process AND ALGORITHM
Collect Data
The data is collected using the Matlab app on a phone in the welders glove shown in figure 6. The phone records accelerometer data in the X, Y, and Z axis. This location was chosen as it would pick up motion of the welder's hand while protecting the phone from the heat of the welder.
Figure 6: The orientation of the phone in the welders glove
Pre-Process Data
The data coming in needs to cleaned before it is worthwhile to run a Fourier transform on it. First, the data is converted to an array format to simplify doing operations on it. Next, it is trimmed to a 10 second section of when the user is welding. Lastly the Matlab de-trend function is used to remove any obvious trend (such as a 9.8 m/s/s offset due to gravity). These steps are show respectively in figure 7.
Figure 7: The data on the time domain as it is pre-processed.
Fourier Transform
Figure 8: The untrimmed graph of the Fourier transform of the beginner baseline weld
The data is converted from the time domain to the frequency domain using a discrete Fourier transform (DFT). Specifically, the FFT function in Matlab. The sampling rate of the sensor was 100Hz so the maximum frequency it could measure was 50Hz. The data was shifted to be centered about 0, and scaled on the X axis from -50 Hz to 50 Hz. This is shown by figure 8.
Looking at the data on the frequency domain allows us to identify the major frequencies a welder is moving at as well as the consistency of those frequencies. Figure 9 shows the DFT of welders of varying skill levels from 0 Hz to 5 Hz.
A beginner welder’s frequencies have larger variation, which translates to inconsistent motion and a weaker weld. As shown in figure 9 as the welder gets more experienced, the variation in their frequency decreases and a primary frequency around .7 Hz is visible.
Algorithm
To calculate the skill level of the welder the Fourier transform is trimmed to only contain frequencies from 0 to 5 hz. This was chosen given our expected frequencies for a good welder from the motion model was between .5 and 1 HZ, and in looking at our experimental data there was not motion beyond 5 HZ. Figure 7 shows this range for the three baseline welders.
Then the average weighted distance (where normalized amplitude is weight) from .7 HZ is calculated. .7 HZ was chosen as that was the frequency primary our expert welder moved the MIG gun at. This is shown by equation 1
Equation 1: The weighted distance of a given frequency from ideal frequency
The average of this value for the X, Y, and Z accelerations becomes the quality metric. The range of values shown below came from test data from welders of three different skill levels. Where a welder falls on this range determines their skill and recommendation the system will go.
Beginner: .5
Intermediary: .4
Expert: .25
The algorithm will inform the user where they fall on this scale and advice them to work on consistency of motion and speed to improve their quality score.
Figure 9: The trimmed graph of the Fourier transform of the three baseline welders