CN106597847A - Recurrent-neural-network-based maneuver load controller and controlling method thereof - Google Patents

Abstract

The invention discloses a recurrent-neural-network-based maneuver load controller and a controlling method thereof. A neural network of the maneuver load controller is formed by series connection of a recurrent identification neural network and a recurrent control neural network; and each of the recurrent identification neural network and the recurrent control neural network uses six neurons for formation. In addition, the invention also discloses a controlling method of a recurrent-neural-network-based maneuver load controller. With the recurrent neural network method, the controller has advantages of simple structure, high working performance stability and robustness, and high adaptability. Effective self-adaptive maneuver load retarding of an airplane with variable-mach-number flight can be realized. A design of a recurrent-neural-network-based maneuver load controller of an airplane within a wide mach number range can be realized; and adaptive control of the maneuver load of the airplane on the complicated flight conditions can be realized.

Description

A kind of maneuver load controller and its control method based on recurrent neural network

Technical field

The invention belongs to flying vehicles control technical field, and in particular to a kind of maneuver load control based on recurrent neural network Device processed and its control method.

Background technology

Aircraft can produce larger moment of flexure in maneuvering at wing root, so as to affect the life-span of aircraft, bring peace Full hidden danger.Redistribute can the aerodynamic force of aircraft wing by using maneuver load controller, reach reduction wing root bending moment Purpose.

Traditional maneuver load controller, such as PID cannot realize becoming the Self Adaptive Control of Mach number, work as mission requirements change When, controller parameter needs to choose again, it is impossible to accomplish the control of real-time online, cannot play during aircraft practical flight Effectively effect.Therefore a kind of maneuver load controller is designed, and air maneuver Load alleviation is realized using the controller Method is always those skilled in the art's technical barrier to be solved.

The content of the invention

The present invention is directed to problems of the prior art, discloses a kind of maneuver load based on recurrent neural network Controller and its control method, carry out system identification and maneuver load slow down two stages by recurrent neural network, can Wing root bending moment slows down when realizing that aircraft becomes Mach number, while ensure the maneuver of aircraft can normally complete.

The present invention is achieved in that a kind of maneuver load controller based on recurrent neural network disclosed by the invention, The neutral net of described maneuver load controller recognizes neutral net and a recursion control neutral net string by a recurrence Connection is formed, and respectively constitutes network using 6 neurons.

Further, the basic mathematic model of described neutral net is：

In formula, v_kFor internal net activation, u_jRepresent input vector, w_kjThe weight matrix of synapse is represented,For activation primitive, y_kOutput vector is represented, p is to include being biased in interior neutral net input sum.

The invention also discloses a kind of control method of the maneuver load controller based on recurrent neural network, concrete steps It is as follows：

Step 1, using the normal g-load signal of aircraft, the angle of attack and wing root bending moment signal as input, horizontal tail and aileron Deflection is used as output；Normal g-load aircraft normal g-load and the angle of attack be able to can be passed through by corresponding sensor acquisition, wing root bending moment The mode that foil gauge is pasted at wing root gathers strain signal, and wing root bending moment is calculated in real time.Horizontal tail is deflected through work with aileron Dynamic device is realizing.

Step 2, the parameter and initial weight matrix of neutral net activation primitive are recognized to system identification by adjusting recurrence；

Recurrence identification network is used for the sensor signal of identification system.At the n moment, when recurrence identification network predicts n+1 The signal at quarterAs the input of controlling network；Recurrence identification neutral net is learned online by realtime recurrent algorithm Practise, recognize normal g-load, the angle of attack and the wing root bending moment signal of sensor acquisition.

Step 3, by adjust recursion control neutral net activation primitive parameter and initial weight matrix to normal g-load, The tracking of the angle of attack and the suppression to wing root bending moment.Recursion control neutral net is equally learned online using realtime recurrent algorithm Practise, normal g-load, the angle of attack and the wing root bending moment of neural network identification output are recognized as input signal with recurrence, realize to normal direction Overload, the real-time tracing of the angle of attack and the real-time control to wing root bending moment.

Further, the activation primitive that described recurrence identification neutral net and recursion control neutral net are adopted for：

The quiet activation v of neutral net_jWith input u_iBetween meet：

In formula：w_ijFor weight matrix；v_jFor the quiet activation of neutral net, u_jFor neutral net input vector.

Further, described step 2 is specially：

2.1, recurrence identification network carries out on-line study, calculation error gradient, recurrence identification god by realtime recurrent algorithm The error function of Jing networks is：

In formula, E^idFor object function, nm is sensor sum,It is the error of real sensor signal and identification signal.

2.2, recognize neutral net is recognized n-th in recurrence according to the error of real sensor signal and identification signal Individual time step, expression formula is：

Recurrence identification neutral net neuron is output as：

In formula,For the output of neuron, nid is neutral net input sum.

2.3, the design parameter of recurrence identification network, i.e. E are made by identification process^idMinimum is reached, its detailed process is：

In formula, η^idFor the learning rate of neutral net.

Further, described step 3 is specially：

3.1, recursion control neutral net is similar with recurrence identification neural network structure, recursion control neutral net neuron Output expression formula be：

Object function is：

Error expression between system reality output and desired output is：

The minimum of object function, need to be iterated to weight matrix：

Target function gradient is represented by：

In formula, nm is the output sum of recurrence identification network；Ni is the output sum of recursion control network, and γ is controlled to adjust The parameter that device processed is activated only；e^coFor the error between system reality output and desired output；η^coFor the study of recursion control network Rate.Recursion control neutral net still carries out on-line study using realtime recurrent algorithm.

The present invention is relative to the beneficial effect of prior art：Maneuver load control of the present invention based on recurrent neural network Utensil processed has adaptivity, and the maneuver load that can realize the aircraft for carrying out change Mach number slows down；By using recurrence god The method of Jing networks, with simple structure, stable work in work is good, robustness is high, self adaptation is good the characteristics of, can to become horse The aircraft of conspicuous number flight carries out efficient adaptive maneuver load to be slowed down；Present invention achieves aircraft is based in wide range of Mach numbers The design of the maneuver load controller of recurrent neural network, realize aircraft under complicated flying condition maneuver load it is adaptive Should control.

Description of the drawings

Fig. 1 is maneuver load controller architecture schematic diagram of the present invention based on recurrent neural network；

Fig. 2 is the schematic diagram that the present invention realizes air maneuver Load alleviation；

Fig. 3 is the system identification stage control deflecting facet schematic diagram of the present invention；

Fig. 4 is the maneuver load deceleration phase horizontal tail deflection schematic diagram of the present invention；

Fig. 5 is the system identification result schematic diagram of the present invention；

Fig. 6 is the Load alleviation result schematic diagram of the present invention；

Fig. 7 is the change Mach number Load alleviation schematic diagram of the present invention；

Fig. 8 is the closed loop control deflecting facet schematic diagram of the present invention.

Specific embodiment

The present invention provides a kind of maneuver load controller and its control method based on recurrent neural network, to make the present invention Purpose, technical scheme and effect it is clearer, clearly, and referring to the drawings and give an actual example to the present invention further specifically It is bright.It should be understood that described herein be embodied as, only to explain the present invention, being not intended to limit the present invention.

Fig. 1 is maneuver load controller architecture schematic diagram of the present invention based on recurrent neural network.One recurrence of controller Identification neutral net and a recursion control neutral net are formed by connecting, and respectively constitute network using 6 neurons, neutral net Basic mathematic model is：

Recurrence identification network is used for the sensor signal of identification system.At the n moment, when recurrence identification network predicts n+1 The signal at quarterAs the input of controlling network.

Recurrence identification network carries out on-line study by realtime recurrent algorithm.Realtime recurrent algorithm is that a kind of gradient declines Method, its main algorithm is the error gradient for calculating certain particular expression formula in each time step.Recurrence identification neutral net Error function is：

Real sensor signal recognizes n-th time step that neutral net is recognized with the error of identification signal in recurrence Expression formula be：

Recurrence identification neutral net neuron is output as：

The target of identification process is by each time successive step weight matrix, so that the design ginseng of recurrence identification network Number, i.e. E^idMinimum is reached, its detailed process is：

Recursion control neutral net and recurrence identification neural network structure is similar, recursion control neutral net neuron it is defeated Going out expression formula is：

Object function is：

Error expression between system reality output and desired output is：

The minimum of object function need to be iterated to weight matrix：

Target function gradient is represented by：

Recursion control neutral net still carries out on-line study using realtime recurrent algorithm, and specific derivation process is as follows：

In formula, nco is that recursion control network inputs are total, δ_trIt is kronecker delta, nhi is the number of implicit input, That is the virtual output number of recursion control network-feedback.

Fig. 2 is to realize air maneuver Load alleviation based on the maneuver load controller of recurrent neural network using the present invention Schematic diagram, the input of controller includes aircraft normal g-load, the flying angle of aircraft, wing root bending moment, and output includes that horizontal tail is deflected Angle, aileron movement angle.Air maneuver load is realized based on the maneuver load controller of recurrent neural network using the present invention Slow down and be divided into two stages：1st, system identification；2nd, maneuver load slows down.Specific embodiment is as follows：

Controller parameter is designed under Mach 2 ship 0.6.

The system identification stage lasts 10s, and horizontal tail is with aileron initial input as shown in figure 3, recurrence identification neural network parameter It is as follows：

Neural network learning rate：η^id=0.5

Activation primitive is：

Recurrence identification neutral net optimum initial weight matrix is obtained by numerical simulation.System identification result such as Fig. 5 institutes Show, it can be seen that recurrence identification network output can quickly and accurately tracing system open loop output.

Maneuver load deceleration phase lasts 10s, and an open loop input horizontal tail is deflected, as shown in Figure 4.Recursion control nerve net Network parameter is as follows：

Neural network learning rate：η^co=0.02

Activation primitive is：

Recursion control neutral net optimum initial weight matrix is obtained by numerical simulation.System control result such as Fig. 6 institutes Show.It can be seen that in Mach 2 ship 0.6, this can make the wing based on the maneuver load controller of recurrent neural network Root bending moment ratio does not adopt controller to reduce by 26%, and normal g-load and the angle of attack be consistent when not adopting controller.

After the completion of maneuver load controller parameter design based on recurrent neural network, parameter is applied to into other Mach numbers Under, and the parameter lambda related to Mach number is introduced, as recurrence identification neutral net initial weight matrix and recursion control nerve net The multiplier of network initial weight matrix, λ is as follows with the relation of Mach number Ma：

λ=0.29352-1.96925₁Ma+10.0211Ma²-8.10263Ma³

For aircraft, the maneuvering in the range of Mach number 0.3 to Mach number 0.69 is suitable for this parameter, is shown in Fig. 7 Maneuver load slows down situation under each Mach number, it can be seen that wing root bending moment highest can slow down 49%.Under different Mach number horizontal tail with Aileron movement is as shown in Figure 8.

The above is preferred embodiment of the invention, and here description of the invention and application are illustrative；It should be pointed out that For those skilled in the art, some changing can also be made under the premise without departing from the principles of the invention Enter, these improvement also should be regarded as protection scope of the present invention.

Claims (6)

1. a kind of maneuver load controller based on recurrent neural network, it is characterised in that described maneuver load controller Neutral net recognizes neutral net by a recurrence and a recursion control neutral net is in series.

2. the maneuver load controller based on recurrent neural network according to claim 1, it is characterised in that described passs Return identification neutral net be with the basic mathematic model of recursion control neutral net：

$v_k=\underset{j=1}{\overset{p}{\Σ}}{w}_{kj}{u}_j$

In formula, v_kFor internal net activation, u_jRepresent input vector, w_kjThe weight matrix of synapse is represented,For activation primitive, y_kRepresent Output vector, p is to include being biased in interior neutral net input sum.

3. a kind of control method of the maneuver load controller based on recurrent neural network, it is characterised in that comprise the following steps that：

Step 1, using the normal g-load signal of aircraft, the angle of attack and wing root bending moment signal as input, the deflection of horizontal tail and aileron As output；

Step 2, the parameter and initial weight matrix of neutral net activation primitive are recognized to system identification by adjusting recurrence；

Step 3, by adjusting the parameter and initial weight matrix of recursion control neutral net activation primitive to normal g-load, the angle of attack Tracking and the suppression to wing root bending moment.

4. the control method of a kind of maneuver load controller based on recurrent neural network according to claim 3, it is special Levy and be, the activation primitive that described recurrence identification neutral net and recursion control neutral net are adopted for：

The quiet activation v of neutral net_jWith input u_iBetween meet：

$v_j\left(n\right)=\underset{i=1}{\overset{p}{\Σ}}{w}_{ij}\left(n\right)\·u_i\left(n\right)$

In formula：w_ijFor weight matrix, v_jFor the quiet activation of neutral net, u_jFor neutral net input vector.

5. the control method of a kind of maneuver load controller based on recurrent neural network according to claim 3 or 4, its It is characterised by, described step 2 is specially：

2.1, recurrence identification network carries out on-line study, calculation error gradient, recurrence identification nerve net by realtime recurrent algorithm The error function of network is：

$E^{\mathrm{id}}\left(n\right)=\frac{1}{2}\underset{j=1}{\overset{\mathrm{nm}}{\mathrm{\Σ}}}{\left[e_j^{\mathrm{id}}\left(n\right)\right]}^2$

In formula, E^idFor object function, nm is sensor sum,It is the error of real sensor signal and identification signal；

2.2, according to the error of real sensor signal and identification signal when recurrence recognizes n-th that neutral net recognized Spacer step, expression formula is：

$e_j^{id}\left(n\right)=y_j\left(n\right)-{\hat{y}}_j\left(n\right)$

Recurrence identification neutral net neuron is output as：

$v_j^{id}\left(n\right)=\underset{i=1}{\overset{nid}{\Σ}}{w}_{ij}^{id}\left(n\right)\·u_i^{id}\left(n\right)$

In formula,For the output of neuron, nid is neutral net input sum；

2.3, the design parameter of recurrence identification network, i.e. E are made by identification process^idMinimum is reached, its detailed process is：

$w_{ij}^{id}(n+1)=w_{ij}^{id}\left(n\right)+{\mathrm{\Δw}}_{ij}^{id}\left(n\right)$

${\mathrm{\Δw}}_{kl}^{id}\left(n\right)=-{\mathrm{\η}}^{id}\frac{\∂E^{id}\left(n\right)}{\∂w_{kl}^{id}\left(n\right)}$

In formula, η^idFor the learning rate of neutral net.

6. the control method of a kind of maneuver load controller based on recurrent neural network according to claim 5, it is special Levy and be, described step 3 is specially：

Recursion control neutral net carries out on-line study, the output of recursion control neutral net neuron using realtime recurrent algorithm Expression formula is：

Object function is：

$E^{co}\left(n\right)=\frac{1}{2}\underset{j=1}{\overset{nm}{\Σ}}{e}_j^{{\mathrm{co}}^2}\left(n\right)+\frac{1}{2}\mathrm{\γ}\underset{j=1}{\overset{ni}{\Σ}}{\mathrm{\δ}}_j^{{\mathrm{co}}^2}\left(n\right)$

Error expression between system reality output and desired output is：

$e_j^{co}\left(n\right)=y_j^{des}(n+1)-{\hat{y}}_j^{id}(n+1)$

Weight matrix is iterated：

${\mathrm{\Δw}}_{tv}^{co}\left(n\right)=-{\mathrm{\η}}^{co}\frac{\∂E^{co}\left(n\right)}{\∂w_{tv}^{co}\left(n\right)}$

Target function gradient is represented by：

$\frac{\∂E^{co}\left(n\right)}{\∂w_{tv}^{co}\left(n\right)}=\underset{j=1}{\overset{nm}{\Σ}}{e}_j^{co}\left(n\right)\·\frac{\∂e_j^{co}\left(n\right)}{\∂w_{tv}^{co}\left(n\right)}+\mathrm{\γ}\underset{j=1}{\overset{ni}{\Σ}}{\mathrm{\δ}}_s^{co}\left(n\right)\·\frac{\∂{\mathrm{\δ}}_s^{co}\left(n\right)}{\∂w_{tv}^{co}\left(n\right)}=E_1^{\′}\left(n\right)+\mathrm{\γ}\·E_2^{\′}\left(n\right)$

In formula, nm is the output sum of recurrence identification network；Ni is the output sum of recursion control network, and γ is regulation controller The parameter of net activation；e^coFor the error between system reality output and desired output；η^coFor the learning rate of recursion control network.