1、Experimental Analysis of the Infl uence of Olfactory Property on Chemical Plume Tracing Performance Shunsuke Shigaki1, Kei Okajima2, Kazushi Sanada3and Daisuke Kurabayashi4 AbstractIn this study, we investigated the relationship between the performance of chemical plume tracing (CPT) and odor sensin

2、g property. Tracking of chemical plumes plays an important role because it facilitates the identifi cation of an odor source. Conventional research has focused on the development of CPT algorithms, whereas the infl uence of the performance odor sensors on CPT performance has not been investigated. T

3、herefore, in this study, we fi rst compared the olfactory characteristics of an insect (silkworm moth, Bombyx mori), which has high CPT performance and the characteristics of an artifi cial odor sensor. In particular, we focused on and compared the recovery time of the two types of sensors, which pl

4、ays an important role in the acquisition of odor information. As a result, it was determined that the recovery time for the insect olfactory sensor was 10 times faster than that of the artifi cial odor sensor. We also experimentally evaluated the effect of this difference in the recovery time on CPT

5、 performance. CPT experiments using silkworm moths and a robot revealed that there was a correlation between the CPT performance and sensor recovery time. As such, it was demonstrated that it is necessary to improve not only the algorithm but also the sensor recovery time improve the CPT performance

6、. Index TermsChemical Plume Tracing, Electroantenno- gram, Olfactory Property, Bio-inspired robotics. I. INTRODUCTION Biological organisms acquire information about their ex- ternal environment using various sensory organs and respond according to this information. Among them, olfactory sens- ing ha

7、s numerous advantages in terms of sustainability and diffusibility compared to physical quantities transmitted by waves such as light and sound. Therefore, it is extensively used as a tool for exploring and communicating in numerous areas. However, current robotic systems are not well-adapted to usi

8、ng olfactory sensing because chemical (odor) has high uncertainty due to airfl ow, temperature, and humidity. However, the use of olfactory sensing in a robotic system has high engineering value because of the potential application to life rescue (e.g., by sensingtenone1) and dangerous object discov

9、ery (e.g., by sensing thionyl chloride2). The odor source localization problem refers to the identifying of an odor source based on the odor information in the air 3. Odor source localization is roughly divided into three 1Shunsuke Shigaki is with the Department of System Innovation, Osaka Universit

10、y, Japanshigakiarl.sys.es.osaka-u.ac.jp 2Kei Okajima is with the Department of Mechanical Engineering, Mate- rials Science, and Ocean Engineering, Yokohama National University, Japan okajima-kei-hxynu.jp 3Kazushi Sanada is with the Division of Systems Research, Yokohama National University, Japansan

11、ada-kazushi-snynu.ac.jp 4Daisuke KurabayashiiswiththeDepartmentofSystems andControlEngineering,TokyoInstituteofTechnology,Japan dkurairs.ctrl.titech.ac.jp stages 45: (1) fi nding a plume, (2) tracing the chemical plume, and (3) identifying the odor source. In this research, we focus on chemical plum

12、e tracing (CPT) of stage (2) which dominantly and intermittently uses olfactory sensing . In the case of stage (2), there are many search algorithms based on bio-inspired or statistical methods from previous studies 678910. In the literature 1112, they comprehen- sively examined using simulation and

13、 an actual machine and they found that the difference in the search strategy affects CPT performance. Moreover, it was reported that the mechanism of intaking odorants is closely related to CPT performance 22. However, the infl uence of odor sensors with different characteristics on the search perfo

14、rmance has not been investigated. In particular, there are almost no studies that directly compare the difference between living organisms and artifacts with regard to olfactory sensing. Therefore, in this study, we compare the odor sensor of an organism (insect) with artifi cial odor sensors and in

15、vestigated whether the difference in characteristics affects CPT performance. II. PROBLEMSTATEMENT In this study, we investigate the difference in the CPT per- formance caused by differences in the response performance of an odor sensor and identify the sensor characteristics that are suitable for o

16、lfactory sensing in a robotic system. Although insects have few neurons compared to mammals, they are well adapted to using olfactory sensing and can readily locate feeding sites and mating partners. However, a robotic system encounters diffi culty when using olfactory sensing because odorants have

17、low temporal/spatial resolu- tion and high uncertainty. Therefore, in this investigation, we consider the extent of the difference between an insect and a machine at the receptor (sensor) level and the effect on the CPT performance. In addition, we performed a CPT experi- ment using moth and an auto

18、nomous robot and investigated the relationship between odor sensor characteristics and CPT performance. A male silkworm moth was used (Bombyx mori) as the insect for comparison with the artifi cial sensors. These insects were used because of a one-to-one relation between the stimulus input (pheromon

19、e) and the behavior output 13. The body of the silkworm moth is approximately 30 mm long and 10 mm wide. It has two antennae with a length of approximately 6 mm on its head, and it detects bombykol, which is a pheromone released by female silkworm moths to attract mates 14. After emergence, a moth d

20、oes not perform any spontaneous behavior, including eating and drinking. IEEE Robotics and Automation Letters (RAL) paper presented at the 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) Macau, China, November 4-8, 2019 Copyright 2019 IEEE (a) Mating dance for locatin

21、g the female Surge Loop Zig-zag Wait Surge (Forward-moving) Loop (Left/Right turn) Time course Stimulus Zig-zag (Left/Right turn) (b) State transition diagram Fig. 1.Female exploratory behavior of male silkworm moth. When a moth detects the female sex pheromone, it performs a ”mating dance” to locat

22、e the female by moving about as shown in Fig. 1 22. The mating dance can be divided into three states: surge, zig-zag, and loop 15. By repeating three behaviors in response to the pheromone stimulus, the male silkworm moth can locate the female 15. For these reasons, we employed bombykol in the expe

23、riments with the silkworm moth. We kept the silkworm moths in a 16 h:8 h light:dark photoperiod at 26C and 50%-60% relative humidity using an incubator (MIR-154, Sanyo Electric Co., Ltd., Japan) . Moreover, we used adult male silkworm moths within 3-5 d after eclosion. Furthermore, all experiments w

24、ere conducted from 9 am to 5 pm because moths are diurnal insects and the serotonin in their brains is highest during this period 16. III. COMPARISON OF PERFORMANCE BETWEEN OLFACTORY SENSING OF AN INSECT AND AN ARTIFICIAL SENSOR A. Experimental setup and design In this chapter, we investigate the di

25、fference in response characteristics between insect olfactory sensing(silkworm moth) and an artifi cial odor sensor. A silkworm moth detects odors and pheromones via olfactory nerve cells on its antennae. After the electrical signal generated by the olfactory cells of the antennae is transmitted to

26、the brain of the silkworm moth, the odor is identifi ed. An approach that involves the use of an elec- troantennogram (EAG) that was proposed by Schneider et al. in the mid-1950s was used to ascertain whether the insect detected an odor. The EAG method involves measurement of voltage change with res

27、pect to an odor stimulus by inserting electrodes at both ends of an insects antenna as shown in Fig. 2 17. The EAG signal was recorded on a PC using an amplifi er to enhance the very weak biological signals that were detected. With reference to 18, we used an amplifi er with a bandpass fi lter opera

28、ting in the range 0.1 - 300 Hz and a 50 Hz band stop fi lter. The detected electrical signals were amplifi ed by a factor of 250. In this study, the EAG Antenna Electrode Electrode 1 mm Fig. 2.Insertion of electrodes into the antenna. Air compressor Absorbent cotton Activated carbon DW Sensor 10 mm

29、Arduino Uno DAQ Glass pipe Fig. 3.Odor stimulation device for sensor response measurement. response is compared with the response from an artifi cial odor sensor. We utilized two alcohol sensors as artifi cial olfactory sensors: MQ-3(Henan Hanwei Electronics Co.,Ltd., China) and MiCS5524 (SGX Sensor

30、tech Ltd., Switzerland). The operating principle of both sensors is the same and chemical detection is achieved by measuring the devices resistance. Thesethreetypesofsensors(EAG,MQ-3,and MiCS5524) were set in the sensor attachment position of the experimental setup shown in Fig. 3. The distance betw

31、een the discharge port made out of a glass pipe and the sensor was set to 10 mm and an odorous substance was deposited on fi lter paper in the glass pipe. The pheromone (Bombykol, 100 ng)was deposited during EAG measurement and ethanol (10 %, 50 l)was used with the MQ-3 and MiCS 5524 sensors. An odo

32、r stimulation device 19was used which consisted of an air compressor (MAS-1, AS ONE, Osaka, Japan), three gas washing bottles (absorbent cotton, activated carbon and distilled water), a fl ow meter (RK1710-AIR, KOFLOC, Kyoto, Japan) and a solenoid valve (VT307, SMC Corporation, Tokyo, Japan). As sho

33、wn in Fig. 3, the air exhausted from the air compressor passed through absorbent cotton, activated carbon and distilled water, adjusted to a constant fl ow rate of 1.0 L/min using a fl owmeter, and fi nally exposed to the sensor. In this experiment, we measured the output value of the sensor when ex

34、posed to a one-shot odor stimulus (open valve duration: 200 ms), and analyzed the characteristics of the response. The output value of the sensor was recorded using a PC at a sampling rate of 1 kHz via an A/D converter (USB-6215, National Instruments, USA). We performed ten repetitions of the experi

35、ments for each sensor. 5 Time s 01015 Normalized value 0 1 0.5 Model MQ-3 0 1 0.5 0 1 0.5 Normalized valueNormalized value (a) EAG (b) MQ-3 (c) MiCS5524 Model EAG Model MiCS5524 Antenna Electrode Electrode 1 mm Fig. 4.Sensor response to a one-shot odor stimulus. B. Experimental results The response

36、of each sensor to a one-shot odor stimulus is shown in Fig. 4. Fig. 4 (a)(c) shows the results for the EAG, MQ-3 and MiCS5524, respectively. The red line in Fig. 4 represents the mean value of the sensor output, and the grey one represents the standard deviation (n = 10). The vertical and horizontal

37、 axes of Fig. 4 represent the normalized value of the sensor output value and the elapsed time, respectively. The release timing of the odor stimulus is represented by a triangle ” in Fig. 4. In this study, we investigated the infl uence of the difference between the rise time and the recovery time

38、of the sensor value on the CPT performance against the odor stimulus to normalize the amplitude of each sensor. To quantitatively evaluate the response of the sensors, we modeled their response waveform based on a previous study 20 and calculated the rise and recovery time. According to 20, the rise

39、 and recovery time can be determined by calculating the time constant because the response of the odor sensor s rise and recovery can be represented by a fi rst- order lag system. The period over which the sensor output value rises immediately after the odor stimulus is applied is defi ned as the ri

40、sing phase, and the period until the output returns to the baseline after reaching a peak point is defi ned as the recovery phase. The response waveform of the sensor can be formulated separately into the rising and recovering phases as shown in Equ. (1). R(t) = r(1 exp(tts r ) + r, (ts t tp) dexp(t

41、tp d ) + d, (tp t) (1) where and represent the gain and offset, respectively, and tsand tprepresent the rise start time and the peak point time, respectively. We defi ne the rise start time as the time when the differential value of the sensor response is a maximum. rand drepresent the time constant

42、s of the rise and the recovery phase, respectively. The results obtained by calculating the coeffi cients of the model (Equ. (1) based on the method of least squares to fi t the sensor experimental values are represented by the black lines in Fig. 4. In addition, Table I shows the results obtained f

43、or calculation of the time constants of the rising and the recovery phase of EAG, MQ-3, and MiCS5524. TABLE I COMPARISON OF THE TIME CONSTANT OF EACH SENSOR. EAGMQ-3MiCS5524 rs0.008450.02170.00955 ds0.20713.11.96 Based on the results, EAG has the smallest time constant in each of the rise and the re

44、covery phase, and MQ-3 had the largest. Focusing on the time constant of the recovery phase, it was determined that the time constant of EAG and MiCS 5524 is different by a factor of approximately 10 times different, and the difference between the time constant of MiCS5524 and MQ-3 is about 10 times

45、. As the recovery time gets slower, the sensor value tends to saturate; therefore, it is diffi cult to accurately acquire high-frequency odor information 19, 20. For this reason, it is expected that odor information will be lost if the sensor with a long recovery time is used as the olfactory sensor

46、 in the robot. However, the investigator needs to acquire high-frequency odor information because the odor stimulus frequency is correlated to the distance to the odor source 21. Hence, we investigate the effect of this difference in the recovery time on the CPT performance via an experimental study

47、. IV. EXPERIMENTS FORCPTPERFORMANCE EVALUATION A. Confi guration of Autonomous Search Robot In this section, we describe the confi guration of an au- tonomous search robot for mounting the artifi cial sensor. For the autonomous search robot, we employ a palm-sized mobile robot platform that was used

48、 in a previous study as shown in Fig. 5(a) 8. The robot is a differential two-wheel 160 mm 180 mm Intake system (a) Autonomous search robot(b) Male silkworm moth Antennae 5 mm Wings Fig. 5.The autonomous search robot and the male silkworm moth. type robot and has an intake system that can install tw

49、o odor sensors. The silkworm moth exhibited high localization performance using the airfl ow generated by the fl apping of its wings 22. For this reason, we introduced a mechanism for the robot to obtain odor substances by itself by mounting a small fan (NMBMAT 1608KL-04W-B59, CWC Groups, USA) on th

50、e intake system. The suction airspeed of this system was 0.37 0.021 m/s, which is the same value as the suction airspeed caused by the fl apping of the wings of the silkworm moth (5(b) 23. Moreover, the weak point of the artifi cial odor sensor is that it cannot detect high- frequency odor stimuli b

51、ecause of its long recovery time. It was reported that the time resolution can be improved by identifying the sensor with an auto-regressive using an exogenous input (ARX) model for this problem 19. Hence, the presence or absence of an odorous substance can be determined by applying the ARX model to

52、 both the MQ-3 and MiCS5524. In addition, we implemented the silkworm moth-inspired CPT algorithm 8 in the robot. Fig. 6 shows a fl owchart of the algorithm. As shown in this fi gure, the silkworm moth-inspired CPT algorithm transits to the surge state during the detection of an odor stimulus in any

53、 behavioral state. The control period of the robot was 50 Hz and the linear and angular velocities were 26 mm/s and 1 rad/s, respectively, based on the behavioral data of the silkworm moth. B. CPT experiment fi eld We evaluated the CPT performance from the viewpoint of the search success rate and se

54、arch time. Experiments were conducted using a silkworm moth and a robot as a searcher to investigate the relationship between the sensor performance and the CPT performance. Fig. 7 shows a schematic diagram of the CPT experimental fi eld. The odor source was placed at the origin (x, y) = (0, 0) m, a

55、nd the searcher starts the search from the position (x, y) = (1, 0.3) m. The searcher initiates the search with an initial heading angle of = /6 rad. It is necessary to perform an exploratory maneuver (zig-zag or loop) at least once because the searcher needs to move to the line above the odor sourc

56、e to determine its position. In this CPT experiment, if the searcher reached a radius of 0.1 m from the odor source Start Wait Detect odor Surge Zigzag Loop Yes No Detect odor Fig. 6.Flowchart of the silkworm moth-inspired CPT algorithm. /6 rad x y 1 m 0.3 mOdor source Goal area (0.1) 0.5 m Fan Star

57、t position Wind speed: 0.68 m/s Odor (1.0 L/min) 0.083Hz Fig. 7. Outline of CPT experiment fi eld. within 600 s, the search is considered to be successful. Bombykol (1000 ng) and ethanol (99.5 %) were used as the odor source and an electric fan (PCF-C15, IRISO- HYAMA INC., Japan) was placed above th

58、e source to allow it to diffuse. The wind fl ow environment was generated by the fan. It is possible to operate the head swing of the fan so that a more turbulent environment can be created. The swing angle of the fan and the swing frequency were /2 rad and 0.1 Hz, respectively. We set the wind spee

59、d generated by the electric fan at the odor source to approximately 0.68 m/s. As an odor source in the CPT experiment, we employed the odor stimulation device described in Chapter I II. Thereby, odor discharge can be precisely controlled using a computer. The discharge frequency was set in this CPT experiment to 1 Hz (open: 0.2 s, close: 0.8 s) because it was determined that the female silkworm moth discharges pheromones at approximately 1 Hz 24. We conducted 20 trials at each of the experimental conditions and we ventilated the experimental room for 10 minutes per t