Optimizing IMU-Based Locomotion Mode Detection System Using Taguchi Method

Automatic detection of the locomotion mode is crucial in many applications, such as designing controller for exoskeleton robots and other assistive devices. Among the available methods and sensors for detecting locomotion type, inertial measurement units (IMU) have found increasing applications, due to their small size, ease of use, needing no special preparation, and portability. This study performs a statistical analysis to optimize the placement of IMU sensors and determine the minimum number of inputs required for the automatic classification of locomotion type. The k-nearest neighbor algorithm (kNN) is chosen as the classifier and a Taguchi U16 orthogonal experimental design is used to evaluate the effect of different factors on the accuracy of the classification. Quaternion data, acceleration magnitude, and gyroscope magnitude from thigh-mounted and shank-mounted IMUs are considered as the input factors in the Taguchi design. A simple cost function is used to incorporate both the classification error and the number of inputs. The implemented cost function includes a weighting term to identify the optimal number of inputs based on the required level of accuracy. When considering equal weights for the classification error and the number of inputs, the Taguchi analysis ranks the quaternion data, acceleration magnitude, and gyroscope magnitude as the 1st, 2nd, and 3rd order of importance, respectively. For our specific application, a weighting value of 0.6 is used to give slightly more emphasis to the accuracy of the locomotion mode detection. The optimal combination reduces the number of inputs from 12 to 6, while maintaining an acceptable accuracy level of 76.4 percent. This is only slightly lower than the 79.5 percent accuracy achieved when all of the 12 inputs are used. Furthermore, results indicated that the shank-mounted IMU had a greater impact on the classifier’s accuracy compared to the thigh-mounted IMU.