Neural processing-type displacement sensor employing multimode waveguide

394

Neural Processing-Type Displacement Sensor Employing Multimode Waveguide Shigeki Aisawa, Kazuhiro Noguchi, and Takao Matsumoto Abstract-A novel neural processing-type displacement sensor, consisting of a multimode waveguide and a neural network, is demonstrated. This sensor detects displacement using changes in the interference output image of the waveguide. The interference image is directly processed by a three-layer perceptron neural network. This Sensor has a feature so that the environmental change, such as the intensity fluctuation, and change of the temperature can be followed by training the neural network. Experimental results show that the sensor has a resolution of 1 Ctm I. INTRODUCTION

C

OHERENT light propagating through a multimode fiber creates a speckle pattern in the far field which is caused by interference among the fiber’s propagating modes. The speckle pattern changes in fibers with a limited number of modes have been used to create pressure, strain, and vibration sensors [1]-[3]. In these sensors, both the signal and reference light beam go through the same path. This feature brings high stability to these sensors in comparison to two-arm sensors such as the Mach-Zender type; however, this also brings lower sensitivity because these sensors use the interference of LP,,, and LP,, modes which has a small difference of propagating constants. Since such sensors detect only the light intensity of the interference pattern, the instability by intensity fluctuation of the light source is also brought. In this letter, a novel method of displacement sensing suitable for measuring fine displacements is presented. In this method, changes in the coupling between a light beam and a multimode waveguide caused by the object displacement are transformed into interference image changes. A neural network (NN) is trained to recognize the different output interference patterns. An experiment in which the displacement sensor is used to measure one-dimensional displacement is described.

Detector Array

Lens

I-

Multimode Waveguide Neural Network

Fig. 1. Setup of the displacement sensor.

condition, the output interference image also changes. Although the output interference image has information on light beam displacement, the relation between the interference image and the displacement to be measured is very complicated. An NN has a pattern recognition and training function. We apply these functions to the displacement sensor. Since an NN has a function of nonlinear mapping, and can be trained to recognize patterns [4], [5] after training, the NN can be used to determine the displacement in accordance with the differences in the interference image. Since displacement is measured not by using the light intensity directly but by using the output intensity distribution in this method, the intensity fluctuation of the light source is not affected by the sensor. Although the output interference images would be changed by environmental perturbation such as temperature, the instability can be eliminated by changing the parameter of the NN with the training function. Fig. 1 shows a system to measure horizontal displacement employing a one-dimensional multimode waveguide in the horizontal direction and a one-dimensional detector array. The vertical displacement sensor can also be constructed by using a one-dimensional multimode waveguide in the vertical direction. It may be easily expanded to a two-dimensional displacement sensor, by employing a two-dimensional multimode waveguide and a two-dimensional detector array.

II. WNCPLE

III. EXPERIMENT

The arrangement of the neural processing-type displacement sensor is shown in Fig. 1. A light beam from an object to be measured is launched into a multimode waveguide which has a lot of propagating modes. The power of the light beam is distributed among the propagating modes, and an interference image appears on the output end of the waveguide. According to changes of the light beam launching Manuscript received January 29, 1991. The authors are with the NTT Transmission Systems Laboratories, 1-2356 Take, Yokosuka-shi Kanagawa-ken, 238-03 Japan. IEEE Log Number 9144507.

To investigate the principle of this sensor, a preliminary experiment was carried out using a ridge-type multimode waveguide (core width = 50 pm, core thickness = 7 pm, waveguide length = 48 mm, core’s refractive index = 1.468, A = 0.002). An output facet of a single-mode fiber was used as the object to be measured. An He-Ne laser light ( A = 0.63 pm) was fed into the single-mode fiber (A = 0.03, core radius = 4.7 pm, fiber length = 40 m). The fiber facet was ground flat and air coupled to the multimode waveguide, and it was attached to a three-dimensional translation stage. The output far-field image of the waveguide was detected by a

1041-1135/91 /O4OO-0394$01.OO 0 1991 IEEE

395

AISAWA et al.: NEURAL PROCESSING-TYPE DISPLACEMENT SENSOR

Fiber

T

f

v

._ * h

I * CI

1

64

128

Detector Number Fig. 2. Output image versus displacement.

Displacement @m)

Fig. 4. Output of the neural network versus displacement (after training).

.. . .

, I

..

I I

I I I

**I

I ,

Input Layer

. .

, ..,

Output Layer

(b)

Fig. 3. Model of the neural network. (a) Model of neuron. (b) Model of the neural network.

one-dimensional detector array. In the experiment, a vidicon camera was used as the detector array. The light intensity distribution was sampled with the vidicon camera, and fed into the NN. The function of the NN was performed by a computer. Fig. 2 shows the one-dimensional far-field images detected with the detector array which had 128 sampling points when the fiber was displaced 0.2 pm steps. The origin of the displacement was set to the center of the waveguide. Positive values for the displacement was measured for the right side from the center of the waveguide. The output image changed even with very slight displacement of the fiber. The output image at the origin was not symmetrical because of the rough surface at the waveguide facet. There was enough difference for the NN to distinguish each pattern. The function of the proposed sensor was completed by training the NN to recognize the patterns. Far-field images for displacements ranging from - 8 to +9 pm in 1-pm steps were used to create the training patterns. The neural network model consisted of neurons and links as shown in Fig. 3(a). Neurons are considered as the processing elements in the neural network. Each neuron sums all inputs and passes the result through an activation function f which typically employs hard-limiting, threshold logic, or sigmoid function. The NN used in the experiment used the three-layer perceptron structure shown in Fig. 3(b) [6]. The unit numbers of the

input, hidden, and output layer were 32, 20, and 18, respectively. The 32 input signals for the NN were made by averaging the intensity of 4 adjoining detector elements. The back propagation training process was used [6].One training cycle modified all weight and offsets with all the training data. A total of 500 training cycles was used. The initial values of the weights and the thresholds were randomly set. We used one-dimensional vector signals as the teaching signals during training. The teaching signals ensured that only the output layer neuron, which corresponded to the input displacement, output a value of one, and all other neurons showed a zero. It was assumed that the training process was completed when the sum-squared error between the training data and output data dropped beneath 0.5. After the training process for the 18 input images was finished, we tested the operation of displacement sensor. Fig. 4 shows the output of the NN versus fiber displacement. Since the NN trained in 1-pm steps, the NN output is shown as a step function. The result shows that the NN output completely agrees with the actual displacement of the fiber. Although the resolution of this sensor was 1 pm in this experiment, higher resolution can be achieved by enlarging the NN elements and by optimizing he waveguide size. In this experiment, the functions of the NN and the detector array were realized by a neural simulator and a vidicon camera, respectively. This restricted the sensor’s operation speed. However, with the high-speed detector arrays and the NN-LSI’s that have been developed 171, [ 8 ] , very high-speed sensors can be developed. IV. CONCLUSION A high-resolution displacement sensor using a neural network and multimode waveguide was proposed. An experiment that measured one-dimensional displacement of an opti-

IEEE PHOTONICS TECHNOLOGY LETTERS, VOL. 3, NO. 4, APRIL 1991

396

cal fiber was carried out. The result showed that displacement sensing to a resolution of 1 p m is possible with the proposed method.

[4]

ACKNOWLEDGMENT The authors thank A. Sugita for supplying the multimode waveguide, and also wish to thank Dr. S. Shimada and Dr. H. Ishio for their continuous encouragement.

REFERENCES [l] K. A. Murphy, M. S . Miller, A. M. Vengsarkar, and R. 0. Claus, “Elliptical-core two-mode optical-fiber sensor implementation methods,’’ J. Lightwave Technol., vol. 8, pp. 1688-1696, 1990. [2] H. J. Lee, 0. Moonsu, and K. Yongduk, “Two mode fiber-optic ring interferometer as a sensor,” Opt. Lett., vol. 15, pp. 198-200, 1990. [3] B. D. Duncan, B. W. Brennan, and R. 0. Claus, “Intermodal pattern

[5]

[6] [7]

[8]

modulation in optical fiber modal domain sensor systems: Experimental result,” Proc. SPIE, vol. 986, Int. Opt. Eng., pp. 186-193, 1988. B. P. Yuhas, M. H. Goldstein, and T. J . Sejnowski, “Integration of acoustic and visual speech signals using neural networks,” ZEEE Commun. Mag., pp. 65-71, Nov. 1989. Y. Le Cun, L. D. Jackel, B. Bosen, J. S . Denker, H. P. Graf, I. Goyon, D. Henderson, R. E. Howard, and W. Hubbard, “Handwritten digit recognition: Application of neural network chips and automatic learning,” IEEE Commun. Mag., pp. 41-46, Nov. 1989. D. E. Rummelhert, J. L. McCelland, and PDP Research Group, Parallel Distributed Processing. Cambridge, MA; MIT, vol. 1 , 1986. M. Koga and T. Matsumoto, “A novel optical WDM multiplexer consisting of a simple optical multimode guide and an electrical neural network,” IEEE Photon. Technol. Lett., vol. 2, pp. 487-489, 1990. M. Holler, S . Tam, H. Castro, and R. Benson, “An electrically trainable artificial neural network with 1024 floating gate synapses, in PTOC.ZJCNN’89, vol. 2 , 1989, pp. 191-196.