Multi-axial analysis of vibration with IMU-based sensors for activity monitoring

Activity monitoring is an essential tool to measure the physical activity of living beings, which is used to evaluate wellbeing, detect dangerous events and observe behavior for research. Recently, vibration-based activity monitoring has gained increased research attention, as its sensing principle enables non-contact monitoring without line-of-sight. The method analyzes activity-induced vibrations of objects in the vicinity of living beings to estimate their physical activity. A promising use case is large-scale animal monitoring in research, where previous methods had limited practical success due to poor scalability or accuracy. However, previous vibration-based activity monitoring methods solely relied on commercial geophones which offer high sensitivity but are unsuited to this use case due to limited bandwidth, a large physical size and a higher cost.
This thesis presents a method for vibration-based activity monitoring that utilizes a compact six-axis vibration sensor based on an inertial measurement unit (IMU) to enable monitoring of mice in the home-cage monitoring (HCM) scenario. HCM refers to methods to monitor home-cages, which are transparent containers that are employed at research facilities for long-term housing of rodents. The proposed method measures activity-induced vibrations of the home-cage with a small IMU-based vibration sensor and analyzes the vibration to extract activity-related information and classify physical activity. A major challenge is the low amplitude of activity-induced vibrations, which has resulted in an insufficient signal-to-noise ratio (SNR) in a preliminary study where an IMU was directly attached to a cage for vibration measurement. For this reason, this thesis proposes the novel tuned-beam IMU vibration sensing device that is able to measure activity-induced vibrations with an excellent SNR. The tuned-beam IMU measures vibration in six axes using the accelerometers and gyroscopes of a commercial IMU, and integrates a beam-shaped support structure into the PCB of the sensor. The beam structure is designed to oscillate in resonance with the cage's vibration, which magnifies the measured vibration amplitude and substantially increases the SNR. The beam geometry is optimized in a transient structural finite element analysis (FEA) and finely tuned with an experimental procedure in the assembled state. Additionally, a sensor fusion algorithm is presented that fuses signal components across the IMU's sensors in the wavelet space to reduce sensor noise. It combines correlated signals of accelerometers and gyroscopes that are generated by certain normal modes of the beam structure. Furthermore, a robust classification algorithm is developed that classifies multi-axial vibration sequences of variable length. It analyzes the sequences with multi-level discrete wavelet transformation (MLDWT) to extract sequential time-frequency features. These features are then classified by a convolutional neural network (CNN) -- long short-term memory (LSTM) classification network to predict the activity that had generated the vibration. The network is designed to extract local and long-term patterns of the vibration that are predictive of the activity class. The proposed method is complemented by a camera-based reference system that is used to label the vibration sequences. It uses a commercial behavior analysis software and a custom post-processor that corrects errors which are a result of the challenging home-cage environment. The proposed method is verified in an experimental study where data is collected over a week. The method is able to predict activity with high accuracy and is able to monitor long-term activity with comparable accuracy to the reference method while using low-cost hardware. In summary, the proposed method presents a high-performance cost-efficient method that has a high impact on scalable monitoring methods for the HCM use case. This enables automated evaluation of wellbeing, optimization of caretaking procedures and procurement of unbiased long-term behavioral data for research, which are all in high demand.