor, different sensors may identify different results of the same target, including conflicting results, so the multi-sensor target identification result has high uncertainty and conflict, in the information fusion, the redundant information of multi-sensor target is integrated, the uncertainty and high conflict problem of information source is solved, the multi-source heterogeneous sensor information fusion target identification is realized, and the identification confidence of the sensor to the target is improved.

[0005] (II) Technical scheme

[0006] In order to solve the above technical problems, the present application provides a data-driven hierarchical recursive weighted sensor information fusion identification method, comprising the following steps:

[0007] Step 1: modeling the identification results of the unknown target by multiple sensors using evidence theory, and performing discount calculation on the obtained evidence on the power set of the identification framework, to finally obtain n evidences output by the sensor;

[0008] Step 2: calculating the similarity of any two evidences to obtain a distance matrix between the evidences, and performing evidence clustering using k-means algorithm according to the distance between the evidences to obtain evidence class clusters;

[0009] Step 3: obtaining the inter-class conflict distribution of the two evidence class clusters by Bootstrap sampling method, and calculating the conflict degree of the evidence class clusters based on the weighted weights obtained by the maximum entropy formula based on the ordered weighted average idea;

[0010] Step four: for all evidences in the evidence cluster, calculate the average of the information entropy of all evidences based on their vector forms;

[0011] Step five: calculate the weight of the evidence cluster according to the average information entropy in the evidence cluster and the conflict degree between clusters;

[0012] Step six: repeat steps two to five for each evidence cluster until each cluster contains only one evidence, and based on the obtained hierarchical recursive evidence cluster weights, obtain the weight of each of the n evidences through layer-by-layer multiplicative aggregation, and calculate the average evidence of the n evidences according to the weight distribution;

[0013] Step seven: fuse the obtained average evidence with itself n-1 times according to the Dempster combination rule to obtain the fusion result of all sensor information;

[0014] Step eight: convert the evidence on the power set space after fusion to the recognition framework, obtain the probability distribution of the fusion recognition result based on the game probability conversion, and select the class with the maximum probability as the final recognition result of the multi-sensor information fusion.

[0015] (Three) beneficial effects

[0016] The data-driven hierarchical recursive weighted sensor information fusion recognition method provided by the above technical solution has the following advantages:

[0017] (1) can effectively handle the conflict between multiple evidences. The classical evidence theory cannot handle the fusion problem of highly conflicting evidences, and for this, the present application provides a new evidence conflict measurement scheme, which can more effectively measure the direct conflict degree of different evidence groups.

[0018] (2) can improve the fusion recognition confidence of the target. For the target recognition results from multiple sensors, the fusion result given by the present application has lower uncertainty, and the final result is more explicit.

[0019] (3) can realize efficient fusion of heterogeneous sensor information. The present application does not depend on the original data of the underlying sensor in the implementation process, and the fusion process is efficient and rapid with low complexity. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The hierarchical recursive weighted sensor information fusion recognition flowchart of the present application. DETAILED DESCRIPTION

[0021] In order to make the purpose, content and advantages of the present application clearer, the specific embodiments of the present application are further described in detail below in combination with the drawings and examples.

[0022] The embodiment utilizes multi-source heterogeneous sensor information to provide a data-driven hierarchical recursive weighted information fusion recognition method, and the recognition confidence of the target after fusion processing is obtained by inputting different information of multiple sensors. The embodiment needs to fuse multi-source sensor information, and highly conflicting information needs to be processed in the fusion process. Therefore, the key of the whole invention is to measure the conflict degree of sensor information, calculate the weight of each evidence, and perform hierarchical recursive weighted fusion.

[0023] The data-driven hierarchical recursive weighted sensor information fusion recognition method of the embodiment includes the following processes:

[0024] Step one: the multi-sensor fusion recognition system performs type recognition on the unknown target T, and the possible types of the target T are t1, t2, …, t s The system is equipped with n sensors S i , i = 1, …, n, and the recognition accuracies of the sensors are R i , i = 1, …, n. For the target to be recognized T, the recognition result of the sensor S i is {t i , p i}, i = 1, …, n, wherein t i represents that the sensor S i recognizes the type of the target T as t i , and p i is the probability that the target T is recognized by the sensor S i as type t i .

[0025] Let the recognition framework be Ω = {t1, t2, …, t s}, and the power set of Ω be The recognition result of the sensor S i for the unknown target T is {t i , p i}, and let represent the set of s-1 target types in the recognition framework Ω except t i . The above recognition result is modeled according to the evidence theory as follows.

[0026]

[0027] According to the recognition accuracy R i of the sensor S i , the above evidence is discounted on the power set 2 Ω of the recognition framework:

[0028]

[0029] The above formulas (1) and (2) are combined and represented as:

[0030]

[0031] The physical meaning of formula (3) is that, for the recognition result of sensor S i , according to its recognition accuracy R i , it is discounted to obtain the evidence m i ; in m i , the confidence of the target T being of type t i is R i ·p i , the confidence of the target T being of one of the types in type 2 is R i ·(1-p i ), and the confidence of the target T being of one of the types in type 2 Ω is 1-R i .

[0032] Step two: for the obtained n evidences m1,..., m n , express them as vectors in the power set space, i = 1,..., n. For the obtained n evidence vectors, calculate the cosine similarity between any two evidence vectors according to the following formula:

[0033]

[0034] Based on the similarity of any two evidences according to formula (4), define the distance between any two evidences m i , m j :

[0035] d ij = 1-cosθ (5)

[0036] Thus, the distance matrix between the n evidences m1,..., m n is obtained:

[0037]

[0038] Based on the distance matrix Dis, use the k-means algorithm to cluster the evidences. Set the number of clustering clusters to c = 2, and denote the obtained clustering cluster result

[0039] Step three: use the Bootstrap sampling method to obtain the inter-class conflict distribution of the evidence clusters Cluster1 and Cluster2. Set the sampling number q = 100000 times, perform random sampling with replacement, and each time extract one evidence from the clusters Cluster1 and Cluster2, denoted as and The distance between them is calculated as follows:

[0040]

[0041] where D is 2 n ×2 n matrix, |A i ∩A j | represents the number of identical elements in set A i and A j , |A i ∪A j | represents the number of elements in the union set of A i and A j , is the vector form of the evidence and .

[0042] After q sampling, q distance values are obtained The maximum, 3 / 4 quantile, 1 / 2 quantile, 1 / 4 quantile, and minimum of the q distance values are calculated, respectively denoted as d1, d2, d3, d4, and d5. In addition, the variance of the q distance values is calculated, denoted as σ d .

[0043] The main idea of the evidence cluster conflict degree calculation method is: 1) using ordered weighted average to aggregate the five conflict values d1, d2, d3, d4, and d5, without other parameters, which is a purely data-driven way; 2) the weighted weight is obtained by the maximum entropy formula, taking the variance σ d as the orness degree of the weight distribution, so that the weight distribution reflects the influence of the dispersion of the distance value distribution.

[0044] Based on the idea of ordered weighted average, the conflict degree between the evidence cluster Cluster1 and Cluster2 is calculated according to the following formula:

[0045]

[0046] where the weight w k , k = 1, 2, …, 5, is obtained by the following maximum entropy formula:

[0047]

[0048] Step four: for all evidence in the evidence cluster Cluster1, based on its vector form, the average value of the information entropy of all evidence is calculated:

[0049]

[0050] Similarly, the average information entropy of all evidences in Cluster2 is calculated as follows:

[0051]

[0052] Step five: The weights η1, η2 of Cluster1 and Cluster2 are calculated by integrating the conflict degree between clusters and the average information entropy of evidence clusters, as follows:

[0053]

[0054]

[0055] Step six: The process of steps two to five is repeated for each evidence cluster until each evidence cluster contains only one evidence. Based on the obtained weights of evidence clusters in the hierarchical recursion, the n evidences m1, …, mn of the sensor output are obtained by multiplicative aggregation layer by layer. n Respective weights The n weights satisfy the normalization condition

[0056] For all n evidences m1, …, mn, the average evidence is calculated according to the weight distribution n

[0057]

[0058] Step seven: The obtained average evidence m is fused with itself n-1 times according to the Dempster combination rule to obtain the fusion result m of all sensor information.

[0059]

[0060] wherein the Dempster combination rule is The calculation process is as follows:

[0061] m1, m2 are two independent evidences on the recognition framework Ω, and the combined evidence of m1 and m2 is

[0062]

[0063] wherein

[0064]

[0065] Step eight: The evidence m in the power set space 2 Ω after fusion is converted to the recognition framework Ω, and the formula is as follows:​​​

[0066]

[0067] For the obtained evidence m, the probability distribution of the fusion recognition result is obtained based on the game probability conversion

[0068]

[0069] |A| represents the number of elements in set A. The class t with the maximum BetP(t) probability is selected max As the final recognition result of multi-sensor information fusion,

[0070]

[0071] Among them, the recognition confidence of class t max is BetP(t max ).

[0072] The above only describes the preferred embodiments of the present application, and it should be noted that for those skilled in the art, without departing from the technical principles of the present application, a number of improvements and modifications can be made, and these improvements and modifications should be considered as the protection scope of the present application.

Claims

1. A data-driven hierarchical recursive weighted sensor information fusion recognition method, characterized in that, The method comprises the following steps: Step one: the recognition result of the unknown target by the multi-sensor is modeled by using the evidence theory, and the obtained evidence is calculated on the power set of the identification framework to obtain n evidences of the sensor output; Step two: the similarity of any two evidences is calculated to obtain a distance matrix between the evidences, and the k-means algorithm is used to cluster the evidences according to the distance between the evidences to obtain evidence clusters; Step three: the Bootstrap sampling method is used to obtain the inter-class conflict distribution of the two evidence clusters, the weighted weight obtained by using the maximum entropy formula is used to calculate the conflict degree of the evidence clusters based on the ordered weighted average idea; Step four: the average value of the information entropy of all the evidences in the evidence cluster is calculated based on the vector form of the evidences; Step five: the weight of the evidence cluster is calculated according to the average information entropy in the evidence cluster and the conflict degree between the clusters; Step six: steps two to five are repeated for each evidence cluster until each cluster contains only one evidence, and the weight of each evidence is obtained by using the hierarchical recursive evidence cluster weight obtained by multiplicative aggregation layer by layer, and the average evidence of the n evidences is calculated according to the weight distribution; Step seven: the obtained average evidence is fused with itself n-1 times according to the Dempster combination rule to obtain the fusion result of all the sensor information; Step eight: the evidence on the power set space after fusion is converted to the identification framework, the probability distribution of the fusion recognition result is obtained based on the game probability conversion, and the class with the maximum probability is selected as the final recognition result of the multi-sensor information fusion.

2. The data-driven hierarchical recursive weighted sensor information fusion recognition method of claim 1, wherein, In step one, a multi-sensor fusion recognition system is used to identify the type of unknown target T, and the possible types of target T are t1, t2, …, t s One of the types; the system is equipped with n sensors S i , i = 1, …, n, and the recognition accuracy of each sensor is R i , i = 1, …, n; for the target to be identified T, the recognition result of sensor S i is {t i , p i}, i = 1, …, n, where t i represents that sensor S i recognizes the type of target T as t i , and p i is the probability that target T is recognized by sensor S i as type t s . Let the recognition frame be Ω = {t1, t2, …, t s} and let the power set of Ω be Given a sensor S i The recognition result of the unknown target T is {t i ,p i} and let denote the set of s-1 target types in the recognition frame Ω other than t i The above recognition result is modeled by the evidence theory as follows: According to sensor S i Recognition accuracy R i Regarding the above evidence in the power set 2 of the identification framework Ω The discount is calculated above: The above formulas (1) and (2) are combined and represented as: The physical meaning of formula (3) is that, for the recognition result of sensor S i , the recognition accuracy R i is discounted to obtain the evidence m i ; in m i , the confidence of the target T being of type t i is R i ·p i , the confidence of the target T being of one of types 1 is R i ·(1-p i ), and the confidence of the target T being of one of types 2 Ω is 1-R i .

3. The data-driven, hierarchical, recursively weighted sensor information fusion recognition method of claim 2, wherein, In step two, for the obtained n pieces of evidence m1, …, mn n , which is represented as a vector in the power set space, For the obtained n pieces of evidence vectors, the cosine similarity between any two pieces of evidence vectors is calculated as follows: The cosine similarity between any two pieces of evidence vectors is calculated as follows: Based on the similarity of any two evidences, the distance between any two evidences m i ,m j is defined as: d ij = 1 - cos θ (5) Thus, we obtain the distance matrix between n pieces of evidence m1,..., mn: n Thus, we obtain the distance matrix between n pieces of evidence m1,..., mn: Based on the distance matrix Dis, the evidence is clustered by using the k-means algorithm; the number of clusters of clustering is set to c=2, and the clustering cluster result obtained is denoted as 4. The data-driven hierarchical recursive weighted sensor information fusion recognition method of claim 3, wherein, In step three, Bootstrap sampling method is used to obtain the inter-class conflict distribution of evidence clusters Cluster1 and Cluster2; let the sampling number q = 100000 times, perform random sampling with replacement, and each time extract one evidence from clusters Cluster1 and Cluster2, respectively, and mark them as and The distance between them is calculated as follows: Where D is 2 n ×2 n The matrix, |A i ∩A j | represents set A i With A j The number of identical elements in |A i ∪A j | represents set A i With A j The number of elements in the union set. As evidence and The vector form; After q samplings, q distance values are obtained The maximum, 3 / 4 quantile, 1 / 2 quantile, 1 / 4 quantile, and minimum of the q distance values are calculated, denoted as d1, d2, d3, d4, and d5, respectively. In addition, the variance of the q distance values is calculated, denoted as σ d ;​ Based on the ordered weighted average idea, the conflict degree of the evidence cluster Cluster1 and Cluster2 is calculated according to the following formula: where the weights w k k = 1,2,...,5, are obtained from the following maximum entropy equation:

5. The data-driven, hierarchical, recursively weighted sensor information fusion recognition method of claim 4, wherein, In step four, the average value of the information entropy of all the evidences in the evidence cluster Cluster1 is calculated based on the vector form of the evidences: Similarly, the average value of the information entropy of all the evidences in the evidence cluster Cluster2 is calculated:

6. The data-driven, hierarchical, recursively weighted sensor information fusion recognition method of claim 5, wherein, In step five, the weight η1, η2 of the evidence cluster Cluster1 and Cluster2 is calculated by comprehensively considering the conflict degree between the clusters and the average information entropy of the evidence cluster, and the method is as follows:

7. The data-driven hierarchical recursive weighted sensor information fusion recognition method of claim 6, wherein, In step six, the process of steps two to five is repeated for each evidence class cluster until each evidence class cluster contains only one evidence; based on the resulting hierarchical recursive evidence class cluster weights, the n evidences m1,...,m n the respective weights the n weights satisfy a normalization condition For all n pieces of evidence m1,...,m n , the average evidence is computed according to their weight distribution ​ 8. The data-driven, hierarchical, recursively weighted sensor information fusion recognition method of claim 7, wherein, In step seven, the obtained average evidence m is fused with itself n-1 times according to the Dempster combination rule to obtain the fusion result m of all the sensor information: where the Dempster combination rule The calculation proceeds as follows: m1, m2 are two independent pieces of evidence on the recognition framework Ω, and the combined evidence of m1 and m2 is is: Wherein, 9. The data-driven, hierarchical, recursively weighted sensor information fusion recognition method of claim 8, wherein, In step eight, the fused power set space 2 Ω The evidence m on the distribution is transformed into the recognition frame Ω, as follows: For the obtained evidence m, the probability distribution of the fusion recognition result is obtained based on the game probability conversion |A| represents the number of elements in set A; select the class t with the maximum probability BetP(t) max As the final recognition result of multi-sensor information fusion, wherein the class t max has an identification confidence BetP(t max ).

10. An application of the data-driven hierarchical recursive weighted sensor information fusion recognition method according to any one of claims 1-9 in the field of information fusion technology.

Images