Weft Knit Strain Sensors for Human Motion Capture

Weft knit strain sensors are textile sensors that measure motion at the joints of a human body. These sensors are formed by the weft knitting of conductive yarn and can be implemented as wearable devices that resemble normal clothing. The conductive yarn is a tactile yarn with electrical properties formed from a composite of stainless steel and polyester filaments. Although, it can be mechanically manipulated to create a woven or knitted fabric that measures strain, weft knitting is the optimal method because of its elastic structure. Due to the novelty of this research area, there are a lot of gaps that prevent the wide-scale adoption of this sensing mechanism. This thesis aims to address two of these gaps. Firstly, the lack of a model that accurately simulates the electromechanical behaviour of structural variations of a weft knit strain sensor. Secondly, the lack of a wearable device framework that illustrates the design of the sensor, its implementation in a wearable device and the processing of its acquired data using machine learning algorithms.

Firstly, an electromechanical model that simulates the behaviour of plain knit sensor with only conductive yarn is proposed. The length resistance are obtained from the loop and interlocking angles of the conductive loops in the sensor and the contact resistance is derived using a novel algorithm. The model was validated by a tensile test performed on sensors with the simulated knitting parameters. It was shown that the simulation results agreed with the experimental results. In particular, the proposed model has a lower percentage error in comparison to previous studies. Furthermore, the effect of changes in the loop and interlocking angles on the piezoresistivity of the sensor is simulated.

Thereafter, this model is applied on a novel sensor configuration comprising of a conductive yarn and a non-conductive elastomeric yarn. The simulation results provide a very accurate representation of the empirical piezoresistivity of the sensor. Subsequently, we create a wholly textile data glove by knitting its support structure and its weft knit strain sensor in a single manufacturing process using WholeGarment technology. The data glove measures motion at the interphalangeal joints in a human hand. The consistency of the glove is verified and classical machine learning algorithms are applied to classify the data acquired using a robotic hand.

In addition, deep learning is evaluated in a grasp classification using the weft knit data glove on human participants. A convolutional neural network (CNN) algorithm is proposed to classify the grasp type from the acquired data. Classical machine algorithms are also used to classify the data to provide a comparative performance. The results illustrate that the CNN algorithm achieved a higher accuracy than the classical machine algorithms in the classification scenarios.

Finally, the effect of miss and tuck stitches on the piezoresistivity of the sensors is investigated. Miss and tuck stitches affect structural properties such as the length, width and extension of a knit fabric and therefore, may affect the piezoresistivity of a weft knit strain sensor. By adapting the electromechanical model to geometrical properties of miss and tuck stitches, several sensor configurations comprising of varying amounts of tuck or miss stitches are simulated. Subsequently, a tensile test is performed on knitted sensors with the simulated properties. The simulation results generally agree with the experimental results. Moreover, it is observed that increases in miss and tuck stitches decrease the initial and mean resistance of the sensor. In addition, the results show that increasing the percentage of tuck stitches in the sensor increases the linearity of a weft knit strain sensor.