Unmanned trunking communication network performance evaluation method based on neural network

Abstract

Aiming at the communication network problem of an airborne communication module of a large-scale unmanned cluster, the invention provides an airborne communication network evaluation method based on a neural network, and the method comprises the following steps: selecting an evaluation index; determining the weight of each evaluation index; standardizing each index value; constructing an output network score; and sample division. A core module of the scheme is a neural network model, specifically, the neural network model receives related data from an onboard network as input, carries out calculation and processing learning through a hidden layer and weight parameters of the network, and finally outputs a prediction result of a network evaluation condition; when the method is applied to an unmanned cluster flight task, communication faults and abnormal conditions can be found in time, timely feedback and decision support are provided, and task execution efficiency and safety are improved.

Description

Technical Field

[0001] The present invention relates to the technology of unmanned cluster field, and relates to an onboard communication network performance evaluation method, and specifically to an onboard communication network performance evaluation method for large-scale unmanned clusters. Background Art

[0002] With the continuous advancement of technology, unmanned aerial vehicle (UAV) swarms are playing an increasingly important role in search and rescue, logistics, and other fields. Unmanned aerial vehicle (UAV) swarms require real-time communication with ground control centers or other equipment when performing their missions. However, when UAVs fly in complex environments, network communication quality can be affected by interference, congestion, or weak signals. This can lead to communication delays, data loss, or connection interruptions, hindering the normal execution of UAV missions. Therefore, it is necessary to develop a method to evaluate the quality of UAV network communication to improve communication performance and stability.

[0003] Currently, the main methods for evaluating the status of in-flight communication networks are the Analytic Hierarchy Process (AHP) and the Ideal Point Method. The AHP is a quantitative analysis method for multi-criteria decision-making that breaks down objectives into different criteria and sub-criteria, forming a hierarchical structure. Within each layer, the relative importance of criteria is determined using pairwise comparisons, conducted through expert judgment or questionnaires. A scale is used to express the relative importance of two criteria, resulting in relative weights and a comprehensive evaluation. The Ideal Point Method is a commonly used multi-objective optimization method that identifies performance indicators and their ideal states. The indicators are normalized and the degree of deviation from the ideal state is determined by calculating the distance between each performance indicator and the ideal point. A weighted summation is then used to comprehensively evaluate network performance using matrix calculations. The evaluation results of these two methods depend on expert judgment and data quality, and their over-reliance on quantitative analysis of evaluation indicators limits their flexibility.

[0004] Currently, research on the use of neural networks to evaluate the effectiveness of mobile communication systems is relatively limited. However, there are methods for evaluating the effectiveness of military communication systems using fuzzy neural networks, BP neural networks, ANPs, and cloud models. These methods can analyze network conditions more dynamically and efficiently. Therefore, we consider designing a neural network-based performance evaluation method for unmanned swarm communication networks. By using the neural network to learn from large amounts of data to build a model, we can handle the complex interactions of various parameters of the unmanned swarm, thereby improving the real-time efficiency and accuracy of network evaluation. Summary of the Invention

[0005] The technical problem to be solved by the invention is to provide a method for evaluating the performance of an unmanned cluster communication network based on a neural network, which specifically includes the following steps:

[0006] Step 101, selecting evaluation indicators;

[0007] The evaluation indicators constructed include: data transmission rate, information update rate, packet loss rate, communication bit error rate and point-to-point link delay. The explanation and calculation methods of the specific indicators are as follows;

[0008] (1) Data transmission rate

[0009] Data transmission rate refers to the amount of data transmitted per unit time. The data transmission rate R data The calculation formula is as follows:

[0010]

[0011] Among them, B total Indicates the total number of bytes of data transmitted during the transmission time, T total Represents data transmission time;

[0012] (2) Information update rate

[0013] The information update rate is the number of data packets updated per second, expressed as R update The calculation formula is as follows:

[0014]

[0015] Among them, N pack Indicates the number of received data packets, T update Indicates update time;

[0016] (3) Communication bit error rate

[0017] The communication bit error rate is expressed as BER;

[0018]

[0019] Among them, Byte wrong Indicates the number of erroneous bytes in the data packet received during the test time, Byte trans Indicates the number of bytes in the data packet sent during the test time;

[0020] (4) Packet loss rate

[0021] The packet loss rate is expressed as R loss express;

[0022]

[0023] Among them, N trans Indicates the number of packets sent, N rec Indicates the number of packets received;

[0024] (5) Point-to-point link delay

[0025] The transmission time of information from the source to the receiving point is represented by RT;

[0026] RT=T rec -T trans (5)

[0027] Among them, T rec Indicates the receiving time of the receiver, T trans Indicates the sending time of the sender;

[0028] Step 102, determining the weight of each evaluation indicator;

[0029] Given R data 、R update ,BER,R loss The weights ω corresponding to the five indicators of RT are 0.3, 0.1, 0.2, 0.2 and 0.2 respectively, and the sum of the weights is 1;

[0030] Step 103, standardize each indicator value;

[0031] Normalize the five indicators of data transmission rate, information update rate, communication bit error rate, packet loss rate and point-to-point link delay; the positive indicators are standardized using the following formula, with X r Represents the normalized data:

[0032]

[0033] Among them, X represents the actual value of the indicator, X max Indicates the maximum value of the data, X min Indicates the minimum value of the data.

[0034] Negative indicators are normalized using the following formula, with X l express:

[0035]

[0036] Step 104, constructing an output network score;

[0037] The network score is calculated by formula (8), assuming there are n indicators in total, and the weighted sum is X sum express;

[0038]

[0039] Among them, X r / l represents the standardized index, represents the i-th standardized index, ω iRepresents the weight of the i-th indicator; for the existing data set, use formula (8) to construct its theoretical score, that is, the label; assume that the maximum value of the data label in the form of weighted sum is Sum Maxweighted , get the score of the data set through formula (9), and use G n express:

[0040]

[0041] Step 105, sample division;

[0042] In the subsequent model training, a part of the data in the sample set S is used as the training set, and the rest of the data is used as the test set. In the network training and learning phase, the network weight parameters are initialized and the training set data is input into the network for training until the target requirements are met or all samples are trained.

[0043] By collecting various indicators of the drone swarm communication network, the data is standardized using the methods in steps 103 and 104, and labels are obtained as their theoretical scores to form a sample set S. The form of the i-th data in the sample set S is:

[0044]

[0045] in, BER i , RT i , They represent the data transmission rate, information update rate, communication bit error rate, packet loss rate, point-to-point link delay and label value of the i-th sample data in the sample set respectively.

[0046] In one embodiment of the present invention, in step 103, the positive indicator includes R data 、R update ,Negative indicators include packet loss rate, communication bit error rate and point-to-point link delay.

[0047] In another embodiment of the present invention, in step 104, the weight coefficient used is such that the label value obtained is between 0 and 1. In actual applications, if the weight is changed, or manual scoring is used as the data label, the label value may be greater than 1, and the theoretical score needs to be scaled to ensure that the score is within 100 points.

[0048] In a specific embodiment of the present invention, in step 105, 70% of the data in S is used as a training set, and the remaining 30% of the data is used as a test set.

[0049] In another embodiment of the present invention, the neural network structure design specifically includes the following steps:

[0050] Step 201, input layer design;

[0051] The number of neurons in the input layer is 5, corresponding to the five indicators of the unmanned cluster communication network: data transmission rate, information update rate, communication bit error rate, communication packet loss rate, and point-to-point link delay;

[0052] Step 202, hidden layer design;

[0053] The hidden layer is one or more layers located between the input layer and the output layer. The hidden layer consists of 10 neurons, and the neurons in each layer are connected in a fully connected manner.

[0054] In the design of the hidden layer, the choice of activation function and optimizer also determines the performance of the neural network; the activation function is ReLU; the loss function is defined to measure the difference between the predicted output of the model and the true label. The loss function is defined as follows:

[0055]

[0056] Where n represents the sample size of the dataset used for training, represents the predicted value of the model, that is, the true value of the network score in the dataset, y represents the true value of the label class, and MSE represents the loss function defined as the mean square error between the true value and the predicted value. The Adam optimizer is used to adjust the model parameters to minimize the loss function.

[0057] Step 203, output layer design;

[0058] The output layer is responsible for generating the output results of the network; the number of neurons in the output layer is 1, and its output result is the prediction score of the network.

[0059] In yet another embodiment of the present invention, neural network training and use specifically include the following steps:

[0060] Step 301: input training sample data;

[0061] The real-world training data is fed into the model. The input data includes five indicators: data transmission rate, information update rate, communication bit error rate, communication packet loss rate, and point-to-point link delay.

[0062] Step 302: Sample data preprocessing

[0063] Sample data preprocessing is divided into data cleaning and data standardization;

[0064] The 3σ principle is used for data cleaning. The mean μ of each column in the data set is calculated one by one, and the standard deviation σ of the data is calculated. Each indicator is traversed. If the value point is more than three standard deviations away from the mean of its column, that is, outside the range of (μ-3σ, μ+3σ), the data point is considered an outlier and the data is deleted from the data set.

[0065] Data normalization sets the data transmission rate R data , information update rate R update , communication bit error rate BER, communication packet loss rate R loss The point-to-point link delay RT is normalized to the range of (0,1), where the data transmission rate and information update rate are normalized using formula (6), and the communication bit error rate, communication packet loss rate, and point-to-point link delay are normalized using formula (7); the theoretical score value is calculated using formula (8) as a label, and G is normalized using formula (9). n Processed to the range of (0,100);

[0066] Step 303: model training;

[0067] After the model receives the input data, it gets the predicted value The label is y, and the loss function is calculated:

[0068]

[0069] During the model training process, the model training is completed when the loss function reaches the minimum value; the predicted value obtained by the model It is determined by each neuron in the hidden layer and is expressed as:

[0070]

[0071] where ω i ,b i are the weight coefficient and bias coefficient of the i-th neuron in the hidden layer, respectively. Only one hidden layer is used, and the gradient value of the loss function with respect to each parameter is calculated using formula (14):

[0072]

[0073] Where x is any weight coefficient or bias coefficient; for each parameter x, the parameter is updated using formula (15):

[0074]

[0075] in, x(n) represents the value of parameter x at the nth iteration, and α is the learning rate;

[0076] The model is trained until the loss function stops decreasing after 10 consecutive iterations, or until the specified number of training rounds is reached;

[0077] Step 304, model verification and use;

[0078] Use the test set to evaluate the model performance during training; input the i-th data into the network to get the predicted value Calculate its difference from the true value y i The absolute error e i :

[0079]

[0080] Calculate the mean absolute error of the test data:

[0081]

[0082] When the mean absolute error reaches the allowable range, the generalization ability of the model is verified;

[0083] The five indicators of the unmanned cluster communication network: data transmission rate, information update rate, communication bit error rate, packet loss rate, and point-to-point link delay, are input into the model, and the model can return the performance evaluation score of the current communication network. By continuously obtaining the indicators of the unmanned cluster communication network, the model can quickly return the score, thereby realizing real-time performance evaluation of the communication network.

[0084] This invention provides a neural network-based method for evaluating the performance of unmanned cluster communication networks. This method leverages the advantages of neural networks' strong generalization, high accuracy, and optimized cluster network design, providing strong support for unmanned cluster communication network performance evaluation. The generalization capability of neural networks enables accurate prediction and control of untrained samples. In unmanned cluster communication network performance evaluation, this means that neural networks can quickly and accurately evaluate new network scenarios or parameter configurations. Compared to traditional evaluation methods, neural network-based evaluation methods offer higher accuracy and stability. By constructing a multi-layer feedforward neural network and training it with experimental data, precise evaluation of communication network performance can be achieved, and the evaluation results are more stable and reliable. Furthermore, the neural network evaluation results can provide valuable support for the design and optimization of unmanned cluster communication networks. By comparing performance under different network parameter configurations, the optimal configuration solution can be selected to improve the performance and efficiency of cluster networking. This invention can rapidly perceive real-time changes in the communication network status during unmanned cluster mission execution, accurately and timely evaluating the communication status of drones. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 shows sample data set acquisition;

[0086] Figure 2 shows the neural network design module;

[0087] Figure 3 Shows the neural network model training flow chart. DETAILED DESCRIPTION

[0088] The present invention will be described in detail below with reference to the accompanying drawings.

[0089] The present invention proposes a neural network-based unmanned cluster communication network performance evaluation method.

[0090] Figure 1 Get a flow diagram for a sample dataset, including:

[0091] Step 101: Select evaluation indicators.

[0092] In order to effectively extract indicators for training, the evaluation indicators are optimized and combined. The evaluation indicators constructed in the present invention include: data transmission rate, information update rate, packet loss rate, communication bit error rate and point-to-point link delay. The explanation and calculation method of the specific indicators are as follows.

[0093] (1) Data transmission rate

[0094] Test the data transmission rate between UAVs and between UAVs and ground stations, including platform and system status messages, target platform motion status messages, etc., to determine the data transmission speed. The data transmission rate refers to the amount of data transmitted per unit time. The data transmission rate R data The calculation formula is as follows:

[0095]

[0096] Among them, B total Indicates the total number of bytes of data transmitted during the transmission time, T total Represents the data transmission time.

[0097] (2) Information update rate

[0098] Test the information update rate between drones and between drones and ground stations, including platform and system status messages, target platform motion status messages, etc., to determine the speed of information update. That is, the number of data packets updated per second, expressed as R update The calculation formula is as follows:

[0099]

[0100] Among them, N pack Indicates the number of received data packets, T update Indicates update time.

[0101] (3) Communication bit error rate

[0102] By comparing the ratio of the number of sent data and received data between the UAV and the ground station that are greater than a certain threshold during the test time, the communication bit error rate is expressed as BER.

[0103]

[0104] Among them, Byte wrong Indicates the number of erroneous bytes in the data packet received during the test time, Byte trans Indicates the number of bytes in the data packet sent during the test time.

[0105] (4) Packet loss rate

[0106] The channel transmission quality is reflected by comparing the difference between the amount of packets sent and received between the UAV and the ground station in unit time. The packet loss rate is expressed as R loss express.

[0107]

[0108] Among them, N trans Indicates the number of packets sent, N rec Indicates the number of received data packets.

[0109] (5) Point-to-point link delay

[0110] Information real-time assessment is the analysis of timestamps. The timestamp of each message is checked to determine the transmission time of the information from the source to the receiving point, expressed as RT. Long delays may indicate network congestion or insufficient processing capacity.

[0111] RT=T rec -T trans (5)

[0112] Among them, T rec Indicates the receiving time of the receiver, T trans Indicates the time when the sender sends the message.

[0113] Step 102: Determine the weight of each evaluation indicator.

[0114] In the present invention, it is necessary to determine the indicator weights to generate the training set data of the network. Based on expert experience and consideration of the actual communication network status, given R data 、R update ,BER,R loss The weights ω corresponding to the five indicators RT are 0.3, 0.1, 0.2, 0.2 and 0.2 respectively, and the sum of the weights is 1. The network training involved in the present invention takes the above weights as an example.

[0115] Step 103: standardize the values ​​of each indicator.

[0116] Before data is transferred to the network, it is necessary to process data of different dimensions to the same scale. In this invention, the five indicators of data transmission rate, information update rate, communication error rate, packet loss rate and point-to-point link delay are normalized. For positive indicators, such as R data 、R update , use the following formula to standardize, with X r Represents the normalized data:

[0117]

[0118] Among them, X represents the actual value of the indicator, X max Indicates the maximum value of the data, X min Indicates the minimum value of the data.

[0119] For negative indicators (such as packet loss rate, communication bit error rate and point-to-point link delay), the formula can be used for normalization, using X l express:

[0120]

[0121] Step 104: construct an output network score.

[0122] Theoretically, the network evaluation score can be calculated by formula (8), assuming there are n indicators in total, and the weighted sum is X sum express.

[0123]

[0124] Among them, X r / l represents the standardized index, represents the i-th standardized index, ω i Represents the weight of the i-th indicator. For an existing data set, use formula (8) to construct its theoretical score, i.e., label. For the weight coefficient adopted by the present invention, the theoretical score obtained is between 0 and 1. In actual application, if the weight is changed, or manual scoring is used as the data label, the theoretical score, i.e., the label value, may be greater than 1. The theoretical score needs to be scaled to ensure that the score is within 100 points. Assume that the maximum value of the data label in the form of weighted sum is Sum Maxweighted , get the score of the data set through formula (9), and use G n express:

[0125]

[0126] Step 105: Sample division.

[0127] By collecting various indicators of the drone swarm communication network, the data is standardized using the methods in steps 103 and 104, and labels are obtained as their theoretical scores to form a sample set S. The i-th data in the sample set S is in the form of.

[0128]

[0129] in, BER i , RT i , They represent the data transmission rate, information update rate, communication bit error rate, packet loss rate, point-to-point link delay and label value of the i-th sample data in the sample set respectively.

[0130] The present invention uses 70% of the data in S as a training set and the remaining 30% of the data as a test set. During the network training and learning phase, the network weight parameters are initialized and the training set data is input into the network for training until the target requirements are met or all samples are trained.

[0131] Figure 2 Design modules for neural networks, including:

[0132] Step 201: input layer design.

[0133] The input layer is the first layer of the neural network, responsible for receiving the input data parameter part. The number of neurons in the input layer should match the number of input items. In the present invention, the number of neurons in the input layer is 5, corresponding to the five indicators of the unmanned cluster communication network: data transmission rate, information update rate, communication bit error rate, communication packet loss rate, and point-to-point link delay.

[0134] Step 202: hidden layer design.

[0135] A hidden layer is one or more layers located between the input and output layers. Its function is to perform linear transformations and feature extraction on the input data. The number of neurons and layers in the hidden layer depends on the complexity of the problem and the characteristics of the data. Increasing the number of hidden layers and neurons can improve the network's fitting ability, but it also increases the computational load and the risk of overfitting. In the present invention, the hidden layer consists of 10 neurons, and the neurons in each layer are connected using a fully connected method.

[0136] In addition to the number of hidden layers and the connection method, the choice of activation function and optimizer in the design of the hidden layer also determines the performance of the neural network. In this invention, the activation function is ReLU (Rectified Linear Unit, ReLU). ReLU is an activation function widely used in neural networks. Its main function is to introduce nonlinear factors in the hidden layer of the neural network. The main advantages of the ReLU function are its simplicity and computational efficiency. The loss function is defined to measure the difference between the predicted output of the model and the true label. The loss function is defined as follows:

[0137]

[0138] Where n represents the sample size of the dataset used for training, represents the model's predicted value, i.e., the true value of the network score in the dataset, y represents the true value of the label class, and MSE represents the loss function defined as the mean squared error between the true value and the predicted value. The optimizer used in this paper is an algorithm used to adjust model parameters to minimize the loss function. The Adam optimizer is used in this paper.

[0139] Step 203: output layer design.

[0140] The output layer is the final layer of the neural network and is responsible for generating the network's output. Since the number of neurons in the output layer is typically determined by the task type and the output dimension, in this paper, the number of neurons in the output layer is 1, and its output is the network's predicted score.

[0141] Figure 3 Flowchart for training neural network models, including:

[0142] Step 301: input training sample data.

[0143] The data of the training set is passed into the model. The data here are the real collected values. The input data in this invention includes five indicators: data transmission rate, information update rate, communication bit error rate, communication packet loss rate, and point-to-point link delay.

[0144] Step 302: Sample data preprocessing

[0145] Sample data preprocessing is divided into data cleaning and data standardization. In the present invention, since some indicators of the cluster communication network vary widely and the probability of extreme values ​​is small, the 3σ principle is used for data cleaning. The 3σ principle is an important statistical concept and is often used to identify and process outliers in data sets. This principle is based on the characteristics of the normal distribution and points out that in a normally distributed data set, approximately 99.73% of the data values ​​will fall within the range of the mean plus or minus three standard deviations. Specifically, in the present invention, the mean μ of each column in the data set is calculated one by one, and the std() function of the Python pandas library is called to calculate the data standard deviation σ. Each indicator is traversed. If the numerical point is more than three standard deviations away from the mean of the column in which it is located, that is, it is outside the range of (μ-3σ, μ+3σ), then the data point is regarded as an outlier and the data is deleted from the data set.

[0146] Data normalization is the process of normalizing the five indicators in the training data set. In this step, the data transmission rate R data , information update rate R update , communication bit error rate BER, communication packet loss rate R loss The point-to-point link delay RT is normalized to the range of (0,1), wherein the data transmission rate and information update rate are normalized using formula (6), and the communication bit error rate, communication packet loss rate, and point-to-point link delay are normalized using formula (7). In the present invention, the theoretical score value needs to be calculated as a label, and the formula given in (8) is used for calculation, and G is converted to n Processed to the range (0,100).

[0147] Step 303: model training.

[0148] After the model receives the input data, it gets the predicted value The label is y, and the loss function is calculated:

[0149]

[0150] During the model training process, the model training is completed when the loss function reaches the minimum value. The predicted value obtained by the model It is determined by each neuron in the hidden layer and can be expressed as:

[0151]

[0152] Among them, ω i ,b i are the weight coefficient and bias coefficient of the i-th neuron in the hidden layer, respectively. The present invention uses only one hidden layer and uses formula (14) to calculate the gradient value of the loss function with respect to each parameter:

[0153]

[0154] Where x is any weight coefficient or bias coefficient. For each parameter x, use formula (15) to update the parameter:

[0155]

[0156] Where x(n) represents the value of parameter x at the nth iteration, and α is the learning rate.

[0157] The model is trained until the loss function stops decreasing after 10 consecutive iterations, or until the specified number of training rounds is reached.

[0158] Step 304: Model verification and use.

[0159] Use the test set to evaluate the model performance during training. Input the i-th data into the network to get the predicted value Calculate its difference from the true value y i The absolute error e i :

[0160]

[0161] Calculate the mean absolute error of the test data:

[0162]

[0163] When the mean absolute error reaches the allowable range, the generalization ability of the model is verified. The model can not only make good predictions for the data in the training set, but also make good predictions for new data outside the training set.

[0164] During the model's use, the original values ​​of five unmanned cluster communication network metrics—data transmission rate, information update rate, communication bit error rate, packet loss rate, and point-to-point link latency—are input into the model, which then returns a performance evaluation score for the current communication network. By continuously acquiring metrics from the unmanned cluster communication network, the model can quickly return scores, enabling real-time performance evaluation of the communication network.

Claims

1. A method for evaluating the performance of an unmanned cluster communication network based on a neural network, characterized in that: The specific steps include: Step 101, selecting evaluation indicators; The evaluation indicators constructed include: data transmission rate, information update rate, packet loss rate, communication bit error rate and point-to-point link delay. The explanation and calculation method of the specific indicators are as follows; (1) Data transmission rate Data transmission rate refers to the amount of data transmitted per unit time. The data transmission rate R data The calculation formula is as follows: Among them, B total Indicates the total number of bytes of data transmitted during the transmission time, T total Represents data transmission time; (2) Information update rate The information update rate is the number of packets updated per second, expressed as R update The calculation formula is as follows: Among them, N pack Indicates the number of received data packets, T update Indicates update time; (3) Communication bit error rate The communication bit error rate is expressed as BER; Among them, Byte wrong Indicates the number of erroneous bytes in the data packet received during the test time, Byte trans Indicates the number of bytes in the data packet sent during the test time; (4) Packet loss rate The packet loss rate is expressed as R loss express; Among them, N trans Indicates the number of packets sent, N rec Indicates the number of packets received; (5) Point-to-point link delay The transmission time of information from the source to the receiving point is represented by RT; RT=T rec -T trans (5) Among them, T rec Indicates the receiving time of the receiver, T trans Indicates the sending time of the sender; Step 102, determining the weight of each evaluation indicator; Given R data , R update ,BER,R loss The weights ω corresponding to the five indicators of RT are 0.3, 0.1, 0.2, 0.2 and 0.2 respectively, and the sum of the weights is 1; Step 103, standardizing each index value; The five indicators of data transmission rate, information update rate, communication bit error rate, packet loss rate and point-to-point link delay are normalized; the positive indicator is standardized using the following formula, with X r Represents the normalized data: Among them, X represents the actual value of the indicator, X max Indicates the maximum value of the data, X min Indicates the minimum value of data; Negative indicators are normalized using the following formula, with X l express: Step 104, constructing an output network score; The network score is calculated by formula (8). Assume there are n indicators in total, and the weighted sum is X sum express; Among them, X r / l represents the standardized index, represents the i-th standardized index, ω i represents the weight of the ith indicator; for the existing data set, use formula (8) to construct its theoretical score, that is, the label; assume that the maximum value of the data label in the form of weighted sum is Sum Maxweighted , the score of the data set is obtained by formula (9), and G n express: Step 105, sample division; In the subsequent model training, part of the data in the sample set S is used as the training set, and the rest of the data is used as the test set. In the network training and learning phase, the network weight parameters are initialized, and the training set data is input into the network for training until the target requirements are met or all samples are trained. By collecting various indicators of the drone group communication network, the data is standardized using the methods in steps 103 and 104, and labels are obtained as their theoretical scores to form a sample set S; the i-th data in the sample set S is in the form of; in, They respectively represent the data transmission rate, information update rate, communication bit error rate, packet loss rate, point-to-point link delay and label value of the i-th sample data in the sample set.

2. The method for evaluating the performance of an unmanned cluster communication network based on a neural network as claimed in claim 1, characterized in that: In step 103, the positive indicators include R data , R update ,Negative indicators include packet loss rate, communication bit error rate and point-to-point link delay.

3. The unmanned cluster communication network performance evaluation method based on neural network as claimed in claim 1, characterized in that: In step 104, the weight coefficient used results in a label value between 0 and 1. In actual applications, if the weight is changed, or manual scoring is used as data labels, the label value may be greater than 1, and the theoretical score needs to be scaled to ensure that the score is within 100 points.

4. The method for evaluating the performance of an unmanned cluster communication network based on a neural network as claimed in claim 1, characterized in that: In step 105, 70% of the data in S is used as a training set, and the remaining 30% of the data is used as a test set.

5. The method for evaluating the performance of an unmanned cluster communication network based on a neural network as claimed in claim 1, characterized in that: The neural network structure design specifically includes the following steps: Step 201, input layer design; The number of neurons in the input layer is 5, corresponding to the five indicators of the unmanned cluster communication network: data transmission rate, information update rate, communication bit error rate, communication packet loss rate, and point-to-point link delay; Step 202, hidden layer design; The hidden layer is one or more layers located between the input layer and the output layer. The hidden layer consists of 10 neurons, and the neurons in each layer are connected in a fully connected manner. In the design of the hidden layer, the choice of activation function and optimizer also determines the performance of the neural network; ReLU is selected as the activation function; the loss function is defined to measure the difference between the predicted output of the model and the true label. The loss function is defined as follows: Where n represents the sample size of the dataset used for training. represents the predicted value of the model, that is, the true value of the network score in the data set, y represents the true value of the label class, and MSE represents the loss function defined as the mean square error between the true value and the predicted value; the Adam optimizer is used to adjust the model parameters to minimize the loss function; Step 203, output layer design; The output layer is responsible for generating the output results of the network; the number of neurons in the output layer is 1, and its output result is the prediction score of the network.

6. The method for evaluating the performance of an unmanned cluster communication network based on a neural network as claimed in claim 1, characterized in that: Neural network training and use specifically include the following steps: Step 301, input training sample data; The real data of the training set is passed into the model. The input data includes five indicators: data transmission rate, information update rate, communication bit error rate, communication packet loss rate, and point-to-point link delay; Step 302: Sample data preprocessing Sample data preprocessing is divided into data cleaning and data standardization; The 3σ principle is used for data cleaning. The mean μ of each column in the data set is calculated one by one, and the standard deviation σ of the data is calculated. Each indicator is traversed. If the value point is more than three standard deviations away from the mean of the column, that is, it is outside the range of (μ-3σ, μ+3σ), then the data point is regarded as an outlier and the data is deleted from the data set. Data standardization sets the data transmission rate R data , information update rate R update , communication bit error rate BER, communication packet loss rate R loss The point-to-point link delay RT is standardized to the range of (0,1), where the data transmission rate and information update rate are standardized using formula (6), and the communication bit error rate, communication packet loss rate, and point-to-point link delay are standardized using formula (7); the theoretical score value is calculated using formula (8) as the label, and G is converted into n Processed to the range of (0,100); Step 303, model training; After the model gets the input data, it gets the predicted value The label is y and the loss function is calculated: During the model training process, when the loss function reaches the minimum value, the model training is completed; the predicted value obtained by the model It is jointly determined by each neuron in the hidden layer and is expressed as: where ω i ,b i are the weight coefficient and bias coefficient of the i-th neuron in the hidden layer, respectively. Only one hidden layer is used, and the gradient value of the loss function with respect to each parameter is calculated using formula (14): Where x is any weight coefficient or bias coefficient; for each parameter x, the parameter is updated using formula (15): Among them, x(n) represents the value of parameter x at the nth iteration, and α is the learning rate; The model training stops when the loss function does not decrease for 10 consecutive iterations, or when the specified number of training rounds is reached; Step 304, model verification and use; Use the test set to evaluate the model performance during training; input the i-th data into the network to get the predicted value Calculate its difference from the true value y i The absolute error e i : Calculate the mean absolute error of the test set data: When the mean absolute error reaches the allowable range, the generalization ability of the model is verified; The five indicators of the unmanned cluster communication network: data transmission rate, information update rate, communication bit error rate, packet loss rate, and point-to-point link delay, are input into the model with their original values. The model can then return the performance evaluation score of the current communication network. By continuously acquiring the indicators of the unmanned cluster communication network, the model can quickly return the score, thereby realizing real-time performance evaluation of the communication network.

Images