Method for locating abnormal temperature event of distributed optical fiber

A method of locating an abnormal temperature event of a distributed optical fiber, comprising the steps of:

• (a) generating a training dataset by:
• (a.1) placing a sensing optical fiber of a fiber-optic temperature sensing system in a constant temperature water tank, and presetting a plurality of background temperatures and a plurality of abnormal temperatures, wherein the fiber-optic temperature sensing system acquires an array with a length of L×2 each time, each column of data corresponds to a location information on the sensing optical fiber, and L number of location information constitute a location sequence of length L;
• (a.2) collecting multiple times at each of the background temperatures, and obtaining one background temperature array each time and a total of A number of background temperature arrays;
• collecting multiple times at each of the abnormal temperatures, and obtaining one abnormal temperature array each time and a total of B number of abnormal temperature arrays;
• (a.3) normalizing the background temperature arrays and the abnormal temperature arrays to obtain normalized background temperature arrays and normalized abnormal temperature arrays respectively, and normalizing the location sequence to obtain a sequence P;
• (a.4) constructing A number of A-type arrays by constructing one L×3 array with the normalized background temperature arrays and the sequence P respectively;
• constructing B number of B-type arrays by constructing one L×3 array with the normalized abnormal temperature arrays and the sequence P respectively;
• (a.5) dividing each of the B-type arrays into a plurality of intervals with a length M, selecting a segment in each of the intervals randomly as an abnormal temperature range, replacing data at one corresponding location of the A-type arrays by data in the abnormal temperature range to form one replacement sample for each of the B-type arrays;
• (a.6) dividing all the replacement samples and the A-type arrays evenly into training samples having a size of M×3, wherein all the training samples constitute the training data set;
• (b) setting labels for each training sample with a label length M, if a j-th column of the particular training sample refers to the data in the abnormal temperature range, the j-th bit of the label is set as 0, otherwise set as 1, j=1˜M;
• (c) constructing a convolutional neural network which comprises a multi-layer convolutional network and a fully connected layer, wherein the multi-layer convolutional network is formed by multiple convolutional layers, with an input size M×3, an output size 512, and an output size of a fully connected layer M;
• and processing training by sending the training samples into the convolutional neural network, and using the labels as expected output to obtain the convolutional neural network model;
• d) utilizing the fiber-optic temperature sensing system on a testing object for measurement, and collecting one L×2 temperature array, which is normalized and constructs into an L×3 array with the sequence P, and dividing the L×3 array evenly into a plurality of M×3 sub-arrays;
• (e) sending the sub-arrays into the convolutional neural network model to obtain an output feature with a length M, which is mapped to the 0-1 interval to obtain a mapped feature, and then carrying out a binarization process to obtain a binary feature;
• (f) for each of the binary feature, performing an offset in an x direction to get an offset feature for each offset, calculating a cosine similarity between the offset feature and a first row of data in its corresponding sub-array;
• (g) obtaining a location of the abnormal temperature event of the sub-array on the sensing fiber by identifying the offset feature with a largest cosine similarity and searching a location of the offset feature in the sequence P;
• and
• (h) repeating steps (e)-(g) for the rest of the sub-arrays to obtain the location of abnormal temperature events of all sub-arrays on the sensing optical fiber.