An online evaluation method for UAV formation performance based on incremental learning

Abstract

The present invention discloses an online evaluation method for the performance of a UAV formation based on incremental learning, comprising: generating different evaluation index values ​​of the formation at the th moment according to different attribute values ​​of the formation UAVs at the th moment; using a performance evaluation model based on a confidence rule base, performing performance evaluation according to these different evaluation index values, and obtaining the performance evaluation result of the formation at the th moment; in order to overcome the shortcomings of traditional parameter training methods that they take a long time to train when encountering new sample data and lack flexibility, the present invention introduces an incremental learning method to achieve parameter optimization of the performance evaluation model based on the confidence rule base. The training set used in the th round of incremental learning contains highly representative historical samples and the th round of newly added samples, which effectively reduces the training samples while maintaining excellent model accuracy. The advantage of the present invention is that it reduces the complexity of model parameter updates and gives the model the ability to continuously learn on sample data collected online.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle flight control, and in particular relates to an online evaluation method for unmanned aerial vehicle formation performance based on incremental learning. Background Art

[0002] At present, the role of UAV technology is becoming increasingly important, especially in some special missions. UAV formations improve the efficiency and flexibility of mission execution through the collaborative work of multiple aircraft, but they also face challenges brought by complex mission execution environments and strict mission requirements. In this context, UAV formation performance evaluation technology is particularly critical. It can effectively monitor the collaborative performance of the formation and detect potential faults in time to improve operational reliability. It is also the basis for optimizing formation performance and provides strong technical support for the long-term and high-efficiency mission execution of UAV formations.

[0003] Performance evaluation methods are mainly divided into analytical model-based methods and artificial intelligence-based methods. Although analytical models have been widely used in traditional scenarios, they rely too much on precise mathematical modeling and are difficult to cope with complex situations. In contrast, artificial intelligence methods can learn and adapt from data by utilizing data analysis and machine learning techniques, thereby significantly improving the ability to handle complex scenarios.

[0004] Drone formations have the characteristics of strong uncertainty in the operating environment and complex and changeable data features. Currently, most existing AI-based performance evaluation methods use offline learning, which requires retraining on the entire data set when data features change, resulting in high model training costs and insufficient flexibility. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides an online evaluation method for UAV formation performance based on incremental learning.

[0006] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] The present invention provides an online evaluation method for UAV formation performance based on incremental learning, comprising:

[0008] Get the first The first attribute value and the The second attribute value of the moment;

[0009] According to the obtained first attribute value and the second attribute value, the first The first evaluation index value at the moment and the The second evaluation index value at the moment;

[0010] A performance evaluation model based on a confidence rule base is used. The first evaluation index value at the moment and the The performance of the UAV formation is evaluated by the second evaluation index value at the time Performance evaluation results at each moment;

[0011] The performance evaluation model based on the confidence rule base is obtained by training the initial performance evaluation model based on the confidence rule base using online sampled sample data and incremental learning. The training set used in the round of incremental learning contains the target history samples and the The target historical sample is the newly added sample of the previous round. The most representative of all training samples used in round 1 incremental learning samples, is a positive integer greater than or equal to 1, is a positive integer greater than 1.

[0012] Compared with the prior art, the present invention has the following beneficial effects:

[0013] The present invention utilizes a performance evaluation model based on a confidence rule base, and can effectively track the performance level of the drone formation according to the first attribute value and the second attribute value of the drone formation. When training the performance evaluation model, the present invention trains the model parameters online based on the incremental learning technology. Compared with the traditional machine learning training method, this method does not need to retrain the entire model when encountering new training data, which can reduce the training consumption of the model and give the model continuous learning ability and more dynamic adaptability; in addition, the present invention uses some of the most representative samples in the historical samples for model training in each incremental learning to maintain the accuracy of the model on the learned data. In this way, the model can solve the problem of model effect degradation caused by the fluctuation of incremental data quality while learning new data, and slow down the "catastrophic forgetting" phenomenon inherent in incremental learning. Therefore, the present invention can reduce the complexity of model parameter updating and improve model performance.

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of an incremental learning-based online evaluation method for UAV formation performance provided by an embodiment of the present invention;

[0016] Figure 2 is a schematic diagram of an exemplary parameter training process based on incremental learning provided by an embodiment of the present invention;

[0017] Figure 3 is a schematic diagram of the trajectory of a UAV formation under an exemplary condition without faults and interference provided by an embodiment of the present invention;

[0018] Figure 4 is an exemplary schematic diagram of the trajectory of a UAV formation under the condition of considering failure and interference provided by an embodiment of the present invention;

[0019] Figure 5 is a schematic diagram of a monitoring curve of a first evaluation index value of an exemplary UAV formation under different performance states provided by an embodiment of the present invention;

[0020] Figure 6 is a schematic diagram of a monitoring curve of a second evaluation index value of an exemplary UAV formation under different performance states provided by an embodiment of the present invention;

[0021] Figure 7 A schematic diagram of the evaluation accuracy change curves of different training methods provided by the examples of the present invention;

[0022] Figure 8 A schematic diagram comparing the total amount of training samples and the total training time for different training methods provided in the examples of the present invention. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0024] Figure 1 is a flow chart of an incremental learning-based online evaluation method for UAV formation performance provided by an embodiment of the present invention, such as Figure 1 As shown, the method includes:

[0025] S101. Get the first and last position of each drone in the drone formation. The first attribute value and the The second attribute value of the moment.

[0026] S102: Generate the first attribute value and the second attribute value of the drone formation according to the obtained first attribute value and the second attribute value. The first evaluation index value at the moment and the The second evaluation index value at the time.

[0027] S103, using a performance evaluation model based on a confidence rule base, according to The first evaluation index value at the moment and the The performance of the UAV formation is evaluated by the second evaluation index value at the moment The performance evaluation result at the moment; wherein, the performance evaluation model based on the confidence rule base is obtained by training the initial performance evaluation model based on the confidence rule base using the sample data of online sampling and the incremental learning method, wherein, the first The training set used in the round of incremental learning contains the target history samples and the The target historical sample is the newly added sample of the previous round. The most representative of all training samples used in round 1 incremental learning samples, is a positive integer greater than or equal to 1, is a positive integer greater than 1.

[0028] It should be noted that the first The first round of incremental learning is Batch parameter training.

[0029] In the present invention, the drone formation is a leaderless drone formation, that is, there is no drone as a formation leader in the drone formation. The size of the drone formation is ,Moreover, the performance of the UAV formation is affected by unknown internal faults and external interference, is a positive integer greater than 1. drones (also called unmanned aerial vehicles The kinematic characteristics of ) can be expressed by formula (1):

[0030] (1);

[0031] in, The value range is 1 to , is the Cartesian coordinate, representing the drone Real-time location, yes The first derivative of yes The first derivative of represents the rate of change. Indicates drone Ground speed, For influence Control input of Indicates drone The heading angle, For influence control input. For Related to the response time constant.

[0032] Will Decomposing in the x-axis and y-axis directions, we can obtain the following expressions for heading angle and ground speed:

[0033] (2);

[0034] By taking the derivative of formula (2), we get formula (3):

[0035] (3);

[0036] make represents the virtual input vector, With the actual input vector The conversion relationship between them can be obtained from formula (1) to formula (3), which is specifically expressed as:

[0037] (4);

[0038] So, As the state vector, As control input, build a drone taking into account internal faults and external disturbances The state space expression of , we get formula (5):

[0039] (5);

[0040] in, is the state vector, is the input vector, is the output vector, is the actuator fault vector, is the sensor fault vector, Represents external interference. , , , and are all constant matrices of suitable dimension. yes The first derivative of .

[0041] In the embodiment of the present invention, in order to achieve output consistency between the drones in the drone formation, the drones The following control protocol is adopted:

[0042] (6);

[0043] In formula (6), Represents the coefficient corresponding to the formation communication topology. Able to receive drones When sending status information, The value is 1, otherwise, The value of is 0. It's a drone The corresponding preset reference vectors represent the desired relative positions between the drones to maintain the formation shape. It's a drone The corresponding preset reference vector; is the preset reference vector shared by the UAV formation. and All are preset gain matrices to ensure consistency. It's a drone The output vector of .

[0044] In the present invention, the drone The first attribute value at the moment is The ground speed value at the moment, the UAV The second attribute value of the moment is The heading angle value at the moment.

[0045] In the present invention, the above S102 is implemented by the following steps:

[0046] S1021. Determine the average value of the absolute value of the first attribute value obtained, and obtain the first The first evaluation index value at the moment.

[0047] For example, for a fleet size of The drone formation, the first The first evaluation index value at the moment The expression of is formula (7a): (7a).

[0048] S1022. Determine the average value of the absolute value of the acquired second attribute value, and obtain the first The second evaluation index value at the time.

[0049] For example, for a fleet size of The drone formation, the first The second evaluation index value at the moment The expression of is formula (7b): (7b).

[0050] In the present invention, the initial performance evaluation model based on the confidence rule base includes Rules and Preset performance evaluation results , among which, The rule contains an initial rule weight ,and The preset performance evaluation results correspond to Initial result confidence , and the input The evaluation index values ​​correspond to Initial attribute weights , and the input The evaluation index values ​​correspond to Preset index reference value ; The value range is 1 to ; and are all positive integers greater than 1. For the initial performance evaluation model based on the confidence rule base, Rule weights in rules , Result Confidence and attribute weights are all preset initial values ​​given by experts based on relevant field knowledge. Accordingly, after the initial performance evaluation model based on the confidence rule base is trained, the performance evaluation model based on the confidence rule base includes Rules and Preset performance evaluation results , among which, A rule contains a rule weight ,and The preset performance evaluation results correspond to Confidence of results , and the input The evaluation index values ​​correspond to Attribute weight , and the input The evaluation index values ​​correspond to Preset index reference value For the performance evaluation model based on the confidence rule base obtained after training, Rule weights in rules , Result Confidence and attribute weights are all optimized values ​​obtained after learning. It should be noted that In the rules , are the same. For example, The four preset performance evaluation results of "good", "better", "poor" and "bad" can be represented respectively.

[0051] Exemplarily, the performance evaluation model based on the confidence rule base Rules It can be expressed as:

[0052] (8);

[0053] in, , , …, Indicates input An evaluation index value.

[0054] The performance evaluation model based on the confidence rule base of the present invention, and its parameter learning process based on incremental learning can be described as the following steps:

[0055] S001. Obtain an initial performance evaluation model based on a confidence rule base.

[0056] S002, in the process of During the round of incremental learning, the first The new samples of the round are obtained The performance evaluation model based on the confidence rule base is obtained after rounds of incremental learning; When it is 1, The performance evaluation model based on the confidence rule base obtained after a round of incremental learning is the initial performance evaluation model based on the confidence rule base.

[0057] Here, each sample is a set of evaluation index values ​​consisting of a first evaluation index value and a second evaluation index value. Each round of new samples is obtained by online observation of the UAV formation.

[0058] S003, Once Upon a Time Select the most representative training samples used in the round of incremental learning samples as the target historical samples.

[0059] S004, will The sample set consisting of the newly added samples of the round and the target historical samples is used as the The training set used in the round of incremental learning.

[0060] S005, adopt The training set used in the round of incremental learning is The performance evaluation model based on the confidence rule base obtained after the round of incremental learning is trained to obtain the first After rounds of incremental learning, the performance evaluation model based on the confidence rule base is obtained. This iteration is repeated until the preset termination condition is met to obtain the performance evaluation model based on the confidence rule base.

[0061] Here, the preset termination condition may be that the number of iterations reaches a preset number of iterations, or that the evaluation accuracy of the model reaches a preset accuracy threshold, etc. It can be set according to actual needs, and the present invention is not limited to this.

[0062] For example, when training the model, the Projection Covariance Matrix Adaptation Evolutionary Strategies (P-CMA-ES) can be used to optimize the model parameters (i.e., train the model) according to the determined training set. In this process, the parameters should satisfy the constraints of the following formula (9) to maintain the physical meaning of the parameters:

[0063] (9);

[0064] For example, the objective function when training the model can be expressed as:

[0065] (10);

[0066] in, express function, represents the number of correct evaluations, Represents the number of incorrect evaluations. is the parameter vector to be optimized.

[0067] Specifically, in the During the round of incremental learning, the most representative The samples are obtained by using a highly representative historical sample selection algorithm, which includes the following steps:

[0068] S1, move forward All training samples used in the round of incremental learning (i.e. training samples) are input into In the performance evaluation model based on the confidence rule base obtained after rounds of incremental learning, a one-to-one correspondence is obtained. model outputs; among them, the performance evaluation model based on the confidence rule base has The model output corresponding to a sample is Contains the confidence level that the sample belongs to each preset performance evaluation result ; and is a positive integer greater than 1, is the default value and Less than .

[0069] S2. For Each sample in the training samples is calculated based on the true probability distribution of the sample belonging to each preset performance evaluation result, the model output corresponding to the sample, and the preset weight coefficient , determine the uncertainty coefficient of the sample ; Among them, the sample belongs to The true probability distribution of the preset performance evaluation results represents the true label of the sample.

[0070] Specifically, for a sample, the entropy of the evaluation result of the sample is determined according to the confidence level contained in the model output of the sample. ; According to the true probability distribution of each preset performance evaluation result, the sample belongs to , and the model output corresponding to the sample, determine the cross entropy of the evaluation result of the sample ; Use preset weight coefficients , the entropy of the evaluation result of this sample and cross entropy Sum and get the uncertainty coefficient of the sample .

[0071] For example, the uncertainty coefficient for a sample is The expression of is formula (11):

[0072] (11);

[0073] in, Represents the model output of this sample The sample contained in The confidence level of the preset performance evaluation results , The value range is 1 to , Indicates that the sample represented by the true label belongs to Preset performance evaluation results The true probability distribution of represents the entropy of the evaluation result of the sample, Represents the cross entropy of the evaluation result of this sample.

[0074] S3, according to the uncertainty coefficient ,from Select from the training samples The most representative samples samples.

[0075] Exemplary, most representative The first Samples It can be expressed as:

[0076] (12);

[0077] in, Indicates The uncertainty coefficient of the sample. is a preset fixed value, so that the computing resources consumed by parameter training for each batch can be stabilized.

[0078] For example, when selecting the most representative sample according to the uncertainty coefficient, a greedy search method may be used for selection.

[0079] For example, Figure 2 Figure 1 is a schematic diagram of the parameter training process based on incremental learning. Figure 2 As shown, when the model is trained online, the initial model parameters are first obtained, that is, the initial model is first obtained. In the first round of incremental learning, the P-CMA-ES method is used to train the model according to the first round of newly added samples to obtain the first round of incremental learning results (that is, the model obtained after the first round of incremental learning). After that, the highly representative historical sample selection algorithm is used to select the target historical samples from the newly added samples in the first round. After that, the selected target historical samples and the obtained second round of newly added samples and the P-CMA-ES method are used to continue training the model obtained after the first round of incremental learning to obtain the second round of incremental learning results. After that, the highly representative historical sample selection algorithm is used to select the target historical samples from all the training samples used in the first and second rounds of incremental learning. After that, the target historical samples selected this time and the obtained third round of newly added samples and the P-CMA-ES method are used to continue training the model obtained after the second round of incremental learning to obtain the third round of incremental learning results. This cycle is repeated until the optimization stop condition is reached, and the optimized model parameters can be obtained. Incremental learning is a technology that allows a model to learn from newly added samples and update existing parameters. It does not need to retrain the entire model when encountering new training data, thereby greatly saving computing resources and training time, and enabling the model to achieve efficient dynamic adaptation in complex environments with limited resources. At the same time, the present invention adopts a playback-based strategy to select a small number of highly representative historical samples to participate in model training before each training to maintain the accuracy of the model on the learned data. Using this method, while the model is learning new data, the problem of model effect degradation caused by fluctuations in the quality of incremental data can be solved, while mitigating the "catastrophic forgetting" phenomenon inherent in incremental learning.

[0080] In the present invention, the above S103 can be implemented by the following steps:

[0081] S1031, determine the first evaluation index value and the second evaluation index value respectively The matching degree between the corresponding preset index reference values ​​contained in the rules is obtained to obtain the matching degree matrix .

[0082] For example, the matching matrix The expression is:

[0083] (13);

[0084] in, Indicates Evaluation index value With The corresponding rule The matching degree between the preset index reference values, The value range is 1 to .

[0085] Specifically, It can be calculated by the following recursive relationship: (14);

[0086] in, Indicates The result of the recursive calculation is: represents the Kronecker product operation, The number of rows is A column vector with all elements having the value 1. Indicates The number of preset index reference values ​​for the input evaluation index values, The value of is , is the product operator. express and between the preset index reference values The matching degree vector , represents the transpose symbol, The preset index reference value is The rules contain The corresponding reference values ​​of all preset indicators. For example, when Greater than The first of the preset indicator reference values Preset index reference value hour, and The degree of match between The expression of is formula (15a), when Less than The first of the preset indicator reference values Preset index reference value hour, and The degree of match between The expression of is (15b):

[0087] (15a);

[0088] (15b);

[0089] in, The value range is 1 to .

[0090] S1032, according to the matching matrix , The rule weights and attribute weights contained in the rule determine the The activation level of the rule .

[0091] For example, The expression is as follows:

[0092] (16);

[0093] in, For the The rule weight of the rule.

[0094] S1033, according to Rules and The activation degree of the rules is calculated using the evidence reasoning algorithm. The fusion of the rules gives The preset performance evaluation results correspond to A confidence level.

[0095] For example, The confidence level corresponding to the preset performance evaluation results The expression is as follows:

[0096] (17);

[0097] in, Indicates The confidence level corresponding to the preset performance evaluation results, The value range is 1 to , is the normalization factor, is the activation weight, Indicates The activation level of the rule, Indicates The rule contains the The confidence level of the result corresponding to the preset performance evaluation result, Indicates The rule contains the The confidence level of the result corresponding to the preset performance evaluation result.

[0098] S1034, will A preset performance evaluation result corresponding to the maximum value of the confidence levels is used as the first Performance evaluation results at the moment .

[0099] For example, The expression is as follows:

[0100] (18);

[0101] The present invention utilizes a performance evaluation model based on a confidence rule base, and can effectively track the performance level of the drone formation according to the first attribute value and the second attribute value of the drone formation. When training the performance evaluation model, the present invention trains the model parameters online based on the incremental learning technology. Compared with the traditional machine learning training method, this method does not need to retrain the entire model when encountering new training data, which can reduce the training consumption of the model and give the model continuous learning ability and more dynamic adaptability; in addition, the present invention uses some of the most representative samples in the historical samples for model training in each incremental learning to maintain the accuracy of the model on the learned data. In this way, the model can solve the problem of model effect degradation caused by the fluctuation of incremental data quality while learning new data, and slow down the "catastrophic forgetting" phenomenon inherent in incremental learning. Therefore, the present invention can reduce the complexity of model parameter updating and improve model performance.

[0102] Hereinafter, the technical effects of the present invention will be further described in combination with simulation experiment data.

[0103] Before building the performance evaluation model, sample data is first collected based on the mathematical simulation results. , preset ground speed , preset heading angle , and the communication topology of the UAV formation is fully connected, that is, . Set the sampling interval to 0.01 seconds. , the remaining parameters in formulas (1), (5), and (6) are set as follows:

[0104] ;

[0105] ;

[0106] ;

[0107] .

[0108] When the internal fault and external interference in formula (6) are both 0, the trajectory of the UAV formation is as follows: Figure 3 On this basis, by gradually adjusting the coefficient floating range to increase the impact of faults and interference, the attribute monitoring data of the drone formation under four different performance states of "good", "better", "poor" and "bad" can be obtained. The running trajectory of the drone formation in an embodiment considering faults and interference is shown as follows Figure 4 As shown. According to the observed simulation data, the UAV formation in this embodiment reaches the consistency condition after 40 seconds, so samples should be collected from this moment for model verification. Accordingly, according to the synthesis formula provided by formula (7), the first evaluation index value and the second evaluation index value monitoring curve for performance evaluation from 40 seconds to 140 seconds are obtained, as shown in Figures 5 and 6 As shown. The monitoring data is discretized. Specifically, the continuous monitoring information is divided into 1 second sampling intervals, and finally 100 samples of each preset performance evaluation result, totaling 400 samples, are obtained for model training and verification. The training set and the verification set are divided in a ratio of 6:4. The training set will participate in the incremental learning process, and the model training will be carried out in batches. Then, the samples in the verification set will be used to verify the effectiveness of the method provided by the present invention.

[0109] After the above data collection steps are completed, an initial performance evaluation model based on a confidence rule base is created based on the relevant field knowledge provided by experts. Let the preset index reference value set of the first evaluation index value be , the preset index reference value set of the second evaluation index value is Next, let the preset performance evaluation result utility value set be The initial value settings of the confidence rule base parameters are shown in Table 1.

[0110] Table 1

[0111]

[0112] Thus, the initial performance evaluation model is established and the model learning optimization step is entered. The model is trained with 12 batches of incremental learning using a training sample set of 240. For each batch of incremental learning, the number of new samples used for each preset performance evaluation result (i.e., each performance state) is set to 5, the number of the most representative historical samples is set to 10, and the weight coefficient is set to It is set to 1. A total of 30 rounds of random experiments were conducted and the average evaluation accuracy was taken to verify the effectiveness of the method provided by the present invention.

[0113] Test results such as Figures 7 and 8As shown, the evaluation accuracy of the model under the incremental learning training method stabilizes and reaches 96.7% after the fifth batch of incremental training, which proves the effectiveness of the method provided by the present invention. Compared with the traditional offline learning training method, the method provided by the present invention reduces the sample size by 64%, so the training time is also greatly reduced. Although offline learning has a higher average accuracy, the parameter training method based on incremental learning proposed in the present invention has more advantages in resource-constrained environments due to its higher efficiency and lower computational overhead.

[0114] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0115] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0116] In the specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality of components. Certain measures are recorded in different embodiments, but this does not mean that these measures cannot be combined to produce good effects.

[0117] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. An online evaluation method for UAV formation performance based on incremental learning, characterized in that: include: Obtain the first attribute value at time t and the second attribute value at time t of each drone in the drone formation; Generate a first evaluation index value at time t and a second evaluation index value at time t of the UAV formation respectively according to the obtained first attribute value and the second attribute value; Using a performance evaluation model based on a confidence rule base, performing performance evaluation according to the first evaluation index value at the t-th moment and the second evaluation index value at the t-th moment, and obtaining a performance evaluation result of the UAV formation at the t-th moment; The performance evaluation model based on the confidence rule base is obtained by training the initial performance evaluation model based on the confidence rule base using online sampled sample data and an incremental learning method, wherein the training set used in the s-th round of incremental learning includes target historical samples and newly added samples in the s-th round, and the target historical samples are the g most representative samples of all training samples used in the previous s-1 rounds of incremental learning, s is a positive integer greater than or equal to 1, and g is a positive integer greater than 1; The performance evaluation model based on the confidence rule base includes L rules and N preset performance evaluation results, wherein the kth rule includes a rule weight, N result confidences corresponding to the N preset performance evaluation results, M attribute weights corresponding to the input M evaluation index values, and M preset index reference values ​​corresponding to the input M evaluation index values; the value of k is 1 to L; N and L are both positive integers greater than 1; Among them, all the training samples used in the first s-1 rounds of incremental learning are Gs training samples, and the most representative g samples required for the s-th round of incremental learning are obtained by the following method: The Gs training samples are respectively input into the performance evaluation model based on the confidence rule base obtained after the s-1th round of incremental learning, and Gs model outputs are obtained in one-to-one correspondence; wherein the model output corresponding to one sample includes the confidence level that the sample belongs to each preset performance evaluation result; Gs is a positive integer greater than 1, g is a preset value and g is less than Gs; For each sample in the Gs training samples, determining the entropy of the evaluation result of the sample according to the confidence level included in the model output of the sample; Determine the cross entropy of the evaluation result of the sample according to the true probability distribution of the sample belonging to each preset performance evaluation result and the model output corresponding to the sample; wherein the true probability distribution of the sample belonging to N preset performance evaluation results represents the true label of the sample; According to the preset weight coefficient η, the entropy and cross entropy of the evaluation result of the sample are summed to obtain the uncertainty coefficient of the sample in, The expression is: Wherein, P(n) represents the confidence level that the sample contained in the model output of the sample belongs to the nth preset performance evaluation result, n is 1 to N, Q(n) represents the true probability distribution that the sample belongs to the nth preset performance evaluation result, H(P) represents the entropy of the evaluation result of the sample, and H(P,Q) represents the cross entropy of the evaluation result of the sample; From the Gs training samples, the Gs uncertainty coefficients corresponding to each other In the equation, select the g smallest uncertainty coefficients The g smallest uncertainty coefficient The one-to-one corresponding g samples are used as the most representative g samples.

2. The method for online evaluation of UAV formation performance based on incremental learning according to claim 1 is characterized in that: The drone formation is a leaderless drone formation, the first attribute value is a ground speed value, and the second attribute value is a heading angle value.

3. The method for online evaluation of UAV formation performance based on incremental learning according to claim 1 is characterized in that: The step of respectively generating a first evaluation index value at the tth moment and a second evaluation index value at the tth moment of the UAV formation according to the obtained first attribute value and the second attribute value comprises: Determine an average value of the absolute values ​​of the first attribute values ​​obtained to obtain a first evaluation index value of the UAV formation at the tth moment; An average value of the absolute values ​​of the acquired second attribute values ​​is determined to obtain a second evaluation index value of the UAV formation at the tth moment.

4. The method for online evaluation of UAV formation performance based on incremental learning according to claim 1 is characterized in that: The performance evaluation model based on the confidence rule base is obtained by the following method: Obtaining an initial performance evaluation model based on a confidence rule base; When performing the s-th round of incremental learning, the newly added samples of the s-th round are obtained online, and the performance evaluation model based on the confidence rule base obtained after the s-1-th round of incremental learning is obtained; when s is 1, the performance evaluation model based on the confidence rule base obtained after the s-1-th round of incremental learning is the initial performance evaluation model based on the confidence rule base; Select the most representative g samples from the training samples used in the previous s-1 rounds of incremental learning as the target historical samples; Using the sample set consisting of the newly added samples in the s-th round and the target historical samples as the training set used in the s-th round of incremental learning; The training set used in the s-th round of incremental learning is used to train the performance evaluation model based on the confidence rule base obtained after the s-1-th round of incremental learning to obtain the performance evaluation model based on the confidence rule base obtained after the s-th round of incremental learning. This iteration is performed until a preset termination condition is met to obtain the performance evaluation model based on the confidence rule base.

5. The method for online evaluation of UAV formation performance based on incremental learning according to claim 1 is characterized in that: The performance evaluation model based on the confidence rule base is adopted to perform performance evaluation according to the first evaluation index value at the t-th moment and the second evaluation index value at the t-th moment, and obtain the performance evaluation result of the UAV formation at the t-th moment, including: Determine the matching degree between the first evaluation index value and the second evaluation index value and the corresponding preset index reference values ​​contained in the L rules, respectively, to obtain a matching degree matrix; Determining the activation degree of the kth rule according to the matching degree matrix, the rule weight and the attribute weight; According to the L rules and the activation degrees of the L rules, adopting an evidential reasoning algorithm to fuse the L rules, and obtaining N confidence levels corresponding to the N preset performance evaluation results; A preset performance evaluation result corresponding to the maximum value of the N confidence levels is used as the performance evaluation result of the UAV formation at the tth moment.

6. The method for online evaluation of UAV formation performance based on incremental learning according to claim 5 is characterized in that: The expression of the confidence level corresponding to the nth preset performance evaluation result is as follows: Among them, βn(t) represents the confidence level corresponding to the n-th preset performance evaluation result, μ(t) is the normalization factor, ωk(t) is the activation weight, ωk(t) represents the activation level of the k-th rule, βn′,k represents the result confidence level corresponding to the n'-th preset performance evaluation result contained in the k-th rule, and βn,k represents the result confidence level corresponding to the n-th preset performance evaluation result contained in the k-th rule.

Images