FPGA Implementation of Adaptive Neuro-Fuzzy Inference Systems Controller for Greenhouse Climate

Transcription

1 (IJACSA) Internatonal Journal of Advanced Computer Scence and Applcatons, FPGA Implementaton of Adaptve Neuro-Fuzzy Inference Systems Controller for Greenhouse Clmate Charaf eddne LACHOURI Electroncs department Unversty Badj Mokhtar Annaba, Algera Khaled MANSOURI Electroncs department Unversty Badj Mokhtar Annaba, Algera Assa BELMEGUENAI Electroncs research Laboratory Unversty 20 August 1955 Skkda, Algera Mohamed mourad LAFIFI Electroncs department Unversty Badj Mokhtar Annaba, Algera Abstract Ths paper descrbes a Feld-programmable Gate Array (FPGA) mplementaton of Adaptve Neuro-fuzzy Inferences Systems (ANFIS) usng Very Hgh-Speed Integrated Crcut Hardware-Descrpton Language (VHDL) for controllng temperature and humdty nsde a tomato greenhouse. The man advantages of usng the HDL approach are rapd prototypng and allowng usage of powerful synthess controller through the use of the VHDL code. The use of hardware descrpton language (HDL) n the applcaton s sutable for mplementaton nto an Applcaton Specfc Integrated Crcut (ASIC) and Feld tools such as Quartus II 8.1. A set of sx nputs meteorologcal and control actuators parameters that have a major mpact on the greenhouse clmate was chosen to represent the growng process of tomato plants. In ths contrbuton, we dscussed the constructon of an ANFIS system that seeks to provde a lngustc model for the estmaton of greenhouse clmate from the meteorologcal data and control actuators durng 48 days of seedlngs growth embedded n the traned neural network and optmzed usng the backpropagaton and the least square algorthm wth 500 teratons. The smulaton results have shown the effcency of the mplemented controller. Keywords Neuro-Fuzzy; ANFIS; VHDL; FPGA; Quartus; ASIC I. INTRODUCTION Under greenhouse producton, the clmate control s a tool used for yeld crop manpulaton that maxmzes the entrepreneural benefts. Once the objectves that optmze crop growth and development are defned, the control engneer must desgn and mplement automatc control systems that make possble to obtan a maxmum crop yeld at mnmum producton costs. In ths sense, control engneerng has undergone a consderable development. Researchers have used many control technques n dfferent felds, from the conventonal or classc strateges [proportonal ntegral dervatve (PID) control, cascade], artfcal ntellgence (AI) (fuzzy control, neural networks and genetc algorthms), advanced control technques (predctve control, adaptve), to robust control strateges, non-lnear and optmal control. Specfcally, they have been appled n the area of greenhouse clmate control [1][2][3]. Conventonal control technques are dffcult to mplement n greenhouse systems due to ther mult-varable and non-lnear nature. Where nterrelatons between nternal and external varables are complex (nonlnear physcal phenomena that govern these systems dynamcs are complcated). Ths provdes justfcaton for the use of ntellgent control technques as a good alternatve. In ths way, fuzzy logc as part of AI technques s an attractve and well-establshed approach to solvng control problems [4]. We were brought to develop a Neuro-Fuzzy control of the nternal humdty and nternal temperature of the greenhouse. Ths last characterzes the operaton of the complex system that the greenhouse consttutes. The dentfcaton that s n the center of ths step s a process of search for a mathematcal representaton that mnmzes the varatons of the real system compared to the modeled system. The development of the plant s nfluenced manly by the envronmental, clmatc varables. The greenhouse, whch s a closed crcle n whch the clmatc varables can be controlled, consttutes the deal medum for the control of the plants growth. The greenhouse must not only create the favorable condtons of the plants growth, but t must moreover be able to ensure certan flexblty n the calendar of producton: precocty and spreadng out of the calendar. To carry out ths objectve a robust model usng the Artfcal Neural Networks and the fuzzy logc can be well adapted to control the nonlnear comportment of greenhouse clmate accurately s more than necessary [5]. For the mplementaton of agrcultural technologes (nnovatons n control systems, remote montorng, nformaton management), robustness, low-cost and real-tme capabltes are needed. In ths sense, feld programmable gate arrays (FPGAs) proved as a good opton for greenhouse technology development and mplementaton, because FPGAs allow fast development of prototypes and the desgn of complex hardware systems. These devces are used n many real applcatons [6]. Through FPGAs, rapd tests, modfcatons accomplshment, up-dates usng sngle software modfcatons and an effectve producton cost (relaton performance-prce s very favorable) are obtaned. In the same sense, reducton n development and commercalsaton tme s P age

2 (IJACSA) Internatonal Journal of Advanced Computer Scence and Applcatons, accomplshed. On the other hand, for neuro-fuzzy control mplementaton, whch based on software or hardware, FPGAs are an alternatve that keep both benefts, hardware speed and software flexblty. Research made about these devces has experenced an enormous development, n the academc feld as well as n the ndustral area. There s a great number of contrbutons about FPGAs applcatons n dfferent felds [7][8][9]. Also, there are some contrbutons reported about hardware mplementatons of neuro-fuzzy control [10]. Moreover, problems of dgtzed neuro-fuzzy control have been studed [11]. The approach proposed here s focused on greenhouse technologes development, based on AI technques, partcularly fuzzy logc, cascaded wth a feedforward a neural network, and system-on-a-chp (SoC) applcatons usng FPGA technology, wth the purpose of obtanng complete engneerng solutons on a sngle Integrated crcut. In our case, an ntellgent SoC was developed to carry out the perfect functonalty for the greenhouse clmate control due to an ANFIS system that seeks to provde a lngustc model for the estmaton of greenhouse clmate from the meteorologcal data and control actuators durng 48 days of seedlngs growth embedded n the traned neural network and optmzed usng the backpropagaton and the least square algorthm wth 500 teratons. II. NEURONAL METHODS IN THE FUZZY SYSTEMS In a conventonal fuzzy nference system, the number of rules s decded by an expert who s famlar wth the system to be modeled. In ths partcular case study the rules generated by an agrculture expert and the number of membershp functons assgned to each nput s chosen from real data. Ths s carred out by examnng the desred and real nput-output data. Ths stuaton s much the same as ANN s. In ths secton ANFIS topology and the learnng method used for ths neural network are presented. Both neural network and fuzzy logc are model-free estmators and share the common ablty to deal wth the uncertantes and nose. It s possble to convert fuzzy logc archtecture to a neural network and vce versa [12]. Ths makes t possble to combne the advantages of neural network and fuzzy logc [13-14]. Layer 1: Every node n n ths layer s a square node wth a node functon 1 o = µ A ( x ) (1) Where x s the nput node, and A s the lngustc label (Mnmum, Moderate, Maxmum) assocated wth ths node 1 functon. In other words, o s the membershp functon and t specfes the degree to whch the A gven x satsfes the quantfer A. Usually we choose µ A ( x) to be bell shaped wth maxmum equal to 11, moderate equal to 00 and mnmum equal to 10, such as (2) 1 µ A ( x) = 2 b x c 1+ a Where {a, b, c } s the parameter set. As the values of these parameters change, the best bell-shaped functons vary accordngly, thus exhbtng varous forms of membershp functons on lngustc label A. In fact, any contnuous and pecewse dfferentable functons, such as commonly used trapezodal or trangular-shaped membershp functons are also qualfed canddates for node functons n ths layer. Parameters n ths layer are referred to as premse parameters. Layer 2: Every node n ths layer s a crcle node labeled whch multples the ncomng sgnals and sends the product out. For nstance, w = µ A(x)* µ A(y), = 1,..., 40 Each node output represents the frng strength of a rule (In fact, other T-norm operators that perform generalzed AND can be used as the node functon n ths layer). Layer 3: Every node n ths layer s a crcle node labeled N. The th node calculates the rato of the th rule s frng strength to the sum of all rules frng strengths: w w =, 1,...,40 w w = 1 40 For convenence, outputs of ths layer are called normalzed frng strengths. Layer 4: Every node n ths layer s a square node wth a node functon O = wf= w(p x+ q y+ r ) 4 Where w s the output of layer 3, and (p, q, r ) s the parameter set. Parameters n ths layer wll be referred to as consequent parameters. Layer 5: The sngle node n ths layer s a crcle node labeled Σ that computes the overall output as the summaton of all ncomng sgnals. wf 5 O1 = overalloutput = w f = w (6) Thus we have constructed an adaptve network whch s functonally equvalent to a fuzzy nference system [14-15]. The hybrd algorthm s appled to ths archtecture. Ths means that, n the forward pass of the hybrd learnng algorthm, functonal sgnals go forward up to fourth layer and the consequent parameters are dentfed by the least squares estmaton. In the last backward and the premse parameters are updated by the gradent descent [14]. A. ANFIS Predctve Archtecture Usng a gven nput/output data set, the ANFIS method constructs a fuzzy nference system (FIS) whose membershp functon parameters are tuned (adjusted) usng ether a backpropagaton algorthm alone, or n combnaton wth a least squares type of method. Ths allows fuzzy systems to learn from the data they are modelng. FIS Structure s a network-type structure smlar to that of a neural network, whch maps nputs through nput membershp functons and assocated parameters, and then through output membershp functons and assocated parameters to outputs [16]. (3) (4) (5) P age

3 (IJACSA) Internatonal Journal of Advanced Computer Scence and Applcatons, In our case ANFIS s a four-layer neural network that smulates the workng prncple of a fuzzy nference system. The lngustc nodes n layers one and four represent the nput and output lngustc varables, respectvely. Nodes n layers two are term nodes actng as membershp functons for nput varables. Each neuron n the thrd layer represents one fuzzy rule, wth nput connectons representng precondtons of the rule and the output connecton representng consequences of the rules. Intally, all these layers are fully connected, representng all possble rules. Sx feature varables, nternal temperature, nternal humdty, external temperature, external humdty, global radaton and wnd speed, are selected as nputs of the ANFIS. Three membershp functons (Mfs) are assgned to each lngustc varable. The suggested ANFIS model s shown n Fg. 1. Then an error for each data par s found. If ths error s larger than the threshold value, update the premse parameters usng the gradent decent method as the followng (Qnext=Qnov+ηd, where Q s a parameter that mnmzes the error, η the learnng rate, and d s a drecton vector). The process s termnated when the error becomes less than the threshold value. Then the checkng data set s used to compare the model wth actual system. A lower threshold value s used f the model does not represent the system. Fg. 2, shows the unform fallng of the value of testng error ETest wth the number of teratons durng the testng process for the ANFIS confguraton wth trangular Mf and wth gaussan Mf. The smallest error of testng (ETest) s reached at teraton 145 (trangular Mf) and at teraton 107 for Gaussan Mf. It can be seen n the Fg. 2, that error converges not to zero but to 12% and 2%. Ths s caused by the presence of some contradctng examples n the tranng and testng set. Fg. 1. ANFIS model structure of greenhouse clmate B. ANFIS Modelng, Tranng and Testng ANFIS modelng process starts by obtanng a data set (nput-output data) and dvdng t nto tranng, testng and checkng data sets. Tranng data consttutes a set of nput and output vectors. The data s normalzed n order to make t sutable for the tranng process. Ths was done by mappng each term to a value between 00, 01 and 10 usng the Mn, moderate and Max method. Ths normalzed data was utlzed as the nputs (nternal clmate condtons and meteorologcal data) and outputs (actuators condtons) to tran the ANFIS. In other words, two vectors are formed n order to tran the ANFIS. Input vector = [nternal temperature, nternal humdty, external temperature, external humdty, global radaton and wnd speed]. The output vector = [Ventlatng and heatng]. The ANFIS regsters the nput data only n the numercal form therefore the nformaton about the control actuators, nternal and external clmate of the greenhouse must be transformed nto numercal code. The tranng data set s used to fnd the ntal premse parameters for the membershp functons by equally spacng each of the membershp functons. A threshold value for the error between the actual and desred output s determned. The consequent parameters are found usng the least-squares method. Fg. 2. Decrease of error durng the testng process for the ANFIS confguraton wth Trangular Mf and wth Gaussan Mf Tranng of the ANFIS can be stopped by two methods. In the frst method, ANFIS wll be stopped to learn only when the testng error s less than the tolerance lmt. Ths tolerance lmt would be defned at the begnnng of the tranng. It s obvous that the performance of the ANFIS that s traned wth lower tolerance s greater than ANFIS that s traned wth hgher tolerance lmt. In ths method the learnng tme wll change wth the archtecture of the ANFIS. The second method to stop the learnng s to put constrant on the number of learnng teratons. In our study, the ANFIS archtecture s stopped to learn after 500 tranng teratons. III. NEURO-FUZZY CLIMATE CONTROLLER As s already known from neuro-fuzzy prncples, a neuro fuzzy controller acts as a non-lnear system capable of mplementng expert reasonng for computaton of the control values. Indeed, a neuro fuzzy controller whch s defned by a set of lngustc rules and fuzzy sets were traned by neural network and optmzed usng the back-propagaton and the least square algorthm s able to compute approprate values P age

4 (IJACSA) Internatonal Journal of Advanced Computer Scence and Applcatons, for greenhouse actuators (heatng, ventlatng) takng nto account nformaton data from the system for control proposes. In the expermental greenhouse, the temperature s controlled by means of heaters, whle the humdty s controlled ndrectly wth the ventlatng ndex regulaton. That affects on the temperature and the humdty. Usng the physcal model, a complete system smulator s shown n Fg. 3. Wth ths smulator, a frst experment was carred out usng a conventonal controller (on-off) wth a dead band of 2 C; ths s based on a heatng system that s actvated or deactvated when the error exceeds the fxed regulaton range. The humdty depends on the nternal ar temperature and the ventlatng ndex. Ths varable s regulated by wndows openng n the greenhouse accordng to the wnd speed measurements. In ths case, a multple nputs, multple outputs (MIMO) non-lnear controller for temperature regulaton was used. A MIMO neuro fuzzy controller can be dstrbuted n several multple nputs, sngle output (MISO) controllers keepng the same performance. These controllers are ndependent and can be executed n parallel, whch s advsable for the clmate controller mplementaton n a FPGA. The Neuro Fuzzy Controller has sx nput varables and two output varables, characterzed by three fuzzy sets n the unverse of dscourse. Input varables are nsde and outsde temperature (T, Text), nsde and outsde humdty (H, Hext), global radaton (Gr) and wnd speed (Ws). Membershp functons sets and ther approprate modfcatons were obtaned followng a test and error strategy by makng exhaustve smulatons n Matlab untl reachng a good performance through a careful tunng. Fg. 4, shows an example of a membershp functons set for the nput. For ths one, three lngustc varables were used (MIN, mnmum; MOY, medum; MAX, maxmum). The set of fuzzy rules to develop the controller for each varable has been obtaned from the expert grower. For tunng the fuzzy rules as well as for membershp functons sets a tral-and-error strategy (manual tunng) was used, ths s modfyng control rule sets untl we reachng a good performance of the controller by usng the ANFIS edtor (smulaton system). Each possble lngustc value of nputs s assgned to a consequental acton. Fg. 3. The model of the greenhouse control system Fg. 4. Membershp functon of nternal temperature IV. DESIGN AND HARDWARE IMPLEMENTATION The Neuro Fuzzy Controller shown n Fg. 5, has been mplemented on an FPGA. The hardware platform used s the Altera DE2 development and educaton board that s based on the Altera Cyclone II EP2C35F672C6 FPGA. Fg. 5. Neuro Fuzzy Controller In order to mplement our applcaton effectvely the desgn s broken down nto modules. A. The de-multplexer component The system should accept multple nputs wth 8-bts n total of 48-bts. In order to reduce the number of pns used n FPGA we have made a de-multplexng as shown n Fg. 6, t has one nput of 8-bts and three selecton lnes, n order to learn at each clock pulse one nput and he settles t nto a buffer. After sx clock top t wll acqure all nputs. At the seventh clock pulse t delvers the enable sgnal and the values of multples nputs to the rest of the system. B. The Fuzzer module In ths secton we have realzed sx blocks, where each block s ntended for one of membershps functons. The example of such block s presented n the fgure t used for the external temperature gven n Fg. 7. The blocks transformed the numercal data to three lngustc varables (MIN, MOY, MAX). For easy mplementaton and as we have three cases two bts are used to materalze these case as follows (mn => 10, moy => 00 and max = > 11) P age

5 (IJACSA) Internatonal Journal of Advanced Computer Scence and Applcatons, V. RESULTS AND DISCUSSION The next step s the smulaton of the desgn to llustrate how t works. Fg. 9, shows the global smulaton tmng obtaned by Quartus II verson 8.1 SJ Web edton. Data are the nput values nformaton, (T_nt, T_ext, H_nt, H_ext, W_s and R_g) are the values of the deferent parameters, (fuz_tnt, fuz_text, fuz_hnt, fuz_hext, fuz_ws and fuz_rg) are the resultant of all the membershp functons. Cmd-H-W s the fnally output value represent the ventlatng and heatng. Fg. 6. The de-mulplxer component Fg. 7. The fuzzer module C. The command module The followng operaton s the order of the ventlatng and the heatng. Ths component shown n Fg. 8, admts at the nputs the varous decsons for the multple nputs and t wll computng the rules of our FIS structure obtaned by Matlab Fuzzy Logc Toolbox. To reduce the use of the hardware resource, fnte state machne (FSM) s adopted to model ths computng process. Fnally t wll transform the lngustcs values on the bnary values. Fg. 9. Neuro Fuzzy Controller Quartus II smulaton The table I shows the strong smlarty between the results obtaned by Matlab Fuzzy Logc Toolbox envronment and those obtaned n Fg. 9. It shows the best operaton of all modules. We can also see how the transformaton of these data from the lngustc values to numercal values. Synthess of fuzzy neural network on FPGA: We have mplemented the desgn usng the DE2 board, contan Cyclone II 2C35Altera FPGA devce, EP1C6Q240. The prncpal features of Cyclone II EP2C35 FPGA are as follows: Logc elements. 105 M 4K RAM blocks. 483,840totalRAMbts. 35 embedded 1818 multplers. Four PLLs. 475 user I/O pns. Fg. 8. The command component P age

6 (IJACSA) Internatonal Journal of Advanced Computer Scence and Applcatons, TABLE I. COMPARISON OF THE RESULTS GIVEN BY MATLAB FUZZY LOGIC TOOLBOX AND THOSE OBTAINED WITH QUARTUS II Fg. 10. RTL schematcs of neuro-fuzzy controller VI. CONCLUSION The current work focuses on the applcaton of neurofuzzy control of a greenhouse nternal clmate. It successfully demonstrated the performance through co-smulaton by usng ANFIS and ModelSm. Ths mplementaton accurately reproduces the theoretcal behavor of the system, thus s ready to be used. The future work wll be destned to mprove the desgn of our work ncludng the number and the type of nputs membershp functons. REFERENCES [1] Lees M J; Taylor J; Chota A; Young Z S; Chalab Z S(1996). Desgn and mplementaton or a proportonal-ntegral plus (PIP) control system for temperature, humdty and carbon doxde n a glasshouse. Acta Hortculturae, 406, [2] Hanan J(1998). Greenhouses: Advanced Technology for Protected Hortculture, 1st Edn. CRC Press, New York, USA. [3] Tap F(2000). Economcs-based optmal control of greenhouse tomato crop producton. PhD Thess, Wagenngen Agrcultural Unversty, The Netherlands. [4] Lee C C(1990). Fuzzy logc n control systems: fuzzy logc controller (parts I and II). Transactons on Systems, Man, and Cybernetcs,20, [5] Draou B(1994). Charactersaton and analyses hydrous assessment thermo of a hortcultural greenhouse. In stu dentfcaton of the parameters of a dynamc model, Thess of Doctorate of the unversty of Nce Sopha Antpols, France, [6] Vega-Rodrguez M A; Sanchez-Perez J M; Gomez-Puldo J A (2004). Specal ssue on FPGAs: applcatons and desgns. Mcroprocessors and Mcrosystems,28, [7] Al L; Sdek R; Ars Ishak; Al A M; Suparjo B S(2004). Desgn of a mcro-uart for SoC applcaton. Computers and Electrcal Engneerng,30, [8] Romero-Troncoso R; Herrera-Ruz G; Terol-Vllalobos I; Jauregu- Correa J C (2004). FPGA based on-lne tool breakage detecton system for CNC mllng machnes. Mechatroncs,14, [9] Mendoza-Jasso J; Ornelas-Vargas G; Castan eda-mranda R; Ventura- Ramos E; Zepeda-Garrdo A; Herrera-Ruz G (2005). FPGA-based realtme remote montorng system. Computers and Electroncs n Agrculture, 49(2), [10] Dpal L.Gakwad; Prabha Kaslwal (2013). FPGA Based Crtcal Patent Health Montorng Usng Fuzzy Neural Network. Internatonal Journal of Scentfc & Engneerng [11] Dhananjay E. Upasan (2010). FPGA mplementaton of ntellgent clmate control for greenhouse. Internatonal Journal of Computer Applcatons ( ). [12] H. C. Cho, and K. S. Lee, Adaptve control and stablty analyss of nonlnear crane systems wth perturbaton, J. Mech. Sc. Techn., vol. 22, pp , [13] M. Y. El. Ghoumar, H. J. Tantau, and J. Serrano, Nonlnear constraned MPC: real-tme mplementaton of greenhouse ar temperature control, Comput. Elect. Agrc., vol. 49, pp , [14] F. Lafont, and J. F. Balmat, Fuzzy logc to the dentfcaton and the command of the multdmensonal systems, Internatonal J. Comput. Cognton, vol. 2, pp , [15] F. Fourat, and M. Chtourou, A greenhouse control wth feed-forward and recurrent neural networks, Smulaton Modelng Pract. and Theory, vol.15, pp , [16] J.S.R. Jang ANFIS: Adaptve-Network-Based Fuzzy Inference System, IEEE Trans. Systems, Man, Cybernetcs, 23(5/6): , P age