Distributed Cooperative Fault Diagnosis Method for Internal Components of Robot Systems

Hitoshi Kono, Musab Obaid Alhammadi, Yusuke Tamura, Atsushi Yamashita & Hajime Asama
International Journal of Intelligent Systems and Applications in Robotics (IJRA), Volume (8) : Issue (1) :
2017 1
Distributed Cooperative Fault Diagnosis Method
for Internal Components of Robot Systems
Hitoshi Kono kono@robot.t.u-tokyo.ac.jp
Department of Precision Engineering
The University of Tokyo
Tokyo 113-8656, Japan
Musab Obaid Alhammadi musab.alhammadi@kustar.ac.ae
Department of Electrical and Computer Engineering
Khalifa University
Abu Dhabi 127788, United Arab Emirates (UAE)
Yusuke Tamura tamura@robot.t.u-tokyo.ac.jp
Atsushi Yamashita yamashita@robot.t.u-tokyo.ac.jp
Hajime Asama asama@robot.t.u-tokyo.ac.jp
Abstract
Robot systems have recently been studied for real world situations such as space exploration,
underwater inspection, and disaster response. In extreme environments, a robot system has a
probability of failure. Therefore, considering fault tolerance is important for mission success. In
this study, we proposed a distributed cooperative fault diagnosis method for internal components
of robot systems. This method uses diagnostic devices called diagnosers to observe the state of
an electrical component. These diagnosers execute each diagnosis independently and in parallel
with one another, and it is assumed that they are interconnected through wireless
communication. A fault diagnosis technique was proposed that involves gathering the diagnosis
results. Further, computer simulations confirmed that the distributed cooperative fault diagnosis
method could detect component faults in simplified fault situations.
Keywords: Fault Detection, Distributed Cooperative System, Internal Component, Robot
System.
1. INTRODUCTION
Robot systems have recently been deployed in many real world situations. Among other
applications, mobile robot inspection systems for disaster-stricken areas, underwater inspection,
and space satellites have been developed. The benefits of robot systems, particularly rescue
robots, have been demonstrated in extreme environments [1], [2]. Herein, robot system refers to
a mobile robot, such as rescue robots. These systems can decrease the risk associated with
dangerous work and increase work efficiency. However, it is difficult to understand the state of an

2017 2
extreme environment. A robot system has a probability of failure because extreme environments
are dangerous not only for humans but also for robot systems [3].
In this context, discussion of fault tolerance is important to understand the reliability of robot
systems. To date, many types of fault tolerance methodologies have been considered. For
example, model-based, signal-based, and model-free methods have been proposed [4], [5].
These approaches assumed that fault tolerance was considered when the robot system was
developed. However, many robot systems are already in operation worldwide, and it is not easy
to implement fault tolerance capabilities for these robot systems. Discussion regarding attachable
fault tolerances, specifically fault detection method and diagnosis for currently operational robot
systems, is important.
Previous research on general fault detection and diagnosis methods have focused on centralized
architecture, and the systems required a large number of sensors [6]–[8]. Moreover, in the model-
based method, there is a need to design dynamic and environmental models of robot systems
before operation, and real-time calculation is necessary. Therefore, previous work has the
following drawbacks:
 System architecture is centralized. If the main computer stops working, the fault diagnosis
function becomes invalid as well.
 Systems need environmental models, robot kinematic models, and real-time calculation.
 It is difficult to apply to currently-operated robot systems.
To increase the reliability of a robot system, a novel fault detection and diagnosis methodology
that considers the above drawbacks must be developed. To do so, our research focused on the
following features:
 Decentralized system architecture
 Model-free fault detection and diagnosis methodology
 Attachable system for currently operating robot systems
In this study, a distributed cooperative fault diagnosis method was proposed. This method
focused on fault diagnosis of electrical components in the robot system’s body because robot
systems typically consist of various electrical components. Although robot systems also have
mechanical components, this research considered only electrical components, such as
embedded computers, motor controllers, and motor drivers. To realize distributed fault diagnosis,
the implementation of a small diagnostic device, called a diagnoser, in every component was
proposed.
The rest of the paper is organized as follows. Section 2 describes the previous approaches.
Section 3 is an overview of the proposed distributed cooperative fault diagnosis method. Section
4 provides details about the conditions of the computer simulation experiments and results.
Section 5 contains concluding remarks.
2. PREVIOUS APPROACH TO FAULT TOLERANCE
Until now, numerous fault detection and diagnosis methods have been proposed. However, fault
detection and diagnosis are elements of fault tolerance technology. It is difficult to define each
technology independently, and the techniques overlap one another. To utilize previous research,
attention was not only paid to fault detection and diagnosis but also to related research on the
fault tolerance of a system.
2.1 Faults in Robot Systems
Okina et al. defined the fault of computer systems as the “difference between realizable and
required function” [6]. A robot system has outputs, such as action and behavior, based on inputs,
such as tele-operated commands and environmental information. In this study, based on the

2017 3
definition of the fault by Okina et al., the fault of the robot system was interpreted as that
occurring when the robot system gives an unexpected response to the input commands or
information.
On the other hand, a robot system is constructed from many types of electrical and mechanical
components; the robot system cannot be down in all components’ broken cause with exception of
explortion and radiation effect. Therefore the fault of a robot system is defined such as broken of
the some components. In detail, based on the fault definition by Okina, the fault state of a robot
system emerges when there are unexpected outputs from the internal components in the robot
system.
In addition, as the basic approach, it was assumed that the faulty component was the electrical
component in the robot body without sensors, such as a camera, encoder, or measurement
sensor for environmental information, because input information from the environment was
unknown or unexpected based on the above fault definition.
2.2 Typical Approach
As mentioned above, many kinds of fault detection and diagnosis methods for robot systems
have been discussed, such as model-based, signal–based, and model-free approaches. There
are traditional techniques for fault detection and fault diagnosis [4], [5]. For example, outside the
robot system domains, Lefebvre proposed an on-line fault diagnosis method using partially
observed petri nets [9]. A petri net is effective for the representation of the state of a system;
however, this method requires a model of the system, as is the case with model-based approach.
Okina et al. proposed a signal-based fault diagnosis method [6]. However, this type of system is
centralized and requires voltage and current values when the robot system is fault-free.
Moreover, this system needs invasive sensors, such as current sensors, to observe the current
state.
In the model-free approach, fault tolerance utilizing learning algorithms is a representative
technique. Liu et al. proposed a system using the credit assignment fuzzy cerebellar model
articulation controller (FCA-CMAC) neural network for unmanned underwater vehicles [10]. The
model-free approach does not require designing environmental and robot models. However, the
learner is constructed with a centralized architecture. If the main computer equipped with learning
mechanisms is broken, the fault diagnosis function is also inoperative.
2.3 Multi-agent Approach
In the fault tolerant domain, Parker proposed ALLIANCE, which is a multi-robot approach for
redundant robot systems [11]. ALLIANCE realizes redundancy using cooperative control of teams
of robots. It can detect the faults of one robot, and allow other robots take over the lost function to
complete the required task. However, it cannot diagnose the cause of each fault. Christensen et
al. proposed fault-tolerant swarms of robots [12]. This research involved a multi-agent approach
and focused on the methodology and protocols for the recovery of robot swarms. However, it also
cannot diagnose the cause of the robot fault.
2.4 Computer Simulation Approach
In recent years, run-time fault detection methods based on the comparison of simulated and real
robot behavior have been proposed [13][14]. This is similar to the model-based approach
because environmental and robot models are designed before running the system. This method’s
approach to the run-time mobile robot fault detection leverages high performance computing
resources. However, this method needs highly accurate environmental models, robot kinematics,
and high accuracy observation of robot motion. It is difficult to implement in complicated real-
world robot systems.

2017 4
2.5 Distributed Approach To Fault Diagnosis
In the fault diagnosis of computer domains, distributed approaches have been proposed. For
example, adaptive distributed system-level diagnosis (Adaptive-DSD) was proposed by Bianchini
et al. [15]. This methodology has still been discussed and applied in recent years [16]. Figure 1
(a) and (b) show the concept of Adaptive-DSD. For example, a computer network system that has
six computers connected by communication is shown in Figure 1 (a). Black arrows indicate
corresponding relationship of diagnosis. In this case, in a fault-free situation, computer works
as an autonomous distributed network. If is broken, can detect the fault of . Then,
obtained information about the fault of is shared using the communication among the
computers, and changes the diagnosis target and communication path. In doing so, the
system can retain the communication network and diagnostic function. After that, if is broken
(Figure 1 (b)), can also detect the fault of , and changes the communication path to .
This concept may contribute to this research; however, a concrete diagnosis method is not
defined, and Adaptive-DSD is assumed for usage in computers.
(a) (b)
FIGURE 1: A concept of Adaptive-DSD. (a) shows that one computer is broken. (b) shows that two
computer nodes are broken, and the computer prior to the broken computer changes the diagnosis target to
maintain the communication loop.
3. PROPOSED METHOD
To determine the fault tolerance of robot systems, one must go beyond previous research and the
three keywords indicated in Section 1, which are “decentralized,” “model-free,” and “attachable.”
Hence, in this study, a distributed cooperative fault diagnosis method (DCFD) is proposed.
For this method, the following assumptions were made. A diagnostic device called a diagnoser
observes the state of a component, such as its input-output signals, to detect a fault. Based on
Adaptive-DSD, the result of the diagnosis can be shared using the wireless communication
capabilities in each diagnoser. Finally, all diagnosers can obtain the overall diagnosis results.
3.1 Model of Component
In this study, for basic research, we assumed a simplified model for each internal electrical
component, as shown in Figure 2. Component has an input and output (Figure 2), and
outputs the same input value as in the following equation:
(1)
FIGURE 2: Simplified model of internal electrical component for simulation.

2017 5
3.2 Faults in Robot Systems
Diagnosers can observe the input and output signals of their corresponding components (Figure
2). Before diagnosis, diagnosers learn the input and output signals of the corresponding
components using a learning algorithm. When diagnoser diagnoses , observes the
signals of its corresponding component and compares the observed signal with its knowledge of
the expected signal, as shown in the following equation.
(2)
Here, is the input signal of the component. is the output signal of component. is the
expected output of the component, which is calculated from by the diagnoser using the learned
input-output signal pair data. If the expected output value, , is different from the observed
output value, , the component is in a fault condition. In contrast, if is equal to ,
the component is fault-free.
3.3 Distributed Cooperative Fault Diagnosis
The component architecture shown in Figure 3 was assumed. In this figure, three components
and three diagnosers are in the robot body, and the components are connected in series.
Diagnoser is connected to component . To detect the fault in a component, a diagnoser
compares the observed signal and expected output signal, which was pre-calculated by learned
data. If the observed signal and learned data are different, the component is faulty. When is
broken, other diagnosers (e.g. , ) observe the state of using an Adaptive-DSD
technique [14]. To diagnose the fault of the diagnoser, neighboring diagnosers use signals, such
as the heartbeat of the broken diagnoser. Moreover, the corresponding component, , is
referred to as a hidden component when the diagnoser, , is broken. This is called a hidden
component fault diagnosis problem (Figure 4). The DCFD method can estimate the state of
hidden components using communication and cooperation with neighboring diagnosers. This
method contributes to the detection of multiple faults, i.e. malfunction of components and
diagnosers at the same time.
FIGURE 3: Left side figure shows the specified model of a diagnoser with a component, and right side figure
shows the simplified model.
In Figure 5, Equation 2 is modified as follows to detect the fault using observed input signal, ,
and output signal, .
(3)

2017 6
FIGURE 4: An example architecture of electrical components and diagnosers in a robot system.
indicates a component, and is a diagnoser. An arrow indicates a correspondence relationship of fault
diagnosis.
FIGURE 5: Hidden component fault diagnosis problem.
4. COMPUTER SIMULATION
Simulation experiments were carried out to confirm the effectiveness of DCFD in a computer-
simulated environment. In this computer simulation, the three simple components and diagnoser
model were examined under the following fault conditions.
4.1 Experimental Conditions
4.1.1 Multiple Fault Situations
The components, diagnosers, and their connections were set as shown in Figure 4. In this
experiment, it was assumed that each diagnoser obtained the knowledge of the input-output
signal pairs from learning before the experiment. In the initial state of the experiment, all
components and diagnosers were fault-free. Experimental procedures were as follows:
1) All components and diagnosers run with DCFD.
2) After 60 seconds, breaks down.
3) After the breakdown of , confirm that can detect the fault of using DCFD
4) After 120 seconds, breaks down resulting in a hidden component fault diagnosis problem.
5) Confirm that can estimate the fault of .
6) Finally, confirm that the DCFD method can estimate multiple faults.
The proposed system executed the diagnosis and presented the results at the end of
experimental operation. A component outputs the same value when it obtains an input signal. The
input signal, , alternated randomly every 0.5 seconds. Each diagnoser diagnosed the
state of the component once every second.
In this experimental setup, diagnosis results were not deterministic with one diagnosis because
when the input signal of the component was ‘0’, the output signal was ‘0’ regardless of whether
the component was broken or not. In other words, a diagnoser could detect the fault of
components only when the input signal was ‘1’. In response to this problem, the likelihood of
faults was adopted to increase the accuracy of fault diagnosis. Fault likelihood is given by .

2017 7
is the number of observed faults, and is the number of diagnoses performed. Here, is set to
50.
4.1.2 Three Types of Faults
In the above experimental condition, it was confirmed that DCFD could detect both component
and diagnoser faults at the same time. However, various fault behaviors exist in the actual
components. In these experimental conditions, three fault types were set as the component fault
states (Figure 6). The following fault types were adopted:
 Fault condition A: stable fault with output ‘0’
 Fault condition B: stable fault with output ‘1’
 Fault condition C: intermittent fault
Here, a “stable fault” means that the component output for any input is permanently 1 or 0. On the
other hand, an “intermittent fault” alternates between the fault state and the fault-free state
randomly. Note that fault condition A is the same fault condition as the above experiment.
The observation targets for the result are and . Other experimental settings are the same as
those described in the previous experiment.
FIGURE 6: Expected output and three types of faults.
4.2 Results
4.2.1 Multiple Faults
The results for the diagnosis are shown in Figure 7. The fault likelihood of increases and
converges to ‘1’ in 60 seconds. This result means that can detect faults of when is
broken. In Figure 7, the fault likelihood of increases after 120 seconds; however, this result of
does not exhibit convergence to one value compared with the likelihood of . The main cause
of this phenomenon is that sometimes a diagnoser can detect a fault-free state of regardless of
whether is broken or not. However, these results indicate effectiveness in solving the hidden
component diagnosis problem by using a suitable threshold value.

2017 8
FIGURE 7: Results of experimental condition ‘’Multiple faults‘’. Comparison of change of fault likelihood of
diagnoser no.2 and component no.2.
4.2.2 Three Type of Faults
The results for the likelihood transition of the three fault types are shown in Figure 8. Note that the
result of fault condition A is the same as that of the previous experimental condition. The
likelihood of all types of fault conditions increases at 120 seconds, and the likelihood of fault
conditions A and B becomes greater than 0.5. However, in fault condition C, the likelihood
increases more gradually, and the value does not reach 0.5.
FIGURE 8: Results of experimental condition “Three type of faults”. Comparison of change of fault likelihood
in three types of faults.
4.3 Discussion
4.3.1 Accuracy of Fault Estimations
The results for the incorrect estimation case are shown in Figure 9. In Figure 9, the likelihood of
fault in the fault-free component is increased in and . The main cause of this incorrect
estimation is a delay between the input and output. In this experiment, a propagation delay of 0.1
seconds was designed between the input and output signals of the component.

2017 9
In this case, and detect the difference between the input and output signals of their
respective corresponding components. Therefore, the threshold value, , should be set as >
0.
FIGURE 9: Results of experimental condition ‘’Multiple faults‘’. Comparison of change of fault likelihood of
diagnoser no.2 and component no.2.
4.3.2 Threshold for Fault Detection
The experimental results indicate that the DCFD can estimate the faults of and . However,
the transition of fault likelihood from the DCFD can also be determined, and it is necessary to
define a suitable threshold value to detect the fault of a component.
The fault likelihood of the diagnoser becomes ‘1’, and it is easy to estimate the fault because
diagnosers estimate the state of other diagnosers based on direct communication, such as heart
beat. On the other hand, the change in likelihood of components is unstable. Therefore, a low
threshold value should be set to estimate the fault. In fault conditions A and B, fault behavior is
simpler than that in fault condition C. Therefore, the threshold value can be set between 0.4 and
0.6. However, in fault condition C, transition of likelihood is lower than other fault conditions. In
this case, to estimate the fault of fault condition C, the threshold value should be set at 0.1.
The threshold depends on the system architecture and signal pattern. To apply this threshold to
an actual robot system requires a discussion on determining threshold values on a case by case
basis.
4.3.3 Limitations of DCFD
Experimental results indicate that DCFD can estimate the fault of as a hidden component
when is broken. Moreover, if , , or is broken, DCFD can detect the fault using
cooperation among surviving diagnosers. However, DCFD cannot estimate the fault state of a
component located at the outer edges of a system when there are multiple faults. For example, in
Figure 4, the fault states of and cannot be determined when the diagnosers and are
broken because the surviving diagnoser cannot observe the signals or .
5. CONCLUSION
A DCFD method was proposed to diagnose the state of the internal electrical components in
robot systems. Simulation experiments were performed under simplified conditions. The
experimental results suggest that the proposed method can detect both diagnoser and hidden

2017 10
component malfunction. This result indicates that the DCFD method has the potential to solve the
hidden component fault diagnosis problem.
In future, as described in Section 4.3.3, investigation of an estimation method for faults occurring
in components located at the outer edge of a system is required. Once complete, the
effectiveness of the DCFD method will be demonstrated in more complex and dynamic situations.
In the current simulations, components and diagnosers are represented by simplified
experimental models. Future experiments will involve extending this method to actual robot
systems.
6. ACKNOWLEDGEMENT
This work was funded by Tough Robotics Challenge, ImPACT Program of the Council for
Science, Technology and Innovation (Cabinet Office, Government of Japan), and was supported
by JSPS KAKENHI Grant-in-Aid for Research Activity Start-up (15H06102).
7. REFERENCES
[1] R.R. Murphy. “Trial by fire,” IEEE Robotics & Automation Magazine. vol. 11, pp50-61, 2004.
[2] K. Nagatani, S. Kiribayashi, Y. Okada, S. Tadokoro, T. Nishimura, T. Yoshida, E. Koyanagi
and Y. Hada. “Redesign of rescue mobile robot Quince,” in Proc. 2011 IEEE International
Symposium on Security, and Rescue Robotics (SSRR), 2011, pp. 13-18.
[3] J. Carlson and R.R. Murphy. “How UGVs physically fail in the field,” IEEE Transactions on
Robotics, vol. 21, pp. 423-437, 2005.
[4] Z. Gao, C. Cecati and S.X. Ding. “A survey of fault diagnosis and fault–tolerant techniques–
Part I: Fault diagnosis with model-based and signal-based approaches,” IEEE Transactions
on Industrial Electronics, vol. 62, pp. 3757-3767, 2015.
[5] Z. Gao, C. Cecati and S.X. Ding. “A survey of fault diagnosis and fault tolerant techniques -
Part II: Fault diagnosis with knowledge–based and hybrid/active approaches,” IEEE
Transactions on Industrial Electronics, vol. 62, pp. 3768-3774, 2015.
[6] S. Okina, K. Kawabata, T. Fujii, Y. Kunii, H. Asama and I. Endo. “Self-diagnosis system of an
autonomous mobile robot using sensory information,” Journal of Robotics and Mechatronics,
vol. 12, pp. 72-77, 2000.
[7] I. Eski, S. Erkaya, S. Savas and S. Yildirim. “Fault detection on robot manipulators using
artificial neural networks,” Robotics and Computer-Integrated Manufacturing, vol. 27, pp.
115-123, 2011.
[8] M. Schlechtingen and I.F. Santos. “Comparative analysis of neural network and regression
based condition monitoring approaches for wind turbine fault detection,” Mechanical systems
and signal processing, vol. 25, pp. 1849-1875, 2011.
[9] D. Lefebvre. “On-line fault diagnosis with partially observed petri nets,” IEEE Transactions on
Automatic Control, vol. 59, no. 1919-1924, 2014.
[10] Q. Liu, D. Zhu and S.X. Yang. “Unmanned underwater vehicles fault identification and fault-
tolerant control method based on FCA-CMAC neural networks, applied on an actuated
vehicle,” Journal of Intelligent & Robotic Systems, vol. 66, pp. 463-475, 2012.
[11] L.E. Parker. “ALLIANCE: An architecture for fault tolerant multirobot cooperation,” IEEE
Transactions on Robotics and Automation, vol. 14, pp. 220-240, 1998.

2017 11
[12] A.L. Christensen, R.O. Grady and M. Dorigo. “From fireflies to fault-tolerant swarms of
robots,” IEEE Transactions on Evolutionary Computation, vol. 13, pp. 754-766, 2009.
[13] A.G. Millard, J. Timmis and A.F. Winfield. “Run-time detection of faults in autonomous mobile
robots based on the comparison of simulated and real robot behavior,” in Proc. 2014
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2014), 2014,
pp. 3720-3725.
[14] Z. Gao, J. Jia, J. Xie, W. D. Toh, P. Lin, H.Lyu, D. Julyanto, C.S. Chin and W.L. Woo,
“Modelling and simulation of a 12-cell battery power system with fault control for underwater
robot, ” in Proc. of 2015 IEEE 7th International Conference on Cybernetics and Intelligent
Systems (CIS) and IEEE Conference on Robotics, Automation and Mechatronics (RAM) ,
2015, pp.261-267.
[15] RP. Bianchini Jr and RW. Buskens. “Implementation of online distributed system-level
diagnosis theory,” IEEE Transactions on Computers, vol. 41, pp.616-626, 1992.
[16] S. Kelkar and R. Kamal. “Adaptive fault diagnosis algorithm for controller area network” IEEE
Transactions on Industrial Electronics, vol. 61, pp. 5527-5537, 2014.

Distributed Cooperative Fault Diagnosis Method for Internal Components of Robot Systems

More Related Content

What's hot (14)

Similar to Distributed Cooperative Fault Diagnosis Method for Internal Components of Robot Systems (20)

Recently uploaded (20)

Distributed Cooperative Fault Diagnosis Method for Internal Components of Robot Systems