Research Works

Some of my research work during undergraduate study

Prediction of Silicon Content in Molten Iron Based on Time Series Interpolation-Attention Mechanism

1. Designed a prediction method and software of silicon content in molten iron based on time series interpolation-attention mechanism

2. Implemented downsampling and interpolation processing in time dimension towards input variable to obtain the input variables within a certain period and corresponding silicon content as the training set

3.Built GRU network of time-space attention mechanism and inputted training sample to self-adaptively assigned the attention weight to the process variables to obtain the prediction model

Numerical Analysis of temperature field with PINN

1. Solved the transient problem of heat conduction using physics-based machine learning approach in the forward direction, and analyzed the accuracy and complexity of the physics-based machine learning approach compared to traditional methods

2. Solved the inverse problem of heat conduction using physics-based machine learning method by performing parameter identification and changing different identification parameters

3. Designed several cases with different conditions for the forward problem of heat conduction, verified the correctness of the physically guided neural network approach by using the conventional method and the physically guided neural network method

4. Designed three cases with different parameter identification requirements for the inverse problem of heat conduction, and verified the feasibility of the physically guided neural network for solving the inverse problem of heat conduction

DanceVis: toward better understanding of online cheer and dance training

1. Undertook data preprocessing and front end building task

2. Designed a comprehensive visualization system with characteristics of objective evaluation, fine-grained analysis, high efficiency, high accuracy and clear training paths

3. Divided the overall cheerleading dance performance of one student into nine dimensions, and tracked the individual performance change through the dimension scores of in-class and after-class

4. Classified trainees and further built group portraits which help proposed training paths for each group through nonlinear dimensionality reduction and clustering method

5. Used human pose estimation method to automatically analyze videos so as to improve the analysis efficiency and obtain individual global performance curves

6. Invited experts to conduct with DanceVis, and demonstrated the usability of the system through expert interviews

Link:

Article