Confidence normalization method for target recognition attributes

Abstract

The invention discloses a confidence normalization method for target recognition attributes. The method comprises the following steps: converting discrete membership into confidence; the continuous membership degree is converted into confidence; and calculating the comprehensive identification confidence of the multi-source sensor. The confidence degree of multiple sensors under the same dimension is calculated through discrete and continuous membership degrees of sensor recognition results, then the comprehensive recognition confidence degree of the multi-source sensor system is calculated, generation of a decision-level fusion strategy of the system can be supported, and the accuracy rate of attribute comprehensive recognition results is improved.

Description

Technical Field

[0001] The invention relates to a multi-means cross-domain early warning network system, in particular to a confidence normalization method for target recognition attributes. Background Art

[0002] The original concept of confidence in probability and mathematical statistics is the probability that the true parameter falls within the estimation interval when the parameter is estimated. In recognition, the confidence that is generally understood to be meaningful refers to the prior probability that the specific recognition result given by the recognition module is correct, that is, when a certain recognition module identifies and reports that the model of a certain aircraft target is F15, it can be considered that it is indeed likely to be F15. In the case of the inability to perfectly understand the theoretical process of detection and processing, the specific distribution of this probability cannot be accurately calculated. Objectively, the distribution of this probability can only be estimated, and this estimation behavior must be based on a large number of sample observations to be meaningful.

[0003] In a cross-domain early warning network system composed of multiple sensors, different sensors will independently generate recognition results for target attributes such as friend or foe, military or civilian, target category, and target aircraft model, and the confidence of the target recognition attribute is an important basis for the system to adjust the fusion strategy at the target attribute decision level. At present, most sensors have a built-in reporting parameter of "confidence", but in engineering implementation, this parameter is not the same concept as the real physical "confidence" defined above. These reported parameters can only qualitatively show how much confidence the sensor module has in its reported results, rather than statistics. In this article, it is named membership. Usually, membership is mainly derived from manual experience and has a certain guiding significance for the credibility of target recognition results, but its number itself is only a relative value, and the absolute value has no physical meaning, and there is no direct correspondence between it and the confidence.

[0004] In order to achieve effective comprehensive recognition of target identification attributes by multi-source heterogeneous sensors in the cross-domain early warning network system, it is necessary to combine the membership reported by the sensor, compare a large number of membership values ​​with the true value of its identification result based on statistical principles, normalize and calculate the confidence of multi-source heterogeneous sensors, establish the confidence distribution of target identification attributes in the same dimension, and provide parameter support for the generation of conflict resolution strategies in comprehensive identification and the confidence calculation of the multi-sensor comprehensive identification system. Summary of the invention

[0005] In view of the problems existing in the prior art and the comprehensive identification task of target attributes in the cross-domain early warning network system, the present invention proposes a confidence normalization method for multi-source heterogeneous sensors. The confidence of multiple sensors in the same dimension is calculated by the discrete and continuous membership of the sensor identification results, and then the comprehensive identification confidence of the multi-source sensor system is calculated. It can support the generation of system decision-level fusion strategy and improve the accuracy of comprehensive attribute identification results.

[0006] The purpose of the present invention is achieved through the following technical solutions.

[0007] A confidence normalization method for target recognition attributes comprises the following steps:

[0008] Step 1: Convert discrete membership degree to confidence degree;

[0009] Step 2: Continuous membership is converted into confidence;

[0010] Step 3: Calculation of confidence of comprehensive identification of multi-source sensors.

[0011] The step 1 is based on the measured data set, counting the number of samples of each membership degree in the set, and the number of correct samples when the corresponding membership degree appears, and estimating the accuracy of the samples under each membership degree by dividing the frequency.

[0012] The step 2 defines the confidence corresponding to the membership as the density function of the probability distribution of membership-confidence, and estimates the confidence corresponding to the membership by calculating the recognition accuracy of samples under different membership values.

[0013] In the step three, when each sensor has a precisely defined confidence level, the confidence level of the comprehensive recognition result is obtained by calculating the conditional probability under the current recognition results of each sensor and the confidence level.

[0014] In step 3, the recognition results of each sensor are assumed to be A1, A2, ..., An, their confidences are P1, P2, ..., Pn, and the comprehensive recognition result is A. According to the Bayesian formula, we have:

[0015]

[0016] in It is the complement of A, that is, not A. Since the results of each sensor are independent, we can get:

[0017]

[0018] in is the marginal probability of recognition result A, When Ai is equal to A, it is equal to Pi. When Ai is not equal to A, it is 1-Pi, which is the probability that the sample is mistakenly identified as Ai.

[0019] Compared with the prior art, the advantages of the present invention are: the present invention solves the problem of attribute conflict in the system decision-level fusion process through normalized calculation of the target recognition attribute membership reported by multi-source sensors such as radar, radar reconnaissance, communication reconnaissance, and optoelectronics, realizes the confidence calculation of the system's comprehensive recognition results of target attributes, and improves the accuracy of the system's target attribute recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of the present invention.

[0021] Figure 2 This is a corresponding table of membership degree and confidence level of the present invention. DETAILED DESCRIPTION

[0022] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0023] This method is to use confidence normalization technology for a system composed of multi-source heterogeneous sensors. First, the discrete and continuous membership of each sensor unit recognition result is converted into confidence, and then the confidence of the multi-source sensor comprehensive recognition system is calculated based on the confidence of each sensor unit, solving the problems of inconsistent confidence dimensions and difficulty in formulating conflict resolution strategies in the process of comprehensive recognition of target attributes. In order to make the purpose, technical solution and advantages of this application clearer, the application is further described in detail below.

[0024] (1) Membership-confidence conversion of single sensor recognition results

[0025] Discrete membership is usually filled in by artificial rules and is divided into different levels of values. In the current cross-domain early warning network system, the membership of the passive reconnaissance equipment identification results mostly belongs to discrete membership. For example, the confidence level of the radar reconnaissance sensor for its reported results is divided into 1 to 5 levels, corresponding to 0.2, 0.4, 0.6, 0.8 and 1 respectively.

[0026] In order to convert discrete membership into confidence, it is necessary to select a sufficiently large sample set, count the number of samples of each membership in the set, and the number of correct samples when the corresponding membership appears, and estimate the accuracy of samples under each membership by dividing the frequency. Still taking the radar reconnaissance sensor as an example, the example is as follows:

[0027]

[0028] A1~A5 and B1 to B5 are actual statistical values.

[0029] This method is simple and easy to operate, has low requirements for data preparation, and has a clear mathematical meaning. It is easy to further imagine that the accuracy of the sensor's judgment of different targets is different, which will be particularly prominent within a limited discrete level. For example, with the same membership degree of 0.8, some models may be slightly different from others. Therefore, further, each recognition result of the mine detection (note that it is the recognition result rather than the true value) can be separately counted as above, so as to obtain a more detailed confidence value:

[0030]

[0031] It should be pointed out that this method is an estimation of the distribution and is only accurate when the sample size is sufficient. When actually collecting data, only a portion of the recognition result samples may meet the assumption that the sample size is sufficient, and only these results can be converted to independent confidence. For the types of recognition results with insufficient sample size, all such types can be classified as other categories and a unified confidence conversion table can be calculated. This is actually using the average value of all samples to estimate the confidence of each different result. As time goes by, when the sample size gradually increases, some recognition results in other categories can be separated out separately.

[0032] (2) Continuous membership degree converted to confidence degree

[0033] Continuous membership usually appears in the recognition results of machine learning models, and currently only appears in radar recognition results. The continuous membership for each sample is usually not equal to that of other samples, so the conversion method of discrete membership cannot be directly used.

[0034] Considering that the continuous membership is usually calculated by some normalization method as the decision value for model classification, and still has the function of decision value after normalization, the present invention makes the following requirements for it: the membership is positively correlated with the confidence in principle. The loss function used by some machine learning models is to take the type with the smallest loss as the output. At this time, the normalized loss value must be inverted (i.e. 1-the loss value) before it can be reported as the membership to ensure that the membership is positively correlated with the confidence in principle.

[0035] After ensuring that the continuous membership meets the above conditions, the confidence f(x) corresponding to the membership x can be defined as the density function of the probability distribution of the membership-confidence. In order to estimate f(x), it is necessary to establish a sufficiently large sample set T and clarify the membership of each sample and whether the recognition result is correct. At this time, when the membership x is obtained during the recognition process, as long as the recognition accuracy of all samples in T with membership near x is calculated, a more accurate estimate of the confidence corresponding to the membership x can be obtained.

[0036] In actual operation, considering that the sample set T is large, recalculating the confidence of each independent x will cause a large processing delay. It is proposed to solve this problem by quantifying x, that is, calculating the confidence corresponding to the membership value of a certain number of quantiles in advance, and approximating the membership obtained during the recognition process by the corresponding membership of the nearest quantile. In this way, only a table lookup ( Figure 2 ) can quickly get the corresponding confidence. We can further consider using mathematical coding in information theory to take more detailed quantiles in the interval where the confidence changes faster to improve the approximate accuracy.

[0037] Although it is similar to estimating the confidence of different recognition results separately in discrete membership, it is also possible to obtain independent confidence for different recognition results in continuous membership, but because the number of points is far more than the artificially given interval of discrete membership, the demand for data volume is greater. Considering the actual effective sample rate in radar recognition, it is temporarily decided not to subdivide the recognition results of continuous membership.

[0038] (3) Confidence calculation of comprehensive identification of multi-source sensors

[0039] When each sensor has a precisely defined confidence level, we can try to calculate the confidence level of the comprehensive recognition result. The confidence level of the comprehensive recognition result is the prior probability that the comprehensive recognition result is correct, which can be obtained by calculating the conditional probability under the current recognition results of each sensor and the confidence level.

[0040] Assume that the recognition results of each sensor are A1, A2, ..., An, their confidences are P1, P2, ..., Pn, and the comprehensive recognition result is A. According to the Bayesian formula, we have:

[0041]

[0042] in It is the complement of A, that is, not A. Since the results of each sensor are independent, we can get:

[0043]

[0044] in is the marginal probability of recognition result A, When Ai is equal to A, it is equal to Pi. When Ai is not equal to A, it is 1-Pi, that is, the probability that the sample is misclassified as Ai. Therefore, P(A) can be calculated in this way. More generally, when some recognition results have independent confidence estimates, Expand to get a more accurate result. The formula is relatively complicated and will not be described here.

[0045] By observing the formula, we can see that the value of P(A) cannot be determined by only using the recognition results of each sensor as input. P(A), as the marginal probability of the recognition result A, has nothing to do with the observation results of the sensor, but only with the objective situation of its actual appearance. Therefore, it is necessary to calculate the probability of various aircraft models appearing in the airspace after accumulating more data. Only after obtaining this probability value can the accuracy of comprehensive recognition be calculated.

Claims

1. A confidence normalization method for target recognition attributes, characterized in that The following steps are involved: Step 1: Convert discrete membership degree to confidence degree; Step 2: Continuous membership is converted into confidence; Step 3: Calculation of confidence of comprehensive identification of multi-source sensors.

2. A method for normalizing the confidence of target recognition attributes according to claim 1, characterized in that The step 1 is based on the measured data set, counting the number of samples of each membership degree in the set, and the number of correct samples when the corresponding membership degree appears, and estimating the accuracy of the samples under each membership degree by dividing the frequency.

3. A method for normalizing the confidence of target recognition attributes according to claim 1, characterized in that The step 2 defines the confidence corresponding to the membership as the density function of the probability distribution of membership-confidence, and estimates the confidence corresponding to the membership by calculating the recognition accuracy of samples under different membership values.

4. A method for normalizing the confidence of target recognition attributes according to claim 1, characterized in that In the step three, when each sensor has a precisely defined confidence level, the confidence level of the comprehensive recognition result is obtained by calculating the conditional probability under the current recognition results of each sensor and the confidence level.

5. A method for normalizing the confidence of target recognition attributes according to claim 4, characterized in that In step 3, the recognition results of each sensor are assumed to be A1, A2, ..., An, their confidences are P1, P2, ..., Pn, and the comprehensive recognition result is A. According to the Bayesian formula, we have: ； in It is the complement of A, that is, not A. Since the results of each sensor are independent, we can get: ； in is the marginal probability of recognition result A, When Ai is equal to A, it is equal to Pi. When Ai is not equal to A, it is 1-Pi, which is the probability that the sample is mistakenly identified as Ai.

Images