A UAV decision-making evaluation method in complex environments

Abstract

The invention discloses a decision-making evaluation method for unmanned aerial vehicles in a complex environment, which collects various flight data indicators of unmanned aerial vehicles; distinguishes between directly quantifiable indicators and indicators that cannot be directly quantified; forms standardized indicator data; builds LSTM neural networks and MF-LSTM neural networks; calculates the importance of various flight data indicators of the unmanned aerial vehicle to be evaluated; applies exponential importance impact to the flight data indicators of the unmanned aerial vehicle to be evaluated, and uses the trained LSTM neural network and MF-LSTM neural network to predict the development trend of the overall flight performance indicators of the unmanned aerial vehicle and various flight data indicators. The invention can effectively predict the future performance status of unmanned aerial vehicles in the current environment, and can be used to guide the deployment of unmanned aerial vehicle parameters.

Description

Technical Field

[0001] The present invention belongs to the fields of machine learning and neural networks, and particularly relates to a decision-making evaluation method for unmanned aerial vehicles. Background Art

[0002] Drones (UAVs) are unmanned aerial vehicles controlled by radio remote control and self-contained programmable devices. They play a vital role in both national defense and the national economy. Decision-making and evaluation of UAV flight performance involves evaluating and predicting various UAV indicators based on limited UAV flight data to assess the overall flight performance of the UAV in its current and future states.

[0003] Drone flight performance primarily encompasses speed, altitude, flight distance, takeoff, and landing. Drone flight control is a key issue in drone research. However, drones are subject to various interferences during flight, which can severely impact flight quality. This is due to the high uncertainty of the drone's environment. Furthermore, drones struggle to reliably predict their current state and adjust its behavior accordingly. Therefore, predicting drone flight performance can guide parameter deployment and improve flight performance.

[0004] Currently, quantitative research methods for evaluating drone flight status decisions include time series methods, neural network methods, and others. However, the data required to predict drone flight performance is complex, requiring more than simple single-variable, small-sample data, making traditional time series prediction methods inadequate.

[0005] Due to the limited performance of traditional algorithms and the uncertainty of specific UAV deployment environments, drone decision-making and evaluation in complex environments are difficult to achieve. Therefore, the deployment of UAVs in these complex environments presents difficulties. Furthermore, UAVs have a large number of performance indicators, some of which are not directly quantifiable. Quantifying, evaluating, and predicting these numerous indicators, and establishing coupled relationships between these data to evaluate the overall flight performance of UAVs, are of great research significance. Furthermore, neural networks can fit nonlinear systems, making it necessary to conduct research based on neural networks. Summary of the Invention

[0006] In order to solve the technical problems mentioned in the above background technology, the present invention proposes a UAV decision evaluation method in a complex environment, which can effectively predict the future performance of the UAV in the current environment.

[0007] In order to achieve the above technical objectives, the technical solution of the present invention is:

[0008] A method for evaluating drone decision-making in a complex environment includes the following steps:

[0009] (1) Collect various flight data indicators of the UAV during flight;

[0010] (2) Distinguish between indicators that can be directly quantified and those that cannot be directly quantified;

[0011] (3) For indicators that cannot be directly quantified, discrete coding and grey correlation methods are used to construct comprehensive indicators;

[0012] (4) Normalize and interpolate directly quantifiable indicators and comprehensive indicators to form standardized indicator data;

[0013] (5) Build a single-input-single-output LSTM neural network and train each standardized indicator data in batches;

[0014] (6) Build a multi-input-single-output MF-LSTM neural network to train the many-to-one mapping between various standardized indicator data and the overall flight performance indicators of the UAV;

[0015] (7) Use the TOPSIS algorithm to calculate the importance of each flight data indicator of the UAV to be evaluated;

[0016] (8) The flight data indicators of the UAV to be evaluated are subjected to an exponential importance shock. Based on the flight parameter indicators under the shock, the trained LSTM neural network and MF-LSTM neural network are used to predict the development trends of the UAV's overall flight performance indicators and various flight data indicators.

[0017] Furthermore, in step (2), the difference between the indicator that can be directly quantified and the indicator that cannot be directly quantified lies in whether the indicator can be intuitively expressed in time series using valid numbers.

[0018] Furthermore, the specific process of step (3) is as follows:

[0019] (301) Discrete coding:

[0020] For indicators that cannot be directly quantified, a discrete integer set is selected to encode them. The values ​​from small to large represent the development trend of the indicator from good to bad. Random numbers in the range of [0,1] are introduced for enhancement:

[0021]

[0022] In the above formula, MValue i is the encoded index, rand(0,1) represents a random number in the range [0,1], e is a natural constant, m i It is an indicator that cannot be directly quantified;

[0023] (302) Grey correlation:

[0024] Define the grey correlation reference sequence W0 = {w0(k)}, and the comparison sequence Z i ={z i (k)}, k represents the total amount of single data, subscript i represents a certain type of indicator, w0(k) is the reference sequence parameter, z i (k) is the comparison sequence parameter; the reference comparison error and the average error are calculated:

[0025] Δ i (k)=|W0-Z i |

[0026]

[0027] In the above formula, Δ i (k) is the reference comparison error, Δ v (k) is the mean error, m is the number of categories;

[0028] Calculate the resolution coefficient:

[0029]

[0030] Calculate the grey relational coefficient:

[0031]

[0032] In the above formula, when 2≤ε(k)≤3, ρ=2ε(k); when ε(k)<2, ρ=0.8; when ε(k)=0, ρ=0.5;

[0033] (303) Calculation of comprehensive index:

[0034]

[0035] In the above formula, mean(ξ i (k)) means to find all ξ i The average value of (k).

[0036] Furthermore, in step (4), the normalization method is as follows:

[0037]

[0038] In the above formula, NValue i is the normalized data, n i is the data before normalization, std means to find the variance, and mean means to find the mean.

[0039] Furthermore, in step (4), the interpolation method is a cubic spline interpolation method.

[0040] Furthermore, the optimization method of the LSTM neural network is the Adam algorithm.

[0041] Furthermore, the specific process of step (7) is as follows:

[0042] (701) Let the flight data index set X = {x1, x2, ..., x N} and the weight set of each indicator w={w1,w2,…,w N}, the following table N is the number of indicators, each indicator x i are all M-dimensional column vectors, then x ij Represents the index x i The elements of the jth row are homogenized to obtain:

[0043]

[0044] In the above formula, A is x ij The maximum value that can be obtained, a is x ij The minimum value that can be obtained;

[0045] (702) x ij 'Standardize:

[0046]

[0047] Determine the best and worst solutions:

[0048]

[0049]

[0050] In the above formula, For the optimal solution, is the worst solution;

[0051] (703) Calculate the distance between the evaluation object and the optimal solution and the worst solution:

[0052]

[0053] In the above formula, is the distance between the evaluation object and the optimal solution, is the distance between the evaluation object and the worst solution; calculate the importance coefficient:

[0054]

[0055] In the above formula, C i is the importance coefficient.

[0056] Furthermore, in step (8), the method of applying an exponential importance impact to the flight data indicators of the drone to be evaluated is as follows:

[0057]

[0058] In the above formula, b i ′ is the flight data index after impact, b i is the flight data index before impact, e is a natural constant, and mean represents the average value.

[0059] The beneficial effects brought about by adopting the above technical solution are:

[0060] 1. This invention overcomes the problems of low prediction accuracy, poor reliability, susceptibility to overfitting / underfitting, and poor data robustness of traditional machine learning algorithms; it has strong model portability and is suitable for predicting data of various types of time series changes;

[0061] 2. This invention can quantitatively analyze the importance of UAV flight indicators, providing a reliable theoretical basis for the deployment of various parameters during UAV flight. For economic indicators that cannot be directly quantified, this invention adopts a clever quantification method and explicitly expresses them, allowing the use of data that cannot be expressed numerically on UAVs.

[0062] 3. The present invention can clearly display the development trends of the overall flight performance and various flight index data of the UAV before and after deployment, and show the expected development comparison before and after deployment; establish the relationship between the overall flight performance index and various flight indicators, and use the multi-input-single-output method to improve the prediction accuracy of the overall flight performance index. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a flow chart of the method of the present invention;

[0064] Figure 2 To construct a comprehensive indicator diagram for indicators that cannot be directly quantified;

[0065] Figure 3 It is a characteristic diagram of LSTM neural network;

[0066] Figure 4 This is a schematic diagram of the chain form of repeated neural network modules of the LSTM neural network;

[0067] Figure 5 Repeating diagram of chain modules of LSTM neural network;

[0068] Figure 6 Schematic diagram of the characteristics of the MF-LSTM neural network. DETAILED DESCRIPTION

[0069] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0070] This paper designs a UAV decision-making evaluation method in complex environments, such as Figure 1 As shown, the steps are as follows:

[0071] Step 1: Collect various flight data indicators of the drone during flight;

[0072] Step 2: Distinguish between indicators that can be directly quantified and those that cannot be directly quantified;

[0073] Step 3: For indicators that cannot be directly quantified, discrete coding and grey relational methods are used to construct comprehensive indicators;

[0074] Step 4: Normalize and interpolate the directly quantifiable indicators and comprehensive indicators to form standardized indicator data;

[0075] Step 5: Build a single-input-single-output LSTM neural network and train each standardized indicator data in batches;

[0076] Step 6: Build a multi-input, single-output MF-LSTM neural network to train the many-to-one mapping between various standardized indicator data and the overall flight performance indicators of the UAV;

[0077] Step 7: Use the TOPSIS algorithm to calculate the importance of each flight data indicator of the drone to be evaluated;

[0078] Step 8: Apply an exponential importance shock to the flight data indicators of the UAV to be evaluated. Based on the flight parameter indicators under the shock, use the trained LSTM neural network and MF-LSTM neural network to predict the development trend of the UAV's overall flight performance indicators and various flight data indicators.

[0079] In this embodiment, specifically, in step 2, the difference between the indicator that can be directly quantified and the indicator that cannot be directly quantified lies in whether the indicator can be intuitively represented in time series using valid numbers.

[0080] In this embodiment, specifically, Figure 2 As shown, the process of step 3 is as follows:

[0081] 301. Discrete coding:

[0082] For indicators that cannot be directly quantified, a discrete integer set {0, 1, 2, 3} is selected to encode them. The values ​​from small to large represent the development trend of the indicator from good to bad. Random numbers in the range of [0, 1] are introduced for enhancement:

[0083]

[0084] In the above formula, MValue iis the encoded index, rand(0,1) represents a random number in the range [0,1], e is a natural constant, m i It is an indicator that cannot be directly quantified;

[0085] 302. Grey correlation:

[0086] Define the grey correlation reference sequence W0 = {w0(k)}, and the comparison sequence Z i ={z i (k)}, k represents the total amount of single data, subscript i represents a certain type of indicator, w0(k) is the reference sequence parameter, z i (k) is the comparison sequence parameter; the reference comparison error and the average error are calculated:

[0087] Δ i (k)=|W0-Z i |

[0088]

[0089] In the above formula, Δ i (k) is the reference comparison error, Δ v (k) is the mean error, m is the number of categories;

[0090] Calculate the resolution coefficient:

[0091]

[0092] Calculate the grey correlation coefficient:

[0093]

[0094] In the above formula, when 2≤ε(k)≤3, ρ=2ε(k); when ε(k)<2, ρ=0.8; when ε(k)=0, ρ=0.5;

[0095] 303. Calculate comprehensive indicators:

[0096]

[0097] In the above formula, mean(ξ i (k)) means to find all ξ i The average value of (k).

[0098] In this embodiment, specifically, the normalization method is as follows:

[0099]

[0100] In the above formula, NValue i is the normalized data, n iis the data before normalization, std means to find the variance, and mean means to find the mean.

[0101] The interpolation method is a cubic spline interpolation method.

[0102] In this embodiment, specifically, the structure of the LSTM neural network is as follows: Figure 3-5 As shown in the figure, I(t) is the input layer parameter, O(t) is the output layer parameter, and h t is the previous hidden layer output, x t is the hidden layer input, C t is the cell state, σ is the activation function, and tanh is the hyperbolic tangent function. For LSTM networks, a gradient descent method is used, which propagates from the input layer through the hidden layers to the output layer, and then from the output layer through the hidden layers back to the input layer, gradually correcting the connection weights. The optimization strategy is the stochastic gradient descent method (Adam). The Adam algorithm is a stochastic objective function optimization algorithm based on single-step, adaptive estimation of low-order moments. Compared to the classic stochastic gradient descent method, it can more efficiently update network parameters (weights, biases).

[0103] like Figure 6 As shown in Figure 1, the MF-LSTM (Multiple Features Long Short-Term Memory) network is an improved version of the LSTM neural network. It is suitable for learning multiple related features, each of which can share weights during the learning process. It can more efficiently establish multiple input data and the relationship between multiple inputs and a single output. The calculation and solution methods are similar to those of the LSTM neural network.

[0104] In this embodiment, specifically, the process of step 7 is as follows:

[0105] 701. Let the flight data index set X = {x1, x2, ..., x N} and the weight set of each indicator w={w1,w2,…,w N}, the following table N is the number of indicators, each indicator x i are all M-dimensional column vectors, then x ij Represents the index x i The elements of the jth row are homogenized to obtain:

[0106]

[0107] In the above formula, A is x ij The maximum value that can be obtained, a is x ij The minimum value that can be obtained;

[0108] 702, x ij 'Standardize:

[0109]

[0110] Determine the best and worst solutions:

[0111]

[0112]

[0113] In the above formula, For the optimal solution, is the worst solution;

[0114] 703. Calculate the distance between the evaluation object and the optimal and worst solutions:

[0115]

[0116] In the above formula, is the distance between the evaluation object and the optimal solution, is the distance between the evaluation object and the worst solution; calculate the importance coefficient:

[0117]

[0118] In the above formula, C i is the importance coefficient.

[0119] In this embodiment, specifically, the method of applying an exponential importance impact to the flight data indicators of the drone to be evaluated is as follows:

[0120]

[0121] In the above formula, b i ′ is the flight data index after impact, b i is the flight data index before impact, e is a natural constant, and mean represents the average value.

[0122] The embodiments are only for illustrating the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A UAV decision evaluation method in a complex environment, characterized in that: The following steps are involved: (1) Collect various flight data indicators of the UAV during flight; (2) Distinguish between indicators that can be directly quantified and those that cannot be directly quantified; (3) For indicators that cannot be directly quantified, discrete coding and grey correlation methods are used to construct comprehensive indicators; (4) Normalize and interpolate the indicators that can be directly quantified with the comprehensive indicators to form standardized indicator data; (5) Build a single-input-single-output LSTM neural network and train each standardized indicator data in batches; (6) Build a multi-input-single-output MF-LSTM neural network to train the many-to-one mapping between various standardized indicator data and the overall flight performance indicators of the UAV; (7) The TOPSIS algorithm is used to calculate the importance of each flight data indicator of the UAV to be evaluated; (8) The flight data indicators of the UAV to be evaluated are subjected to an exponential importance shock. Based on the flight parameter indicators under the shock, the trained LSTM neural network and MF-LSTM neural network are used to predict the development trend of the UAV's overall flight performance indicators and various flight data indicators; The specific process of step (3) is as follows: (301) Discrete coding: For indicators that cannot be directly quantified, a discrete integer set is selected to encode them. The values ​​from small to large represent the development trend of the indicators from good to bad. Random numbers in the range of [0,1] are introduced for enhancement: In the above formula, MValue i is the encoded index, rand(0,1) represents a random number in the range [0,1], e is a natural constant, and m i It is an indicator that cannot be directly quantified; (302) Grey correlation: Define the grey correlation reference sequence W0 = {w0(k)}, and the comparison sequence Z i ={z i (k)}, k represents the total amount of single data, subscript i represents a certain type of indicator, w0(k) is the reference sequence parameter, z i (k) is the comparison sequence parameter; calculate the reference comparison error and the average error: Δ i (k)=|W0-Z i | In the above formula, Δ i (k) is the reference comparison error, Δ v (k) is the average error, m is the number of categories; Calculate the resolution factor: Calculate the grey correlation coefficient: In the above formula, when 2≤ε(k)≤3, ρ=2ε(k); when ε(k)<2, ρ=0.8; when ε(k)=0, ρ=0.5; (303) Calculation of comprehensive index: In the above formula, mean(ξ i (k)) means to find all ξ i The average value of (k).

2. The method for evaluating the decision-making of unmanned aerial vehicles in a complex environment according to claim 1, characterized in that: In step (2), the difference between the indicator that can be directly quantified and the indicator that cannot be directly quantified lies in whether the indicator can be intuitively expressed in time series using valid numbers.

3. The method for evaluating the decision-making of unmanned aerial vehicles in a complex environment according to claim 1, characterized in that: In step (4), the normalization method is as follows: In the above formula, NValue i is the normalized data, n i is the data before normalization, std means to find the variance, and mean means to find the mean.

4. The method for evaluating the decision-making of unmanned aerial vehicles in a complex environment according to claim 1, characterized in that: In step (4), the interpolation method is a cubic spline interpolation method.

5. The method for evaluating the decision-making of unmanned aerial vehicles in a complex environment according to claim 1, characterized in that: The optimization method of the LSTM neural network is the Adam algorithm.

6. The method for evaluating the decision-making of unmanned aerial vehicles in a complex environment according to claim 1, characterized in that: The specific process of step (7) is as follows: (701) Let the flight data index set X = {x1, x2, ..., x N } and the weight set of each indicator w={w1,w2,…,w N }, in the following table, N is the number of indicators, and each indicator x i are all M-dimensional column vectors, then x ij Indicates the index x i The elements of the jth row of , and their attributes are homogenized to obtain: In the above formula, A is x ij The maximum value that can be obtained, a is x ij The minimum value that can be obtained; (702) x ij 'Standardize: Determine the best and worst solutions: In the above formula, For the best solution, is the worst solution; (703) Calculate the distance between the evaluation object and the optimal solution and the worst solution: In the above formula, is the distance between the evaluation object and the optimal solution, is the distance between the evaluation object and the worst solution; calculate the importance coefficient: In the above formula, C i is the importance coefficient.

7. The method for evaluating the decision-making of unmanned aerial vehicles in a complex environment according to claim 6, characterized in that: In step (8), the method of applying an exponential importance impact to the flight data indicators of the drone to be evaluated is as follows: In the above formula, b i ′ is the flight data index after impact, b i is the flight data index before impact, e is a natural constant, and mean represents the average value.

Images