1、 While these kinds of inspections are important to prevent ecologicaldisasters and risks for the people working around the plant, these are veryexpensive because scaffolding is often required and can be very dangerous Fig. 1. Typical operating environment and the Alicia3 robotfor technicians that ha
2、ve to perform these inspections. Moreover, for safetyreasons, plant operations must be stopped and the tank must be emptied,cleaned and ventilated when human operators are conducting inspections. InFig. 1(a) and 1(b) typical environments for climbing robots are shown. Figure1c shows the Alicia3 robo
3、t prototype while attached to a concrete wall duringa system test.2 System Description The Alicia II system (the basic module for the Alicia3 system) is mainlycomposed by a cup, an aspirator, two actuated wheels that use two DC motorswith encoders and gearboxes and four passive steel balls with clea
4、rance toguarantee plain contact of the cup to the wall. The cup can slide over a wallby means of a special sealing that allows maintaining a suitable vacuum insidethe cup and at the same time creating the right amount of friction with respectsystem weight and a range of a target wall kind. The struc
5、ture of the Alicia II module, shown in Fig. 2, currently comprisesthree concentric PVC rings held together by an aluminums disc. The biggerring and the aluminums disc have a diameter of 30 cm. The sealing system isallocated in the first two external rings. Both the two rings and the sealing are Fig.
6、 2. Structure of the Alicia II moduledesigned to be easily replaceable, as they wear off while the robot is running.Moreover the sealing allows the robot passing over small obstacles (about 1 cmheight) like screws or welding traces. The third ring (the smallest one) is usedas a base for a cylinder i
7、n which a centrifugal air aspirator and its electricalmotor are mounted. The aspirator is used to depressurize the cup formed bythe rings and the sealing, so the whole robot can adhere to the wall like astandard suction cup. The motor/aspirator set is very robust and is capable of working in harshen
8、vironments. The total weight of the module is 4 Kg. The Alicia3 robot is made with the three modules linked together by meansof two rods and a special rotational joint. By using two pneumatic pistons it ispossible to rise and to lower each module to overcome obstacles. Each modulecan be raised 15 cm
9、 with respect to the wall, so obstacles that are 1012 cmheight, can be easily overcame. The system is designed to be able to stayattached using only two cups while the third, any of the three, is raised up.The total weight of the system is about 20 Kg.3 Electro-Pneumatic System Model By using this k
10、ind of movement and sealing method, it is possible, due tounexpected small obstacles on the surface, to have some air leakage in thecup. This leakage can cause the internal negative pressure to rise up and inthis situation the robot could fall down. On the other side if the internalpressure is too l
11、ow (high p), a very big normal force is applied to thesystem. As a consequence, the friction can increase in such a way to notallow robot movements. This problem can be solved by introducing a controlloop to regulate the pressure inside the chamber to a suitable value to sustainthe system. The consi
12、dered open loop system and the most easily accessiblesystem variables has been schematized in Fig. 3; in this scheme the first blockincludes the electrical and the mechanical subsystem and the second blockincludes the pneumatic subsystem. The used variables are the Motor voltagereference (the input
13、signal that fixes the motor power) and the Vacuum level(the negative pressure inside the chamber). Fig. 3. The open loop system consideredFig. 4. I/O variable acquisition scheme Since it is very difficult to have a reliable analytical model of that system,because of the big number of parameters invo
14、lved, it has been decided toidentify a black box dynamic model of the system by using input/outputmeasurements. This model was designed with two purposes: to compute asuitable control strategy and to implement a simulator for tuning the controlparameters. An experimental setup was realized, as repre
15、sented in Fig. 4, by usingthe DS1102 DSP board from Dspace in order to generate and acquire theinput/output variables. Since the aspirator is actuated by an AC motor, apower interface has been realized in order to translate in power the referencesignal for the motor coming from a DAC channel of the
16、DS1102 board. Theoutput system variable has been measured with a piezoresistive pressuresensor with a suitable electronic conditioning block and acquired with oneanalog input of the DS1102. The software running on the DSpace DSP board,in this first phase simply generates an exciting motor voltage re
17、ference signal(pseudo random, ramp or step signals) and acquires the two analog inputswith a sampling time of 0.1 s, storing the data in its internal SRAM. Typical Input/Output measurements are represented in Fig. 5 and Fig. 6.In order to obtain better results in system modeling, the relationship be
18、tweenInput and Output needs to be considered as non-linear. A NARX model hasbeen used is in the form of (1), where f is a non linear function 8, 9. y(k) = f(u(k), u(k 1), . . . ; y(k 1), y(k 2), . . .) (1)To implement this kind of non-linearity, some trials have been done usingNeuro-Fuzzy and Artifi
19、cial Neural Network (ANN) methodologies. Once thatmodel has been trained to a suitable mean square error, it has been simulatedgiving it as input the real input measurement only (infinite step predictor) 8.So (1) can be modified in order to obtain (2). y(k) = f(u(k), u(k 1), . . . ; y(k 1), y(k 2),
20、. . .) (2)In (2), 4y is the estimated system output. In order to compare the simulationresults, a number of descriptor has been defined and used. Among these aremean error, quadratic mean error and some correlation indexes. A first setof simulation for both methodologies has been done to find out th
21、e best I/Oregression terms choice.3.1 Neuro-Fuzzy IdentificationUsing this kind of methodology, the best model structure was found to be inthe form of (3). y(t) = f(u(t), y(t 1) (3)Once the best model structure has been found, some trials have beenperformed modifying the number of membership functio
22、ns. The best results,comparing the indexes described above, have been obtained with 3 functionsand in Fig. 7 the simulation results has been reported. The structure of theNeuro-Fuzzy model is the ANFIS-Sugeno 10.3.2 ANN Identification Using this kind of methodology, the best model structure was foun
23、d to be inthe form of (4). y(t) = f(u(t), u(t 1), u(t 2), y(t 1), y(t 2) (4) A single layer perceptron network has been used. The training algorithm is the standard LevenbergMarquardt. Once the best model structure has been found, some trials have been performedmodifying the number of hidden neurons
24、. The best results, comparingthe indexes described above, have been obtained with 7 hidden neurons andin Fig. 8 the simulation results has been reported. From a comparison between the two models and their related indexes, itcan be seen that the Neuro-Fuzzy model has best approximation performanceand
25、 use less input information. In the next section, this model will be used assystem emulator to tune and test the required regulator.4 Pressure Control Algorithm Once a system model has been obtained, a closed loop configuration like thatin Fig. 9, has been considered. The target of the control algorithm is to regulate the internal vacuum levelto a suitable value (from some trials, it was fixed to about 10 kPa) to sustainthe whole system and its payload; the maximum steady state error allowedwas fixed