J. Cent. South Univ. (2016) 23: 808-816

DOI: 10.1007/s11771-016-3127-8

Actuator fault diagnosis of autonomous underwater vehicle based on improved elman neural network

SUN Yu-shan(孙玉山)^{1, 2}, LI Yue-ming(李岳明)^{1, 2}, ZHANG Guo-cheng(张国成)^{1, 2},

ZHANG Ying-hao(张英浩)^{1, 2}, WU Hai-bo(吴海波)^{1, 2}

1. Science and Technology on Underwater Vehicle Laboratory, Harbin Engineering University, Harbin 150001, China;

2. College of Shipbuilding Engineering, Harbin Engineering University, Harbin 150001, China

 Central South University Press and Springer-Verlag Berlin Heidelberg 2016

Abstract: Autonomous underwater vehicles (AUV) work in a complex marine environment. Its system reliability and autonomous fault diagnosis are particularly important and can provide the basis for underwater vehicles to take corresponding security policy in a failure. Aiming at the characteristics of the underwater vehicle which has uncertain system and modeling difficulty, an improved Elman neural network is introduced which is applied to the underwater vehicle motion modeling. Through designing self-feedback connection with fixed gain in the unit connection as well as increasing the feedback of the output layer node, improved Elman network has faster convergence speed and generalization ability. This method for high-order nonlinear system has stronger identification ability. Firstly, the residual is calculated by comparing the output of the underwater vehicle model (estimation in the motion state) with the actual measured values. Secondly, characteristics of the residual are analyzed on the basis of fault judging criteria. Finally, actuator fault diagnosis of the autonomous underwater vehicle is carried out. The results of the simulation experiment show that the method is effective.

Key words: autonomous underwater vehicle; fault diagnosis; thruster; improved Elman neural network

1 Introduction

Underwater vehicles have been applied in many fields all around the world, such as marine research, scientific investigation, ocean development and underwater engineering. For the broad prospects in maritime research and ocean development, the technology of underwater vehicles has been an important and positive study field among the world’s powerful countries [1-3]. Underwater vehicles also have a place in the fields of underwater information acquisition, precision strikes and asymmetric intelligence wars.

Because an underwater vehicle’s working condition is usually complex and its loss will cost a lot of financial damages, an underwater vehicle should have the ability of automated diagnosis for fault detection, identification and recovery in order to deal with the emergency situation. The automated diagnosis for fault detection of an underwater vehicle’s control system is an important embodiment of its intelligence [4-6].

The appearance of the neural network provides a new train of thought and method to solve the issue of fault diagnosis, especially for the nonlinear systems and complex systems. The fault diagnosis based on the analytical model faces the practical difficulties of establishing mathematical model, while the fault diagnosis based on knowledge becomes an important and feasible method. The characteristics of the neural network, such as the I/O nonlinear mapping features, the distribution of information storage, parallel processing and global collective action, especially the high degree of self-organization and self-learning ability make it an effective mean of fault diagnosis, and the neural network has been successfully used in many practical systems [7].

The tradition identification method has significant limitation to a dynamic nonlinear system like an underwater vehicle’s because its movement is multi- degree of freedom and strongly nonlinear, besides that, the modification during development and the frequent changes of the payload during the practical process alter the balance of its shape and weight, which makes the characteristics of underwater vehicle’s movement modify accordingly. Due to neural network’s ability which can approximate any nonlinear mapping with any degree of accuracy; its inherent learning ability which can reduce the uncertainty and increase the generalization ability to adapt to the environmental varieties and its structure of the distributed information storage and process which makes it has the nature of fault tolerance, adopting the dynamic neural network and making it study the motion data through certain learning algorithm data to recognize movement characteristics of an underwater vehicle, which can be used as an identifier for an adaptive controller, an virtual sensor or the basic information for fault diagnosis. And that is to use neural network to establish the motion model of an underwater vehicle in order to solving the corresponding problems.

Pang et al [8] introduced an algorithm, which is applied to build the dynamic model of an AUV, based on the character that hidden layer wavelet function of wavelet networks can adjust scale factor and shift factor to affect output of networks via comparing the output of neural network with the real state value under the condition of fault to build detection rules for the thruster fault diagnosis. Wang and DING [7] applied an improved wavelet neural network to movement modeling and made the actuator fault diagnosis of an underwater vehicle by comparing the model output and the actual measured values, which is tested by simulation experiment. Wu [9] had also tried to apply wavelet neural network to actuator fault diagnosis of underwater vehicles. Zhang and CHUN [10] proposed a fault detection and isolation method of underwater vehicle thrusters based on observer, and the fault reconstruction approach with the combination of RBF neural networks is also introduced to restructure the thruster fault from the neural network output. Zhu and SUN [11] tried to apply the fuzzy cerebellar model articulation controller (FCA-CMAC) neural network information fusion cerebellar model in the thruster fault identification and diagnosis for open-frame underwater vehicles,and carried out pool experiment.

Antonelli et al [12] adopted the support vector machine (SVM) for offline training to improve the underwater vehicle’s model and used a radial basis function (RBF) network for actuator fault diagnosis. The fault diagnosis and fault-tolerant control system (FDAS), which is invented by machinery research center of Wales University, for the open-frame underwater vehicles “URIS” (University of Girona) and “FACLON” (SeaEye Marine ltd.) was composed of two subsystems: fault tolerant control subsystem (FAS) and fault diagnosis system (FDS) based on self-organizing maps and fuzzy clustering analysis used fault detection unit (FDU) corresponding to each thruster to test its state. FDS is a kind of hybrid and on-line method that has nothing to do with the model. In the training phase, FDS obtained self-organizing mapping feature according to the data from tests and in fault detection phase, the type of fault is determined by comparing the position of the feature vector in the mapping table [13]. Researchers from Italian navy institute of automation have applied the extended kalman filter (EKF) and the sliding mode observer (SMO) to fault diagnosis for a ROV called “Roby 2” which is based on the approximate hydrodynamic model. They have also tried to use neural network in fault diagnosis for underwater vehicles [14]. Researchers from National Oceanography Centre in Southampton, Emits et al [15], and EZEKIEL et al [16] have also developed automatic fault detection and execution monitoring methods for AUV missions, and try to applied them on Autosub 6000 AUV.

2 An improved elman neural network

2.1 model

Due to the existence of the feedback from the output variable to the input, recursive network (also called feedback network) contains time delay network in its variables and is the real dynamic network system. Compared with the static neural network, recursive network does not need to presuppose the system order and has opened up a very promising area for the dynamic system identification and control. Because of the inherent feedback structure of dynamic recursive neural network, just single-layer network can express complex dynamic systems as well as the dynamic process of approximation system. Common recursive neural networks are the Hopfield net, Elman network, Jordan, RMLP (recurrent multi-layer perceptron) network and so on.

Basic Elman network is added a layer “connection” on the base of BP network, which will be achieved by linking memory on a moment of the hidden layer state with the current moment input as the input of hidden layer, the equivalent of state feedback. The input are weighted sum of each layer. The transfer function of hidden layer as a nonlinear function, still generally as the Sigmoid function, the output layer is a linear function or nonlinear function, and link layer as the linear function, the network structure is shown in fig. 1.

The improved Elman network structure shown in fig. 2, which is introduced to the motion model structure of AUV. By comparing figs. 1 and 2, we can see the difference is that a self-feedback connection of fixed gain α is added in the connection unit x_C of the improved Elman network. therefore, the output of the connection unit x_C at k time is equal to the output of the hidden layer at (k-1) time plus α times the output of the connection unit at (k-1) time, that is x_C,l(k)=ax_C,l(k-1)+x_l(k-1), l=1, 2, …, n, among them x_C,l(k)  and x_l(k-1)  are respectively the output of the number l connecting unit and the number l hidden layer unit, while α is the self-connection feedback gain. In addition, considering that the feedback of each layer neurons will affect the signal processing ability of the network and increase the feedback of the output layer node, the improved Elman network has faster learning convergence speed and generalization ability compared with basic Elman network, it has stronger recognition for the high order nonlinear system.

Fig. 1 structure of Elman network

The nonlinear state space expression described by the improved Elman network is

       (1)

                   (2)

                    (3)

                            (4)

Fig. 2 Improved Elman network

2.2 algorithm

In this work, it mainly use an improved dynamic back-propagation learning algorithm for network learning and training, the inter-layer weights adjustment method is as follows.

Consider the following overall error of the objective function:

                                   (5)

where

The connection weights from the hidden layer towards the output layer W³

               (6)

order  so

        (7)

The connection weights from the input layer towards the hidden layer W²

                (8)

Also, order  so

      (9)

Familiar with the connection weights from the structural layer unit towards the hidden layer W¹, and

 ;               (10)

Ignore the dependencies between  connection weights  so

and

so

The formula actually constitutes a dynamic recurrence relationships of gradient

Similarly, for W⁴:

                 (11)

                (12)

Owing to

Therefore, an improved Elman neural network dynamic back-propagation learning algorithm can be summarized as follows:

      (13)

   (14)

 (15)

                    (16)

where η₁, η₂, η₃, η₄, are learning step-lengths of W¹, W², W³ and W⁴, respectively.

3 Motion modeling and Identification AUV based on Neural Network

In this work, the research AUV equipped with eight thrusters. According to the applying thrust function, thrusters can be divided into four groups: vertical main-thrusters, horizontal main-thrusters, vertical and lateral thrusters, each part is made up of two thrusters. Main thrusters adopt ducted thrusters: the maximum thrust of horizontal main-thrusters can reach 600 N and the angle of the axis and longitudinal axis is 13°. The maximum thrust of vertical reach 210 N and angle of the axis and longitudinal axis of the AUV is 26°. Vertical and lateral thrusters both adopt channel thrusters. With the increase of the AUV surge speed, thrust deduction factor dramatically increased. To save energy, four channel thrusters closed in high-speed. This work mainly discusses four main thrusters on fault diagnosis.

We apply the improved Elman network shown in fig. 3 to the underwater vehicle modeling. The autonomous underwater vehicle is equipped with the Doppler velocimeter for measurement of three linear velocities, the compass for measurement of three angles and 8 thrusters (including two horizontally arranged main thrusters, two vertically arranged main thrusters, a lateral thruster and a vertical thruster at the stem and the stern). According to the sensors and actuators actually configurated on the underwater vehicle, input and output of the network are set as follows:

where u, v, w, p, q, r are underwater vehicle’s surge speed, sway speed, heave speed, roll angular velocity, pitch angle velocity, and angular rocking bow. r₁, r₂, … and r₈ are Underwater vehicle’s eight thruster voltage command.

Fig. 3 AUV actuator arrangement

The number of hidden layer nodes takes the value of 28, the learning rate η₁=η₂=η₃=η₄=0.02, the momentum factor takes the value of 0.5. Before putting the system into use, we should train the neural network. In order to accelerate the convergence speed and reduce the training time, we mainly use experiments of fore, fixed-point uniform direct flights, direct for training the network. Trained improved Elman network can better simulate the underwater vehicles movement. Figure 4 shows the path curve of one target search at 1 m deep underwater during the sea trials in Penglai. Figure 5 shows the identification results of surge velocity, sway velocity and yaw angle.

Fig. 4 result of search in marine area

Fig. 5 Identification result:

4 Underwater vehicles actuator fault detection simulation

Working state detection of the actuator is actually to judge if the actuator can provide the normal needs of thrust. As an underwater vehicles works in complex marine environment, force created by the actuator will also change with the environment (e.g., current). Due to the influence of many factors, such as the modeling error, the measurement noise and the external disturbance, precise modeling of the actuator is very difficult. The appearance of the neural network technology has provided a new solution for fault detection and diagnosis problems. The characteristics of the neural network, such as the I/O nonlinear mapping features, the distribution of information storage, parallel processing and global collective action, especially the high degree of self-organization and self-learning ability make it an effective method and means for fault diagnosis and the neural network technology has been successfully used in many practical systems.

4.1 Fault diagnosis strategy and process

Figure 6 shows fault diagnosis process based on improved Elman neural network. AUV collects sensor information, such as carrier’s attitude and speed. Motion controller calculates and distributes of thrust according to control algorithm based on control target instructions and motion sensor information, calculates each thruster’s thrust voltage, at the same time controls the output information transmission to the improved Elman neural network. We can calculate residual information by comparing outputs, which based on improved Elman neural network model and sensor information, use logical rules to judge residual and get thruster fault diagnosis conclusion.

Fault diagnosis rules are listed in table 1. Horizontal main-thruster’s fault diagnosis can be executed by yaw angle residual and other residual. Residual of surge velocity and yaw angle will exceed threshold value when a horizontal main-thruster occurs faults: the yaw angle will turn minus when occurs left main-thruster fault while turn positive when occurs right main-thruster fault. Vertical main-thruster’s fault diagnosis can be executed by analyzing pitch angle residual and heave velocity residual. Both pitch angle residual and heave velocity residual will exceed threshold value when a vertical main-thruster occurs faults: the pitch angle will turn positive minus when occurs upper main-thruster fault while turn minus when occurs under main-thruster fault. If surge velocity, yaw angle and other general residual appears alone, we need to be further diagnosis.

Fig. 6 Flowchart of AUV thrusters based on improved Elamn netural network

Table 1 Fault mode table

4.2 Simulation test

We used the improved Elman neural network for propeller fault detection and diagnosis of underwater vehicles. Network structure of 8×28×6 was adopted, network input and output were as mentioned in section 3, the learning rate η₁=η₂=η₃=η₄=0.02, the momentum factor took the value of 0.5. Before putting the system into use, we carried on data training such as the uniform fore, fixed-point direct flights, such as training data. It can be seen from figs. 3 and 4 that this network model can well simulate the movement of the underwater vehicles. By comparing the output of the underwater vehicles model with the actual measured values, we get 6 groups of residual. Then, we can infer whether there is a thruster fault and in which thruster the fault occurs through analysing the residual which contains a large number of fault information of the underwater vehicles. In order to reduce the influence of environment noise on the residual as far as possible, this paper adopts the following way to analyze the residual information: firstly, a set of residuals in a fixed time interval is recorded and the maximum and minimum value are removed; secondly, the average value of the residual is calculated. If the residual is greater than the given threshold (threshold and the length of time interval is set in advance through a lot of experiments and experience), we can consider a failure.

We conducted related simulation test, figs. 7(a)-(c) give the simulation test results when the left main propulsion failures. According to fig. 7, we can see there is no fault at the beginning and failure occurs on the left main propulsion from the 420 s. It can be seen from the diagram that the output (estimate) of the improved Elman neural network is in line with the actual measured value in the absence failure and a deviation of the surge velocity estimate appears when the main propulsion failures. With the accumulation of time, the deviation will become larger to a stable value which is fixed under a steady speed. The deviation of the shake bow angle also appears and becomes larger. When the residual error exceeds the set threshold, we can determine fault occurs. Residuals of the surge velocity can determine a main failure, according to shake bow Angle we can determine the residual error of plus or minus is left on the fault. Figure 8 shows the residual error curve when the vertical main thruster failures from the 170 s began to change. Deviation of the surge velocity estimate as well as the heave velocity estimate appear when the main propulsion failures. With the accumulation of time, the deviation will become larger to a stable value which is fixed under a steady speed on the surface of the surge velocity and will appear deviation, and accumulate over time, the deviation will be more and more big to estimate to a stable value, there is a fixed deviation under the steady speed; And trim angle deviation, but the main failure underwater vehicles will trim residual is positive, the main failure underwater vehicles trim residuals is negative, it can be judged whether the main fault is the main fault.

4.3 Sea trial

The AUV system has conducted its sea trial to test and verify the relevant performance indicators. The experiment contents include: motion control, underwater navigation, fault diagnosis, underwater topography detection, motion target following etc. Figure 9 is thefault diagnosis results in an experiment. During this experiment, the AUV sailed from static to uniform velocity -1 m/s. After AUV stable, at 160 time step, we gave right main-thruster a zero voltage to simulate the right main-thruster occurs a fault. From zero to 159 time step as shown in Fig. 9, there were no faults occurred, and surge velocity residual information approaching zero value. But when main-thrusters occurred faults, bias would appear in surge velocity residuals, and cumulative bias gradually increase with time, finally stable around a certain deviation while yaw angle residual increased andyaw angle fore drift to the right.

Fig. 7 Fault diagnosis based on inproved Elman network (left main-thruster fault)

Fig. 8 Residuals:

From zero to 53 time step, within control adjustment stage, the AUV sailed from static to uniform velocity. The surge velocity residual exceeded its threshold value while yaw angle residual nearly zero. When faults occurred, the surge velocity residual and yaw angle residual both would exceed their threshold value. We can judge AUV is in control adjustment stage or Fault diagnosis according to these principles, so as to effectively prevent misjudgment.

Fig. 9 Autonomous diagnosis results as horizontal-main right thruster’s fault:

5 Conclusions

1) Aiming at the characteristics of the underwater vehicle which has uncertain system and modeling difficulty, this work puts forward an improved Elman neural network and apply it to the underwater vehicles motion modeling. the residual is calculated by comparing the output of the underwater vehicles model (estimation in the motion state) with the actual measured values, and actuator fault diagnosis of the autonomous underwater vehicle is carried out.

2) Through increasing self-feedback connection with fixed gain in the unit connection, the improved Elman network has improved the state feedback of the network and network dynamic reflection ability.

3) Improved Elman neural network conveys feedback to the input layer from the output layer, so it can reflect the output characteristics of the system. Compared with the basic neural network, the improved neural network has faster learning convergence speed and generalization ability, as well as stronger recognition ability towords high-order nonlinear systems.

References

[1] XU Yu-ru, PANG Yong-jie, Gan Yong, SUN Yu-shan. AUV-state-of-the-art and prospect [J]. CAAI Transactions on Intelligent Systems, 2006, 1(1): 9-16. (in Chinese)

[2] SUN Yu-shan, WAN Lei, GAN Yong, WANG Jian-guo, JIANG Chun-meng. Design of motion control of dam safety inspection underwater vehicle [J]. Journal of Central South University, 2012, 19(6): 1522-1529.

[3] Kalwa J. Underwater inspection using the AUV sea cat [C]// Proceedings of 7th Security Research Conference. Bonn, Germany, 2012: 311-314.

[4] WAN Lei, YANG Yong, LI Yue-ming. Actuator fault diagnosis of automatic underwater vehicle using gaussian particle filter [J]. Journal of Shanghai jiao tong University, 2013, 47(7): 1072-1076. (in Chinese)

[5] LIU Qian, ZHU Da-qi. Fault diagnosis and fault tolerant control algoriythm of thrusters of unmanned underwater vehicles [J]. Journal of system simulation, 2010, 22(1): 96-101. (in Chinsese)

[6] Choi J K, Kondo H, Shimizu E. Thruster fault-tolerant control of a hovering AUV with four horizontal and two vertical thrusters [J]. Advanced Robotics, 2014, 28(4): 245-256.

[7] WANG Li-rong, DING Kai. Actuator fault diagnosis of autonomous underwater vehicle based on wavelet neural network [J]. Journal of System simulation, 2007, 19: 206-209. (in Chinese)

[8] Pang Yong-jie, Fang Shao-ji, Wang Li-rong. Thruster fault diagnosis of autonomous underwater vehicle based on least disturbance wavelet neural network [J]. Shipbuilding of China, 2008, 49(2): 94-100. (in Chinese)

[9] Wu Jian. Research on multi-faults diagnosis technology of thruster and sensor for AUV [D]. Harbin: Harbin Engineering University, 2011. (in Chinese)

[10] Zhang Ming-jun, Chun Zhen-zhong. Thruster fault detection, isolation and reconstruction for autonomous underwater vehicle [J]. Journal of Nanjing University of Aeronautics & Astronautics, 2011, 43(s): 142-146.

[11] Zhu Da-qi, Sun Bing. Information fusion fault diagnosis method for unmanned underwater vehicle thrusters [J]. IET Electrical Systems in Transportation, 2013, 3(4): 102-111.

[12] Antonelli G, Caccavale F, Sansone C, Villani L. Fault diagnosis for AUVs using support vector machines [C]// 2004 IEEE International Conference on Robotics and Automation. New Orleans, USA, 2004: 4486-4491.

[13] Edin O, Geoff R. Thruster fault diagnosis and accommodation for open-frame underwater vehicles [J]. Control Engineering Practice, 2004(12): 1575-1598.

[14] Corradini M L, Monteriu A, Orlando G. An actuator failure tolerant control scheme for an underwater remotely operated vehicle [J]. IEEE Transactions on Control Systems Technology, 2011, 19(5): 1036-1046.

[15] Emits J, Dearden R, Pebody M. Automatic fault detection and execution monitoring for AUV missions [J]. IEEE Journal of Oceanic Engineering, Monterey, CA, United states, 2010:1-10.

[16] Ezekiel J, Lomuscio A, Molnar L. Verifying fault tolerance and self-diagnosability of an autonomous underwater vehicle [C]// 22nd International Joint Conference on Artificial Intelligence. Barcelona, Catalonia, Spain, 2011: 1659-1664.

(Edited by DENG Lü-xiang)

Foundation item: Project(2012T50331) supported by China Postdoctoral Science Foundation; Project(2008AA092301-2) supported by the High-Tech Research and Development program of China

Received date: 2015-01-21; Accepted date: 2015-06-11

Corresponding author: SUN Yu-shan, Associate Professor, PhD; Tel: +86-451-82519733; E-mail: sunyushan@hrbeu.edu.cn