TW200934197A - Real-time control system of dynamic petri recurrent-fuzzy-neural-network and its method - Google Patents

Abstract

This invention focused on the reveal of real-time control system of dynamic Petri recurrent-fuzzy-neural-network (DPRFNN) and its method. It is composed of a DPRFNN, a controlled plant, an on-line parameter tuning module, and a learning-rate update module. In the DPRFNN, the concept of a Petri net (PN) and the recurrent frame of internal feedback loops are incorporated into a traditional fuzzy neural network (FNN) to alleviate the computation burden of parameter learning and to enhance the dynamic mapping of network ability. Moreover, the supervised gradient descent method is used to develop the on-line training algorithm for the DPRFNN control. In order to guarantee the convergence of system control errors, analytical methods based on a discrete-type Lyapunov function are proposed to determine varied learning rates for DPRFNN.

Description

200934197 214 Weight value between network output layer and network rule layer 215 Network output signal 8. If there is a chemical formula in this case, please disclose the chemical formula that best shows the characteristics of the invention: Nine, invention description: [Technical field of invention] The technical field involved in the "dynamic dispatching fuzzy back-type fuzzy neural network real-time control system and its method" mainly includes fuzzy control, neural network, nonlinear control and intelligent control; according to the above technology, To develop an instant control system and its method, the real-time control system takes the fuzzy neural network as the core, joins the network pie layer and the recursive structure, and forms a dynamic dispatched fuzzy neural network. The online parameter adjustment module and the learning rate update module enable the controlled body to achieve the desired control. [Prior Art] In the field of traditional control, the accuracy of the dynamic mode of the control system is the most important factor affecting the performance of the control. The more detailed the information of the system dynamics, the more precise control can be achieved; however, for nonlinear systems, It is often difficult to correctly describe the dynamics of the system, so research scholars can easily use various methods to simplify the system dynamics to achieve control purposes, but not ideal. In other words, traditional control theory has strong and powerful control ability for clear systems, but it is powerless for systems that are too complicated or difficult to describe accurately. Therefore, many researchers try to deal with such control problems with fuzzy mathematics 5 200934197. . Since the development of fuzzy mathematics by Zadeh scholars, it has greatly contributed to the control of ambiguous systems. Since the Seventies, there have been some practical fuzzy controllers. The fuzzy control is robust. There is no need for precise mathematical models, strong approximation capabilities, and the use of human experience to establish fuzzy rules... and so on [1]. Reference [2] proposes to use the fuzzy map and the new map measurement in the _ self-propelled vehicle. References m issued a human body silk paste controller ❹ 1 # · ^ self-propelled vehicle device ' and through the Yarapno (LyapunGV) stability theory to prove its pride. Reference [4] is designed as an instant fuzzy control architecture and uses an infrared sensor to enable the self-propelled vehicle to achieve the desired function. References [Learning the use of fuzzy logic. It is able to follow the behavior of human thinking, enabling the self-propelled car to follow a specific path. Although these techniques can be used to construct control systems in the same way as human behavior, it is quite difficult to achieve appropriate fuzzy rules to achieve good control performance. The Neural Network (NN) is a model that mimics the development of human brain activity. As far as the network architecture is concerned, it consists of many simple and interconnected Processing Elements, ie Neurons (Neur〇ns); in terms of network functions, it is a new state produced by biological models. Information processing and calculation methods. In recent years, many people have studied the use of neural networks in system identification or dynamic system control [6]-[8]. Reference [7] uses a neural network-like approach to study how to avoid obstacles during self-propelled vehicle tracking. Reference [8] develops a simple neural network to conceal self-propelled vehicles without the need to use speed information for self-propelled vehicles. Although the 6 200934197 neural network has a strong function approximation capability, the network parameter values usually have to undergo long-term offline (Off-Line) training to achieve good control, since the initial network parameter values have not been trained. As a result, the network parameters do not have corresponding values, so the initial control performance is generally poor. In recent years, reference [9]_[11] proposes that the Recurrent Neural Network (RNN) achieves a fast correspondence capability. Its performance is significantly superior to that of a neural network, but due to the neural network. It is composed of neurons, so in a more complex system, it is easy to cause the problem of too large a network. Nowadays, combining fuzzy control and neural networks has become a very popular research topic [12]-[14]. Fuzzy Neural Network (FNN) can combine the advantages of fuzzy control and neural network to achieve good performance. The network architecture is simpler than the neural network, and the corresponding fuzzy rules can be used. The theory of neural network learning is obtained. In addition, references [15], [16] propose a recurrent Fuzzy Neural Network (RFNN) architecture, because the recursive structure can speed up the network's corresponding ability, generally speaking The control performance of the back-fuzzy neural network is superior to that of the fuzzy neural network. On the other hand, after the introduction of Petri Net (PN), many people began to study in different fields [Π], [18], and reference [19] proposed the Paifu fuzzy neural network (Petri). The architecture of the Fuzzy Neural Network (PFNN) is applied to linear induction motors. The main concept is to add the mechanism of the Pai Cui operation to the fuzzy neural network to achieve the effect of reducing the computational complexity. Back to the structure, so the control performance is slightly inferior to the recursive 7 200934197 fuzzy class neural network. Remarks: References [1] L. X. Wang, A Course in Fuzzy Systems and Control. New Jersey: Prentice-Hall, 1997.

[2] H. Seraji and A. Howard, ^Behavior-based robot navigation on challenging terrain: a fuzzy logic approach/* IEEE Trans. Robot. Automat., vol. 18, no. 3, pp. 308-321, 2002 .

[3] SX Yang, H. Li, MQH Meng, and PX Liu, ^An embedded fuzzy controller for a behavior-based mobile robot with guaranteed performance, 5, IEEE Trans. Fuzzy Syst., vol. 12, no. 4, Pp. 436-446, 2004.

[4] T. H. S. Li, S. J. Chang, and W. Tong, uFuzzy target tracking control of autonomous mobile robots by using infrared sensors, IEEE Trans. Fuzzy Syst., vol. 12, no. 4, pp. 491-501, 2004.

[5] G. Antonelli, S. Chiaverini, and G. Fusco, t$A fuzzy-logic-based approach for mobile robot path tracking, 5, IEEE Trans. Fuzzy Syst., vol. 15, no. 2, pp. 211-221, 2007.

[6] O. Omidvar, and D. L. Elliott, Neural Systems for Control. Academic Press, 〇 1997.

[7] SX Yang and MQH Meng, ^Real-time collision-free motion planning of a mobile robot using a neural dynamics-based approach,” / 5th Five-Year NeuralNetw., vol. 14, no. 6, pp. 1541- 1552, 2003.

[8] T. Das, IN Kar, and S. Chaudhury, "Simple neuron-based adaptive controller for a nonholonomic mobile robot including actuator dynamics,5, Neurocomputing, vol. 69, no. 16-18, pp. 2140- 2151, 2006.

[9] FJ Lin, RJ Wai, WD Chou and SP Hsu, ^Adaptive backstepping control using recurrent neural network for alternating induction motor drive,55 IEEE Trans. Ind. Electron., vol. 49, no. 1, pp. 134- 146,2002. 8 200934197 [10] RJ Wai, C. M. Lin and Υ· F. Peng, “Adaptive hybrid control for linear piezoelectric ceramic motor drive using diagonal recurrent CMAC network," IEEE Trans. Neural Netw., vol . 15, no. 6, pp. 1491-1506, 2004.

[11] SJ Yoo, YH Choi, and JB Park, ^Generalized predictive control based on self-recurrent wavelet neural network for stable path tracking of mobile robots: adaptive learning rates approach, 5* IEEE Trans. Circuit Syst., vol. 53 , no. 6, pp. 1381-1394, 2006.

[12] C. T. Lin, and C. S. George Lee, Neural Fuzzy Systems, New Jersey: Prentice-Hall, 1996.

[13] FJ Lin, RJ Wai, and CC Lee, MFuzzy neural network position controller for ultrasonic motor drive using push-pull DC-DC converter, M IEE Proc. Control Theory Appl., vol. 146, no. 1, pp. 99-107, 1999.

[14] C. Ye, NHC Yung, and D. Wang, MA fuzzy controller with supervised learning assisted reinforcement learning algorithm for obstacle avoidance/* IEEE Trans. Syst., Man, Cybern. B, vol. 33, no. Pp. 17-27, 2003.

[15] FJ Lin, RJ Wai, KK Shyu and TM Liu, ^Recurrent fuzzy neural network control for piezoelectric ceramic linear ultrasonic motor drive, M IEEE Trans. Ultrason., Ferroelect., Freq. Cont., vol. 48, no. 4, pp. 900-914, 2001.

[16] R. J. Wai, “Total sliding-mode controller for PM synchronous servo motor drive using recurrent fuzzy neural network/* IEEE Trans. Ind. Electron. t vol. 48, no. 5, pp. 926-944, 2001.

[17] R. David and H. Alla, “Petri nets for modeling of dynamic systems-A survey, 5, Automatica, vol. 30, no. 2, pp. 175-202, 1994.

[18] V. R. L. Shen, ^Reinforcement learning for high-level fuzzy petri nets, 55 IEEE Trans. Syst., Man, Cybern. B, vol. 33, no. 2, pp. 351-362, 2003.

[19] RJ Wai and CC Chu, ^Motion control of analog induction motor via petri fuzzy-neural-network, IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 281-295, 2007. 200934197 SUMMARY OF THE INVENTION The block diagram of the dynamic dispatching fuzzy neural network real-time control system is shown in the first figure, which includes 100 dynamic dispatched fuzzy neural network, 101 controlled body, and 102 online parameter adjustment. Module and 103 learning rate update module, the input of the 100 dynamic dispatching fuzzy-type neural network is the error signal and its differentiation, and the control signal output is obtained through the network operation, and the 100 dynamic dispatching fuzzy type The number of inputs and outputs of the neural network is determined by the state of the controlled body and the control signal. Taking the first picture as an example, the input is e, and the sum of ~= and its differential, where X and less are the controlled body state and ^ is the corresponding control command, and the output control signals are 〇1 and 〇2. The 101 controlled body is any device to be controlled, usually input as a control signal, and the output is the state of the controlled body measured by the sensor, and the controlled body state and the corresponding control should be controlled. The subtraction becomes an error signal, and the error signal and its differential are transmitted to the 100 dynamic dispatched fuzzy neural network to form an instant control system; the 102 online parameter adjustment module is based on the learning rate (%, Ά, %) , error signal and its differential © (έ, Α) to adjust the variation of the dynamic dynamic recursive fuzzy neural network parameters (△ <, Δ^_, Δ5 /, Δα /); the learning rate of 103 The update module is based on the Discrete-type Lyapunov stability theory based on the 100 dynamic dispatched fuzzy neural network parameters («, «), error signal () and its differential (4)'. The updated value of the learning rate. The dynamic dispatching fuzzy-like neural network 100 dynamic dispatched fuzzy-like neural network architecture as shown in the "second map" total 200934197 is divided into five layers, including 201 network input layer, 202 network attribution function layer, 208 network distribution layer, 210 network rule layer and 213 network output layer, in addition to the 202 network attribution function layer, 204 network recursive structure is added, and the signal transmission flow table of each layer is as follows The first layer is the 201 network input layer, which transmits 200 network input signals; ¢, . (/ = 1,..., ", ·) directly to the 202 network attribution function layer. The second layer is the 202 network attribution function layer, and the input of each 203 network attribution function neuron is the last 206 network attribution function layer output multiplied by 205 network recursive weight value, and this time plus 200 network input signal, can be expressed as (1) r / (8) = a («) + 〆 (π - ΐχ (1) where η represents discrete time, α 丨 represents 205 network recursive structure weight value, 〆 represents The last 206 network attribution function layer output, and ζ-1 is 207 time delay unit. In general, there are many kinds of custom attribution functions, including triangle function, trapezoidal function, inverted clock function, Gaussian function, Etc., the attribution function used in the present invention is a Gaussian function, which can be expressed as (2)® (10)^/[«^.(rf)] = expM.(r/)] (2) where exp[.] represents an index Function (Exponential Function), < and 彳〇· = 1,..., ",;··· = 1,.··, ') respectively represent the center point and width of the Gaussian function, ~· for each The number of statement variables for a 200 network input signal. 206 The third layer is the 208 network dispatch layer, in which the network attribution function layer output is transmitted to the fourth layer according to the formula (3). "41, ♦(10) Team* [ο, μ1[ηβί] {τΙ)]<άΛ (3) 200934197 The center represents the preset threshold and ί/ is the 209 transfer valve. The fourth layer is the 210 network rule layer, and the 211 network rule layer output is the product of the corresponding output of the 208 network Pai layer, which can be expressed as (4) (4) ΥΙ^βMlinetjiri)], tj =\ i=\ 0 , ί/=0 where (A: = 1,···, ') represents the 211 network rule layer output; < is the weight value between the 212 network layer and the network rule layer, Set to 1 for 210 neural network in the rule layer

The total number of yuan. The fifth layer is the 213 network output layer. The sum of the 215 network output signals in this layer is the product of the 211 network rule layer output and the weight value between the 212 network layer and the network rule layer. (5) Formula y〇=zl<(^k (5) where w丨 represents the weight value between the 214 network output layer and the network rule layer; where 0 = 1,···, ".) represents the 215 network Output signal, ". The total number of neural units in the 213 network output layer. Second, the online parameter adjustment module 102 online parameter adjustment module, mainly by the Supervised Gradient Descent Method (Supervised Gradient Descent Method) to obtain the variation of the parameter value in the 100 dynamic dispatched fuzzy neural network, the 102 The amount of change in all parameter values of the online parameter adjustment module is generated by partial differentiation of the desired parameter by a specific energy function. The partial differential calculation uses the chain rate. Since the order of partial differential is transmitted back to the internal nodes by the network's 12 200934197 output, it is also called the inverse transfer algorithm. In order to derive the variation of the parameter values of the 100 dynamic dispatching fuzzy neural network, the energy function is defined as follows: E = —[(3⁄4 - ^)2 + {yd - yf + (xd - χ)2 + (yd - ^)2] : (6) = all (e»X) where & is the control command, 弋 and 九 are their differentiation; λ: and j are the controlled body states, and the sum is less differential; 4 =4-λ: and represent the error signal, and < is its Φ differential. In this system, the input of the 100 dynamic dispatched fuzzy-like neural network is A, &, and the sum 4 (ie, ",. = 4); 100 dynamic dispatched fuzzy-like neural network The output is 〇2 (also known as ".= 2"). The inverse transfer algorithm of the 10 2 online parameter adjustment module is described as follows: . 213 In the network output layer, the energy function can be expressed as a partial differential of the 215 network output signal.

dE ❹ dE dex dx dE dey dy dE dex dx dE dey dy dex dx dyQ + dey dy dy0 + dex dx dy0 + dey dy dy0 (7) ^y〇y^y〇x^y0 ydy0 then 214 network output layer and network The amount of change in the weight value between the rule layers can be expressed as

λ 〇 dE

dE (8) where % is the learning rate of the weight value between the 214 network output layer and the network rule layer. 214 The weight value update between the network output layer and the network rule layer can be expressed as (9) < 0 +1)= < (8) + △< (8) 13 200934197 The energy function pairs 211 network rule layer output partial differential dE ^y〇

Here w double to · nei / A) 倨 micro-division dE _dE do 0 (net ! (^)) λ dnetjir ^) L do. Ice _ {dMf(netj(rJ:)) dnet.^) , '/ = 1 t 208 Network Pai Cui (11) 0 , tj =〇© Then m/, 彳 and α/ can be expressed as dE dy r dnetj{rjt)^ dy0 dnetj(rJ.) l > (12) Δ^/=-^^Γ = ^ dmj dE - Department. "I dnet.(rJi)^ lsds{ Vs dy0dnet.{rj.) l dsI > ^IsPj 2{rJ.-mi)2 ~W~ (13)

dE sign dyn Ί r Οηβί^Ϋ nadal % dy0 dnet.(rJ.) l dai J (14) €> where I, I and heart represent the heart point and width of the Gaussian function and the 2〇5 network (15 (16) (17) Regressive learning rate of structural weight values. The center point and width of the Gaussian function and the updated value of the 205 network recursive structure weight value can be expressed as mj (η +1) = mj («) + Am/ (η) sj (η +1) = sj (η) + Asj (η) aj (η +1) = aj (ή) + Δα/ («) Since the general system contains uncertain factors, the system sensitivity is changed/doing, 200934197/yes, δλ/ also and office/office. It is not easy to decide, and the Intelligent Identifier can be used to estimate the sensitivity of the system, but it requires a lot of calculations. To overcome this problem and increase the rate of online learning, the sensitivity of the system of the present invention is approximated by a sign function as follows: dx Δχ, / = sgn ^ - Sgn ^y〇Ay〇) dy τ—= sgn (Αν, r = sgn V ^ Y〇dx i test"Δχ" / = sgn <Ay〇) ί ΔνΝ r —= sgn do. = SgQ V where sgn(·) represents the symbol function jc(n) - x{n -1) y(n)-y{n-\) , y〇{n)-y0{n-\)) (18) ❹ χ{ή) - x(n -1) where (")-fan ("-i) y{n)-y{n-\) y〇{n)-y0{n~^). The rate update module selects different parameter learning rates to have a significant impact on the performance of the network. 'In order to be able to effectively train the parameters of the 100 dynamic dispatched fuzzy-like neural network ο value', the present invention utilizes the discrete Rialp The Discrete-type Lyapunov stability theory obtains the updated value of the learning rate so that the error signal converges. According to the formula (6), the discrete variation of the Discrete-type Lyapunov function is ^(n) ^E(n^\)-E(n) Therefore, the energy function can be expressed by (8) and (12)-(14) as £(« + 1) = £(η) + Δ£(η) 15 200934197 dwk dE(n) ❹ /=1 j^\ 〇mi U,i <E(n) + E(n) + E(n) E(n) 1 4 E(n) 1 nm 4 Five (8) 1 Ns 4 five (8) 1 % 4 E(n) ti tfl /=1 j-\ o=l ί±ί y=i dEjn) dya \ ^y〇daj \2 Δα/] 8E(n) dxf dy0 dE(n ) dx( dy0ax,D/ dE(n) dxt dya dx{ dy0 da{ \2 (20) ❹ where, Am/, As/, and Δα/ respectively represent 214 network loss The amount of change in the weight value between the egress layer and the network rule layer, the change in the center point and width of the Gaussian function, and the change in the weight value of the 205 network recursive structure; Η indicates the absolute value. 100 Dynamic Pai recursive The learning rate of the fuzzy neural network is designed to be five (8) % 4 〇 =! k=l dE(n) dya five (8) \2 + € (21) 4 lit /=1 j=\ 〇=1

Vs dE(n) dxf dy0 dx( dy0 dmj m (22) + € 4 dxi

V + ε (23) 16 % = (24) 200934197 Five (8) 4 mi

dE{n) dxt dy0 dx( dy0daj J + ε where s is a constant constant. According to the (21)-(24) learning rate design, let dE{n) dy0 4 Sy〇dw°k % i A,-1Σ - m \2 dE(n) dxt 4 dx( dy0dmj 1 dE(n) dxt dy0 , 2 dE{n) dxt dyc 4 5 (8). 3a/

Va 巧 nJ %-ΣΣΣ \2 <1 (25) Therefore, five (" + 1" < five (") can be obtained, and then five (") > 0 and Δ^ (η) are obtained according to the formula (6) ) < 0, so the error signal will gradually converge. [Embodiment] One embodiment of the "dynamic dispatching fuzzy back-type fuzzy neural network real-time control system and method thereof" is a schematic diagram of a self-propelled vehicle device used in a tracking automatic control system of a self-propelled vehicle. As shown in the "third diagram", two steering wheels and one supporting wheel are disposed on the vehicle body, and the steering wheels are respectively controlled by two independent DC motors and are parallel to the axle. The support wheel is a passive freewheel that can be controlled at any angle with the steering wheel. In the figure, the distance between the two steering wheels is shown, and the diameter of the steering wheel is expressed as 2r; the point C in the figure is the centroid position of the self-propelled car; the point in the figure is the axis of the wheel and the axis of the wheel passing through the C point. At the intersection of the line, the P point indicates that the self-propelled vehicle is in the position of the coordinate system 17 200934197; in the figure, _, V} is the global coordinate system, and the position of the self-propelled vehicle in the global coordinate system can not be expressed as P = [M Vf, They represent the horizontal and vertical axes of the global coordinates; respectively, {P'X, Y} in the figure is the local coordinate system, that is, the coordinate system with the P point as the origin; 0 is the relative coordinate of the global and the local coordinates. Angle, and the starting angle is measured from the 17th axis. Assuming that the wheel of the self-propelled vehicle only rotates and does not produce a side shift, that is, the direction in which the self-propelled vehicle moves is perpendicular to the axle, the action constraint of the self-propelled vehicle can be expressed as T>cos0-!isin0 = 〇μ ( 26) The schematic diagram of the self-propelled vehicle movement is shown in the “fourth figure”. The Q and Q in the middle of the vehicle represent the position of the self-propelled vehicle from the sampling time „, to %+1; if the sum indicates the global coordinate + u-axis and The displacement amount of the v-axis; ΑΘ indicates the rotation amount of the self-propelled vehicle angle; G, r, and & represents the radius of rotation of the Q point moving to the Q' point, that is, the distance from the origin to the left wheel; The distance between the point and the point Q; ◊ is the distance from the origin Ο to the right wheel; and 4, “and 4 is the arc length of the corresponding radius of rotation; ΔΘ can be expressed as ❹ _ At=milk I) Vr +V, r〇=-L~^~bVr'V, (27)(28) Its A Δί represents the sampling time interval, and ^ represents the left wheel speed and the right wheel catching maximum limit setting respectively; The discrete dynamic equation of the self-propelled vehicle in the global coordinate system can be expressed as <ns +1) = U(ns) + ro{sin[^(Ks) + )] - sin^C^)} v(ns + 1) = v(Wi) + r〇{c〇s^(Mi) - cos ^ ) + A0(ns)]} (29) where the meaning represents the sampling time. If the control position of the self-propelled vehicle in the global coordinates is 18 200934197 〆 = [& , the controller must be designed with the global coordinates by the transformation matrix Cos^ sin0 sinΘ -cosθ ·( cos0 sin0 _ «_ ud _yd. sin0 -cos0 - ·. Λ. T-1 is converted into local coordinate system / whistle w, ie (30) ❹ Ο 'The main one in this system' The purpose of the control is to find a suitable control signal so that the self-propelled vehicle can achieve the instantaneous trajectory tracking control. In order to achieve this control target, the tracking error signals defining the X-axis and the 轴-axis are respectively ten, and the #local coordinate system is the lining. Weaving secrets __ Heshe wheel speed (8) and right wheel speed (〇 to manipulate the self-propelled car to achieve the purpose of path tracking control. In order to enable the self-propelled car to achieve the tracking effect in different reference paths (control orders) t> » Juice 501 self-propelled car dynamic Pai Cui recursive fuzzy neural network, which combines 2〇8 network sent green layer to receive 2〇4 clear recursive structure in the biography (10) __ via the network, its instant The control system block diagram is shown in the "fifth figure", which contains the lion coordinate conversion. 501 self-propelled car dynamic dispatched back-type fuzzy neural network, 5〇2 self-propelled vehicle 503 self-propelled car line parameter adjustment module and 5〇4 self-propelled car learning rate update module. The 500 coordinate converter Turn the global coordinates into local coordinates ·, the training from the dynamic dynamics of the Cui recursive model _ God, _ road using the invention (9) dynamic Pai Cui recursive fuzzy neural network, according to chasing (four) poor job And its differential (U) finds the appropriate left-wheel speed (8) and right-wheel speed (8) to operate the self-propelled vehicle; the 5〇2 self-propelled vehicle moves the self-propelled vehicle according to the left and right wheel speeds, and uses the global coordinates and The car's position indicates the attitude of the self-propelled car; the 5〇3 self-propelled car line parameter adjustment module adopts 200934197 with the 102 online parameter adjustment module of the present invention, according to the learning rate (heart, t, %, %), tracking error The signal (ϋ) and its differential (1 stone) are used to adjust the variation of the parameters of the 501 self-propelled vehicle dynamic dispatching fuzzy-type neural network (△河,^/,碎;,么〇; the 504 self-propelled car The learning rate update module adopts the 1〇3 learning rate update module of the present invention, according to the 501 self-propelled car dynamics The Cui recursive fuzzy neural network parameters (river, calendar/, 5/, 〇, tracking error signal (3⁄4^) and its differential (3⁄43⁄4) are used to obtain the updated value of the learning rate. In order to verify the instant control performance of the present invention, a conventional network frame control system is also proposed to compare its performance. The parameters of the self-propelled device can be expressed as r = 0.0925m; 6 = 0.167m; vmax = 0.4 m/s (31) In order to show that the 501 self-propelled vehicle dynamic dispatched fuzzy neural network has superior performance, compare three different network structures, including fuzzy neural network-(FNN), Recursive fuzzy neural network (RFNN) and Paifu fuzzy neural network (PFNN), and use the same 503 self-propelled vehicle line parameter adjustment module, 504 self-propelled vehicle learning rate update module, wheeled signal And the output signal. The input signals are Tracking ® Error Signal (€, 5) and its derivative (忑, Jy); the output signals are the left wheel speed (V,) and the right wheel speed (vr). The initial value of the network parameter is the previously trained value. The previously trained parameter value adopts the previously described 5〇3 self-propelled vehicle line parameter adjustment module and the 504 self-propelled vehicle learning rate update module' when it is satisfied. Control the value of the performance, and then set the parameter value of this time to the initial value of the next execution. The carcass used in this experiment was visual C++, written on a personal computer of the Pentium IV; the model of the self-propelled car was Pi〇neer, manufactured by MobileRobots. 200934197 Development board is Hitachi H8S; frequency 44·2368ΜΗζ; 32bitRISC; 32k RAM; 128k FLASH; self-propelled car and computer connection using wireless network transmission module; wheels are controlled by 12V DC motor using PWM signal; each motor A 128count/mm sensor is used for position feedback. In this experiment, two reference paths (control commands) are selected to test the performance of the control system. One is an 8-shaped trajectory whose expression ❹ is as follows: cos( cos( 2πί π Τ 2 , π 2πί '2 Τ 2πί π Τ 2 Μ 2πί ), 0 < ί < 40 ), 40 < ί < 80 ), 80 < ί < 120 2 Τ yd 1 + sin ( Τ 1 2) {π 2πίΛ 〔2 Τ (2πί Τ 2) , π 2πί 2 Τ (32) The other is a square trajectory whose expression is as follows: ❹ xd O.lt 1 + cos[ 2 1 + cos[ 2π( ί-\0) Τ 2 2_-30) Τ f] 1- 0.10-70) 2π(卜50) -l + cos[ -2 -l + cos[ T 2 2π{ί + \0) π. Τ 2· , 0 < ί <10 , 10 </<30 , 30<ί <50,50<ί <70,70<ί<90,90<ί<110,110<ί<130,130<ί<150,150<ί<160 21 200934197 yd 1 + Sin[

2π(ί-10) ~T f] 1 + 0.10-30) 2π(ί-30) π] 3 + sin[ 4

T 3 + sin[2M 3-0.10-110) 2. π. ~ΐ' , 0 < ί <10 , 10 < / < 30 , 30 < / < 50 , 50 < ί < 70 , 70 <;ί<90,90<ί<110,110<ί<130 (33) l + sin[2;r^ + 1Q)--] ,130<ί<150 Τ 2

0,150</<160 The initial position and angle of the self-propelled vehicle are preset to zero, and the control parameters of the system are expressed as follows: dth=0.l; ^=0.05; Δ/ = 0.05s (34) (34) The control parameters in the equation are a set of parameter values that take into account the preferred performance selected for the possible operating environment. In order to be able to compare the performance of each network structure control system, the mean error value (MSE) is defined here as MSE=4t>K2(«) + e» (35) where Τ' represents the sum of the sampling times. This experiment firstly produces different instantaneous control performances for fuzzy neural networks (FNN) with different numbers of neurons in the attribution function layer. The average error value (MSE) and execution time are shown in the sixth graph. According to the results of the sixth graph, the neurons are the best at seven times. In other words, the network is %=4, '=7, ~=7x7x7x7 = 2401 and =2, compared to other networks. The size is more suitable for self-propelled vehicle path tracking, so the next experiment based on this network, the control system adjustment of the reference 22 200934197 total number of total 2 χ Χ Χ + + + + + + + + + + + + + + + + + + The number of neural network parameter adjustments is based on the preset threshold 4 of equation (7), so the required memory and execution speed are less than that of the fuzzy neural network control system, although more is needed. The number of rules to enhance the tracking response does not create a large amount of computation, which may cause the computer or microprocessor to crash.

The experimental results of all control systems are shown in Figure 7 to Figure 14, including 8-shaped and square path tracking, where Figures 7, 9, 10, and 13 are 8-shaped tracking responses KA, 10, and 10 Two and fourteen are square tracking responses. In each figure, (a) and (ie, represent the X-axis and the tracking response; (4) and (4) represent the ship error of the X-axis and the Y-axis, respectively; (e) and (7) represent the left-wheel speed and the right-wheel speed, respectively; (8) The ritual tracking response is the path tracking error. The experimental results show that the biggest factor affecting the tracking response and robust performance of the immediate control system is the learning ability of the control system itself, and the execution time is the architecture of the network. The average error value (MSE) and execution time of all network structure control systems, the control performance of the recursive fuzzy neural network (RFNN) is significantly better than that of the fuzzy neural network (FNN) because of the recursive fuzzy class. The recursive structure of the neural network (RFNN) increases the corresponding power of the network; although the average error value (MSE) of the Paifu fuzzy-like neural network (PFNN) is slightly larger than that of the recursive fuzzy-like nerve (RFNN), The execution time is much less, because there is a preset threshold value in the Paifu fuzzy neural network (PFNN), so that the too low network attribution function layer rotation and the corresponding change amount do not perform the operation; Three and fourteen with figure By the 12th, 501 self-propelled car dynamics, the average error value (MSE) of the 200934197 is smaller than that of the recursive fuzzy neural network (RFNN), and its execution time is close to that of the pie. The fuzzy fuzzy neural network (PFNN), the result shows that the performance of the 5〇1 self-propelled vehicle dynamic regenerative fuzzy neural network is indeed superior to other network structure control systems. The successful combination of this embodiment The 208 network sent the layer and the 204 network recursively structure the traditional fuzzy-like neural network, and showed the real-time control performance of the 501 self-propelled vehicle dynamic dispatching fuzzy-like neural network on the path tracking. Dynamic Pai Cui recursive © fuzzy neural network can simplify the calculation of the number of rules and increase the corresponding ability of the network. From the experimental results, it is known that the 5〇1 self-propelled vehicle dynamics are sent back to the fuzzy neural network. The performance is 23 42% higher than the fuzzy neural network (FNN): its execution time is also reduced by 95.65%. The main advantages of the "dynamic dispatching back-type chess-like neural network real-time control system and its method" of the present invention are described. As follows: ❿ 1. This month’s dynamic dispatch Fuzzy-like neural network real-time control system and its method" proposes a new dynamic-styled-style fuzzy-like neural network structure, and joins the online parameter Tiaolong group to reach _Lu Qianzhi learning, money (10) learning rate update module makes The error is reduced (4), so the present invention is quite new. 2. The dynamic structure of the dynamic control system and method of the present-day dynamic dispatching chess-like chess-like neural network is proposed. On the device, the experimental results are also available! The raw I is superior to the previous network structure control performance, so the present invention is significantly more advanced than the conventional technology. 24 200934197 3. The present invention "dynamic dispatching The fuzzy neural network real-time control system and its method" can be applied to any 101 controlled body, and can also add 102 online parameter adjustment module and 103 learning rate update module to achieve the required immediate control. Performance, because the invention is quite industrially useful. The present invention has been disclosed in the foregoing preferred embodiments, and is not intended to limit the scope of the invention, and the invention may be modified and modified without departing from the spirit and scope of the invention. The scope is subject to the definition of the scope of the application for the benefit of the application. BRIEF DESCRIPTION OF THE DRAWINGS The first figure shows the block diagram of the dynamic dispatching fuzzy neural network real-time control system of the present invention. The second figure shows the architecture diagram of the dynamic dispatching fuzzy-like neural network of the present invention. The three diagrams show the schematic diagram of the self-propelled vehicle. The fourth diagram shows the movement of the self-propelled vehicle. The fifth diagram shows the block diagram of the self-propelled vehicle. The sixth diagram shows the average error value and execution time of the fuzzy neural network control system. The seven graphs represent the experimental results of the 8-word path tracking fuzzy neural network control system. The eighth graph shows the square path tracking fuzzy neural network control system. 25 200934197 The results of the ninth figure shows the 8-shaped path. Experimental results of the tracking and returning fuzzy-like neural network control system The tenth figure shows the experimental results of the square path tracking and returning fuzzy-like neural network control system. The eleventh figure shows the 8-shaped path tracking. Experimental results of the fuzzy neural network control system 〇 The twelfth figure shows the square path tracking and the acquisition of the fuzzy fuzzy neural network control system. Results The thirteenth figure shows the experimental results of the 8-shaped path tracking mining dynamic dispatching fuzzy fuzzy neural network control system. The fourteenth figure shows the square path tracking mining dynamic dispatching fuzzy fuzzy neural network control The fifteenth graph of the experimental results of the system shows the average error value of the fuzzy neural network, the recursive fuzzy neural network, the sentimental fuzzy neural network and the dynamic dispatched fuzzy neural network control system. And execution time [main component symbol description] 100 dynamic dispatched fuzzy back-type neural network 101 controlled body 102 online parameter adjustment module 103 learning rate update module 26 200934197

200 network input signal 201 network input layer 202 network attribution function layer 203 network attribution function neuron 204 network recursive structure 205 network recursive structure weight value 206 network attribution function layer output 207 time delay unit 208 network Road Pai Cui Layer 209 Transfer Valve 210 Network Rule Layer 211 Network Rule Layer Output 212 Weight between Network Pai Layer and Network Rule Layer 213 Network Output Layer 214 Between Network Output Layer and Network Rule Layer Weight value 215 Network output signal 500 Coordinate converter 501 Self-propelled vehicle dynamic dispatched fuzzy-type neural network 502 Self-propelled vehicle device 503 Self-propelled vehicle line parameter adjustment module 504 Self-propelled vehicle learning rate update module 27

Claims (1)

200934197 X. Application for Patent Park: 1. A dynamic dispatching fuzzy-like neural network real-time control system and method thereof, comprising: a dynamic dispatching fuzzy-like neural network; a controlled body An on-line parameter adjustment module; and a learning rate update module; Φ The dynamic dispatched fuzzy back-type neural network input is an error signal and its differential, and the output is a control signal; the controlled body is controlled as desired The device usually inputs the control signal as a control signal, and the output is the state of the controlled body measured by the sensor, and the controlled body state is subtracted from the corresponding control command to become an error signal, and the error signal is obtained. The differential transmission is transmitted to the dynamic Pai Cui recursive fuzzy neural network to form an instant control system. The online parameter adjustment module adjusts the dynamic dispatching fuzzy neural network based on the learning rate, error signal and its differential. The change of the road parameter; the Q learning rate update module is based on the dynamic dispatched fuzzy neural network parameters, the error signal and its differential, and the discrete type of Rialpno (Discrete-typ) e Lyapunov) The stability theory finds the updated value of the learning rate so that the error signal converges. 2. The dynamic dispatching fuzzy-like neural network real-time control system and method thereof according to the first application of the patent scope, the dynamic dispatching fuzzy-type neural network comprises a network input layer and a network a attribution function layer, a network dispatch layer, a network rule layer and a network output layer; the network input layer directly transmits the network input signal to the network attribution function layer; the network attribution function layer operation The network input layer sends a signal of 28 200934197, and the result is transmitted to the network Pai layer through a specific attribution function operation; the network sends a valve mechanism to determine whether to transmit data to the network rule. The network rule layer output is the product of the output of the corresponding network Pai layer; the output layer of the network output is a network output signal, and each network output signal is the network rule layer output and the network All sums of the product of the weight value between the road and the network rule layer. 3. The dynamic Pai Cui recursive fuzzy neural network of the dynamic dispatching fuzzy neural network real-time control system and method thereof according to the second paragraph of the patent application scope, the network attribution function layer includes a network The path attribution function neuron and a network recursive structure; the input of the network attribution function neuron is the last network attribution function layer output multiplied by the network recursive structure weight value, and the current network input signal is added. The result is output through a specific attribution function; the network recursive structure can use the network attribution function layer output as the information of the next operation in a time delay manner to achieve the capability of quickly and dynamically corresponding to the network. © 4. Dynamic Dispatch-based fuzzy neural network real-time control system and its method, as described in the second paragraph of Patent Application No. 2, the dynamic dispatching fuzzy-like neural network architecture, the network is a The mechanism of the transfer valve, the mechanism can determine whether the output value of the network attribution function layer is higher than a preset threshold, prohibiting the output of the network attribute layer that is too low, and the output of the failed network attribution function layer does not need to be adjusted. The parameters of the attribution function layer can effectively reduce the amount of network operations. 5. For example, the dynamic dispatching fuzzy-type neural network of the first paragraph of the patent application scope 29 200934197 control system and its method, the online parameter adjustment module is mainly based on the Supervised Gradient Descent Method (Supervised Gradient Descent Method) The variation of the parameter values in the dynamic dispatched fuzzy neural network is obtained. The variation of all parameter values of the parameter adjustment module on the line is generated by a specific energy function to differentiate the parameters to be adjusted, and the dynamics are changed. The parameters of the Pai Cui recursive fuzzy neural network to achieve the learning function, so that the instant control system has better control performance. 6. A dynamic dispatching fuzzy-like neural network real-time control system and method thereof, ❹ comprising: a dynamic dispatched fuzzy back-type neural network; a controlled body; and - an on-line parameter adjustment The module; the dynamic sentimentary fuzzy neural network input is an error signal and its differential, and the output is a control signal; the controlled body is a device to be controlled, and usually the controlled body input is a control signal, and the output is The state of the controlled body is measured by the sensor, and the state of the controlled body is subtracted from the corresponding control command into an error signal, and the error signal and its differential are transmitted to the dynamic dispatching fuzzy type neural The network forms an instant control system; the online parameter adjustment module adjusts the variation of the dynamic dispatching fuzzy neural network parameters according to the fixed learning rate, the error signal and the differential. 7. The dynamic dispatching fuzzy-like neural network real-time control system and method thereof according to the sixth patent application scope includes a 200934197 network input layer and a network. a path attribution function layer, a network dispatch layer, a network rule layer and a network output layer; the network input layer directly transmits the network input signal to the network attribution function layer; the network attribution function layer Computing the signal sent by the input layer of the network, and transmitting the result to the network of the network via a specific attribution function operation; the network sends a valve mechanism to determine whether to transmit data to the network rule layer The network rule layer output is the product of the output of the corresponding network send layer; the output layer of the network is the network output signal, and each network output signal is the network rule layer output and the network All sums of the product of the weight value between the road and the network rule layer. 8. For the dynamic Pai Cui recursive fuzzy neural network of the dynamic dispatch control fuzzy neural network instant control system and method thereof, the network • attribution function layer includes a Network attribution function neuron and a network recursive structure; the input of the network attribution function neuron is the last network attribution function layer output multiplied by the network recursive structure weight value, and the current network input signal The result is output through a specific function of the attribute function; the network recursive structure can use the time delay of the network attribution function layer as the information of the next operation to achieve the capability of quickly and dynamically corresponding to the network. 9. The dynamic dispatching fuzzy-like neural network architecture of the dynamic dispatching fuzzy-like neural network real-time control system and method thereof according to the seventh patent application scope, the network dispatched layer is a transmission The mechanism of the valve, the mechanism can determine whether the output value of the network attribution function layer is higher than a preset threshold, and prohibit the too low network attribution function layer to pass through, and the failed network attribution function layer output is not The parameters of the attribution function layer need to be adjusted, which can effectively reduce the amount of network operations. 10. The dynamic dispatching fuzzy neural network real-time control system and method thereof according to the sixth application of the patent scope, the online parameter adjustment module is mainly obtained by the Supervised Gradient Descent Method. The amount of change in the parameter value in the Pai Cui recursive fuzzy neural network. The variation of all parameter values of the parameter adjustment module on the line is generated by the specific energy function to the parameter differentiation of the desired parameter. The value of the parameters in the Cui reversal fuzzy neural network to achieve the learning function makes the instant control system have better control performance. ❹ 32