Context encoder-based fiber sensing anomaly detection
View Patent ↗Aspects of the present disclosure describe an unsupervised context encoder-based fiber sensing method that detects anomalous vibrations proximate to a sensor fiber that is part of a distributed fiber optic sensing system (DFOS) such that damage to the sensor fiber by activities producing and anomalous vibrations are preventable. Advantageously, our method requires only normal data streams and a machine learning based operation is utilized to analyze the sensing data and report abnormal events related to construction or other fiber-threatening activities in real-time. Our machine learning algorithm is based on waterfall image inpainting by context encoder and is self-trained in an end-to-end manner and extended every time the DFOS sensor fiber is optically connected to a new route. Accordingly, our inventive method and system it is much easier to deploy as compared to supervised methods of the prior art.
1. An unsupervised context-encoder method for operating a distributed fiber optic sensing (DFOS) system including a length of optical sensing fiber in optical communication with a DFOS interrogator and anomaly detector, said unsupervised method comprising:
providing the DFOS system wherein the optical sensing fiber is deployed proximate to a roadway;
in an unsupervised manner
operating the DFOS to obtain normal characteristics of the roadway including road traffic and other infrastructure vibration patterns proximate to the sensing fiber; and
determining, in real-time, from inpainted waterfall images generated during operation of the DFOS, unusual patterns in the waterfall images through the effect of a context encoder/decoder derived from data based on a false alarm rate level; and
self-training, in an end-to-end manner, and extending the context encoder/decoder every time the DFOS sensor fiber is optically connected to a new route;
wherein the unusual patterns are indicative of activities that threaten the integrity of the optical sensing fiber;
wherein the context encoder/decoder employs inpainting and is trained by receiving as input generated waterfall images and for each input image, generates two output images, a ground truth output image and a cropped image, the cropped image having a center portion of the input waterfall image set to zero, and a encoder/decoder model produces a reconstruction of the cropped image wherein the accuracy of the reconstruction is assessed by determining a reconstruction error against the ground truth and an abnormal score is derived by aggregating reconstruction errors.
2. The method of claim 1 wherein the optical sensing fiber is a deployed optical fiber that carries telecommunications traffic in addition to any DFOS signals.
3. The method of claim 2 wherein an encoder of the context encoder/decoder includes three layers, each individual layer having convolution, batch normalization, Leaky Relyu activation and convolution with stride for downsampling.
4. The method of claim 3 wherein a decoder of the of the context encoder/decoder includes three layers, each layer including a convolutional transpose operator.