Detection of hazardous leaks from pipelines using optical imaging and neural network
A method of training a learning machine to detect spills of hydrocarbon liquids from pipelines. A neural network is trained by collecting samples of a number of different ground materials as well as a number of liquid hydrocarbons. For each hydrocarbon, a spill is simulated on each ground material. For each of these spills, a thermal camera and a visible light camera are used to capture images. The images from the two cameras are fused, and input to the neural network for classification training. Once the neural network is trained, a system having the two cameras and the neural network can be used to detect actual hydrocarbon spills.
1. A method of training a learning machine to detect spills of hydrocarbon liquids from pipelines, comprising:
collecting samples of a number of different ground materials;
collecting samples of a number of hydrocarbon materials that are or become liquid when spilled from a pipeline;
for each hydrocarbon material, simulating a spill on each ground material, thereby providing a collection of spill samples;
for each spill sample, performing the following steps: using a thermal camera to capture a thermal image of the spill sample; using a visible wavelength camera to capture a red, blue, green (RGB) image of the spill sample; fusing the thermal image and the RGB image, thereby providing a fused image; inputting the fused image to a neural network, comparing the output of the neural network to at least one hydrocarbon classification represented by the fused image, and adjusting the neural network to provide output data that more closely matches the hydrocarbon identifier.
2. The method of claim 1 , wherein the learning machine is a convolutional neural network.
3. The method of claim 1 , wherein the hydrocarbon classification is one or more of the following: mineral oil, diesel, gasoline, crude oil.
4. The method of claim 1 , wherein the output of the neural network is further compared to a non-hydrocarbon classification.
5. The method of claim 1 , wherein the learning machine is a tiered neural network, having one tier for classifying surfaces and another tier for classifying liquids.