In previous work [
11] it is observed that many physical activities only induce very small vibrations
, which are below the noise floor of commercial IMUs. The proposed sensing device aims to amplify these vibrations to improve the detection of activities that only generate very small vibrations.
Section 2.1 and
Section 2.2 outline the generalized model for the oscillation of ambient objects.
Section 2.3 and
Section 2.4 illustrate the FEM simulation to analyze the use case and generate an optimized beam design. In
Section 2.5, the sensor fusion algorithm is presented, which combines accelerometer and gyroscope data to reduce the intrinsic sensor noise. The experimental setup is shown in
Section 2.6, which serves to evaluate the proposed concept.
2.2. Beam Structure for Vibration Amplification
The proposed sensing device employs an additional spring mass damper system situated between the environmental object and the IMU, which is shown in (B) of
Figure 1. By optimizing the parameters
,
and
of this spring system, an amplified oscillation
can be achieved. This setup forms a two-mass-spring damper system. Depending on the scenario, the influence of the sensing device on the environmental object cannot be neglected. Analytical analysis of systems with complex geometry is near impossible and numerical methods are generally used to find solutions. In this work, finite element method (FEM) analysis in Ansys Mechanical [
20] is employed to analyze the oscillating system, which is presented in
Section 2.3 and
Section 2.4.
It is proposed to implement the additional spring mass damper system as a beam structure. One end of the beam is mounted to the ambient object and the IMU is mounted on the other end. Furthermore, the beam is integrated in the PCB of the device and contains the electrical connections. This approach has many benefits, such as ease of manufacturing and implementation in devices. The beam is designed and optimized using FEM simulation, which is described in detail in
Section 2.4. Although not suitable for determining vibration characteristics, the classical equations for the deflection of a beam under load [
21] are helpful in formulating design criteria. The forces that act on the sensing device are shown in
Figure 2.
An activity-induced force pulse
acts vertically on the cage floor, causing a natural oscillation of the cage. As described in
Section 2.3, the cage floor oscillates like a membrane. The displacement
of the cage floor oscillation exerts a force on the sensing device. This force acts as a vertical reaction force
on the base of the beam. It is directly opposed by the inertial force of the beam
and sensor
. In order to calculate the displacement at the end of the beam, where the IMU is located,
is neglected as it is much smaller than
. The tilt angle and displacement are calculated as:
where
is the displacement,
is the tilt angle,
E is the Young’s modulus and
is the moment of inertia [
21]. Equation (
2) further shows that the displacement
and the tilt angle
are linearly correlated. As a six-axis IMU is capable of measuring both variables directly, this circumstance allows for sensor fusion, which is further described in
Section 2.5.
For proof of concept, a first design of the beam is proposed here. Other designs will be explored in the future. This design resembles a bending beam, which is split laterally and folded to form two parallel braces that are shaped around the base. This design shifts the center of mass further to the base of the beam, which reduces the moment generated by the beam. It is shown in
Figure 3 and
Figure 4. The manufacturing of the prototype requires some design restrictions: The beam needs a surface to mount to the cage and one to mount the IMU, where the surfaces have a size of 17 mm × 17 mm each. The braces need a minimum width
w of 3 mm to accommodate the data lines. The interspace in the beam needs to be at least 4 mm wide.
For prototyping purposes, the system on a chip (SoC) of the IMU is not mounted directly to the beam structure but to a small daughter board shown in
Figure 4, which contains the IMU. This daughter board is easily desoldered and mounted on another prototype for testing without damaging the delicate SoC.
2.4. FEM Analysis Beam Design
In order to optimize the design shown in
Figure 3, a large number of parameters and dimensions can be adjusted. Some parameters are already set due to manufacturing, assembly and mounting requirements. The remaining parameters are the length, height, width, Young’s modulus and mass. While the mass is proportional to the overall size, the height and Young’s modulus can only be changed by using a different raw material. Therefore, the main parameters for optimizing the design are the length and width. These are defined in
Figure 3 as the width
w of each of the symmetrical braces and the length
l of the outer brace.
In order to analyze the beam structure in its intended use case scenario, the simulation includes the cage as described in
Section 2.3. The beam structure is mounted in the middle of the cage floor with a spacer to allow for oscillation. The geometry of the sensor board is not modeled since it does not influence the oscillation characteristics, but the mass of the sensor board is simulated with a surface mass. The IMU is represented by a 2 mm × 3 mm × 2 mm rectangle and assembled on the sensor board mount surface. The material characteristics of the beam are set to a standard FR-4 glass-reinforced epoxy laminate material from the Ansys library.
The modal analysis depicts the oscillation characteristics of the beam when mounted on the cage and shows that the beam exhibits many oscillation modes. However, some of them have a similar form of deformation and typically occur in the same frequency range. The oscillation modes are grouped together based on their deformation form. In
Figure 6 four different deformation forms are presented of a beam structure with
mm and
mm.
The first deformation form appears in the range of 99 to 118 and exhibits a bend in the z-axis, which is accompanied by a rotation around the x-axis. The second form shows a sideways movement in the x-axis and occurs between 398 to 426 . The third form shows a twisting deformation around the y-axis and appears in the range of 648 to 745 . The fourth form shows a twisting motion around the x-axis and a back-and-forth movement in the y-direction. It occurs in the range of 878 to 1101 . Beam structures with similar dimensions exhibit the same deformation forms at different frequencies.
A transient structural FEM analysis is performed to evaluate the activity-induced structural vibration of the cage. The simulation is long with a step time of 50 . At 1 ms, a force pulse of is applied to the center of the cage floor. Afterwards, the system continues its dampened oscillation undisturbed. The result of the simulation is the acceleration at the location of the virtual IMU in the x-, y- and z-direction. The amplitude of the oscillation modes is evaluated in the frequency domain by Fourier transforming the output acceleration data.
Based on this simulation setup, a design with maximum vibration magnitude is identified in
Section 3.1 by comparing beams with different
w and
l dimensions to determine optimal dimensions.
2.5. Sensor Fusion Algorithm
The IMU data consist of six signals, which are the three acceleration signals
,
and
and the three angular rate signals
,
and
. In the standard configuration, where the sensor is mounted directly on an object, the different sensor signals of the IMU are not directly correlated and cannot be fused in the time domain e.g., to reduce noise. This is shown in [
11] where statistical information from different sensors is combined to improve classification accuracy. However, the signals could not be combined on a per-sample basis to reduce noise or increase the SNR.
The simulation of the beam structure in
Section 2.4 shows that certain oscillation modes of the beam are composed of displacement in multiple axes. This creates a mechanical coupling between the different signals of the IMU, which enables sensor fusion. Using prior knowledge of the oscillation modes, a sensor fusion algorithm is proposed that combines multiple correlated sensor signals to create a single signal with lower noise. As the SNR varies between signals, a main signal is determined first, which has the highest SNR. The correlated signals are then combined to reduce the noise of the main signal.
The sensor signals are modeled as a vibration
f with additive Gaussian noise with zero mean. The noise of the different signals is uncorrelated, since the intrinsic sensor noise is mainly due to the thermal noise of the proof mass in each sensor. The underlying theory is presented in the following. Two signals of different sensors are denoted as:
where
x and
y are the two measurements,
denotes the corresponding additive noise,
is the variance of noise and
f is the vibration. A weighted estimator
z is constructed to combine the signals
x and
y:
where
is the noise of the estimator
z,
is the noise variance and
c is the weight. By minimizing
in Equation (
7), the optimal value of
c is determined as:
In the ideal case
, the weight is
. Using Equation (
7), the standard deviation of noise of the estimator is calculated as
. For an example with two signals with
, the weight is
and the standard deviation of the estimator is
.
There are three main challenges in fusing IMU data. First, the sensor signals of the IMU, which are supposed to be fused, are correlated but do not have the same amplitude, which is denoted as f above. Therefore, the signals must first be transformed to equalize the amplitude. For this purpose, a linear transformation is proposed as it maintains the normal distribution of the noise. Second, the structural vibration measurement contains many oscillation modes simultaneously and the correlation as well as transformation between the signals depend on the mode. However, since each mode has a different frequency, it is possible to fuse signals by combining data with regard to the frequency. Third, the noise variance is required to calculate the optimal weights and needs to be estimated.
Figure 7 presents an overview of the sensor fusion algorithm. The first step in the sensor fusion algorithm is preprocessing, which involves discrete differentiation of the angular rate signals
,
and
to obtain the angular accelerations
,
and
. The signals are then decomposed using multi-level discrete wavelet transformation (MLDWT) [
23], which is a key element and enables the frequency-dependent fusion of IMU data.
MLDWT is a wavelet-based analysis method for discrete signals, which extracts time-frequency features by decomposing time signals into low-frequency and high-frequency subseries on multiple levels. The MLDWT decomposes a discrete time signal into one approximation coefficient and M detail coefficients. Each coefficient represents time-frequency information of a unique range of frequencies. By fusing the wavelet coefficients of different signals at a specific level, signals can be fused with regard to frequency. The result of the fusion is a single wavelet coefficient. The appropriate mother wavelet for the MLDWT is determined experimentally by minimizing the error of a sample signal with vibrations and the transformed and recomposed version of the same signal. It is chosen to be a Coiflet wavelet with five vanishing points. Furthermore, the number of decomposition levels is , where N is the length of the signal.
In the next step, data fusion is performed for each previously defined mode. The wavelet level
, which contains the frequency component
of an oscillation mode
u, is calculated as:
where
u is the index of the mode,
is the sampling rate and
is the frequency of the oscillation mode. The level
is then used to extract the relevant wavelet coefficients. Fusion at each oscillation mode requires some prior knowledge of the system, which is provided by the fusion array
. Each element of the fusion array contains a vector with six coefficients
to
and the frequency
of the oscillation. The coefficients
to
describe the transformation between the signals. The coefficients are determined experimentally by measuring high amplitude vibrations and numerically estimating the transformation coefficients by minimizing the mean squared error (MSE).
A transformation coefficient of 1 is assigned to the main signal with the highest SNR. As a result, the other signals are transformed to match the amplitude of the main signal. Signals that are not correlated by the beam oscillation have a coefficient of zero and are removed by the Signal Selection Filter. An example for an element of the fusion array is shown in Equation (
24). The transformation of wavelet coefficients is a multiplication with the transformation coefficient:
where
is the wavelet coefficient of a signal from the IMU,
is the corresponding transformation coefficient and
is the transformed value. After transformation, uncorrelated signals are removed by the Signal Selection Filter, which removes signals with a total sum of zero. In the next step, signals are combined by the adaptive combiner. The optimal weights are calculated as:
where
i is the index representing the different signals for fusion,
is the optimal weight for signal
i,
is the standard deviation of noise of signal
i and
I is the total number of signals to fuse. The result is a single wavelet coefficient containing the combined sensor signals. The time signal of the fusion result is obtained using inverse MLDWT. The result is a time signal with a narrow frequency band, as it only represents the oscillation of a single oscillation mode.
The noise standard deviation
of each signal needs to be estimated. The implemented estimator is based on [
24] and estimates the noise variance by calculating the median of the wavelet coefficients in a section where
f is piecewise smooth. This is a section where little or no vibration occurs. The standard deviation of the noise is estimated to be:
where
is the median of the signal
. The analysis of the sensor noise shows that the noise in the accelerometers has a uniform noise density and Gaussian distribution. In this case, the best estimate is calculated by analyzing the finest-scale wavelet coefficient. The noise in the gyroscope signals is unevenly distributed in the frequency domain and increases at higher frequencies. Because the noise is level-dependent, it must be estimated for each wavelet coefficient individually.
In the real-world scenario, piecewise smooth sections need to be identified to update the noise estimates. In the experimental measurements presented here, no piecewise smooth section is available for noise estimation. Therefore, the noise variance is estimated prior using a separate measurement without vibrations. It is provided as the noise matrix:
2.6. Experimental Setup
The proposed concept is validated experimentally by measuring structural vibrations of a husbandry cage using different sensor arrangements. The vibrations are generated with a custom-built force pulse generator. The experimental setup emulates activity-induced structural vibrations with force pulses that act vertically on the cage floor. The setup includes: mice cages of type Zoonlab HRC500 [
22], 200
of wood flocking, an aluminum stand for the cages, a force pulse generator, a heavy aluminum base to reduce external vibrations and several microcontrollers. The components are shown in
Figure 8.
There is a wide selection of commercial multi-axial IMUs available such as the Bosch BMI270 [
25], STMicroelectronics ASM330LHHTR [
26] and the TDK InvenSense ICM20948 [
27]. For this work, the ASM330LHHTR was chosen due to its superior noise characteristics and sampling rate. It has a noise density of 60
/
z for the accelerometers and
dps/
z for the gyroscopes, which is lower than the other IMUs mentioned. Furthermore, it has a maximum sampling rate of 6667 Hz for all sensors.
Two sensor arrangements are evaluated, which are shown in
Figure 9. The first arrangement serves as reference and has an IMU directly mounted in the center of the cage. This location is deemed the optimal location in
Section 2.3. The second cage has an IMU with the proposed beam structure. The beam uses the optimal design that is covered in
Section 3.1. The prototype is milled from a piece of FR-4 PCB material. Both arrangements use the same STMicroelectronics ASM330LHHTR sensor.
Another parameter of the experiments is the location where the force pulses act on the cage. Two excitation locations are defined to investigate whether amplification also occurs when the cage is excited at different locations. The first location is the center of the cage floor. The second location is beside the center and about half the distance towards the short side of the cage. The locations are illustrated in
Figure 8.
2.6.1. Force Pulse Generator
The experiment requires a purpose build force pulse generator, which is able to generate very small force pulses. The force pulse generator uses a magnetic actuator and is driven with a rectangular signal of 12 V. The pulse amplitude is controlled by the duration of the rectangular control signal and is modulated between 0
and 1000
. At 12
, the actuator generates a constant force of
, which amounts to a force pulse range of 0 to
. The force pulse is calculated as:
where
is the duration of power supplied to the actuator and
is the force generated by the actuator. The repeatability error of the experimental setup is explored in
Section 2.6.3.
For comparison, the pulse of a free-falling object can be calculated as when air resistance is ignored. A 25 mouse would therefore create a pulse of when it lands from a height of five millimeter.
2.6.2. Experimental Measurements
The experimental evaluation aims to explore the characteristics of the proposed sensing device compared to a directly mounted sensor for different force pulse strengths and excitation locations. In the experiment, force pulse sweeps are used to investigate the aforementioned characteristics. The individual force pulses are spaced apart by
, which allows each structural vibration to decay before the next starts. Each measurement begins with a starting sequence of four force pulses, after which 100 force pulses are generated at a rate of 2
. The pulses go up in equidistant steps. In this work, several experiments are presented which have different amplitude ranges and location of force pulse application depending on the evaluation. A sample of a measurement is shown in
Figure 10.
2.6.3. Repeatability of Experimental Setup
The repeatability error of the experimental setup is assessed throughout the entire range of force pulse amplitudes. For this, twenty amplitude sweeps are recorded using an IMU without the beam structure. The amplitude sweep consists of 100 pulses from
to 2160
and the force pulses are located in the center. Based on the acceleration data, the vibration magnitude is calculated for each force pulse. Afterwards, the mean and standard deviation are calculated to determine the average magnitude and error. As a measure of magnitude, the energy of the acceleration magnitude
is calculated. The energy is calculated over a period of 0.2 s, which is the longest duration of a decaying vibration. The magnitude and energy are calculated as follows:
where
is the energy at a certain force pulse
i,
N is the number of samples and
is the sampling rate of the sensor. The mean and standard deviation of
are shown in
Figure 11.
The average vibration magnitude in the left plot of
Figure 11 has a flat slope at the beginning that becomes steeper as it progresses. The error depicted in the plot in the middle follows a similar path and increases with the magnitude of the vibration. The relative error shown in the right plot is calculated as the fraction of error and average. It shows a declining trend with a variation of about ± 1%. In the beginning, the values average about 3% and decline to about 1% at the end.
The results illustrate that the experimental setup is capable of generating structural vibrations with a high degree of repeatability. Furthermore, it is observed that the energy of the vibration magnitude does not rise linearly with the force pulse amplitude. Although starting at about 1550 , the average energy appears to enter a quasi-linear region.