ACTIVE HEAT DISSIPATION SYSTEM USING ADAPTIVE RECURRENT WAVELET NEURAL NETWORK CONTROL

Transcription

1 ACTIVE HEAT DISSIPATION SYSTEM USING ADAPTIVE RECURRENT WAVELET NEURAL NETWORK CONTROL Yung-Lung Lee 1, Shou-Jen Hsu 2 and Yen-Bin Chen 3 1 1Department of Power Vehicle and Systems Engineering, Chung Cheng Institute of Technology, National Defense University, Taoyuan County, Taiwan, R.O.C. 2 Electronic Systems Research Division, Chung-Shan Institute of Science and Technology, Taoyuan County, Taiwan, R.O.C. 3 School of Defense Science, C.C.I.T., National Defense University, Taoyuan County, Taiwan, R.O.C dragonlee@ndu.edu.tw; hank1832@gmail.com IMETI 2015 J5020_SCI No. 16-CSME-06, E.I.C. Accession 3892 ABSTRACT This paper proposes a novel cooling control system with the intelligent active technique, which is based on the NI-PXI systems structure combined with the heat pipe cooling chips. In order to further solve the controlling problems of nonlinear heat transfer system, the proposed intelligent system involves the PID control, traditional control, and the adaptive recurrent wavelet neural network controller (ARWNNC) control techniques. The traditional control there exists the undesirable control chattering, The PID control the Response of Control cannot be processed immediately and the input voltage saturation phenomenon, the adaptive recurrent wavelet neural network controller, is employed to approximate the ideal controller, while the corresponding parameters are derived by the gradient steepest descent method, thus being provided with the adaptive real-time control ability. Keywords: thermoelectric cooler; adaptive recurrent wavelet neural network controller. SYSTÈME ACTIF DE DISSIPATION THERMIQUE UTILISANT UNE COMMANDE ADAPTIVE DE RÉSEAUX DE NEURONES RÉCURRENTS D ONDELETTES RÉSUMÉ Cet article propose un système innovateur de commande de refroidissement utilisant une technique active intelligente, basée sur la structure des systèmes NI-PXI combiné avec des puces de refroidissement à caloduc. Afin de résoudre les problèmes du système non-linéaire de transfert de chaleur, le système intelligent proposé implique le contrôle PID, le contrôle traditionnel, et le contrôle par réseau de neurones récurrents d ondelettes. Dans le contrôle traditionnel il y a un cliquetis indésirable de la commande. Le PID ne peut contrôler immédiatement le phénomène de saturation. La commande adaptive de réseaux de neurones récurrents d ondelettes est employée pour se rapprocher du dispositif de commande idéale, tandis que les paramètres correspondants sont obtenus par la méthode de descente du gradient le plus prononcé, donnant ainsila capacité de contrôle en temps réel adaptif. Mots-clés : refroidisseur thermoélectrique; régulateur adaptatif récurrent d ondelettes de réseau neuronal. Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4,

2 NOMENCLATURE x f (x, t) u(t) d(x,t) τ 1 n i φ(ω) = ω exp(( 1/2)ω 2 ) φ i j φ Ti j = φ Ti j (t τ 1 ) temperature of the system non-linear dynamic input of TEC control external jitters time delay number of the input variables Gaussian function derived from its mother wavelet and recurred loop output of mother wavelet node 1. INTRODUCTION In recent years, most of the radar internal multiplexing chip processor with high performance tends to become smaller in size. Microchip device has developed fast. It is a device that is widely used in diverse disciplines, ranging from electronic engineering, medical engineering to military science and technology. However, the rate of chip processing increases greatly compared to the rate of the past. In the radar with that the multiplexing chip with high performance needs to be processed in a narrow closed environment where convection is not allowed. Accordingly, the problem of heat dissipation must be solved in a narrow closed environment where convection is not allowed. In addition, supposing that this research is unable to take advantage of air cooling with restriction, it is important to adjust the rate of heat dissipation promptly. the control architecture has been widely used in some control issues of manipulator to fix the servo system in which complicated environment interact with external force [2]. In addition, based on sliding mode control techniques [1] of the change-able structured system, it is possible to maintain good control performance under the circumstances with a few more jitters where system and environment function as the uncertain factors to this research. In recent years, neural networks (NNs) have been widely used for identification and control of dynamic systems [4]. Many studies have suggested NNs as powerful building blocks for a wide class of complex nonlinear system control strategies when no complete model information is available or even when a controlled plant is considered as a black box. The most useful property of NNs is their approximation ability, in that they can approximate any function with an arbitrary degree of accuracy. Moreover, according to the structure, NNs can be mainly classified as feed-forward neural networks (FNNs) [6]. Some of these online learning algorithms are based on the back propagation learning algorithm [3, 7], and some on the Lyapunov stability theorem [5, 8]. To build up the identification and control of a system, the research shows the input-output relation of a system so that it is believed that this method can help to improve the traditional way of identification of a system and to cut out the weakness found in control. It also allows many scholars to dedicate themselves to studying WNN. Furthermore, it is a way to deal with the non-linear and uncertain properties found in the control system. The paper makes use of the wavelet neural network to construct a system. Hopefully, it can be a way to solve the non-linear heat transfer problem. Finally, a Radar Active Heat Dissipation System is performed to verify the effectiveness of the proposed control scheme. Results verify that the proposed ARWNNC can achieve favorable tracking performance without any chattering phenomenon. 2. SYSTEM FRAMEWORK 2.1. Hardware Framework Based on NI-PXI as the basic framework, this study combined heat pipe and TEC to establish an active cooling system which is divided into four systems: (1) Computer system (NI-PXI-8108, NI-PXI-6229, NI- PXI-43513); (2) Measurement system; (3) Power system (low power supply, high power supply, amplifier 446 Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4, 2016

3 Fig. 1. System hardware. Table 1. The specifications of each hardware. Hardware Manufacturer Type Specification NI-PXI 8108 National Instruments Inter Core 2 Duo NI-PXI-6229 National Instruments DAQ 16 Bit, 250 KS/s, 32AI, 4AO,48DIO NI-PXI-4351 National Instruments 16-Channel, 15S/s/channel Thermocouple Thermcouple Technology K-Type, 30 A, mm, Max 260 High Power Supply EMS Power supply DC 40 V, 25 A Low Power Supply GW Power supply DC 18 V, 5 A Heat Source Hi intelligence 200 W, Dimensions Heat source control system Hi intelligence PID control, Input AC 220 V Amplifier system MSK Output 1 10 A, 0 10 V module); and (4) Cooling system. The system hardware are shown in Fig. 1. The specifications of each hardware are as shown in Table 1. Figure 2 indicates the connections. Temperature sensor measurements are saved as Excel format in the terminal server (NI-PXI-8108). Voltage is output through NI-PXI 6229 to CONTROL I/P connection, passes operational amplifier, provides ±15 V with high power supply, and uses O/P as output voltage connection to control TEC voltage (Fig. 3). The cooling module (Item 8) are shown in Fig. 4, of which the numbers and names of components are the same as those in Table DESCRIPTION OF SMART ACTIVE HEAT DISSIPATION SYSTEM First of all, assuming that the active heat dissipation system and the dynamic formula of its TEC output can be written down as follows, it is assumed that this formula looks like: ẍ(t) = f (x,t) + u(t) + d(x,t). (1) Here, x indicates the temperature of the system while f (x, t) indicating a non-linear dynamic formula, and u(t) is the input of TEC control whereas d(x, t) indicates the external jitters. Supposing that the temperature heat dissipation system and the parameters of its TEC featured as f (x,t) and d(x,t) are unknown functions, it is believed that it is hard to acquire a figure. Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4,

4 Fig. 2. The connections. Fig. 3. Amplification circuit. 448 Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4, 2016

5 Table 2. Cooling module. Item Name Function 1 Heat Source and Heat Distributor Core Plate Link Heat Source 2 Heat Distributor Plate Support Plate Support 3 Heat Distributor Core Plate Core Plate 4 Fan Cooling 5 Floor Support Support 6 Contact Interface Hot Insulation Heat Insulation 7 Side Heat Distributor Plate Absorbs Heat Area 8 Six Heat Pipe Module Transfer Heat 9 Heat Sink Cooling 10 Thermoelectric Cooler (TEC) Coolant Pump 11 Heat Pipe Transfer Heat Fig. 4. Active cooling system Structure of Recurrent Wavelet Neural Network Recurrent wavelet neural network shares the capacity of powerful learning and similarity since it combines with the mathematical analysis to deal with wavelet theory and artificial neural network. Compared to the traditional NN, RWNN has plenty of advantages to be equipped with regional space, fast learning, convergent properties, and non-linear system. Therefore, it is able to provide a better input-output mapping Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4,

6 Fig. 5. Structure chart of a recurrent wavelet neural network (RWNN). relation. Its learning capability is more effective comparing to NN in a general sense. As a result, it is possible that a better performance can be achieved with this application used in system identification and control problem. For the overall structure, the reader is referred to Fig. 5. A RWNN is proposed and depicted in which τ 1 denotes time delay. The signal transmission process and basic function stored in each layer of the network can be briefly depicted as follows: (a) Input layer: For every node i in this layer, the net input and the net output are represented as net (1) i = x (1) i, (2) y (1) i = net (1) i, for i = 1,2,...,n i, (3) where x (1) i represents the ith input to the node of input layer and n i is the number of the input variables. The link weights at this layer are all set as unity. (b) Mother wavelet layer: Each node of this layer has a mother wavelet. The first derivative of a Gaussian function φ(ω) = ω exp(( 1/2)ω 2 ) is selected as a mother wavelet function, which has the universal approximation property [3]. Each node is derived from its mother wavelet. For the jth node of the ith input we obtain net (2) i j = (x(2) i φ Ti j c i j m i j ) = (y(1) i φ Ti j c i j m i j ), (4) ν i j ν i j y (2) i j = φ(net (2) i j ) for j = 1,2,...,n p, (5) 450 Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4, 2016

7 where x (2) i represents the ith input to the node of mother wavelet layer, c i j is the translation factor and ν i j is the dilation factor of the mother wavelet node in the jth term of the ith input variable, m i j is the translation, factor respectively. Each node φ i j is derived from its mother wavelet and recurred loop. Actor r i j is the recurrent weight of the mother wavelet node in the jth term of the ith input variable. For the jth node of the ith input y (2) i j is the output of the mother wavelet node and φ Ti j = φ Ti j (t τ 1 ) is the output of the mother wavelet node and is the output through delay time τ 1 for the node of mother wavelet layer, and n p is the total number of node in the product layer. In the mother wavelet layer, the recurrent feedback is embedded within the network by adding feedback connections through delay. Since the recurrent feedback has a delayed internal feedback loop, it captures the dynamic mapping network to form the RWNN. (c) Product layer: The node in this layer is given by the product of the mother wavelets as follows: ( ) ( ) net (3) j = Π n i i=1 x(3) i = Π n i i=1 y(2) i j = Π n (y (1) i i φ Ti j c i j m i j ) i=1 exp (y(1) i φ Ti j c i j m i j ) 2 ν i j 2νi 2, j (6) y (3) j = net (3) j for j = 1,2,...,n p, (7) where x (3) i represents the ith input to the node of product layer. (d) Output layer: The output node in this layer is labeled as, which computes the output as the summation of all incoming signals net (4) n p o = θ jo x (4) j=1 n i j + i=1 α i y (1) i = n p θ jo y (3) j=1 n i j + i=1 α i x (1) i, (8) y (4) j = net (4) j for o = 1,2,...,n o, (9) where θ jo is the connection weight between the jth product node and oth output node and x (4) j represents the jth input to the node of output layer, y (4) o is the output of the RWNN, n o is the number of the output node. The overall representation of the ith input of x (1) i and the oth output y (4) o is y o = y (4) o = n p θ jo Π n i j=1 j=1 ( (y (1) ) ( i φ Ti j c i j m i j ) exp ν i j u WNNo = y (4) n p o = θ jo y (3) j (x (1) i,m i j,c i j,ν i j ) + j=1 (y(1) i φ Ti j c i j m i j ) 2 2ν 2 i j n i i=1 ) + n i i=1 α i x (1) i, (10) α i x (1) i. (11) The oth output of WNN can be represented and defined by matrix and vectors Θ, c and ν, m to collect all parameters of the connection weight and mother wavelets of RWNN as Θ = [ω 1,,ω Np,α 1 α Ni ] T R n in p, (12) c = [c 11 c ni 1,c 12 c ni 2, c 1np c ni n p ] T R n in p, (13) m = [m 11 m ni,1, m 12 m ni 2,m 1np m ni n p ] T R n i,n p, (14) ν = [ν 11 ν ni 1,ν 12 ν ni 2,ν 1np ν ni n p ] T R n in p. (15) In summary, the outputs of RWNN expresses in a vector notation as u WNN (x,m,c,νθ) = Θ T β(x,m,c,ν), u WNN = [u WNN1,u WNN2,,u WNN n o ] T, x = [x (1) 1,x(1) 2,...,x(1) n i ] T, β = [y (3) 1,y(3) 2,...,y(3) n p ] T. Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4,

8 Fig. 6. Adaptive recurrent wavelet neural network controller of active cooling system Control Design of Adaptive Recurrent Wavelet Neural Network This research proposes a block diagram of adaptive recurrent wavelet neural network control as shown in Fig 6. The aim of control is to find out a way to allow x(t) being able to track the referential signal x d (t). To achieve the purpose of control, it is assumed that the tracking error must be defined at first. Next, a slide plane shall be defined with the following formula: e(t) = x d (t) x(t). (16) s(t) = ė(t) + ke(t). (17) In this formula, we note that k > 0. Assuming that the dynamic functions of controlled system, f (x,t) and d(x,t) are both known functions, it is believed that the best control can be acquired by Substituting Eq. (18) into Eq. (1), we acquire u (t) = f (x,t) d(x,t) + ẍ d kė(t). (18) ë(t) + kė(t) = 0 = ṡ(t). (19) Furthermore, the above formula can acquire a result of lim t e(t) 0 due to a fact that k > 0 to achieve the purpose of control. However, in actual system, f (x,t) and d(x,t) are two unknown non-linear time variance functions. Thus, u (t) in Eq. (18) cannot be clearly defined. Accordingly, this research controls the system to perfect the best control u (t) in a similar way. First of all, Eq. (1) can be written in a simplified way and the following formula is acquired: ẍ(t) = f (x,t) + u(t) + d(x,t) = F(t) + u(t). (20) In this formula, the non-linear time variance function written down as F(t) cannot be clearly acquired in this formula F(t) = f (x,t) + d(x,t). Thus, this research makes use of adaptive recurrent wavelet neural network 452 Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4, 2016

9 Fig. 7. Temperature points distributed measurement. system written down as ˆF(x,c,ν,θ) to estimate approximation. In addition, it adds a collision control u b (t) to this formula conquering the approximation error of adaptive recurrent wavelet neural network system written down as ˆF and actual non-linear time variance system written as F(t). The control by design is stated as follows: u(t) = u RWNN (t) + u b (t). (21) Here, u RWNN (t) indicates the main control and u b (t) stands for the control force that keeps the track in the system on the slide plane. The main control is stated as u RWNN (t) = ˆF(x,c,ν,θt) + ẍ d (t) + kė(t), (22) where ˆF(x,c,ν,θ) = Θ T β(x,c,ν,m). (23) 4. ACTIVE HEAT DISSIPATION SMART CONTROL SYSTEM Radar inner core processor belongs to a component with high precision. It means that the component consists of the radio frequency integrated circuit electronics with high power and high performance. Core processor shall be kept in a narrow closed environment where convection is not allowed. Heat dissipation problem in long distance can be solved in a narrow closed environment where convection is not allowed to fix the chip through simulation. The highest temperature indicates the 20 W that provides thermal source to replace heating value of the chip while the operation continues. The actual best temperature for control is about 75 C. The measuring Curie point distribution is drawn in Fig. 7. In this figure, No. 1 and No. 2 indicate a measuring the surface temperature of thermal source and heat pipe respectively. No. 1 means the surface temperature of measuring heat pipe and heat slug. The designed active heat dissipation smart control system is produced to achieve the performance with good and accurate temperature control. By comparing PID controller to adaptive recurrent wavelet neural network control, this research shows that the PID controller contains relay, temperature switch, and control circuit in this experiment. The temperature is controlled at degrees of no more than 80, 75, and 70 C, respectively. There is Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4,

10 Fig. 8. Traditional analogy circuit controller. an interval of about 1000 seconds for each. In this research, the cooling system is settled with a wind speed at a rate of 2300 rpm at 24 C in room temperature with an input voltage section ranged from 0 10 V 0 10 A, and the time for heating is 4500 seconds. In this way, this research proposes an active heat dissipation smart control system. It can thus achieve the purpose to cool down the temperature with good and accurate control performance through verification. 5. ANALYSIS OF RESEARCH FINDINGS Figure 8 displays the active heat dissipation system applied to a Traditional analogy circuit controller. Figure 9 shows the active heat dissipation system applied to a PID controller. Figure 10 shows the active heat dissipation system applied to a Adaptive Recurrent Wavelet Neural Network Control controller. The temperature was set at 80, 75, and 70 C, respectively. Assuming the actual temperature exceeds the set point, it is suggested that this research starts to provide voltage to cool down the temperature. It happens that voltage resulted in a frequent switch mode. It turns out to be a saturation voltage phenomenon. In addition, the response to control voltage reaction is not dealt with properly at once. Or perhaps it may end up with an over-control phenomenon, a decrease in voltage due to providing voltage, or instability due to power supply for system use. Figure 10 displays the active heat dissipation system application to adaptive recurrent wavelet neural network control. The temperature was set at 80, 75, and 70 C, respectively. Assuming the actual temperature exceeds the set point, it is suggested that this research is capable of giving orders to tracking through good searching device. In addition, as far as the cooling effect is concerned, nothing happens to this control compared to the Adaptive Recurrent Wavelet Neural Network Control controller since no frequent jitters are detected while the switching of the control voltage input occurred. It is possible to increase the heat dissipation effect according to chip thermal source by rendering adaptive, adjustable heat dissipation speed. It can achieve the purpose to reach a good and accurate temperature control performance to deal with the heat dissipation problem in long distance where convection is not allowed. 454 Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4, 2016

11 Fig. 9. PID Controller. Fig. 10. Adaptive recurrent wavelet neural network controller. Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4,

12 6. CONCLUSION This paper is based on NI-PXI system as the basic structure for this research. It combines TEC with heat pipe to work out a new active heat dissipation system so that the heat dissipation problem in long distance in a narrow closed environment where convection is not allowed can be solved. To compare with an experiment of controller (shown in Figs. 8 10), this research adopts the adaptive recurrent wavelet neural network control and the PID controller and Traditional analogy circuit controller. It is proved that this research proposed an active heat dissipation smart control system that achieves the purpose with a good and accurate temperature control performance. Adaptive recurrent wavelet neural network control has a good capability of searching and tracking. In addition, its cooling control performance is far better than a PID controller and Traditional analogy circuit controller. Nothing happens concerning a saturation voltage phenomenon while controlling the input of voltage. It ends up with a result that the chip can increase its heat dissipation effect in a narrow closed environment where convection is not allowed according to the chip thermal source to give adaptive, adjustable heat dissipation speed. It helps to solve the heat dissipation problem in long distance where convection is not allowed. It proves that this system is effective and practical. It has been proved that this system can provide feasibility applied to a radar heat dissipation system with high power. Hopefully, the structure of the control can provide some solution to solve the key problems of heat dissipation techniques development in terms of livelihood, medical treatment, military affairs, and industry in the long run. REFERENCES 1. Billings, S.A. and Wei, H.L., A new class of wavelet networks for nonlinear system identification, IEEE Transactions on Neural Networks, Vol. 16, No. 4, pp , Chen, C.H., Intelligent transportation control system design using wavelet neural network and PID type learning algorithms, Expert Systems With Applications, Vol. 38, No. 6, pp , Da, F., Fuzzy neural network sliding mode control for long delay time systems based on fuzzy prediction, Neural Computing and Applications, Vol. 17, No. 5, pp , Lin, C.M. and Hsu, C.F., Neural-network hybrid control for antilock braking systems, IEEE Transactions on Neural Networks, Vol. 14, No. 2, pp , Sousa, C.D., Hemerly, E.M. and Galvao, R.K.H., Adaptive control for mobile robot using wavelet networks, IEEE Transactions on Neural Networks, Vol. 32, No. 4, pp , Zhai, J.Y., Fei, S.M. and Mo, X.H., Multiple models switching control based on recurrent neural networks, Neural Computing and Applications, Vol. 17, No. 4, pp , Zhang, Q., Using wavelet network in nonparametric estimation, IEEE Transactions on Neural Networks, Vol. 8, No. 2, pp , Zhu, H.A., Hong, G.S. and Poo, A.N., Internal model control with enhanced robustness, International Journal of System Science, Vol. 26, No. 2, pp , Transactions of the Canadian Society for Mechanical Engineering, Vol. 40, No. 4, 2016