US 12,292,947 B1
Emission inversion simulator-based artificial intelligence system using deep learning technology
Wei Tang, Beijing (CN); Yang Li, Beijing (CN); Jian Gao, Beijing (CN); Zhongzhi Zhang, Beijing (CN); Xiaohui Du, Beijing (CN); Yang Yu, Beijing (CN); Xuezhi Dai, Beijing (CN); Jun Xu, Beijing (CN); Shijie Liu, Beijing (CN); Miaomiao Cheng, Beijing (CN); Yunlang Wang, Beijing (CN); and Dazhi Wu, Shenzhen (CN)
Assigned to CHINESE RESEARCH ACADEMY OF ENVIRONMENTAL SCIENCES, Beijing (CN); and SHENZHEN QIANHAI QIMING TECHNOLOGY CO., LTD., Shenzhen (CN)
Filed by CHINESE RESEARCH ACADEMY OF ENVIRONMENTAL SCIENCES, Beijing (CN); and SHENZHEN QIANHAI QIMING TECHNOLOGY CO., LTD., Shenzhen (CN)
Filed on Jul. 22, 2024, as Appl. No. 18/780,327.
Claims priority of application No. 202410121694.8 (CN), filed on Jan. 29, 2024.
Int. Cl. G06F 17/40 (2006.01); G06F 5/01 (2006.01); G06F 17/11 (2006.01)
CPC G06F 17/40 (2013.01) [G06F 5/01 (2013.01); G06F 17/11 (2013.01)] 6 Claims
OG exemplary drawing
 
1. An inversion method for determining a pollution source list based on artificial intelligence and big data, comprising:
acquiring weather data, emission data and concentration data from satellites and ground monitoring data sources, and preprocessing the three types of data utilizing a computer-driven simulator comprising a non-transitory memory and a processor to store data and execute instructions;
obtaining an emission source that makes the highest contribution to the pollutant concentration of any cell with a 3D Convolutional Neural Network (3DCNN) artificial intelligence algorithm utilizing the simulator, and establishing a model of the relationship between pollutant concentration and emission;
utilizing the simulator to analyze the relationship model with an Integrated Gradients method to estimate influences of input emission data on concentrations of specific cells, and obtaining a final list inversion result;
wherein the step of obtaining an emission source that makes the highest contribution to the pollutant concentration of any cell with a 3DCNN artificial intelligence algorithm and establishing a model of the relationship between pollutant concentration and emission comprises:
reducing the dimensions of weather data from 5 to 1 through the processing of two-dimensional convolution layers, and then activating a ReLU;
reshaping the weather data to match the shape of the emission data in three dimensions;
cascading the three-dimensional emission data with the weather data to form data in a shape processed by a set of three-dimensional convolution layers having an activation function;
the inversion method further comprising a step of predicting emission concentration through a recurrent neural network RNN by using a plurality of timestamped record sequences of emission data cascaded with weather data as inputs;
wherein the step of analyzing the relationship model with an Integrated Gradients method to estimate influences of input emission data on concentrations of specific cells comprises:
using “zero feature” state as a baseline;
calculating gradients of model outputs with respect to input features at an actual input and the baseline input, and obtaining the gradients through back propagation in a 3DCNN model;
dividing a path from the baseline to the actual input into N equally-spaced points; for each point in the path, calculating the gradient of model output with respect to the input features; and for each evaluation point, calculating a difference between the gradient at the actual input and the gradient at the baseline input;
wherein the differences indicate how the significance of each feature varies along the path from the baseline to the actual input;
multiplying the gradient differences with corresponding weights in a Gauss-Legendre quadrature formula, and summing up all weighted gradient differences to obtain a final attributable fraction of each feature;
wherein the fractions indicate the degree of contribution made by each feature to the prediction of the model for a given input; and
wherein the attributable fractions are normalized to ensure that a sum of the final attributable fractions of each feature is equal to the difference between the prediction of the model at the actual input and the prediction of the model at the baseline input.