Embed Size (px)
Citation preview
This article was downloaded by: [University of North Carolina]On: 12 November 2014, At: 08:11Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office:Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
International Journal of Systems SciencePublication details, including instructions for authors and subscriptioninformation:http://www.tandfonline.com/loi/tsys20
Development of a system to diagnoseautonomous underwater vehiclesMotoyuki Takai & Tamaki UraPublished online: 26 Nov 2010.
To cite this article: Motoyuki Takai & Tamaki Ura (1999) Development of a system to diagnose autonomousunderwater vehicles, International Journal of Systems Science, 30:9, 981-988, DOI: 10.1080/002077299291859
To link to this article: http://dx.doi.org/10.1080/002077299291859
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”)contained in the publications on our platform. However, Taylor & Francis, our agents, and ourlicensors make no representations or warranties whatsoever as to the accuracy, completeness, orsuitability for any purpose of the Content. Any opinions and views expressed in this publication arethe opinions and views of the authors, and are not the views of or endorsed by Taylor & Francis.The accuracy of the Content should not be relied upon and should be independently verified withprimary sources of information. Taylor and Francis shall not be liable for any losses, actions, claims,proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoevercaused arising directly or indirectly in connection with, in relation to or arising out of the use of theContent.
This article may be used for research, teaching, and private study purposes. Any substantialor systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, ordistribution in any form to anyone is expressly forbidden. Terms & Conditions of access and use canbe found at http://www.tandfonline.com/page/terms-and-conditions
International Journal of Systems Science, 1999, volume 30, number 9, pages 981± 988
Development of a system to diagnose autonomous underwater
vehicles
Motoyuki Takai² and Tamaki Ura²
In the underwater environment, installation of a proper scheme to cope with unexpectedsituations is essential for Autonomous Underwater Vehicles (AUVs) to carry out theirmission out of human reach. This paper proposes a model-based approach of self-diagnosis for AUVs in order to supervise whether the vehicle operates itself in anappropriate way. The proposed self-diagnosis is carried out based on a dynamicsmodel of an AUV and an active mechanism to obtain desirable information. Thedynamics model is constructed by an arti® cial neural network taking advantage of its¯ exible learning capability. When a sensor is found to be defective, dead reckoningusing its corresponding output of the dynamics model can be introduced in an attempt tocope with the defect. The performance of the proposed system was examined by imple-menting it on the `Twin-Burger’, a test-bed AUV. It is shown that the system withoutthe introduction of extra sensors detects failures of onboard sensors and actuators, andthen selects a proper action scheme to minimize the damage to the AUV.
1. Introduction
In order to carry out a mission successfully under unpre-dictable conditions of high pressure and poor visibilityin the sea, it is essential for Autonomous UnderwaterVehicles (AUVs) to have high performance autonomy.This autonomy includes the ability to monitor the hard-ware condition of the vehicle. In general, there are twoways of monitoring the condition of the vehicle: (1) adirect monitoring method that supervises the hardwarecondition using additional sensors; and (2) an indirectmonitoring method which checks the relationshipbetween causes and results as they appear to the vehicle.
Figure 1(1) shows an example of method (1) withrespect to the monitoring of a propeller thruster.Usually, a counter is attached to measure the rotation.In addition to the counter, special sensors to observe thecounter and other monitoring devices might be attachedto realize higher reliability. Although this method isstraightforward, the vehicle’s hardware becomes com-plex and heavy.
An alternative method is illustrated in ® gure 1(2) asan example of indirect monitoring. It can be reasoned
that the AUV is in a faulty condition when the motionsensor indicates that the vehicle is running to the righteven though only the right thruster is running. Thisindirect monitoring can be realized by comparing thepredicted normal motion of the vehicle with the meas-ured motion: i.e. by comparing the outputs of adynamics model of the vehicle with the actual sensoroutputs (Milgram and Dormoy 1994).
In this paper, we propose a model-based diagnosissystem that contains a dynamics model constructedby an arti® cial neural network to predict sensoroutputs induced by the motion of the vehicle. In orderto diagnose in detail, the proposed system containsa mechanism to actively obtain the necessaryinformation for diagnosis. In the case where acertain fault is detected, the system can select anaction to cope with that fault. The performance of theproposed system is examined through tank tests using atest-bed AUV called a `Twin-Burger’ (Fujii et al.1993a, b).
2. S tructure of the self-diagnosis system
2.1. AssumptionsIn order to simplify the system, four assumptions are
made.International Journal of Systems Science ISSN 0020± 7721 print/ISSN 1464± 5319 online Ñ 1999 Taylor & Francis Ltd
http://www.tandf.co.uk/JNLS/sys.htmhttp://www.taylorandfrancis.com/JNLS/sys.htm
Accepted 3 August 1998.
² Institute of Industrial Science, University of Tokyo, 7-22-1Roppongi Minato-ku Tokyo, 106, Japan. Fax: + 81-3-3401-6259;e-mail: [email protected];[email protected]
Dow
nloa
ded
by [
Uni
vers
ity o
f N
orth
Car
olin
a] a
t 08:
11 1
2 N
ovem
ber
2014
(1) The vehicle is in a steady environment: i.e. a changeof the characteristics of the vehicle’s motion dependson its hardware condition only.
(2) At any instance, only one of the onboard sensorsand actuators may be defective.
(3) For a particular sensor or an actuator, the conditionis either normal or totally defective.
(4) The defective condition lasts until the end of thediagnosis.
2.2. Sub-systems of the self-diagnosis systemThe Preopposed self-diagnosis system consists of two
sub-systems.
(1) Model Matching Part (MMP)The MMP includes a dynamics model that represents
the characteristics of the vehicle’s motion. The MMPproduces index values that result from the comparisonof sensor outputs with the corresponding model outputs(cf. ® gure 2).
(2) Diagnosis Part (DP)In the DP, the index values fromthe MMPare used to
identify the defective component of the vehicle and the
next action is how to cope with the defect. When theinformation to identify the defective component is insuf-® cient, the DP elects an appropriate prede® ned controlsequence for the vehicle, for the purpose of collectingnecessary information (cf. ® gure 3).
982 M. Takai and T. Ura
Figure 1. Two ways of self-diagnosis.
Figure 2. S tructure of the Model Matching Part.
Figure 3. Structure of the Diagnosis part.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
orth
Car
olin
a] a
t 08:
11 1
2 N
ovem
ber
2014
2.3. Routine diagnosis and mission diagnosisAs shown in ® gure 4, the proposed system introduces
two diagnosis procedures which are named `Routinediagnosis’ and `Mission diagnosis’ in accordance withthe situation when the diagnosis is carried out.
2.3.1. Routine diagnosis. The Routine diagnosis (Takaiet al. 1995) is designed to be carried out exclusively.For example, the vehicle executes the Routine diagno-sis before starting a mission in order to check thehardware.
In the Routine diagnosis, the vehicle operates basedon one of the prede® ned control sequences called`Diagnosis Motion Sequences (DMSs)’. The DMS is a® xed time history of control commands to the actuators.In order to design suitable DMSs for the self-diagnosis,it is essential to take into account the fact that the char-acteristics of the vehicle’s dynamics have to be measuredby onboard sensors. By carrying out one of the DMSs,the MMP compares the sensor outputs with the modeloutputs and calculates index values of their di� erences.Then, the DP determines whether the vehicle is in anormal condition or not by referring to a prede® nedtable, names a `Pattern table’.
In some cases, execution of only one DMS may beinsu� cient to identify the defective component. Then,the Routine diagnosis selects and carries out anotherDMS to narrow down the doubtful components.
2.3.2. Mission diagnosis. The Mission diagnosis is in-troduced to supervise the hardware condition of anAUV throughout a mission. The MMP continuallycompares the model outputs with the sensor outputsduring the mission, and the DP veri® es whether a faultexists or not. When a fault is detected, the systemselects an appropriate action in order to cope withthe fault: e.g. interrupt the mission or carry out theRoutine diagnosis to identify the fault.
2.4. Structure of the model matching part2.4.1. Dynamics model. In the underwater environ-ment, the characteristics of the AUV’s dynamics areusually nonlinear. Therefore, its mathematical modelshould represent this nonlinearity. Furthermore, thecomputational load of calculating the model outputshas to be minimized because the payload of the AUVis, in general, strictly limited.
Considering these situations, an arti® cial neural net-work is suitable for representing the nonlinear dynamicsof the system, taking advantage of its ¯ exible leaningcapability (Fujii and Ura 1990, 1994). It should beemphasized that the constructed network can beadapted quickly for changes in the vehicle’s dynamicsdue to modi® cations of the vehicle’s hardware: e.g.attaching new sensors.
2.4.2. Structure of the dynamics model. The structureof the dynamics model shown in ® gure 5, is based on
Diagnosis of autonomous underwater vehicles 983
Figure 4. The Routine diagnosis and the Mission diagnosis.
Figure 5. Dynamics model constructed by arti® cial neural
networks.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
orth
Car
olin
a] a
t 08:
11 1
2 N
ovem
ber
2014
an arti® cial neural network which considers the follow-ing ® ve points.
(1) The inputs of the neural network are control com-mands to actuators and sensor outputs which corre-spond to the state variables of the vehicle.
(2) In order to train quickly and re® ne the network, theoutputs of the output layer are the changes of thesensor outputs during a speci® c time interval ¢t(Suto et al. 1994).
(3) In order that the neural network generates outputswithout being a� ected by the failure of a sensor,input state variables to the input layer are giventhrough RC_1 after integrating the outputs of theoutput layer (Ishii et al. 1994). Therefore, the initialvalues of inputs have to be based on the actual out-puts of the sensors.
(4) The neural network should have multiple inputs andmultiple outputs to cover the full degrees of thedynamics’ freedom, because there exist non-trivialinteractions among modes of dynamics.
(5) The neural network has recurrent connections RC_2that hold the past sequences of the vehicle’s statevariables.
The neural network is su� ciently trained by dividingthe process into stages using the Error BackPropagation method (Ishii et al. 1995, Takai et al. 1995).
2.4.3. Model matching process. When the dynamics ofthe vehicle have been changed, the magnitude of thechange can be expressed by introducing the index vec-tor fgg whose elements gi are de® ned in equation (1):
gi(k) =100KRi
k
j=k KjSSi( j) SMi( j) j, (1)
where i denotes the physical quantity to be measured bya sensor, and k the time step. SSi and SMi are sensor andmodel outputs respectively, which correspond to index i.Ri is a constant to normalize the di� erence, and K is asample time to be averaged. The gi is expressed as apercentage.
2.5. Structure of the Diagnosis partThe DPidenti® es the defective component and selects
an appropriate action to cope with the defect using theindex vector fgg obtained from the MMP (cf. ® gure 3).
In order to realize the function of the DP, gi is trans-formed to a binary number bi as de® ned in equation (2):
bi =0 if gi < ti
1 otherwise,(2)
where ti is the threshold value in index i. The thresholdvalues should be determined experimentally in advanceby examining a time history of gi when the componentof the vehicle is defective. Then, fbg is a vector whoseelements are either 0 or 1.
The DPclassi® es the vector fbg on the basis of a tablewhich is named the `Pattern Table’. The Pattern Tabledescribes the correspondence between a vector fbg andafault. For example, in the case of the vehicle shown in® gure 1 with three thrusters and motion sensors forsurging, yawing and heaving velocities, the PatternTable can be determined as shown in ® gure 3. Whenthe vector fbg is f1, 0, 0g of Case 2, it is concludedthat the velocity sensor for surging is defective.
In the Routine diagnosis, the DP classi® es the vectorfbg after a DMS has been executed. There exist, how-ever, some cases where the system cannot identify thedefective component: e.g. in the case of ® gure 3, thesystem can only conclude that the Thruster 0 or 1 maybe defective when the vector fbg is f1, 1, 0g. For thosecases, the DP selects the following DMS to acquire thenecessary information for identi® cation: this recursiveexecution of model matching and diagnosis is named`Active diagnosis’ .
In the Mission diagnosis, the DPcontinually classi® esfbg when the MMP calculates the fgg throughout themission. Even if one element of fbg becomes 1, thesystem stops the mission and moves to an appropriateaction scheme: e.g. dispatching emergency signals, dis-charging emergency ballast, ascending to the surface, orreturning to the operator using S̀ubstituting control’, asdescribed next. When the defective component is notidenti® ed, the Routine diagnosis can be followed.
2.6. Substituting ControlWhen a sensor failure is identi® ed, the sensor outputs
SSi cannot be used in the feedback control loop.However, the outputs of the dynamics model SMi areavailable instead of SSi for control, as shown in ® gure6, disregarding the small estimation error. Con-sequently, the vehicle can proceed with the missionusing SMi on such an occasion. This control scheme isa kind of dead reckoning, and is named S̀ubstitutingControl’.
3. Implementation of self-diagnosis
The proposed system is implemented on an AUV namedthe `Twin-Burger’ (Fujii et al. 1993a, b) as shown in® gure 7. The Twin-Burger is designed as a versatiletest-bed AUV, and is equipped with a parallel pro-cessing computer system and many sensors: e.g. avision sensor, ranging sensors, ¯ ow sensors and a com-pass. The target components of diagnosis are those
984 M. Takai and T. Ura
Dow
nloa
ded
by [
Uni
vers
ity o
f N
orth
Car
olin
a] a
t 08:
11 1
2 N
ovem
ber
2014
involved in controlling surging and yawing motions: i.e.the `Flow sensor’ which measures the longitudinal rela-tive ¯ ow velocity, the `Yaw rate sensor’, and actuators`Thruster 0’ and `Thruster 1’.
3.1. Construction of model matching part3.1.1. Diagnosis Motion Sequences (DMSs). Threekinds of zigzag test sequences are prepared for DMSsas illustrated in ® gure 8. Sequence `ZA’, where bothThruster 0 and 1 are concurrently actuated, is designedto be carried out as the ® rst step in the Routine diag-nosis. In `ZB’ and `ZC’ sequences, only one thruster(Thruster 0 or 1) is actuated.
3.1.2. Construction of dynamics model. An arti® cialneural network is constructed to represent the dy-namics of the Twin-Burger based on the structureshown in ® gure 5. The inputs to the input layer arecontrol commands to each thruster and sensor out-puts. The outputs of the output layer are the di� er-ences of sensor outputs in a speci® c time interval ¢t.Data for training the neural network are sampled byoperating the vehicle in a normal condition accordingto the ZA, ZB and ZC sequences and also with a mis-sion scenario consisting of clockwise and counterclock-wise rectangular-path (5 2.5m) following.
The neural network is trained until its model outputerrors are reduced to about 0.0175m/s for the FlowSensor and 0.0168rad/s for the Yaw Rate Sensor.Figure 9 shows the time history of training data andtrained model outputs of a path following mission. Itcan be seen that the dynamics of the vehicle are wellrepresented by the trained neural network.
3.2. Construction of diagnosis part3.2.1. Threshold values. In order to determine thethreshold values for binary transformation in the DP,the Twin-Burger was operated in defective conditionsby decreasing the power of each thruster. The thresh-
Diagnosis of autonomous underwater vehicles 985
Figure 6. Substituting control.
Figure 7. Top and side view of the Twin-Burger.
Figure 8. Diagnosis Motion Sequences (DMSs).
Dow
nloa
ded
by [
Uni
vers
ity o
f N
orth
Car
olin
a] a
t 08:
11 1
2 N
ovem
ber
2014
old values were determined as shown in table 1. Thethreshold values for the Mission diagnosis are largerthan those for the Routine diagnosis considering thevariety of missions.
3.2.2. Pattern table and active diagnosis. The Patterntable for yaw and surge motions of the Twin-Burger isdetermined as in table 2 by examining the con® gura-tion and the performance of onboard sensors andthrusters. When fbg is calculated as f1, 0g after theZA sequence is carried out, it is concluded from table2 that the Flow Sensor is defective. But when fbg iscalculated as f1, 1g the system only concludes thatThruster 0 or 1 may be defective. In this case, the ZBor ZC sequence should be carried out in the Active
diagnosis in order to identify which is defective. Whenthe fbg is not f0, 0g after the execution of the ZBsequence, it is concluded that the Thruster 0 isdefective because only Thruster 0 is actuated in the ZBsequence. The redundancy of the diagnosis can beincreased by carrying out the ZC sequence in thesame way.
4. Experiment
In this section, we present a Mission diagnosis experi-ment, using the Twin-Burger, which demonstrates thatthe proposed self-diagnosis system can detect a sensorfailure which occurs during a mission. The vehicle fol-lows a rectangular-path (4 2m) keeping 0.7m deep asillustrated in ® gure 10. The Flow Sensor is covered at115s when the vehicle almost ® nishes turning to the
986 M. Takai and T. Ura
Table 1. Threshold values
g value thresholds [%]
Flow Yaw RateSensor Sensor
ZA 8 7Routine diagnosis ZB 8 7
ZC 8 7Mission diagnosis 15 10
Table 2. The pattern table de® ned for the Twin-Burger
State0Ð normal 1Ð abnormal
Flow Yaw Rate DefectiveCase Sensor Sensor component
0 0 0 Nothing1 1 1 Thruster 0 or
Thruster 12 1 0 Flow Sensor3 0 1 Yaw Rate Sensor
Figure 9. Sensor outputs and trained model outputs.
Figure 10. Experiment of the mission diagnosis.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
orth
Car
olin
a] a
t 08:
11 1
2 N
ovem
ber
2014
right and is starting to move forward. The output of thesensor is then zero, as shown in ® gure 11.
At about 130s, b(FlowSensor) becomes 1 because thevalue of g(FlowSensor) is larger than the threshold value15% as in table 1. Then, the DP concludes that thevehicle is not in the normal condition. It takes about15s to detect the fault after its occurrence because theforward velocity of the vehicle is very small when thefault occurs. It is identi® ed that the Flow Sensor isdefective because the calculated fbg = f1, 0g corre-sponds to Case 2 in table 2. After the fault detection,the vehicle proceeds with the mission scenario, based onSubstituting Control. Finally, the vehicle gets to thepredetermined destination even though the FlowSensor is malfunctioning.
The results of this experiment show that the proposedself-diagnosis system can identify the fault and can suc-ceed in completing the given mission scenario based onSubstituting Control.
5. Conclusions
The work described in this paper demonstrates a faultdetection scheme for AUVs based on a dynamics modelconstructed by an arti® cial neural network. Theproposed self-diagnosis system has three signi® cantfeatures: (1) it contains a neural network to representthe nonlinear dynamics of the AUV; (2) it has amechanism to obtain the necessary information by`Active diagnosis’; and (3) in the case where a certainfailure is detected, the system selects an appropriateaction to cope with the defect, including S̀ubstitutingcontrol’. These functions are achieved by the `Routinediagnosis’ and the `Mission diagnosis’ whose structuresconsist of two subsystems: the Model Matching Part(MMP) and the Diagnosis Part (DP).
The performance of the proposed system is examinedthrough tank tests using a test-bed AUV. It is shownthat the system can detect failures of onboard sensorsand actuators without the introduction of extra sensors,
Diagnosis of autonomous underwater vehicles 987
Figure 11. Experimental results of
the mission diagnosis.
Dow
nloa
ded
by [
Uni
vers
ity o
f N
orth
Car
olin
a] a
t 08:
11 1
2 N
ovem
ber
2014
and selects a proper action in order to reduce thedamage to the AUV. It can be concluded that inclusionof the proposed self-diagnosis system in the AUVs’ soft-ware increases its reliability and of emergency manage-ment ability.
References
Fujii, T., and Ura, T., 1990, Development of motion control systemfor AUV using neural nets, Proceedings of IEEE AUV 9̀0,Washington, DC, pp. 81± 86; 1994, Self-generation of neural-netcontroller by training in natural environment. ComputationalIntelligence: Imitating Life, edited by J. M. Zurada, R. J. MarksII, and C. J. Robinson (IEEE Press), pp. 340± 351.
Fujii, T., Ura, T.,Kuroda, Y.,Chiba, H.,Nose, Y., and Aramaki,K.,1993a, Development of a versatile test-bed `Twin-Burger’ towardrealization of intelligent behaviours of autonomous underwatervehicles. Proceedings of IEEE OCEANS 9̀3, Victoria, pp. 186± 191.
Fujii,T.,Ura, T., and Kuroda,Y.,1993b, Mission execution experi-ment with a newly developed AUV The Twin-Burger. Proceedings ofthe Eighth International Symposium on Unmanned UntetheredSubmersible Technology, UUST 9̀3, New Hampshire, pp. 92± 105.
Ishii, K.,Fujii, T., and Ura, T.,1994, A quick adaptation method ina neural network based control system for AUVs. Proceedings ofIEEE AUV 9̀4, Cambridge, pp. 269± 274; 1995, An on-line adapta-tion method in a neural network based control system for AUVs.IEEE Journal of Oceanic Engineering, 20, 221± 228.
Milgram, J., and Dormoy, J. L., 1994, Monitoring a model-basedsensor failure detection system. ISAP 9̀4, International Conferenceon Intelligent System Application to Poser Systems, pp. 135± 142.
Suto,T., and Ura,T.,1994, Self-generation of controller of an under-water robot with neural network. Proceedings of ISOPE 9̀4, pp.362± 365.
Takai, M., Fujii, T., and Ura, T., 1995, A model based diagnosissystem for autonomous underwater vehicles using arti® cial neuralnetworks. Proceedings of the Ninth International Symposium onUnmanned Untethered Submersible Technology, UUST 9̀5, NewHampshire, pp. 243± 252.
988 Diagnosis of autonomous underwater vehicles
Dow
nloa
ded
by [
Uni
vers
ity o
f N
orth
Car
olin
a] a
t 08:
11 1
2 N
ovem
ber
2014
