Deep learning frameworks have become indispensable for modeling complex relationships within multi-dimensional time-dependent data across diverse domains. One useful application of these frameworks is in mechanical systems forces and torques forecasting, such as predicting the lift forces generated by a flapping wing system when it follows a predefined stroke profile. This approach, which we refer to as "forward" dynamics modeling, typically involves learning a mapping from movements to forces that were generated by these movements. This work, however, focuses primarily on a distinct task that we term "inverse" dynamics modeling, where the objective is to understand the motion that generates a given force value. For instance, given a desired lift force, the goal is to find the wing movement that will produce it. This task is particularly useful from a control perspective, as it is akin to specifying the desired force outcome and having a model handle the details of achieving it. We achieve this using deep learning time series and enhance the known time-attention-based sequence-to-sequence network (seq2seq) with a novel neural network layer, the Adaptive Spectrum Layer, which learns weights in Fourier space. We evaluate our model using two datasets: one derived from our proprietary mechanical model and measurement methodology, and another open-sourced. Our model's performance, assessed in terms of Mean Absolute Error (MAE), demonstrates comparable performance with state-of-the-art, and up to 11% improvement in MAE in some cases. The contribution of our work is in the simplification and optimization the control aspects of these complex mechanical systems.
This repo contains all the code that was used for my thesis
- Camera: Image processing for wing tracking in 3D and Euler angles extraction
- Forces: Force sensors signal parser
- DataHandler: Data merging, encoding, and preprocessing
- ML: Multi-dataset multivariate time series framework for inverse physical modeling