CN109946972A - Underwater robot Predictive Control System and method based on on-line study modelling technique - Google Patents
Underwater robot Predictive Control System and method based on on-line study modelling technique Download PDFInfo
- Publication number
- CN109946972A CN109946972A CN201910276540.5A CN201910276540A CN109946972A CN 109946972 A CN109946972 A CN 109946972A CN 201910276540 A CN201910276540 A CN 201910276540A CN 109946972 A CN109946972 A CN 109946972A
- Authority
- CN
- China
- Prior art keywords
- line study
- network
- prediction model
- output
- control
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Feedback Control In General (AREA)
Abstract
Underwater robot Predictive Control System and method based on on-line study modelling technique, it belongs to the movement control technology field of autonomous underwater robot.The present invention solves the problems, such as that the existing face S controller is difficult to obtain optimal control parameter, controller motion control is caused to be affected.The present invention solves the optimal control parameter in finite time-domain by establishing pre- geodesic structure, designs a kind of face prediction S controller based on on-line study model；In the on-line study prediction model link of pre- geodesic structure, establishing parallel organization avoids prediction output from conflicting with the calculating of on-line study, the mode of flexible transition is proposed to reduce prediction model switching bring output jitter, and being added to sliding window and study judgement for reducing on-line study is system bring computational burden, to improve the control performance and autonomous regulating power of AUV motion controller, and enhance its adaptability to marine environment variation.Present invention could apply to the movement control technology fields of autonomous underwater robot.
Description
Technical field
The invention belongs to the movement control technology fields of autonomous underwater robot, and in particular to one kind is based on on-line study mould
The underwater robot Predictive Control System and method of type technology.
Background technique
With the raising of "Oceanic" strategy status, autonomous underwater robot (autonomous underwater in recent years
Vehicle, AUV) importance it is also increasingly prominent.AUV is related to multiple ambits such as computer, control, material, and merges
The multinomial key technology such as advanced design manufacturing technology, the energy and Push Technology, underwater navigation technology and subsurface communication technology.Its
In, movement control technology is the important content of AUV technology, and only AUV has good control performance, can be guaranteed in complexity
Marine environment in smoothly complete job task.
PID control based on PID control is controlled by ratio, passes through integral as one of most widely used control method
Steady-state error is eliminated in control, and uses differential control to weaken overshoot trend.Document [The Research of Fuzzy-PID
Control Based on Grey Predition for AUV] fuzzy control is combined with PID control, for AUV multiple
Bow under miscellaneous marine environment is to control.Document [Depth control for an over-actuated, hover-capable
Autonomous underwater vehicle with experimental verification] propose one kind based on PID
Control system, for overdrive AUV with lower system-computed cost carry out it is deep-controlled.Document [Modeling and
control of autonomous underwater vehicle(AUV)in heading and depth attitude
Via self-adaptive fuzzy PID controller] design parameter self-tuning fuzzy PID controller use fuzzy rule
Pid parameter, and combining adaptive control thought are adjusted, AUV model is established and carries out depth and bow to control.
But PID control substantially belongs to Linear Control, between AUV kinetic characteristic there are many inadaptable places, because
This application effect in the field is still limited.
Fuzzy control is a kind of non-linear control based on fuzzy set theory, Fuzzy Linguistic Variable and fuzzy logic inference
Method processed for certain excessively complexity or is difficult to have preferable control effect with the system that mathematical model describes.Document
[Optimized Fuzzy Control Deisgn of An Autonomous Underwater Vehicle[J]
.Iranian Journal of Fuzzy Systems] devise a kind of roll for AUV of fuzzy controller, bow Xiang Yushen
Degree control, and worked well by emulation experiment access control device.Document [Fuzzy Adaptive Control for
Trajectory Tracking of Autonomous Underwater Vehicle] it is directed to the position of AUV in the horizontal plane
With Attitude Tracking problem, propose a kind of fuzzy adaptivecontroller algorithm, and by Lyapunov's theory prove its convergence with
Stability.Document [An Optimized Fuzzy Control Algorithm for Three-Dimensional AUV
Path Planning] FUZZY ALGORITHMS FOR CONTROL is optimized, based on AUV three-dimensional road in sonar model realization complexity underwater environment
Diameter planning.
In practical applications, the effect of fuzzy control is heavily dependent on fuzzy control rule, input fuzzy variable
The selection of domain and membership function, however these elements rely primarily on professional at present and rule of thumb determine.In addition, with
The refinement of fuzzy rule, control algolithm can also bring larger computational load to computer.
The conventional PD control of collinearity is compared, and the classical face the S control with nonlinear Control curved surface is more applicable for autonomous
The motion control of underwater robot.In addition, fuzzy control needs according to the actual situation to subordinating degree function, fuzzy variable and fuzzy
Numerous Internal Elements such as rule are adjusted, and the classics face S the included control parameter of control is less, can greatly simplify parameter tune
It is had suffered journey, therefore there is stronger practicability.
But in current engineer application, the existing face S controller mainly by designer by experience in a manner of trying to gather Lai
Complete parameter setting and adjustment.The parameter adjustment mode inefficiency, it tends to be difficult to obtain one group of optimal control parameter, even
Sometimes the motion control effects of controller are influenced because parameter setting is improper.
Summary of the invention
The purpose of the present invention is to solve the existing face S controllers to be difficult to obtain optimal control parameter, leads to controller
Motion control the problem of being affected.
The technical solution adopted by the present invention to solve the above technical problem is:
Based on one aspect of the present invention, the underwater robot Predictive Control System based on on-line study modelling technique, institute
Stating Predictive Control System includes prediction model module, feedback compensation module, rolling optimization module and the control of the face S based on on-line study
Molding block, in which:
The prediction model module based on on-line study is used to export the predicted value of AUV state, described to be learned based on online
The prediction model module of habit includes prediction model network submodular and on-line study network submodular；On-line study network
Module obtains new on-line study network for on-line study；The prediction model network submodular is according to new on-line study net
Network updates weight；
The feedback compensation module be used for based on on-line study prediction model module output AUV status predication value into
Row amendment, obtains revised predicted value；
The rolling optimization module is used for according to AUV motion control aim parameter and revised predictor calculation optimum control
Parameter；
The face S control module is according to the deviation and optimum control ginseng between AUV motion control aim parameter and quantity of state
Number is that AUV exports control amount.
Based on another aspect of the present invention: the underwater robot Predictive Control System based on on-line study modelling technique
Control method, method includes the following steps:
Step 1: being utilized respectively training sample to initial predicted using pre-recorded AUV aeronautical data as training sample
Prototype network and initial online learning network carry out off-line training, and the prediction model network for obtaining Predictive Control System is learned with online
Practise network；
Step 2: the learning sample maximum capacity of prediction model network and on-line study network that setting steps one obtain,
Select prediction model network and on-line study network in the learning sample at each moment according to the learning sample maximum capacity of setting；
The state of initial time on-line study network is idle；The condition of on-line study network startup on-line study is set,
In the not up to entry condition of on-line study network, directly exported according to prediction model network as based on the pre- of on-line study
Survey the predicted value of the AUV state of model module output；
When reaching the entry condition of on-line study network, on-line study network is carried out online with the learning sample at current time
Study obtains new on-line study network after on-line study；
According to new on-line study network, prediction model module using flexible transition new mechanism based on on-line study
Prediction output；When the prediction output of the prediction model module based on on-line study is equal to the output of new on-line study network,
The weight of prediction model network is directly replaced with to the weight of new on-line study network；Online learn is reset after the completion of weight replacement
Habit network is idle state；
Step 3: feedback compensation module is exported according to the prediction model module based on on-line study of last moment and AUV
Deviation between reality output is modified the prediction model module output based on on-line study in current time；Worked as
The revised output of prediction model module based on on-line study in the preceding moment；
Step 4: calculating commenting for current control parameter in conjunction with AUV motion control aim parameter and the revised output of step 3
Value, then scanned in solution space according to evaluation of estimate, obtain one group of optimal control parameter k1With k2；
Step 5: the face S control module is according to control parameter k1、k2With AUV motion control aim parameter yinWith quantity of state yout's
Deviation exports control amount u, carries out motion control to AUV using control amount u.
The beneficial effects of the present invention are: a kind of pre- observing and controlling of underwater robot based on on-line study modelling technique of the invention
System and method processed, by establishing by prediction model module, feedback compensation module and rolling optimization module based on on-line study
The pre- geodesic structure of composition solves the optimal control parameter in finite time-domain, designs a kind of prediction S based on on-line study model
Face controller；Output is predicted based on parallel organization in on-line study prediction model link, is established in pre- geodesic structure
Conflict with the calculating of on-line study, proposes the mode of flexible transition to reduce prediction model switching bring output jitter, and
It is added to sliding window and study determines for reducing on-line study to be system bring computational burden, to improve AUV movement
The control performance of controller and autonomous regulating power, and enhance its adaptability to marine environment variation.
Detailed description of the invention
Fig. 1 is the structural representation of the underwater robot Predictive Control System of the invention based on on-line study modelling technique
Figure；
Z in figure-1Represent current time is influenced by time data before；ym' (t+d/t) is t moment on-line study network
Output after on-line study；
Fig. 2 is the schematic diagram of the parallel organization of prediction model network and on-line study network of the invention；
Fig. 3 is distinguished Longitudinal Dynamic Model, off-line learning neural network model and on-line study neural network model
Apply the output response figure obtained with high frequency sinusoidal input signal；
Fig. 4 is distinguished Longitudinal Dynamic Model, off-line learning neural network model and on-line study neural network model
Apply the output response figure obtained with low frequency sinusoidal input signal；
Fig. 5 is distinguished Longitudinal Dynamic Model, off-line learning neural network model and on-line study neural network model
Apply the output response figure obtained with square-wave input signal；
Fig. 6 is in longitudinal velocity control, respectively with off-line learning neural network model and on-line study neural network
Model is as prediction model, the control effect figure of the obtained prediction face S controller；
Fig. 7 is distinguished to kinetic model, off-line learning neural network model and on-line study neural network model bow
Apply the output response figure obtained with high frequency sinusoidal input signal；
Fig. 8 is distinguished to kinetic model, off-line learning neural network model and on-line study neural network model bow
Apply the output response figure obtained with low frequency sinusoidal input signal；
Fig. 9 is distinguished to kinetic model, off-line learning neural network model and on-line study neural network model bow
Apply the output response figure obtained with square-wave input signal；
Figure 10 is in bow into control, respectively with off-line learning neural network model and on-line study neural network model
As prediction model, the control effect figure of the obtained prediction face S controller.
Specific embodiment
Specific embodiment 1: a kind of underwater robot based on on-line study modelling technique described in present embodiment is pre-
Survey control system, the Predictive Control System include prediction model module based on on-line study, feedback compensation module, roll it is excellent
Change module and the face S control module, in which:
The prediction model module based on on-line study is used to export the predicted value of AUV state, described to be learned based on online
The prediction model module of habit includes prediction model network submodular and on-line study network submodular；On-line study network
Module obtains new on-line study network for on-line study；The prediction model network submodular is according to new on-line study net
Network updates weight；
The feedback compensation module be used for based on on-line study prediction model module output AUV status predication value into
Row amendment, obtains revised predicted value；
The rolling optimization module is used for according to AUV motion control aim parameter and revised predictor calculation optimum control
Parameter；
The face S control module is according to the deviation and optimum control ginseng between AUV motion control aim parameter and quantity of state
Number is that AUV exports control amount.
Specific embodiment 2: based on the underwater based on on-line study modelling technique described in specific embodiment one
The control method of people's Predictive Control System, method includes the following steps:
Step 1: being utilized respectively training sample to initial predicted using pre-recorded AUV aeronautical data as training sample
Prototype network and initial online learning network carry out off-line training, and the prediction model network for obtaining Predictive Control System is learned with online
Practise network；
Prediction model network and on-line study network are parallel organization, as shown in Fig. 2, in parallel organization, prediction model
Network and on-line study network use network independent respectively, and prediction output is avoided to conflict with the calculating of on-line study；Just
The structure and weight of prediction model network and on-line study network are all identical when the beginning；
Step 2: the learning sample maximum capacity of prediction model network and on-line study network that setting steps one obtain,
Select prediction model network and on-line study network in the learning sample at each moment according to the learning sample maximum capacity of setting；
The state of initial time on-line study network is idle；The condition of on-line study network startup on-line study is set,
In the not up to entry condition of on-line study network, directly exported according to prediction model network as based on the pre- of on-line study
Survey the predicted value of the AUV state of model module output；
When reaching the entry condition of on-line study network, on-line study network is carried out online with the learning sample at current time
Study obtains new on-line study network after on-line study；
According to new on-line study network, prediction model module using flexible transition new mechanism based on on-line study
Prediction output；When the prediction output of the prediction model module based on on-line study is equal to the output of new on-line study network,
The weight of prediction model network is directly replaced with to the weight of new on-line study network；Online learn is reset after the completion of weight replacement
Habit network is idle state；
Step 3: feedback compensation module is exported according to the prediction model module based on on-line study of last moment and AUV
Deviation between reality output is modified the prediction model module output based on on-line study in current time；Worked as
The revised output of prediction model module based on on-line study in the preceding moment；
Step 4: calculating commenting for current control parameter in conjunction with AUV motion control aim parameter and the revised output of step 3
Value, then scanned in solution space according to evaluation of estimate, obtain one group of optimal control parameter k1With k2；
Step 5: the face S control module is according to control parameter k1、k2With AUV motion control aim parameter yinWith quantity of state yout's
Deviation e (t) exports control amount u, carries out motion control to AUV using control amount u.
As shown in Figure 1, the face S controlling unit is that control object exports control amount within each moment, the closed loop of AUV is realized
Motion control.In each parameter setting beat, pre- geodesic structure solves the optimal control parameter in finite time-domain, realizes the control of the face S
The control parameter of molding block is arranged.
Compared with PID control method, the face the prediction S control based on on-line study model that present embodiment is proposed is used
Nonlinear Control curved surface can better adapt to the strong nonlinearity characteristic of AUV movement.
Compared with existing fuzzy control method, the face prediction S based on on-line study model that present embodiment is proposed
In control, to control effect have larger impact only there are two the face S control parameter k1With k2, and this group of parameter is by controller
It is autonomous to complete setting and adjustment, to effectively increase the engineering efficiency at scene, and ensure the motion control effects of AUV.
Specific embodiment 3: present embodiment is unlike specific embodiment two: the specific mistake of the step 2
Journey are as follows:
The learning sample maximum capacity of prediction model network and on-line study network that setting steps one obtain is N；
Pre- geodesic structure uses individual parameter setting beat, and each parameter setting beat includes N number of moment, it may be assumed that is being predicted
After structure completes primary parameter setting, the face S controlling unit calculates the control amount that N number of moment is completed using this group of parameter, until
Next parameter adjustment beat resets control parameter by pre- geodesic structure.The purpose for the arrangement is that preventing from continually adjusting control ginseng
Number can not effectively improve control effect, and system operations burden is significantly increased.
Then make at each moment only with the AUV aeronautical data in the N-1 moment before current time and current time
For learning sample；I.e. for t moment, the learning sample of prediction model network and on-line study network are as follows: { [u (t-N+1), yout
(t-N+1)],[u(t-N+2),yout(t-N+2)],…,[u(t),yout(t)] }, in which: u (t) is the face t moment S control module
The control amount of output, youtIt (t) is the quantity of state of t moment AUV reality output；
The output y of t moment AUVout(t) be mainly t-1 moment input quantity u (t-1) exercising result, but also contain and be
Unite t-2, t-3 ..., the influence of t-N moment input quantity, therefore, in number of the t moment before within the scope of certain time
According to weight and threshold value inside adjustable neural network, to realize the on-line study of neural network；
The target error function of Neural Network Online study is defined as
Wherein: EiThe error at corresponding i-th moment；Target error E (p) be N number of moment error and；
After t moment, then the recursion renewal learning sample data set in the way of sliding window, i.e., for the t+1 moment,
The learning sample of prediction model network and on-line study network can indicate are as follows: { [u (t-N+2), yout(t-N+2)],[u(t-
N+3),yout(t-N+3)],…,[u(t+1),yout(t+1)] }, and so on；
The state of the initial time of on-line study network is set as idle, and the entry condition of on-line study network is arranged are as follows:
If in continuous N1The deviation of the output of prediction model module and AUV reality output quantity of state in a moment based on on-line study
Mean value is greater than preset value, then it is assumed that marine environment locating for AUV has varied widely, and starts online of on-line study network
It practises；When the error function of on-line study network meet error limitation require (i.e. the error function value of on-line study network be less than etc.
After error threshold ε), terminates on-line study, obtain new on-line study network；
Determined using sliding window and on-line study starting, reduction on-line study is system bring computational burden；
Moreover, the prediction face S controller obtains neural network using last training before on-line study network completes training
Temporarily as prediction model, guarantee the stabilization of controller output.On-line study network needs not participate in the calculating of controller, is absorbed in
It is required in being trained according to current sample data until meeting the limits of error, to improve e-learning speed.Due to it is new
There may be larger differences between line learning network and current predictive prototype network, replace if directlying adopt new on-line study network
Generation former prediction model network may cause the output jitter of controller as new prediction model network, therefore flexible transition is arranged
Mechanism makes prediction model network smoothly complete neural network replacement；
Choose the N after on-line study2A moment is as transitional period, the transitional period interior prediction model based on on-line study
The AUV status predication value of module output are as follows:
ym=σ ym1+θym2 (1)
In formula, ym1With ym2Prediction model module respectively before the on-line study of on-line study network based on on-line study
The output of on-line study network after output and the on-line study of on-line study network, σ and θ are adjustment factor, meet+θ=1 σ；
In transitional period, on-line study network will not temporarily start new training and calculate.In addition, factor sigma is gradually reduced to from 1
0, and coefficient θ then progressively increases to 1 from 0.
As θ=1, indicate online after the output of the prediction model module based on on-line study has transferred to on-line study
The weight of prediction model network, is directly substituted for the weight of new on-line study network by the output of learning network, and resetting is online
Learning network is idle state, and terminates the transitional period.
After transitional period, continue using the output of prediction model network as the prediction model module based on on-line study
Output, until next transitional period starts.
Specific embodiment 4: present embodiment is unlike specific embodiment two: the specific mistake of the step 3
Journey are as follows:
It is exported according to the prediction model module prediction based on on-line study of last moment inclined between AUV reality output
Difference is modified the prediction output of the prediction model module based on on-line study in current time, formula specific as follows:
yp(t+d/t)=ym(t+d/t)+em(t) (2)
em(t)=yout(t-1)-ym(t-1/t-2) (3)
ym(t-1/t-2)=fm[yout(t-2),u(t-1)] (4)
In formula, yp(t+d/t) it indicates to ym(t+d/t) output after being modified；emIt (t) is the correction amount of t moment；ym
(t+d/t) the AUV shape to the t+d moment in predetermined period exported in t moment based on the prediction model module of on-line study is indicated
The predicted value of state；ymIt (t-1/t-2) is that the t-1 moment that the prediction model module based on on-line study is calculated according to historical data is pre-
Measured value；yout(t-1) quantity of state in t-1 moment AUV reality output is indicated；yout(t-2) indicate actually defeated in t-2 moment AUV
Quantity of state out；U (t-1) is the control amount of t-1 moment S face control module output, fm[] is the non-linear letter of neural network
Number.
The purpose of present embodiment is to make the output of pre- geodesic structure closer in real data.
Specific embodiment 5: present embodiment is unlike specific embodiment two: the specific mistake of the step 4
Journey are as follows:
In order to evaluate the control effect of one group of control parameter, rolling optimization link is chosen improvement ITAE criterion and is referred to as performance
Evaluation function is marked to calculate the evaluation of estimate of current control parameter；
The improvement ITAE criterion on the basis of ITAE criterion, introduces overshoot penalty coefficient；Introduce overshoot punishment system
Several purposes is the susceptibility improved to overshoot, to enhance controller to the rejection ability of overshoot；Introduce overshoot penalty coefficient
Performance indicator evaluation function Φ afterwardspExpression formula it is as follows:
In formula, α is overshoot penalty, takes α=1 under non-overshoot state, α > 1, e is taken under overshoot stateΦ(t) when indicating t
The margin of error at quarter, i.e. t moment AUV motion control aim parameter yin(t) and yp(t+d/t) difference；
One group of optimal control parameter k is obtained using SA algorithm again1With k2。
The simulated annealing (Simulated annealing algorithm, SA) of present embodiment derives from solid
Solid, is first heated up to sufficiently high by annealing theory, then it is allowed slowly to cool down.Solid interior particle following temperature rising becomes unordered when heating
Shape, it is interior to increase, and particle is gradually orderly when cooling down, and reaches equilibrium state in each temperature, finally reaches ground state at room temperature,
It is interior to be kept to minimum.
The generation and receiving of simulated annealing new explanation can be divided into following four steps:
It 1) is to generate function by one to generate the new explanation for being located at solution space from current solution；For convenient for it is subsequent calculating and
Receive, reduces algorithm time-consuming, generally select by current new explanation by simply converting the method that can produce new explanation, such as to composition
All or part of element of new explanation is replaced, is exchanged, it is noted that the transform method for generating new explanation determines current new explanation
Neighbour structure, thus have a certain impact to the selection of Cooling -schedule.
It 2) is that calculating is poor with objective function corresponding to new explanation.Because objective function difference is only generated by conversion section,
Incremental computations are preferably pressed in the calculating of objective function difference.It turns out that this is that calculating target function is poor for most of applications
Quickest way.
It 3) is to judge whether new explanation is received, the foundation of judgement is an acceptance criterion, and most common acceptance criterion is
Metropolis criterion: receiving S' as new current solution S if Δ < 0, otherwise using probability exp (- Δ/T) receive S' as
New current solution S.
It 4) is to replace currently solving with new explanation when new explanation is determined receiving, this need to will correspond in current solution generates newly
Conversion section when solution is achieved, while modified objective function value.At this point, current solution realizes an iteration.It can be
Start next round test on the basis of this.And when new explanation is judged as giving up, then continue next round on the basis of former current solution
Test.
Specific embodiment 6: present embodiment is unlike specific embodiment five: described to be obtained using SA algorithm
One group of optimal control parameter k1With k2, detailed process are as follows:
1) original state S is randomly selected, and takes higher initial temperature T0, selecting Markov Chain initial length is L0, initially
Change the number of iterations L=0；
2) new state S ', S ' formula specific as follows is generated after doing random perturbation to current state S:
S'=S+Rand (0,1) ω (6)
In formula, ω is to fixed step size, and Rand (0,1) is random number；
3) the performance indicator Φ of original state S and new state S ' is calculatedp(S) and Φp(S'), increment Delta such as following formula is obtained:
Δ=Φp(S')-Φp(S) (7)
If 4) Δ < 0, S'=S is enabled, is then gone to 6)；
If 5) Δ > 0, random number p is generated, as p < exp (- Δ/T), T is to enable S' when the corresponding temperature of previous iteration
=S, otherwise S is constant；It goes to 6)；
6) L=L+1 is enabled, if L < L0, then return 2)；If L >=L0, then go to 7)；
7) cool down according to the following formula
T=β T (8)
In formula, β is attenuation rate；
8) it checks whether annealing process terminates, if being not finished, enables L=0, turn 2)；If terminating, turn 9)；
9) using current state S as optimal solution, the corresponding control parameter k of rolling optimization module output1With k2, terminate entire
Process.
Specific embodiment 7: present embodiment is unlike specific embodiment two, three, four, five or six: the S
Face control module substitutes the broken line face of entire fuzzy rule base using smooth Sigmoid curved surface, and by adjusting the offset of the face S
Droop is eliminated, toroidal function expression formula is as follows:
In formula, OsIndicate intermediate variable, control output is by taking [- 1,1] after normalized；E withIndicate that control is defeated
Enter, respectively deviation and deviation variation rate, e withEqually pass through normalized；k1With k2Indicate optimal control parameter, respectively
Corresponding deviation and deviation variation rate, take (0 ,+∞)；TmaxIndicate that AUV can be provided maximum thrust (torque)；TcIndicate anti-normalizing
The thrust (torque) of reality output after change, δ are the fixation perturbed force obtained by adaptive mode；
Wherein, the regulation flow process of fixed perturbed force δ is as follows:
It a) is deviation ratioA threshold value is set, is judgedWhether it is less than given threshold to go to step b), otherwise if being less than
It goes to step c)；
B) by the deviation e deposit storage array of the freedom degree, while counter is added 1, and judge that nonce counter is
It is no to reach activation threshold value, if reaching activation threshold value, go to step d), otherwise goes to step c)；
C) storage array first place is removed, and all numerical value below is moved forward one, and counter is subtracted 1, gone to step a)；
D) weighted average for calculating numerical value in storage array, for calculating the offset of AUV motion control output, thus from
The output of adjustment controller is adapted to eliminate fixed control deviation, and storage array and counter are reset, executes subsequent cycle.
Embodiment
The foundation of autonomous underwater robot Controlling model of the invention
Establish following two right-handed coordinate system: first is that fixed coordinate system E- ξ η ζ, is fixed on the earth；Second is that kinetic coordinate system
O-xyz is moved with underwater robot.The optional earth of origin E of fixed coordinate system E- ξ η ζ takes up an official post meaning a bit, and ξ axle position is in level
Face, and forward direction is projected as in horizontal plane with underwater robot base course；η axis is similarly positioned in horizontal plane, by right-hand rule by E ξ
Axis rotates clockwise 90 °；ζ axis is directed toward the earth's core and is positive perpendicular to ξ E η coordinate plane.Fixed coordinate system is given a definition underwater machine
The position vector of device people is [ξ η ζ], and attitude vectors areThe origin O of kinetic coordinate system O-xyz is generally selected in water
Lower robot center of gravity, x, y and z axes pass through O point and are located at Water Plane, cross section and vertical middle section, and forward direction is according to the right side
The regulation of hand system is respectively directed to head end, right side and the bottom of autonomous underwater robot.Kinetic coordinate system is given a definition autonomous underwater machine
The linear velocity vector of device people is [u v w], and angular velocity vector is [p q r].
Assuming that fixed coordinate system is overlapped with kinetic coordinate system, each attitude angle is defined as follows: bow is to angleFor ξ axis and x
In the angle of horizontal plane, right-hand rotation is positive axis；Angle of Trim θ is ξ axis and x-axis in the angle of vertical plane, and tail, which inclines, to be positive；Angle of Heel ψ is
Angle between xOz plane and the vertical plane xO ζ for passing through x-axis, Right deviation are positive.
Position in fixed coordinate system is unified for vector with attitude angleMovement is sat
Linear velocity and angular speed in mark system are unified for vector v=[u v w p q r]T, according to bibliography, [Shi Shengda submarine is grasped
Vertical property] in derivation, autonomous underwater robot kinematics formula is
Matrix J (η)=diag (J is converted in formula1(η),J2(η)), centerline velocities transition matrix is
Angular speed transition matrix is
When due to Angle of Trim θ=± 90 °, transition matrix J2(η) be not significant, therefore is defined to Angle of Trim:
Commonly used underwater human operator model is as follows both at home and abroad
In formula, M is inertial matrix, wherein including additional mass；C (υ) is Coriolis centripetal force matrix, wherein including additional matter
Amount；D (υ) is fluid damping matrix；G (η) is the power and torque vector of gravity and buoyancy；τ be executing agency power and torque to
Amount.
Inertial matrix M=MRB+MA, wherein MRBFor Rigid Mass matrix, such as following formula
In formula, m is quality, and I is inertia item, [xG yG zG] it is center of gravity coordinate under kinetic coordinate system.
For the autonomous underwater robot being fully submerged during navigation in water, additional mass matrix MAInterior each coefficient is
Constant, such as following formula
In formula,WithEtc. being hydrodynamic force derivatives, need the contained model experiment data of AUV and combine to calculate
Hydrodynamics and identification technology etc. obtain.
Coriolis centripetal force Matrix C (v)=CRB(v)+CA(v), wherein CRBIt (v) is rigid body centripetal force matrix, such as following formula
CA(v) coriolis force matrix is
In formula, each coefficient is as follows
Fluid damping matrix D (v)=Dl+Dn(v), wherein DlFor linear damping matrix such as following formula
Dl=-diag { Xu Yv Zw Kp Mq Nr} (20)
Nonlinear dampling matrix Dn(v) it is
Dn(v)=- diag { Xu|u||u| Yv|v||v| Zw|w||w| Kp|p||p| Mq|q||q| Nr|r||r|} (21)
The power of gravity and buoyancy and torque vector g (η) such as following formula
In formula, W is gravity, and B is buoyancy, [xB,yB,zB] it is centre of buoyancy coordinate under kinetic coordinate system
The power and torque vector τ such as following formula of executing agency
τ=[X Y Z K M N]T (23)
In formula, X, Y and Z are three axle thrusts, and K, M and N are three shaft torques.
Practical Project situation carries out following items to the above AUV motion model and simplifies:
(1) setting center of gravity is overlapped with kinetic coordinate system origin；
(2) gravity is equal with buoyancy configuration, and centre of buoyancy is right above center of gravity；
(3) assuming structure has a symmetry, i.e., xGz plane bilateral symmetry and in yGz plane it is symmetrical above and below；
(4) ignore roll motion；
(5) power that executing agency can generate and torque only include longitudinal thrust, vertical thrust, turn bow torque and pitching power
Square.
Further, since AUV six-freedom motion model complexity is higher, to further facilitate controller design, by it
It is decomposed into horizontal plane and vertical plane.
In conclusion to establish AUV Controlling model as follows by the present invention:
Controlling model is in horizontal plane
Controlling model is in vertical plane
The AUV Controlling model and output control amount u that can be established according to the present invention carry out motion control to AUV.
Emulate part
Emulation prepares: to verify the adaptability that there is on-line study prediction model link to change to marine environment, this hair
It is bright under MATLAB environment selection control by the longitudinal velocity that ocean current is affected and bow to control progress emulation experiment.
Firstly, the output response of acquisition Controlling model after adjustment under Setting signal, in this, as sample data to Elman
Neural network carries out off-line training, to obtain neural network model for prediction model link.Then, apply respectively high and low
The input of frequency sinusoidal signal and square-wave signal, Controlling model that the comparison present invention establishes, off-line learning prediction model and online
Learn the response output of prediction model.Then, respectively using offline with on-line study model as prediction model, the face S is predicted in comparison
The control effect of controller, it was demonstrated that the face the prediction S control method based on on-line study model can better adapt to marine environment
Variation.
Wherein parameter setting is as follows:
1) choosing control beat is 0.1s, and it is 3s that parameter, which adjusts beat, and prediction time domain is 8s.
2) face S controlling unit:
According to formula (2), only there are two control parameter k for the face S controller1With k2It needs to set.Enabling initial time is first ginseng
Number predetermined period, i.e., pre- geodesic structure will independently complete the face S control parameter k1With k2Setting without choosing initial value manually.
3) on-line study prediction model link:
Using the neural network after off-line training as prediction model and initial on-line study model；
4) feedback compensation link:
Printenv needs to be arranged.
5) rolling optimization link:
For SA algorithm, initial temperature T is set0=1000, the number of iterations upper limit L0=100, temperature decline coefficient β=
0.9。
In addition, motion control object uses AUV Controlling model established by the present invention in emulation experiment, parameter assignment is shown in Table
1。
1 hydrodynamic force coefficient of table summarizes
Simulation result: longitudinal velocity control
Current speed u is setc=1m/s, the angle of current
Longitudinal Dynamic Model, off-line learning neural network model and on-line study neural network model are applied together respectively
High frequency sinusoidal, low frequency sinusoidal and square-wave input signal, it is as in Figure 3-5 to obtain output response.
Respectively using off-line learning neural network model and on-line study neural network model as prediction model, S is predicted
The control effect of face controller is as shown in Figure 6.
Bow is to control:
Current speed u is setc=0.5m/s, the angle of current
Bow is applied together respectively to kinetic model, off-line learning neural network model and on-line study neural network model
High frequency sinusoidal, low frequency sinusoidal and square-wave input signal, it is as Figure 7-9 to obtain output response.
Respectively using off-line learning neural network model and on-line study neural network model as prediction model, S is predicted
The motion control effects of face controller are as shown in Figure 10.
Simulation analysis: by the comparison of neural network identification result as can be seen that the same addition of the neural network of off-line learning
There are certain output response deviations between the Controlling model of ocean current interference.And after passing through on-line study, neural network prediction model
Output error be obviously reduced.
In terms of control effect, since precision of forecasting model declines, the prediction face the S controller based on off-line learning model
Performance is also declined therewith, or even occurs larger overshoot into control in bow, and is based on on-line study model in contrast
The face prediction S control then have good control effect.
In conclusion on-line study function added by the present invention can effectively correct the response error of off-line model, base
In the prediction face S of on-line study model, control can better adapt to discussed ocean current interference factor.
Above-mentioned example of the invention only explains computation model and calculation process of the invention in detail, and is not to this
The restriction of the embodiment of invention.It for those of ordinary skill in the art, on the basis of the above description can be with
It makes other variations or changes in different ways, all embodiments can not be exhaustive here, it is all to belong to the present invention
The obvious changes or variations extended out of technical solution still in the scope of protection of the present invention.
Claims (6)
1. the underwater robot Predictive Control System based on on-line study modelling technique, which is characterized in that the PREDICTIVE CONTROL system
System includes prediction model module, feedback compensation module, rolling optimization module and the face S control module based on on-line study, in which:
The prediction model module based on on-line study is used to export the predicted value of AUV state, described based on on-line study
Prediction model module includes prediction model network submodular and on-line study network submodular；The on-line study network submodular
New on-line study network is obtained for on-line study；The prediction model network submodular according to new on-line study network more
New weight；
The feedback compensation module is used to repair the AUV status predication value of the prediction model module output based on on-line study
Just, revised predicted value is obtained；
The rolling optimization module is used to be joined according to AUV motion control aim parameter and revised predictor calculation optimum control
Number；
The face S control module according between AUV motion control aim parameter and quantity of state deviation and optimal control parameter be
AUV exports control amount.
2. the controlling party based on the underwater robot Predictive Control System described in claim 1 based on on-line study modelling technique
Method, which is characterized in that method includes the following steps:
Step 1: being utilized respectively training sample to initial predicted model using pre-recorded AUV aeronautical data as training sample
Network and initial online learning network carry out off-line training, obtain the prediction model network and on-line study net of Predictive Control System
Network；
Step 2: the learning sample maximum capacity of prediction model network and on-line study network that setting steps one obtain, according to
Learning sample of the learning sample maximum capacity selection prediction model network and on-line study network of setting at each moment；
The state of initial time on-line study network is idle；The condition of on-line study network startup on-line study is set, not
When reaching the entry condition of on-line study network, directly exported according to prediction model network as the prediction mould based on on-line study
The predicted value of the AUV state of pattern block output；
When reaching the entry condition of on-line study network, on-line study network is learned online with the learning sample at current time
It practises, new on-line study network is obtained after on-line study；
According to new on-line study network, the prediction of the prediction model module using flexible transition new mechanism based on on-line study
Output；It, will be pre- when the prediction output of the prediction model module based on on-line study is equal to the output of new on-line study network
The weight for surveying prototype network directly replaces with the weight of new on-line study network；On-line study net is reset after the completion of weight replacement
Network is idle state；
Step 3: prediction model module output and AUV reality based on on-line study of the feedback compensation module according to last moment
Deviation between output is modified the prediction model module output based on on-line study in current time；When obtaining current
The revised output of prediction model module based on on-line study in quarter；
Step 4: the evaluation of estimate of current control parameter is calculated in conjunction with the revised output of AUV motion control aim parameter and step 3,
It is scanned in solution space according to evaluation of estimate again, obtains one group of optimal control parameter k1With k2；
Step 5: the face S control module is according to control parameter k1、k2With AUV motion control aim parameter yinWith quantity of state youtDeviation
Value e (t) exports control amount u, carries out motion control to AUV using control amount u.
3. the underwater robot forecast Control Algorithm according to claim 2 based on on-line study modelling technique, feature
It is, the detailed process of the step 2 are as follows:
The learning sample maximum capacity of prediction model network and on-line study network that setting steps one obtain is N；
It is then used as and learns only with the AUV aeronautical data in the N-1 moment before current time and current time at each moment
Practise sample；I.e. for t moment, the learning sample of prediction model network and on-line study network are as follows: { [u (t-N+1), yout(t-N+
1)],[u(t-N+2),yout(t-N+2)],…,[u(t),yout(t)] }, in which: u (t) is the output of t moment S face control module
Control amount, youtIt (t) is the quantity of state of t moment AUV reality output；
After t moment, then the recursion renewal learning sample data set in the way of sliding window learns that is, for the t+1 moment
Sample is { [u (t-N+2), yout(t-N+2)],[u(t-N+3),yout(t-N+3)],…,[u(t+1),yout(t+1)] }, with this
Analogize；
The state of on-line study network initial time is set as idle, and the entry condition of on-line study network is arranged are as follows: if even
Continuous N1The mean value of the deviation of the output of prediction model module and AUV reality output quantity of state in a moment based on on-line study is big
In preset value, then start the on-line study of on-line study network；It is wanted when the error function of on-line study network meets error limitation
After asking, terminates on-line study, obtain new on-line study network；
Choose the N after on-line study2A moment is as transitional period, the transitional period interior prediction model module based on on-line study
The AUV status predication value of output are as follows:
ym=σ ym1+θym2 (1)
In formula, ym1With ym2The output of prediction model module respectively before the on-line study of on-line study network based on on-line study with
The output of on-line study network after on-line study network on-line study, σ and θ are adjustment factor, meet+θ=1 σ；
As θ=1, indicate that the output of the prediction model module based on on-line study has transferred to on-line study after on-line study
The weight of prediction model network is directly substituted for the weight of new on-line study network by the output of network, resets on-line study
Network is idle state, and terminates the transitional period.
4. the underwater robot forecast Control Algorithm according to claim 2 based on on-line study modelling technique, feature
It is, the detailed process of the step 3 are as follows:
According to the prediction model module prediction output based on on-line study of last moment and the deviation pair between AUV reality output
The prediction output of the prediction model module based on on-line study in current time is modified, formula specific as follows:
yp(t+d/t)=ym(t+d/t)+em(t) (2)
em(t)=yout(t-1)-ym(t-1/t-2) (3)
ym(t-1/t-2)=fm[yout(t-2),u(t-1)] (4)
In formula, yp(t+d/t) it indicates to ym(t+d/t) output after being modified；emIt (t) is the correction amount of t moment；ym(t+d/
T) indicate t moment based on on-line study prediction model module export in predetermined period the AUV state at t+d moment it is pre-
Measured value；ym(t-1/t-2) the t-1 moment predicted value calculated for the prediction model module based on on-line study according to historical data；
yout(t-1) quantity of state in t-1 moment AUV reality output is indicated；yout(t-2) shape in t-2 moment AUV reality output is indicated
State amount；U (t-1) is the control amount of t-1 moment S face control module output, fm[] is nonlinear function.
5. the underwater robot forecast Control Algorithm according to claim 2 based on on-line study modelling technique, feature
It is, the detailed process of the step 4 are as follows:
It chooses and improves the evaluation of estimate that ITAE criterion calculates current control parameter as performance indicator evaluation function；
The improvement ITAE criterion on the basis of ITAE criterion, introduces overshoot penalty coefficient；After introducing overshoot penalty coefficient
Performance indicator evaluation function ΦpExpression formula it is as follows:
In formula, α is overshoot penalty, takes α=1 under non-overshoot state, α > 1, e is taken under overshoot stateΦ(t) t moment is indicated
The margin of error；
One group of optimal control parameter k is obtained using SA algorithm again1With k2。
6. the underwater robot forecast Control Algorithm according to claim 5 based on on-line study modelling technique, feature
It is, it is described that one group of optimal control parameter k is obtained using SA algorithm1With k2, detailed process are as follows:
1) original state S is randomly selected, and takes initial temperature T0, selecting Markov Chain initial length is L0, initialization iteration time
Number L=0；
2) new state S ', S ' formula specific as follows is generated after doing random perturbation to current state S:
S'=S+Rand (0,1) ω (6)
In formula, ω is to fixed step size, and Rand (0,1) is random number；
3) the performance indicator Φ of original state S and new state S ' is calculatedp(S) and Φp(S'), increment Delta such as following formula is obtained:
Δ=Φp(S')-Φp(S) (7)
If 4) Δ < 0, S'=S is enabled, is then gone to 6)；
If 5) Δ > 0, random number p is generated, as p < exp (- Δ/T), T is to enable S'=S when the corresponding temperature of previous iteration,
Otherwise S is constant；It goes to 6)；
6) L=L+1 is enabled, if L < L0, then return 2)；If L >=L0, then go to 7)；
7) cool down according to the following formula
T=β T (8)
In formula, β is attenuation rate；
8) it checks whether annealing process terminates, if being not finished, enables L=0, turn 2)；If terminating, turn 9)；
9) using current state S as optimal solution, the corresponding control parameter k of rolling optimization module output1With k2, terminate whole flow process.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910276540.5A CN109946972A (en) | 2019-04-08 | 2019-04-08 | Underwater robot Predictive Control System and method based on on-line study modelling technique |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910276540.5A CN109946972A (en) | 2019-04-08 | 2019-04-08 | Underwater robot Predictive Control System and method based on on-line study modelling technique |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109946972A true CN109946972A (en) | 2019-06-28 |
Family
ID=67013891
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910276540.5A Pending CN109946972A (en) | 2019-04-08 | 2019-04-08 | Underwater robot Predictive Control System and method based on on-line study modelling technique |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109946972A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110646783A (en) * | 2019-09-30 | 2020-01-03 | 哈尔滨工程大学 | Underwater beacon positioning method of underwater vehicle |
| CN110647042A (en) * | 2019-11-11 | 2020-01-03 | 中国人民解放军国防科技大学 | Robot robust learning prediction control method based on data driving |
| CN113132905A (en) * | 2021-04-21 | 2021-07-16 | 湖南国天电子科技有限公司 | Autonomous underwater robot wireless sensor network with dynamic nodes |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1718378A (en) * | 2005-06-24 | 2006-01-11 | 哈尔滨工程大学 | S face control method of flotation under water robot motion |
| CN102385316A (en) * | 2011-09-16 | 2012-03-21 | 哈尔滨工程大学 | Deepening controlling method of underactuated automatic underwater vehicle based on neural network back stepping method |
| CN109343350A (en) * | 2018-11-20 | 2019-02-15 | 清华大学 | A kind of underwater robot path tracking control method based on Model Predictive Control |
-
2019
- 2019-04-08 CN CN201910276540.5A patent/CN109946972A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1718378A (en) * | 2005-06-24 | 2006-01-11 | 哈尔滨工程大学 | S face control method of flotation under water robot motion |
| CN102385316A (en) * | 2011-09-16 | 2012-03-21 | 哈尔滨工程大学 | Deepening controlling method of underactuated automatic underwater vehicle based on neural network back stepping method |
| CN109343350A (en) * | 2018-11-20 | 2019-02-15 | 清华大学 | A kind of underwater robot path tracking control method based on Model Predictive Control |
Non-Patent Citations (2)
| Title |
|---|
| 何斌: "DSRV运动控制中参数自寻优典型方法研究", 《中国优秀硕士学位论文全文数据库》 * |
| 孙玉山等: "改进的模拟退火算法在水下机器人S面运动控制参数优化中的应用", 《兵工学报》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110646783A (en) * | 2019-09-30 | 2020-01-03 | 哈尔滨工程大学 | Underwater beacon positioning method of underwater vehicle |
| CN110646783B (en) * | 2019-09-30 | 2021-07-09 | 哈尔滨工程大学 | Underwater beacon positioning method of underwater vehicle |
| CN110647042A (en) * | 2019-11-11 | 2020-01-03 | 中国人民解放军国防科技大学 | Robot robust learning prediction control method based on data driving |
| CN110647042B (en) * | 2019-11-11 | 2022-04-26 | 中国人民解放军国防科技大学 | Robot robust learning prediction control method based on data driving |
| CN113132905A (en) * | 2021-04-21 | 2021-07-16 | 湖南国天电子科技有限公司 | Autonomous underwater robot wireless sensor network with dynamic nodes |
| CN113132905B (en) * | 2021-04-21 | 2022-05-31 | 湖南国天电子科技有限公司 | Autonomous underwater robot wireless sensor network with dynamic nodes |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109901403A (en) | A kind of face autonomous underwater robot neural network S control method | |
| Mohajerin et al. | Multistep prediction of dynamic systems with recurrent neural networks | |
| Zhong et al. | Optimal control for unknown discrete-time nonlinear Markov jump systems using adaptive dynamic programming | |
| Berenji et al. | A hierarchical approach to designing approximate reasoning-based controllers for dynamic physical systems | |
| Lin | Adaptive critic autopilot design of bank-to-turn missiles using fuzzy basis function networks | |
| CN112904728B (en) | Mechanical arm sliding mode control track tracking method based on improved approach law | |
| CN110238839B (en) | Multi-shaft-hole assembly control method for optimizing non-model robot by utilizing environment prediction | |
| CN103592849B (en) | Ship dynamic positioning control method | |
| CN109946972A (en) | Underwater robot Predictive Control System and method based on on-line study modelling technique | |
| CN111176122B (en) | Underwater robot parameter self-adaptive backstepping control method based on double BP neural network Q learning technology | |
| CN113052372B (en) | Dynamic AUV tracking path planning method based on deep reinforcement learning | |
| Fernandez-Gauna et al. | Experiments of conditioned reinforcement learning in continuous space control tasks | |
| Meng et al. | Intelligent control algorithm for ship dynamic positioning | |
| CN113093526A (en) | Overshoot-free PID controller parameter setting method based on reinforcement learning | |
| CN109901622A (en) | A kind of autonomous underwater robot prediction face S control method based on mechanism model | |
| Wang et al. | Adaptive DDPG design-based sliding-mode control for autonomous underwater vehicles at different speeds | |
| CN106569503A (en) | Triggering-type predictive control-based aircraft control method | |
| Chen et al. | Neuroadaptive tracking control of affine nonlinear systems using echo state networks embedded with multiclustered structure and intrinsic plasticity | |
| CN112230545A (en) | AUV course angle control method based on PPGA adaptive optimization PID parameter | |
| Yao et al. | Research and comparison of automatic control algorithm for unmanned ship | |
| Vianna et al. | Neural Network Based Model Predictive Control for an Autonomous Vehicle | |
| Zhao et al. | Consciousness neural network for path tracking control of floating objects at sea | |
| CN110703792B (en) | Underwater robot attitude control method based on reinforcement learning | |
| Cao et al. | A realtime Q-Learning method for unmanned surface vehicle target tracking | |
| Wang et al. | Course tracking control for smart ships based on a deep deterministic policy gradient-based algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |