Unmanned vehicle neural network control method under online model identification

Abstract

The invention relates to the technical field of water surface / underwater unmanned vehicle control, in particular to an unmanned vehicle neural network control method under online model identification. Comprising the following steps of: calculating and establishing a motion model of a water surface / underwater unmanned vehicle through a hydrodynamic theory, or establishing the motion model by adopting a neural network algorithm after real navigation data is acquired and processed in a still water state; a simulation system is established through the model, parameter simulation is controlled based on a model prediction control algorithm, and the horizontal direction is controlled; in the actual navigation process, the simulation control parameters are initially used for control, and online model identification is carried out on dynamic input data based on neural network algorithm model identification so as to correct model parameters; a modeling error is obtained and a feedback compensation control quantity is generated based on a neural network identifier compensation identifier, control parameters are optimized and then substituted into a neural network algorithm, so that the control effect is better, the control error is gradually reduced through iterative learning, and the optimal state is achieved.

Description

Technical Field

[0001] The invention relates to the technical field of surface / underwater unmanned aerial vehicle control, and in particular to an unmanned aerial vehicle neural network control method under online model identification. Background Art

[0002] With the increasing attention paid to the research of unmanned systems at home and abroad, surface / underwater unmanned vehicles, as an important part of intelligent marine equipment at home and abroad, have always been a hot spot in the research of unmanned equipment. Surface / underwater unmanned vehicles refer to intelligent unmanned autonomous platforms that navigate on the surface / underwater and rely on autonomy or remote control. In unmanned mission scenarios, autonomous navigation control capabilities are the basis for all autonomous behaviors of surface / underwater unmanned vehicles, and are also the premise for surface / underwater unmanned vehicles to achieve more complex intelligent behaviors and ensure safe navigation.

[0003] Due to the different tasks and complex working environment of surface / underwater unmanned vehicles, they are inevitably affected by complex external environmental disturbances such as wind / waves / currents during actual navigation, making it difficult to establish an accurate mathematical model of motion and the optimal control parameters. Especially in actual engineering, the mathematical model of motion obtained through hydrodynamic analysis and calculation still deviates from the actual situation. The control parameters can only be adjusted through multiple data analyses, and the control strategy and parameters need to be modified in real time according to different mission environments and speeds for each voyage. In addition, if the parameters are not adjusted to the optimal values ​​at the corresponding speed, the system will always have problems such as deviation or slow convergence speed and large overshoot, which makes the control system poor in adaptability, reduces the accuracy of autonomous navigation, reduces the degree of mission completion, and may even cause unpredictable losses. Summary of the invention

[0004] In order to overcome the deficiencies of the above-mentioned prior art and improve the theoretical application rate in practical applications, the present invention provides an autonomous navigation control method for surface / underwater unmanned vehicles to solve the problems of poor control system dynamic performance, poor control accuracy, poor adaptability, etc. in the autonomous navigation control of traditional surface / underwater unmanned vehicles in complex and unknown ocean environments.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a neural network control method for an unmanned aerial vehicle under online model identification, comprising the following steps:

[0006] 1) Establishing the motion model of surface / underwater unmanned vehicles through hydrodynamic theory calculations; or establishing the motion model of surface / underwater unmanned vehicles through the collection and processing of actual navigation data in static water state using a neural network algorithm;

[0007] 2) Establish a simulation system through the established motion model of the surface / underwater unmanned vehicle, simulate the control parameters based on the model predictive control algorithm, and use it as the initial navigation control state of the surface / underwater unmanned vehicle to control the horizontal direction;

[0008] 3) During the actual navigation process, the surface / underwater unmanned vehicle is initially controlled using simulation control parameters, and then the dynamic input data is identified online through the neural network algorithm model identification to correct the model parameters and dynamically improve the model predictive control effect;

[0009] 4) Based on the neural network identifier, the compensation identifier obtains the modeling error and generates feedback compensation control quantity, optimizes the control parameters online, and substitutes them into the neural network algorithm; through iterative learning, the control error is gradually reduced and reaches the optimal state.

[0010] The motion model of the surface / underwater unmanned vehicle is established by hydrodynamic theory calculation, including the following contents:

[0011] Based on the traditional ship mechanism model, the CFD numerical simulation model test is used to realize the fitting of the maneuverability coefficient of the surface / underwater unmanned vehicle, that is, the motion mathematical model of the unmanned boat is established through the mechanism modeling method based on numerical simulation.

[0012] After the actual navigation data in still water is collected and processed, a motion model is established using a neural network algorithm, including the following steps:

[0013] 1-1) In the direct flight state, the fixed rotation speed of the surface / underwater unmanned vehicle is periodically increased under manual or wireless remote control to obtain navigation data related to the speed;

[0014] The fixed speed is specifically:

[0015] For the surface unmanned boat platform, the fixed speed is set to 500 rpm or 1000 rpm;

[0016] For underwater unmanned vehicles, the fixed speed is set to 200 rpm;

[0017] 1-2) Under various typical speeds, control the surface / underwater unmanned vehicle in a typical Z-shaped or 8-shaped route under manual or wireless remote control;

[0018] 1-3) The collected navigation data of each speed and each typical route are combined with the motion capability of the unmanned boat to remove outliers and normalize them, and the obtained navigation data of each speed and each typical route are used as training samples of the neural network identification model;

[0019] 1-4) Use the neural network model for training and learning, and then obtain the mathematical model of the surface / underwater unmanned vehicle motion under the data sample.

[0020] In step 1-4), the neural network model is used for training and learning, specifically:

[0021] The training samples of the neural network recognition model obtained in step 1-3) are used as input and substituted into the neural network model for training and learning. The weight matrix from the input layer to the hidden layer and from the hidden layer to the output layer is continuously optimized to the optimal value to complete the training and learning of the neural network model.

[0022] The step 2) includes the following contents:

[0023] 2-1) Based on the mathematical model of surface / underwater unmanned vehicle motion obtained by hydrodynamic theory calculation or neural network identification, a simulation system for surface / underwater unmanned vehicle is established;

[0024] 2-2) In the simulation system, a control algorithm based on model prediction is adopted. After continuous rolling optimization and feedback correction, the control parameters are obtained to serve as the initial navigation control state of the surface / underwater unmanned vehicle to control the speed and heading in the horizontal direction.

[0025] The rolling optimization and feedback correction are specifically as follows:

[0026] The rolling optimization is: through local rolling optimization in a limited time domain, the object of optimization is the control parameters of the model prediction method, and the basis of optimization is to obtain the optimal performance index at each moment. The index is set as follows:

[0027]

[0028] Among them, q i and λ i are coefficients, is the softening coefficient and satisfies y(k+i) is the measured system output value at time k+i, y r (k+i) is the expected output value of the system at time k+i, Δu(k+i-1) is the control increment at time k+i-1, N is the maximum prediction length, and M is the control length;

[0029] The feedback correction is: introducing a closed-loop feedback mechanism in the predictive control, taking the detected current state or output information as feedback at each sampling moment, correcting the predicted value at the next moment to make it close to the actual value, and solving the optimal control parameters.

[0030] The step 3) is specifically:

[0031] 3-1) During the actual navigation process, the model predictive control parameters obtained in the simulation system are initially used to control the horizontal motion;

[0032] 3-2) Based on the neural network algorithm model identification, the dynamic input including the navigation data of the surface / underwater unmanned vehicle’s navigation state, heading, speed, rudder angle and interference is subjected to online model identification, and the model parameters are dynamically corrected to improve the model prediction control effect.

[0033] The online model identification is specifically as follows:

[0034] The dynamic input data is used as the input of the neural network model for real-time training and learning. The weight matrix from the input layer to the hidden layer and from the hidden layer to the output layer is continuously optimized until the unmanned aerial vehicle is optimal in this state. The real-time optimal weight matrix is ​​updated to the neural network identification model to complete the online model identification.

[0035] The specific contents of step 4) are as follows:

[0036] 4-1) The neural network identifier obtains a special inverse model with the desired characteristics through the forward model of the back-propagation channel;

[0037] 4-2) The difference between the expected output and the output of the neural network forward model is used to train the inverse model network, and the modeling error of the identifier is compensated by the network identifier to generate a feedback compensation control amount;

[0038] 4-3) Online optimization of the neural network algorithm control parameters is performed and substituted into the neural network algorithm, so that the neural network control effect is gradually improved and the effect of the model predictive control method is reduced;

[0039] 4-4) Through continuous iterative learning, the control error gradually decreases to reach the optimal state, making it have good robustness.

[0040] The present invention has the following beneficial effects and advantages:

[0041] 1. The present invention has strong universality and excellent stability. The autonomous navigation control method of the present invention enables surface / underwater unmanned aerial vehicles to adapt to complex and unknown marine environments, acquire actual navigation data in real time, and adaptively control to adapt to real-time environmental changes through learning, iteration, and optimization. The method has clear logic and improved autonomy, and it also incorporates offline models, preset initial control parameters, model predictive control under online model identification, and dynamic optimization of neural network control parameters. In the initial state and during navigation, it has multiple guarantees and good stability.

[0042] 2. The present invention has good real-time performance and strong robustness. The autonomous navigation control method of the present invention can correct the motion mathematical model and optimize the control parameters in real time during the actual navigation process of the surface / underwater unmanned vehicle, without autonomous dynamic adjustment, and has good real-time performance. The control strategy combines multiple methods such as model predictive control, online model identification, and dynamic parameter adjustment of the neural network identifier, and has strong adaptability and robustness to unknown complex marine environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the model prediction control process of the surface / underwater unmanned vehicle of the present invention based on online neural network model identification;

[0044] Figure 2 It is a schematic diagram of the architecture of the neural network control parameter self-adjustment method of the surface / underwater unmanned vehicle based on online model identification of the present invention. DETAILED DESCRIPTION

[0045] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0046] like Figure 2 The figure shows the architecture of the neural network control parameter self-adjustment method of the surface / underwater unmanned vehicle based on online model identification according to the present invention. Figure 1 A neural network control parameter self-adjustment method for surface / underwater unmanned vehicle under online model identification comprises the following steps:

[0047] Step 1: Establish a motion model of the surface / underwater unmanned vehicle through hydrodynamic theory calculations, or establish a motion model using a neural network algorithm after collecting and processing actual navigation data in static water conditions;

[0048] (1-1) The motion model of surface / underwater unmanned vehicle is established through hydrodynamic theory calculation. Based on the six-degree-of-freedom model of underwater unmanned vehicle and the three / four-degree-of-freedom model of surface unmanned vehicle, the relevant parameters in the model are calculated through CFD simulation, and then the motion model of surface / underwater unmanned vehicle is established based on the hydrodynamic theory calculation.

[0049] (1-2) After the actual navigation data in the still water state is collected and processed, the fixed rotation speed is periodically increased by manual or wireless remote control in the still water state, and the navigation data is recorded. At each typical speed, the surface / underwater unmanned vehicle is controlled by manual or wireless remote control to perform Z-shaped and 8-shaped typical routes. The collected data include: actual navigation heading, speed of the main engine, rudder angle, etc. After the test, the data is normalized and outliers are removed to obtain the training samples of the neural network identification model.

[0050] (1-3) Use neural network algorithm to establish motion model, refer to Figure 1 The actual navigation heading, speed, rotation speed and rudder angle are used as the input parameters of the neural network algorithm, and the heading angular velocity, speed change rate, etc. are output. After learning and training, the motion model of the surface / underwater unmanned vehicle in still water is obtained.

[0051] Step 2: Establish a simulation system through the model, and simulate the control parameters based on the model predictive control algorithm as the initial navigation control state of the surface / underwater unmanned vehicle to control the horizontal direction;

[0052] (2-1) Use this model to build a simulation system, refer to Figure 1 Based on the mathematical model of surface / underwater unmanned vehicle motion obtained by hydrodynamic theory calculation or neural network identification, a simulation system (i.e. project) is programmed to simulate and verify the surface / underwater unmanned vehicle motion calculation, navigation control and other functions.

[0053] (2-2) Based on the control parameter simulation of the model prediction control algorithm, the surface / underwater unmanned vehicle is used as the initial navigation control state to control the horizontal direction. In the simulation system, the control algorithm based on model prediction is adopted. After continuous rolling optimization and feedback correction, the control parameters are obtained to serve as the initial navigation control state of the surface / underwater unmanned vehicle to control the speed and heading in the horizontal direction.

[0054] (2-2-1) Rolling optimization, feedback correction, predictive control is characterized by real-time online rolling optimization. Optimization refers to a local rolling optimization within a limited time domain. At each sampling moment, the optimization performance index only involves the time from that moment to a certain time domain in the future, and when the next sampling moment comes, the optimized time period will roll forward as the time continues to advance. The optimization cannot be completed in one offline operation, but requires repeated and uninterrupted online operations.

[0055] Specifically, rolling optimization is: through local rolling optimization in a limited time domain, the object of optimization is the control parameters of the model prediction method. The basis of optimization is to obtain the optimal performance index at each moment. The index is set as follows:

[0056]

[0057] Among them, q i and λ i are coefficients, is the softening coefficient and satisfies y(k+i) is the measured system output value at time k+i, y r (k+i) is the expected output value of the system at time k+i, Δu(k+i-1) is the control increment at time k+i-1, N is the maximum prediction length, and M is the control length;

[0058] (2-2-2) Introducing a closed-loop feedback mechanism into predictive control can compensate for the impact of various uncertainties on the system to a certain extent. That is, at each sampling moment, the current state or output information is detected, and this information is used as feedback to correct the predicted value at the next moment to make it closer to the actual value, and finally solve the optimal control. This process is called feedback correction.

[0059] Step 3: During the actual navigation process, the simulation control parameters are initially used for control, and then the dynamic input data is subjected to online model identification based on the neural network algorithm model identification to correct the model parameters and dynamically improve the control effect of the model predictive control;

[0060] Among them, input data, input surface / underwater unmanned vehicle navigation state, heading, speed, rudder angle and interference.

[0061] The specific implementation process includes the following steps:

[0062] (3-1) During the actual navigation process, the model predictive control parameters obtained in the simulation system are initially used to control the horizontal plane motion;

[0063] (3-2) Based on the neural network algorithm model identification, the dynamic input including the navigation data of the surface / underwater unmanned vehicle’s navigation state, heading, speed, rudder angle and interference is subjected to online model identification, and the model parameters are dynamically corrected to improve the model prediction control effect.

[0064] Online model identification is as follows: dynamic input data is used as the input of the neural network model, and real-time training and learning are performed to continuously optimize the weight matrix from the input layer to the hidden layer and from the hidden layer to the output layer until it is optimal in the state of the unmanned aerial vehicle, and the real-time optimal weight matrix is ​​updated to the neural network identification model to complete the online model identification.

[0065] Step 4: Based on the neural network identifier, the modeling error of the identifier is compensated and feedback compensation control is generated. The control parameters are optimized online and substituted into the neural network algorithm to gradually improve the control effect and reduce the effect of the model predictive control method. Through iterative learning, the control error is gradually reduced and reaches the optimal state.

[0066] Based on the neural network identifier, the compensation identifier obtains the modeling error and generates feedback compensation control quantity. The control parameters are optimized online and substituted into the neural network algorithm, so that the control effect is gradually improved and the effect of the model predictive control method is reduced. Through iterative learning, the control error is gradually reduced and reaches the optimal state.

[0067] The neural network identifier obtains a special inverse model with desired characteristics through the back propagation channel - the forward model. The difference between the desired output and the output of the neural network forward model is used to train the inverse model network. The network identifier compensates the identifier modeling error and generates feedback compensation control quantity. The control parameters of the neural network algorithm are optimized online and substituted into the neural network algorithm, so that the neural network control effect is gradually improved and the effect of the model predictive control method is reduced. Through continuous iterative learning, the control error is gradually reduced to the optimal state, making it have good robustness.

[0068] The navigation control method of the present invention can make the surface / underwater unmanned vehicle adapt to the complex and unknown ocean environment, obtain the actual navigation data in real time, and adaptively control to adapt to the real-time environmental changes through learning, iteration and optimization. The method has the advantages of clear logic, improved autonomy, strong universality, excellent stability, good robustness, online adaptive adjustment, self-correction control output, and improved autonomous control performance.

[0069] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A neural network control method for an unmanned aerial vehicle under online model identification, characterized in that: The following steps are involved: 1) Establishing the motion model of surface / underwater unmanned aerial vehicles through hydrodynamic theory calculations; or establishing the motion model of surface / underwater unmanned aerial vehicles through the collection and processing of actual navigation data in static water state using a neural network algorithm; 2) Establish a simulation system through the established motion model of the surface / underwater unmanned vehicle, simulate the control parameters based on the model predictive control algorithm, and use it as the initial navigation control state of the surface / underwater unmanned vehicle to control the horizontal direction; 3) During the actual navigation process, the surface / underwater unmanned vehicle is initially controlled using simulation control parameters, and then the dynamic input data is identified online through the neural network algorithm model identification to correct the model parameters and dynamically improve the model predictive control effect; 4) Based on the neural network identifier, the compensation identifier obtains the modeling error and generates feedback compensation control quantity, optimizes the control parameters online, and substitutes them into the neural network algorithm; through iterative learning, the control error is gradually reduced and reaches the optimal state.

2. The unmanned aerial vehicle neural network control method under online model identification according to claim 1 is characterized in that: The motion model of the surface / underwater unmanned vehicle is established by hydrodynamic theory calculation, including the following contents: Based on the traditional ship mechanism model, the CFD numerical simulation model test is used to realize the fitting of the maneuverability coefficient of the surface / underwater unmanned vehicle, that is, the motion mathematical model of the unmanned boat is established through the mechanism modeling method based on numerical simulation.

3. The unmanned aerial vehicle neural network control method under online model identification according to claim 1 is characterized in that: After the actual navigation data in still water is collected and processed, a motion model is established using a neural network algorithm, including the following steps: 1-1) In the direct flight state, the fixed rotation speed of the surface / underwater unmanned vehicle is periodically increased under manual or wireless remote control to obtain navigation data related to the speed; The fixed speed is specifically: For the surface unmanned boat platform, the fixed speed is set to 500 rpm or 1000 rpm; For underwater unmanned vehicles, the fixed speed is set to 200 rpm; 1-2) Under various typical speeds, control the surface / underwater unmanned vehicle in a typical Z-shaped or 8-shaped route under manual or wireless remote control; 1-3) The collected navigation data of each speed and each typical route are combined with the motion capability of the unmanned boat to remove outliers and normalize them, and the obtained navigation data of each speed and each typical route are used as training samples of the neural network identification model; 1-4) Use the neural network model for training and learning, and then obtain the mathematical model of the surface / underwater unmanned vehicle motion under the data sample.

4. The unmanned aerial vehicle neural network control method under online model identification according to claim 1 is characterized in that: In step 1-4), the neural network model is used for training and learning, specifically: The training samples of the neural network recognition model obtained in step 1-3) are used as input and substituted into the neural network model for training and learning. The weight matrix from the input layer to the hidden layer and from the hidden layer to the output layer is continuously optimized to the optimal value to complete the training and learning of the neural network model.

5. The unmanned aerial vehicle neural network control method under online model identification according to claim 1 is characterized in that: The step 2) includes the following contents: 2-1) Based on the mathematical model of surface / underwater unmanned vehicle motion obtained by hydrodynamic theory calculation or neural network identification, a simulation system for surface / underwater unmanned vehicle is established; 2-2) In the simulation system, a control algorithm based on model prediction is adopted. After continuous rolling optimization and feedback correction, the control parameters are obtained to serve as the initial navigation control state of the surface / underwater unmanned vehicle to control the speed and heading in the horizontal direction.

6. The unmanned aerial vehicle neural network control method under online model identification according to claim 5 is characterized in that: The rolling optimization and feedback correction are specifically as follows: The rolling optimization is: through local rolling optimization in a limited time domain, the object of optimization is the control parameters of the model prediction method, and the basis of optimization is to obtain the optimal performance index at each moment. The index is set as follows: Among them, q i and λ i are coefficients, is the softening coefficient and satisfies y(k+i) is the measured system output value at time k+i, y r (k+i) is the expected output value of the system at time k+i, Δu(k+i-1) is the control increment at time k+i-1, N is the maximum prediction length, and M is the control length; The feedback correction is: introducing a closed-loop feedback mechanism in the predictive control, taking the detected current state or output information as feedback at each sampling moment, correcting the predicted value at the next moment to make it close to the actual value, and solving the optimal control parameters.

7. The unmanned aerial vehicle neural network control method under online model identification according to claim 1 is characterized in that: The step 3) is specifically: 3-1) During the actual navigation process, the model predictive control parameters obtained in the simulation system are initially used to control the horizontal plane motion; 3-2) Based on the neural network algorithm model identification, the dynamic input including the navigation data of the surface / underwater unmanned vehicle’s navigation state, heading, speed, rudder angle and interference is subjected to online model identification, and the model parameters are dynamically corrected to improve the model prediction control effect.

8. The unmanned aerial vehicle neural network control method under online model identification according to claim 1 is characterized in that: The online model identification is specifically as follows: The dynamic input data is used as the input of the neural network model for real-time training and learning. The weight matrix from the input layer to the hidden layer and from the hidden layer to the output layer is continuously optimized until the unmanned aerial vehicle is optimal in this state. The real-time optimal weight matrix is ​​updated to the neural network identification model to complete the online model identification.

9. The unmanned aerial vehicle neural network control method under online model identification according to claim 1 is characterized in that: The specific contents of step 4) are as follows: 4-1) The neural network identifier obtains a special inverse model with the desired characteristics through the forward model of the back-propagation channel; 4-2) The difference between the expected output and the output of the neural network forward model is used to train the inverse model network, and the modeling error of the identifier is compensated by the network identifier to generate a feedback compensation control amount; 4-3) Online optimization of the neural network algorithm control parameters is performed and substituted into the neural network algorithm, so that the neural network control effect is gradually improved and the effect of the model predictive control method is reduced; 4-4) Through continuous iterative learning, the control error gradually decreases to reach the optimal state, making it have good robustness.

Images