A target recognition method and device for data and knowledge fusion

Abstract

The present invention proposes a target recognition method for data and knowledge fusion. Considering the complementarity of data and knowledge comprehensively, by obtaining the data to be recognized, as well as the first sample composed of the labeled data obtained from the existing database and the second sample composed of the unlabeled data; successively constructing the first rule premise for the first sample and constructing a data-driven belief rule base by the way of fusing the first rule premise through labels; by repeatedly constructing candidate rule premises for the second sample, selecting the second rule premise from the candidate rule premises by using the mutual information criterion, and providing the second rule premise to the expert, thus constructing a knowledge-driven belief rule base; finally, by using the genetic algorithm to fuse the belief rule base, the rule redundancy and conflict can be effectively reduced. The present invention finds the corresponding rule by comparing the features of the real-time data with the features in the rule base, so as to accurately determine the type of the real-time target.

Description

Technical Field

[0001] The present invention belongs to the technical field of target recognition, and particularly relates to a target recognition method and device for data and knowledge fusion. Background Art

[0002] In a complex environment with strong real-time performance and high confrontation such as a remote warning system, the types of targets increase, and it is necessary to quickly and accurately detect and identify the targets, so as to provide reliable support information for decision-makers. Information fusion has become a key supporting technology in the comprehensive processing of battlefield information. Traditional information fusion mainly performs fusion processing on the measurement data of physical sensors. However, with the continuous enhancement of the concealment of non-friendly targets and countermeasures, traditional fusion methods show certain limitations. People gradually realize that the information provided by people with certain domain knowledge plays an important role in the target recognition process. Therefore, the fusion of data and knowledge can provide an effective means for reliable target recognition.

[0003] Currently, some technologies have tried to achieve the fusion of expert knowledge and data to improve the accuracy of target recognition. These existing technologies can be divided into three methods according to the sequence of knowledge and data participation. That is, first build a model with data, and then correct the model with expert knowledge; first build an initial model according to expert knowledge, and then correct the model parameters with data; build models with data and knowledge respectively, and finally perform model fusion on a unified basis. In these methods, the problem of knowledge acquisition has not been solved yet. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a target recognition method and device for data and knowledge fusion. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0005] The present invention provides a target recognition method for data and knowledge fusion, including:

[0006] Step 1, obtain the real-time data generated by the remote warning system detecting the target, and use the real-time data as the data to be recognized, and obtain the first sample composed of the tagged data and the second sample composed of the untagged data from the existing database;

[0007] Step 2, sequentially construct the first rule premise for the first sample, and through fuzzy partitioning of the first rule premise according to language labels and fusion of rule results, thereby construct a data-driven confidence rule base;

[0008] Step 3: Using the partitioning method of the rule antecedents in the data-driven belief rule base, construct candidate rule antecedents for each second sample, select the second rule antecedents from the candidate rule antecedents using the mutual information criterion, and provide the second rule antecedents to the experts, thereby constructing a knowledge-driven belief rule base;

[0009] Step 4: Use a genetic algorithm to fuse the data-driven belief rule base and the expert-driven belief rule base to obtain a hybrid belief rule base;

[0010] Step 5: Classify the real-time data using the rule antecedents in the hybrid belief rule base to determine the real-time target category to which it belongs.

[0011] Advantages of the present invention:

[0012] The present invention proposes a target recognition method for data and knowledge fusion, comprehensively considering the complementarity of data and knowledge. By obtaining real-time data generated by detecting targets in a remote warning system and using this real-time data as the data to be recognized, as well as obtaining a first sample composed of labeled data and a second sample composed of unlabeled data from an existing database; sequentially constructing first rule antecedents for the first sample and constructing a data-driven belief rule base by fusing the first rule antecedents with labels; by repeatedly constructing candidate rule antecedents for the second sample, selecting the second rule antecedents from the candidate rule antecedents using the mutual information criterion, and providing the second rule antecedents to the experts, thereby constructing a knowledge-driven belief rule base; finally, by using a genetic algorithm to fuse the belief rule base, it is possible to effectively reduce rule redundancy and conflicts and improve the target recognition accuracy. Finally, by extracting the features of the real-time data and comparing them with the features in the rule base to find the corresponding rules, the real-time target category to which it belongs is determined.

[0013] The following will further elaborate on the present invention in detail with reference to the drawings and embodiments. Description of the Drawings

[0014] Figure 1 is a flowchart showing a target recognition method for data and knowledge fusion provided by the present invention;

[0015] Figure 2 is a schematic diagram of a data and knowledge fusion model provided by the present invention;

[0016] Figure 3 is a schematic diagram of the gene encoding provided by the present invention;

[0017] Figure 4 is a characteristic distribution curve diagram of three aircraft targets provided by the present invention;

[0018] Figure 5 is a schematic diagram of the linguistic fuzzy partitioning of the rule antecedent features provided by the present invention. Detailed implementation manners

[0019] The following further describes the present invention in detail with reference to specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0020] Combined with Figure 1 and Figure 2 , the present invention provides a target recognition method for data and knowledge fusion, including:

[0021] Step 1: Obtain real-time data generated by a remote warning system detecting a target, and use the real-time data as data to be recognized, and obtain a first sample composed of labeled data and a second sample composed of unlabeled data from an existing database;

[0022] Step 2: Sequentially construct a first rule premise for the first sample, and by performing fuzzy partitioning on the first rule premise according to language labels and fusing rule results, thereby construct a data-driven confidence rule base;

[0023] Step 3: Use the partitioning method of the rule premise in the data-driven confidence rule base to construct a candidate rule premise for each second sample, and use the mutual information criterion to select a second rule premise from the candidate rule premises, and provide the second rule premise to an expert, thereby construct a knowledge-driven confidence rule base;

[0024] Step 4: Use a genetic algorithm to fuse the data-driven confidence rule base and the expert-driven confidence rule base to obtain a hybrid confidence rule base;

[0025] Step 5: Use the rule premise in the hybrid confidence rule base to classify the real-time data and determine the type of the real-time target.

[0026] Existing data and knowledge fusion target recognition methods ignore the acquisition of knowledge, so it is difficult to achieve a high degree of complementarity between data and knowledge. The present invention proposes a target recognition method that comprehensively considers the complementarity between data and knowledge. First, a data-driven confidence rule target recognition model is constructed based on confidence rules through data. Then, based on an active learning strategy, an expert-driven confidence rule base is obtained. Finally, a genetic algorithm is used to fuse the data-driven confidence rule base and the knowledge-driven confidence rule base to reduce rule redundancy and conflicts and improve the accuracy of target recognition. Finally, the fused rule base can be used to recognize the target to be classified, and the corresponding rule with a label can be obtained by searching in the library, and the conforming rule is determined by comparing eigenvalue during the search.

[0027] In a specific embodiment, step 2 includes:

[0028] Step 21: Construct the rule premise of the first sample based on the eigenvalue of the first sample. For subsequent samples, determine the rule premise of the sample by calculating the membership degree of the sample with respect to other rule premises.

[0029] Step 22: Divide the rule premises of all samples into corresponding linguistic labels according to the index values of the linguistic labels to achieve fuzzy division of the rule premises.

[0030] Step 23: According to the marks carried by the samples, use the belief function theory to first perform intra-class aggregation and then inter-class aggregation on all samples in each linguistic label to obtain the rule results associated with each rule premise, so as to achieve the fusion of rule results.

[0031] Step 24: Form a data-driven belief rule base with all rule conclusions.

[0032] Data-driven belief rule learning:

[0033] In this module, use the labeled data to learn an interpretable and accuracy-traded data-driven belief rule target recognition system, adopt a self-organizing clustering strategy, and learn the rule premises incrementally.

[0034] In a specific embodiment, Step 21 includes:

[0035] Step 211: Construct a rule premise for it according to the eigenvalue of the first sample. First, establish the first rule premise on the basis of the first sample. The form of the rule premise is:

[0036]

[0037] Step 212: Take the eigenvalue of the first sample as the center of the Gaussian membership function in the rule premise in Step 211:

[0038]

[0039] Among them, the initial standard deviation σ0 of the Gaussian membership function is set to a small positive integer;

[0040] Step 213: For the samples newly entered after the first sample, first calculate the membership degree of the sample with respect to the rule premise in Step 211:

[0041]

[0042] Among them,

[0043]

[0044] Step 214, if the membership degree calculated in Step 123 is less than the threshold ε, construct another rule premise for the newly entered sample in the same way as in Step 211;

[0045] Step 215, if the membership degree calculated in Step 123 is greater than the threshold ε, classify the newly entered sample into the rule premise in Step 211, and update the standard deviation and mean of the membership function;

[0046] Among them, the same sample may be classified into multiple rule premises. If the membership degree is less than the threshold (), construct another rule premise for the new sample in the same way; if the membership degree is greater than the threshold, classify the sample into the rule premise, and update the standard deviation and mean of the membership function (the same sample may be classified into multiple premises). The update formulas are as follows:

[0047]

[0048]

[0049] Among them, is the update of the total number of samples assigned to the rule premise A q ;

[0050] Step 216, repeat Steps 212 to 215 until there is at least one corresponding rule premise for all samples.

[0051] In a specific embodiment, Step 22 includes:

[0052] Step 221, determine the number of language labels and the self-index value of the language labels according to requirements;

[0053] Step 222, calculate the index value of each rule premise in the language label according to the standard deviation and mean of the membership function in Step 215; and calculate the distance between this index value and the self-index value of each language label, and determine the language label to which the rule premise belongs according to the distance.

[0054] Step 2221, find the maximum value of the Gaussian membership function mean on each feature dimension of each sample

[0055] Since the premise part of the above rules obtained in 21 is somewhat deviated from human cognition, it is necessary to transform the premise part of the above rules into a rule form that is easy for people to understand. The specific transformation process is as follows:

[0056] Assume that the number of language labels is determined (N L ), such as large, medium, small (N L= 3), and then find the maximum value of the mean of the Gaussian membership function in each feature dimension and the minimum value Calculate the distance (dis p ) between adjacent linguistic labels for each feature:

[0057]

[0058] where the number of linguistic labels is N L ;

[0059] Step 2223, the mean and standard deviation of the Gaussian membership function in each feature for each rule premise are respectively expressed as:

[0060]

[0061]

[0062] where λ represents the overlap parameter of the membership function, selected from [0, 1];

[0063] Step 2224, according to the mean and standard deviation, calculate the index value of each rule premise in the linguistic label, expressed as:

[0064]

[0065] where k * represents the index value of the linguistic label;

[0066] Step 2225, calculate the distance between the index value obtained in Step 2224 and the own index value of each linguistic label, and determine the linguistic label to which the rule premise belongs as the linguistic label with the closest distance.

[0067] In a specific embodiment, Step 23 includes:

[0068] Step 231, fuse the samples with the same class label and belonging to the same rule premise to obtain the within-class fusion result;

[0069] Next, perform rule conclusion learning: To obtain the accurate confidence distribution of the rules, perform within-class fusion and between-class fusion on the calculated membership degrees based on the theory of belief functions. The fusion method is as follows: In the first-level combination, the combination of evidence from the same class is performed using Dempster's rule. Let be the set of samples with class label ω m and belonging to premise A q , and the fusion process is as shown below:

[0070]

[0071] Step 232: Use the discount factor to fuse the within-class fusion results to obtain the between-class combination results;

[0072] At the second layer, further combine the results obtained by within-class combination. Considering the possibility of high conflict, Shafer's discount operation is adopted here to discount unreliable evidence. The fusion process is completed using the discount factor, which is measured by the proportion of samples in the total samples. The discount factor is expressed as:

[0073]

[0074] Based on the discount factor, first perform the Shafer discount operation on the within-class combination results fused in the first layer, and then use the Dempster's rule for combination to obtain the between-class combination results as follows:

[0075]

[0076] where K q is the normalization factor, satisfying

[0077] Step 233: Obtain the rule conclusions associated with the rule premises from the within-class and between-class fusion results; obtain the rule conclusions q associated with the rule premise A are:

[0078]

[0079] Step 234: Form a data-driven belief rule base with the rule conclusions.

[0080] Knowledge-driven belief rule learning:

[0081] In this part, considering the lack of labeled data, the learned belief rules may not cover important regions in the feature space. Therefore, obtaining domain knowledge can supplement the regions that cannot be covered by the data. The key task is to find the rule premises that contribute the most to the current data-driven belief rules and submit them to experts for annotation. The process of this task is similar to active learning, but different from active learning, in this chapter, it is the rule premises that are selected and submitted to experts for annotation, rather than the samples themselves. Therefore, compared with traditional active learning strategies, the method in this chapter can effectively reduce the number of expert participations and improve the efficiency of expert participation.

[0082] First, a set of candidate rule premises that can well cover the feature space needs to be constructed. If there is a series of N' unlabeled available samples x1, x2, …, x N', then these samples can be directly used to construct the candidate rule premises. Otherwise, these unlabeled samples need to be randomly generated in the feature space. To ensure the compatibility between knowledge-driven confidence rules and data-driven confidence rules, the same fuzzy partition is adopted for each feature.

[0083] In a specific embodiment, step 3 includes:

[0084] Step 31, construct a candidate rule premise for each second sample, expressed as:

[0085]

[0086] Step 32, use the data-driven confidence rule base to predict the label of the conclusion of the candidate rule premise.

[0087] N' unlabeled samples can construct N' candidate rule premises without conclusions. However, in this process, it is possible that different samples learn the same candidate rule premise. Therefore, redundant rule premises need to be deleted, and finally Q' different candidate rule premises are obtained.

[0088] After obtaining the rule premises, select the rule premise that contributes the most to the current model and submit it to the expert for annotation. The mutual information criterion is used to select the rule premises. First, use the data-driven confidence rule base to initialize the knowledge-driven confidence rule base to facilitate learning knowledge complementary to the data-driven. Then, use the data-driven confidence rule base to predict the conclusion of the candidate rule premise, and the predicted pseudo-label is:

[0089]

[0090] The confidence rule carrying the pseudo-label is expressed as:

[0091]

[0092] Where

[0093] Step 33, use the confidence rule to select the candidate rule premise that can provide the maximum mutual information for the unlabeled rule premise, expressed as:

[0094]

[0095] Since the first term in formula (18) above is independent of the selection of the rule premise.

[0096] Step 34, transform the selected candidate rule premise.

[0097] Therefore, formula (18) above is equivalent to:

[0098]

[0099] Step 35, the selected candidate rule premise is selected and given to an expert who gives the corresponding confidence distribution conclusion for the expert to mark, and the rule after being marked by the expert is added to the current knowledge-driven rule base

[0100] Step 36, repeat Steps 31 to 35 until the stop condition is reached, and a knowledge-driven confidence rule base is obtained. The stop condition can be reaching a predefined number of marks or the candidate rule premise set being empty. Finally, remove the data-driven confidence rule base during initialization and obtain the final knowledge-driven confidence rule base R K = R K \R D .

[0101] Confidence rule base fusion:

[0102] After obtaining the data-driven confidence rule base and the knowledge-driven confidence rule base respectively, to avoid rule redundancy and rule conflicts and improve the target recognition performance at the same time, a genetic algorithm is used to fuse the two rule bases.

[0103] In a specific embodiment, Step 4 includes:

[0104] Step 41, concatenate the binary encodings of all the data-driven confidence rule bases and the knowledge-driven confidence rule bases to form a chromosome

[0105] Among them, for each rule premise and rule itself in all the data-driven confidence rule bases and the knowledge-driven confidence rule bases, the first p loci are used to encode the rule premise features, 1 indicates that the feature is used, and 0 indicates that the feature is discarded in the rule; the last locus indicates the selection of the rule, 1 indicates that the rule is selected, and 0 indicates that the rule is deleted from the rule base;

[0106] Step 42, fuse the data-driven confidence rule base and the knowledge-driven confidence rule base on the chromosome using a genetic algorithm to obtain a hybrid confidence rule base.

[0107] Reference Figure 3, in gene coding, both the premise features and the rules themselves of the rules adopt binary coding. The selection of rules helps to solve rule conflicts and rule redundancies and makes the rule base more compact. In addition, for each selected rule, deleting the irrelevant features can improve the rule generalization ability and make the rules easier to understand. As shown in the figure, for each rule, the first p sites are used to encode the premise features of the rule ("1" indicates that the feature is used, and "0" indicates that the feature is deprecated in the rule). The last site indicates the selection of the rule ("1" indicates that the rule is selected, and "0" indicates that the rule is deleted from the rule base). Stringing together all the data-driven and knowledge-driven binary coding strings forms a chromosome When initializing the population, to make full use of the initial rules, first set all of the first chromosome to "1", and the remaining n s - 1 chromosomes are randomly generated, where n s is the size of the population.

[0108] In a specific embodiment, before step 42, the data and knowledge fusion target recognition method further includes:

[0109] Step a, construct a fitness evaluation index according to the average number of features in each rule and the number of rules;

[0110] Step b, use the fitness evaluation index as the index for the genetic algorithm to select fitness values, and select chromosomes with the same size as the initial population using roulette wheel selection;

[0111] Step c, pair up and cross the selected parent chromosomes, with a crossover probability of p c ; Mutate the offspring obtained by crossover. For each rule, select two mutation sites, one for the features of the rule premise and the other for the rule itself, and the mutation probability of the two sites is p m ;

[0112] Among them, each chromosome has multiple crossover points, and the number of crossover points is the same as the total number of rule premises;

[0113] Step d, repeat steps a to c until the stop condition is met;

[0114] Step e, decode the best chromosome of the last generation to obtain a hybrid confidence rule base.

[0115] Fitness evaluation: Considering both the recognition accuracy and the interpretability of the model, a suitable fitness evaluation criterion is constructed. The recognition accuracy index acc is defined as the proportion of correctly recognized samples. The interpretability of the model is measured by the number of features (NoF) in the confidence rule base in this section. NoF is a comprehensive index that takes into account both the average number of features in each rule and the number of rules. Since the accuracy index and the interpretability index are in different ranges, the latter is first normalized, and the constructed fitness evaluation index is:

[0116]

[0117] where NoF max represents the total number of features in the initial rule base, and 0 ≤ w acc ≤ 1 is the weight of the accuracy index.

[0118] Genetic algorithm: According to the fitness values, chromosomes of the same size as the initial population are selected using roulette wheel selection. These selected parent chromosomes are crossed in pairs. In this section, multi-point crossover is adopted, and one crossover point is selected for each rule, with a crossover probability of p c . Then, the offspring are mutated. Two mutation sites are selected for each rule (one for the features in the rule premise and the other for the rule itself), and the mutation probability of the two sites is p m .

[0119] For the initialized population, the above fitness evaluation and genetic operation processes are repeatedly executed until the stopping condition (e.g., the maximum number of iterations) is met. By decoding the best chromosome of the last generation, the final hybrid confidence rule base is obtained.

[0120] The present invention provides an object recognition device for data and knowledge fusion, including:

[0121] A data acquisition module, which acquires real-time data generated by a remote warning system detecting a target and uses the real-time data as data to be recognized, and acquires a first sample composed of labeled data and a second sample composed of unlabeled data from an existing database;

[0122] A first construction module, which sequentially constructs a first rule premise for the first sample, and constructs a data-driven confidence rule base by performing fuzzy partitioning on the first rule premise according to language labels and fusing rule results;

[0123] A second construction module, which uses the partitioning method of the rule premise in the data-driven confidence rule base to construct candidate rule premises for each second sample, selects a second rule premise from the candidate rule premises using the mutual information criterion, and provides the second rule premise to an expert, thereby constructing a knowledge-driven confidence rule base;

[0124] The rule fusion module uses the genetic algorithm to fuse the data-driven belief rule base and the expert-driven belief rule base to obtain a hybrid belief rule base;

[0125] The type recognition module classifies the real-time data using the rule premises in the hybrid belief rule base to determine the real-time target type.

[0126] The effectiveness of the present invention has been evaluated and verified in air target recognition.

[0127] (1) Problem description

[0128] In an airspace surveillance system, multiple sensors are commonly used for detection. For example, radar sensors that provide motion features such as speed and acceleration, and infrared sensors that provide shape features. For the sake of demonstration, only three target categories (commercial aircraft ω1, bomber ω2, fighter ω3) are considered in this section. These three categories of targets are determined by three features, namely average speed (AveSpeed), maximum acceleration (MaxAcc), and average length (AveLength).

[0129] During the numerical simulation process, Gaussian conditional density functions are used to generate samples. Appropriate parameters are selected in the samples to ensure P{s min <x<s max} = 0.95, where [s min , s max is the feature interval shown in Table 1. Figure 4 a to Figure 4 c are the feature distribution curves of the three aircraft targets. In the following text, the entire process of the proposed method will be presented in detail. The default experimental parameter settings are as follows: initial standard deviation σ0 = 1, number of membership partitions NL = 3, membership threshold ε = 0.75, number of labels nl = 10, accuracy weight w acc = 0.6, population size n s = 200, crossover probability p c = 0.8, mutation probability p m = 0.01, maximum number of iterations of the genetic algorithm n g = 50.

[0130] Table 1 Feature intervals of the three aircraft

[0131]

[0132] Construction of the data-driven belief rule base:

[0133] Based on Gaussian conditional density functions with the same prior probability, 150 samples were generated. The generated samples were divided into training samples and test samples in a ratio of 4:1. Randomly select 10% of the training samples as labeled samples, and first use the method proposed in the previous chapter to learn the data-driven belief rule base. Among them, Table 2 shows the learned data-driven belief rule base, and the linguistic fuzzy partition of the rule premise features is as Figure 5 shown.

[0134] Table 2 Data-driven belief rule base

[0135]

[0136]

[0137] Knowledge-driven belief rule base construction:

[0138] Next, belief rules will be learned from experts. Since it is difficult to arrange an expert to give labels in the experiment, the remaining 90% of the training samples are used to simulate the process of expert annotation. First, use the parameters of the data-driven fuzzy partition and the remaining 90% of the training samples to construct candidate rule premises. Then, successively find the rule premise that contributes the most to the knowledge-driven belief rule base in each iteration and submit it to the expert for annotation. To simulate the accuracy of expert annotation, in this section, the annotated belief conclusions are obtained by fusing the samples covered by the rule premises. Finally, the obtained knowledge-driven belief rule base is shown in Table 3.

[0139] Table 3 Knowledge-driven belief rule base

[0140]

[0141] Genetic fusion of belief rules:

[0142] After the above two steps, the genetic algorithm is used to fuse the data-driven belief rule base (DBRB) and the knowledge-driven belief rule base (KBRB) to obtain the hybrid belief rule base (HBRB) as shown in Table 4. It can be seen from the finally obtained hybrid belief rule base that 3 rules in the hybrid belief rule base come from DBRD and 5 rules come from KBRB. The hybrid belief rule base also abandons some irrelevant premises of the rules. Compared with DBRB and KBRB, HBRB consists of fewer rules and shorter premises. The performance comparison of HBRB with DBRB and KBRB is shown in Table 5. It can be seen from Table 5 that the hybrid belief rule base not only improves the accuracy, but also has better interpretability (the smaller the NoF, the better the interpretability).

[0143] Table 4 Hybrid Belief Rule Base

[0144]

[0145] Table 5 Performance Comparison of HBRB with DBRB and KBRB

[0146]

[0147] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined.

[0148] Although the present application has been described in conjunction with various embodiments herein, however, in the process of implementing the claimed present application, those skilled in the art can understand and realize other variations of the disclosed embodiments by viewing the accompanying drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "one" does not exclude a plurality of cases.

[0149] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A target recognition method for data and knowledge fusion, characterized in that Including: Step 1: Obtain the real-time data generated by the remote warning system for the detected target, and use this real-time data as the data to be recognized, as well as obtain the first sample composed of the data with tags and the second sample composed of the data without tags from the existing database; Step 2: Sequentially construct the first rule premise for the first sample, and by performing fuzzy partitioning and rule result fusion on the first rule premise according to language tags, thereby construct a data-driven confidence rule base; Step 3: Use the partitioning method of the rule premise in the data-driven confidence rule base to construct candidate rule premises for each second sample, and use the mutual information criterion to select the second rule premise from the candidate rule premises, and provide the second rule premise to the expert, thereby construct a knowledge-driven confidence rule base; Step 4: Use the genetic algorithm to fuse the data-driven confidence rule base and the expert-driven confidence rule base to obtain a hybrid confidence rule base; Step 5: Use the rule premise in the hybrid confidence rule base to classify the real-time data and determine the type of the real-time target.

2. The target recognition method for data and knowledge fusion according to claim 1, wherein Step 2 includes: Step 21: Construct the rule premise of the first sample according to the characteristic values of the first sample. For subsequent samples, determine the rule premise of the sample by calculating the membership degree of the sample to other rule premises; Step 22: According to the index value of the language tag, partition the rule premises of all samples into the corresponding language tags to achieve fuzzy partitioning of the rule premises; Step 23: According to the marks carried by the samples, use the theory of belief functions to first perform within-class aggregation and then between-class aggregation on all samples in each language tag to obtain the rule results associated with each rule premise, so as to achieve the fusion of rule results; Step 24: Form a data-driven confidence rule base with all rule conclusions.

3. The target recognition method for data and knowledge fusion according to claim 1, characterized in that, Step 21 includes: Step 211: Construct a rule premise for it according to the characteristic values of the first sample. The form of the rule premise is: Step 212: Use the characteristic values of the first sample as the center of the Gaussian membership function in the rule premise in Step 211: wherein, the initial standard deviation σ0 of the Gaussian membership function is set to a small positive integer; Step 213: For the samples newly entered after the first sample, first calculate the membership degree of the sample to the rule premise in Step 211: wherein, Step 214: If the membership degree calculated in Step 123 is less than the threshold ε, construct another rule premise for the newly entered sample in the same way as in Step 211; Step 215: If the membership degree calculated in Step 123 is greater than the threshold ε, then classify the newly entered sample into the rule premise in Step 211, and update the standard deviation and mean of the membership function; wherein, the same sample may be classified into multiple rule premises, and the update formula is as follows: Among them, is the update of the total number of samples allocated to rule premise A q ; Step 216: Repeat Steps 212 to 215 until there is at least one rule premise corresponding to all samples.

4. The target recognition method for data and knowledge fusion according to claim 3, characterized in that, Step 22 includes: Step 221: Determine the number of language tags and the self-index value of the language tags according to requirements; Step 222: Calculate the index value of each rule premise in the language label according to the standard deviation and mean of the membership function in Step 215; and calculate the distance between the index value and the self-index value of each language label, and determine the language label to which the rule premise belongs according to the distance.

5. The target recognition method for data and knowledge fusion according to claim 4, characterized in that, Step 222 includes: Step 2221, find the maximum value of the mean of the Gaussian membership function for each feature dimension of each sample and the minimum value Step 2222, calculate the distance dis between adjacent language labels on the features of each sample p : Among them, the number of language tags is N L ; Step 2223: The mean and standard deviation of the Gaussian membership function for each feature in each rule premise are respectively expressed as: where λ represents the overlap parameter of the membership function, which is selected between [0, 1]; Step 2224: Calculate the index value of each rule premise in the language label according to the mean and standard deviation, which is expressed as: where k * represents the index value of the language tag; Step 2225: Calculate the distance between the index value obtained in Step 2224 and the self-index value of each language label, and determine the language label with the closest distance as the language label to which the rule premise belongs.

6. The target recognition method for data and knowledge fusion according to claim 5, characterized in that, Step 23 includes: Step 231: Fuse the samples with the same class label and belonging to the same rule premise to obtain the within-class fusion result; Let be the set of samples with class label ω m and belonging to the premise A q . The fusion process is as follows: Step 232: Use the discount factor to fuse the within-class fusion result to obtain the between-class combination result; The discount factor is measured by the proportion of the sample in the total sample, and the discount factor is expressed as: The between-class combination result is expressed as follows: where K q is a normalization factor that satisfies Step 233, obtain a rule conclusion associated with a rule premise from the in-class and inter-class fusion results; obtain the rule conclusion q associated with the rule premise A as follows: Step 234: Form a data-driven confidence rule base with the rule conclusions.

7. The target recognition method for data and knowledge fusion according to claim 6, wherein Step 3 includes: Step 31: Construct a candidate rule premise for each second sample, which is expressed as: Step 32: Use the data-driven confidence rule base to predict the label of the conclusion of the candidate rule premise, and the predicted pseudo-label is expressed as: The confidence rule carrying the pseudo-label is expressed as: Among them, Step 33: Use the confidence rule to select the candidate rule premise that can provide the maximum mutual information for the unlabeled rule premise, which is expressed as: Step 34: Transform the selected candidate rule premise, which is expressed as: Step 35, the selected candidate rule premise is selected and given to the expert to give the corresponding confidence distribution conclusion so that the expert can mark it, and the rule after being marked by the expert is added to the current knowledge-driven rule base Step 36: Repeat Steps 31 to 35 until the stop condition is reached to obtain the knowledge-driven confidence rule base.

8. The method for target recognition by data and knowledge fusion according to claim 6, characterized in that Step 4 includes: Step 41, concatenate the binary encodings of all data-driven confidence rule bases and knowledge-driven confidence rule bases to form a chromosome Among them, for each rule premise and rule in all data-driven confidence rule bases and knowledge-driven confidence rule bases, the first p positions are used to encode the rule premise features, 1 represents that the feature is used, and 0 represents that the feature is discarded in the rule; the last position represents the selection of the rule, 1 represents that the rule is selected, and 0 represents that the rule is deleted from the rule base; Step 42: Use the genetic algorithm to fuse the data-driven confidence rule base and the knowledge-driven confidence rule base on the chromosome to obtain the hybrid confidence rule base.

9. The target recognition method for data and knowledge fusion according to claim 6, characterized in that Before Step 42, the data and knowledge fusion target recognition method further includes: Step a: Construct a fitness evaluation index according to the average number of features in each rule and the number of rules; Step b: Use the fitness evaluation index as the index for the genetic algorithm to select the fitness value, and select the fitness to use roulette wheel selection to select chromosomes with the same size as the initial population; Step c: Pairwise crossover of the selected parent chromosomes with a crossover probability of p c ; Mutate the offspring obtained from the crossover. For each rule, select two mutation sites, one for the features of the rule premise and the other for the rule itself, with a mutation probability of p for both sites m ; Among them, each chromosome has multiple crossover points, and the number of crossover points is the same as the total number of rule premises; Step d: Repeat Steps a to c until the stop condition is met; Step e: Decode the best chromosome of the last generation to obtain the hybrid confidence rule base.

10. An object recognition device for data and knowledge fusion, characterized in that, Includes: A data acquisition module that acquires real-time data generated by a remote warning system's detection target, uses the real-time data as data to be recognized, and acquires a first sample composed of labeled data and a second sample composed of unlabeled data from an existing database; A first construction module that sequentially constructs first rule premises for the first sample, and constructs a data-driven confidence rule base by performing fuzzy partitioning of the first rule premises according to language labels and fusing rule results; A second construction module that uses the partitioning method of the rule premises in the data-driven confidence rule base to construct candidate rule premises for each second sample, selects second rule premises from the candidate rule premises using the mutual information criterion, and provides the second rule premises to an expert to construct a knowledge-driven confidence rule base; A rule fusion module that uses a genetic algorithm to fuse the data-driven confidence rule base and the expert-driven confidence rule base to obtain a hybrid confidence rule base; A type recognition module that classifies the real-time data using the rule premises in the hybrid confidence rule base to determine the type of the real-time target.

Images