Hanieh Mohammadi

Evolutionary optimization of user intent recognition for transfemoral amputees

Lower-limb prosthetic legs help amputees regain their walking ability. User intent recognition is utilized to infer human gait mode (fast walk, slow walk, etc.) so the controller can be adjusted depending on the detected gait mode. In this paper, mechanical sensor data is collected from an able-bodied subject and used for user intent recognition. Feature extraction, principal component analysis, correlation analysis, and K-nearest neighbor methods are used, modified, and optimized with an evolutionary algorithm for improved performance. The optimized system successfully classifies four different walking modes with an accuracy of 96%.

Hybrid invasive weed/biogeography-based optimization

Abstract We propose a new variant of the ecologically-inspired optimization method known as invasive weed optimization (IWO). The proposed algorithm features three new components that are typically not present in IWO: (1) migration; (2) gradient descent; and (3) mutation. In standard IWO, each individual uses only its own features (that is, independent solution variables) to randomly distribute new seeds over the search space. In other words, there is no sharing of features among individuals. We propose the application of the migration operator from biogeography-based optimization (BBO) to include the feature-sharing capability in IWO. This modification improves the quality of the distributed seeds (that is, new candidate solutions) in the population. To further improve the local search ability of IWO, we propose the use of gradient descent. Mutation is activated under certain conditions to increase the diversity of the population, and escape local optima. We demonstrate the performance of this new hybrid IWO/BBO on a set of single-objective benchmarks, and on a real-world cyber–physical system problem to optimize a user intent recognition system for transfemoral amputees. Hybrid IWO/BBO is compared to standard IWO, BBO, and 10 other optimization algorithms. The Kruskal–Wallis and Wilcoxon signed-rank tests are used to statistically compare the algorithms. The results for hybrid IWO/BBO present promising improvements over standard IWO. For instance, out of 25 benchmarks, hybrid IWO/BBO performs better than IWO on 18 problems with dimension 30. Hybrid IWO/BBO shows competitive performance with comparison to the 10 other state-of-the-art optimization algorithms.

Multi-objective optimization of decision trees for power system voltage security assessment

A method is proposed for online power system voltage security assessment (VSA) using decision trees (DTs). The DT inputs are the data gathered from phasor measurement units (PMUs). The dimensions of the training data are reduced in two ways. First, the number of features is decreased by principal component analysis (PCA). Second, the number of training cases is decreased by correlation analysis. Biogeography-based optimization (BBO) and invasive weed optimization (IWO) are combined with four multi-objective (MO) optimization methods to find the optimum dimensions of the PMU data while minimizing the misclassification rate of the security test. The four MO methods include vector evaluation (VE), nondominated sorting (NS), niched Pareto (NP), and strength Pareto (SP). A systematic comparison of MOIWO and MOBBO is conducted using Pareto front hypervolume and relative coverage. The method is applied to a 66-bus power grid in Iran. The results show that the training data size is reduced by about 98%, and the training time is approximately 200 times faster because of the dimension reduction. The misclassification rates of the DTs are in the range of 4–9%. Hypervolume and relative coverage indicate that VEBBO performs better than the other methods.

State Estimation of an Advanced Rowing Machine Using Optimized Kalman Filtering

This research addresses the problem of state estimation of an advanced rowing machine with energy regeneration. It is assumed that the states of the system, which are position, velocity, and capacitor charge, are measurable. The user force input to the system can be measured by load cells. It is shown that the need for load cells can be eliminated by estimating the force with an unknown-input Kalman filter. The estimated states and the unknown user force input are passed to the controller of the system, which is either an inversion-based controller or a semi-active impedance controller. Two friction models are considered for this system: Coulomb friction, and LuGre friction. The Kalman gains are tuned using an evolutionary algorithm to minimize the standard deviation of the estimation error. The results verify the effectiveness of the proposed approach for simultaneous estimation of the states and the input force. The standard deviation of the state estimation errors are only 10% of their measurement noise. The standard deviation of the input force estimation error is 0.1 N when using an optimized Kalman gain, which is only 25% of the value obtained when using manually tuned gains.