Plastic Fiber Optic Bending Sensor Based On Convolutional Neural Network

Fiber optic bending sensors are widely used and significant in many fields such as architecture,aviation,medicine,flatness monitoring,bending angle measurement of mechanical structure,etc.With the development of material technology,sensors are developing in the direction of precision,sensitivity,intelligence,networking,low cost and easy processing.Optical fiber sensor due to its inherent advantages received extensive attention of the researchers,and sensitization bending type plastic optical fiber sensor,make simple,can judge the bending direction,can effectively increase the fiber bending when the transmission loss of sensitivity and dynamic range,can be applied to distributed fiber sensor,etc.With the development of high-performance computer hardware such as image processor and mobile Internet,the performance of data processing by artificial intelligence technology has been greatly improved.Combining artificial intelligence with optical sensing technology is expected to provide a new type of simple and efficient optical fiber bending sensor.In this paper,a convolutional neural network based on visible light source is proposed to measure the bending angle and direction of plastic fiber at single and multiple points.Based on deep learning,the output light spot pattern carrying optical fiber spatial state information was judged and recognized,and the single point and multi-point bending angle and direction of plastic optical fiber were identified and classified,so as to realize distributed sensing.Among them,distributed optical fiber bending sensor can carry out real-time multi-point measurement,which has a high application value in the field of real-time structural health monitoring of public infrastructure,wearable devices and industrial robot technology.However,the existing distributed bending sensor technology is often complex and costly,which increases the difficulty and cost of such sensor system.Therefore,it is of great significance to study how to further simplify the architecture of distributed optical fiber bending sensing system and improve its sensitivity and cost performance.The main completed working is following:(1)Firstly,the research background,research significance and development status of optical fiber bending sensor are introduced,followed by a detailed description of the key technologies of deep learning.The influence of mode field distribution on the output speckle and the polarity of the sensitivity of side-cast sensitized fiber are studied.(2)A sensitized plastic fiber bending sensor based on convolutional neural network is proposed,and a convolutional neural network for small-scale image data sets is designed on the basis of typical convolutional neural network.Three models are designed by adjusting the size of input data,the number of input channels,the convolution kernel and the number of network layers of the Alex Net model.When the bending angle interval of the plastic fiber is 5°,the classification accuracy reaches 96%,which proves that the bending angle of the side cast sensitized plastic fiber can be well classified and recognized by the neural network,and also confirms the feasibility of this method.(3)Compared with other types of fiber bending sensors,the fiber bending sensor based on intensity variation has low cost and simple manufacturing process,but it lacks the capability of reuse.Existing distributed bending sensing techniques often require complex structure,high cost,and so on the basis of the single point bending angle detection,put forward a kind of based on multi-point bending test of sensitization bending type plastic optical fiber sensor,by changing the parameters of the network,the introduction of batch normalization processing and global average pooling operation DOFBA-Net,build the convolution neural network DOFBA-Net,SVM,decision tree,k-means and other traditional machine learning algorithms were used to carry out classification experiments to verify the recognition of multi-point bending angle and direction.For 19 types of bending from three sensitized regions,the recognition accuracy reached 98%.This experiment will lay a theoretical and experimental foundation for further improving the performance of distributed plastic fiber bending sensor in the future.