Project Title

Fall Sensing

Skills

Software, Fall Sensing, Data Analysis, Data Regression, Sensor Testing

Project

Worked with a team to develop a fall detection system using biometric, collected from wearable sensors, to create a predictive model that can detect falls andpotential injuries in manufacturing workers, using regression models and statistical analysis.

Project Background

Problem Statement

• Slips and falls are one of the most common sources of injury at factory sites
  • In 2022, Fall Protection was OSHA’s top safety violation (5260 citations)
• Falls → broken bones, spinal injuries, sprains, head injuries
  • Most common injuries: fractures and disjointed joints in wrist and ankle
• Slip and fall injuries are leading cause of worker compensation claims – $70 billion annually in the U.S
• OSHA has guidelines for prevention and protection, but no standards for detection, diagnosis, and treatment
• No consensus for definition of a fall: hard to develop thresholds for what constitutes a fall

Project Goal

Our goal was to develop a model that predicts if a wirker has fallen using biometric data, collected from wearable sensors. This model could then later be used to determine the severity of an injury. This system would be useful in situations where a worker falls in a remote location and can't communicate with others that they may be injured and are immobile. Data collected from this model could also able be used to identify what tasks may have a high risk of an individual falling.

Testing Procedure

Our tests were run using five ADAM wearable sensors that were placed on one taller user and one shorter user, as seen below

These sensors were connected wirelessly to an iPad, using the MyRA app, allowing for the collection of biometric data measured by the sensors. For each trial a user stood for about 10 seconds, allowing for the sensors to "zero out" for recalibration. The user would then be asked to bend over five times or fall five times, in a row, on a mat with break intervals between each action, as seen below. Bending trials were ran to collect data to, ensure that a user bending over wouldn't produce false positives that a user has fallen. The user also wore an Apple Watch, during trials, to compare our fall detection system results to the fall detection on an app watch. Following the trials, data collected from the sensor was shared from the iPad to a Google Drive Folder. Ultimately, we chose to focus on the acceleration data generated from the sensors, as the data showed the separability between bending and falling.

Results and Findings

Following analysis, we also found that the separabilty of the acceleration data also varied by sensor, as data collected from the sensors located on the arm tended to have some additional noise and inconsistencies due to how a user may bend over or brace themselves during a fall. We ultimately focused on sensory data collected from the sensor on the user's chest. Below is the acceleration data collected from the chest sensor of a user completing five trails of bending and falling.

Again, The data collected from the sensor placed on the sense seemed to be the most consistent and separable, as it is closest to the person's center of mass, and the data collected from the sensors on the arm can vary based on the orientation of each person's arms. The acceleration for the chest sensor also followed a fairly normal/symmetric distribution throughout the time duration, which allowed us to set an acceleration threshold for our trials, due to the consistency in magnitude and height of falls and bends.

We also tracked the duration of a fall, using a threshold, and identifying tthe duration at which when the acceleration was above the threshold. This allowed for an additional metric to compare falls vs bends or other acceleration spikes. Using this mwthod, we divided the data into equal intervals and determined the total time in each interval that was over a given threshold of 0.5g.

The interval data along with accelerometer data was used for a linear model to classify falls vs non-falls.

Sensor Detection Summary

Performance Summary

Our model had a 0% misclassification rate for falls for both people we ran trials with, however we found a 44% false positive rate for bends. This proved that our model can be fit, but not ideal. Additionally, the Apple Watch fall detection was not triggered during the bending or falling trials.

Areas for Improvements

The following are oportunities for improvement that we found in our process for generating a model for fall detection:

• Build a more precise or adaptive threshold(could be something through an optimzation alogrithmn or a more robust fitting model). This idea would also require more trialing for additional data collection of user's with different body characteristics.
• Utilize additional data metrics captured by the sensors such as posture score, cough index, temperature data and more where we saw small correlation to a person falling, but weren't sure how to best utilize the data alongside the aceleration data.
• Collect more data to observe if there is variation in data between a person's sex, height, and weight. Part of this would also include testing alternative falling and bending motions as we focused on forward motions, so testing backwards slips or more unorthodox falling positions.
• Adding a feature to classify falls that occur based on severity of potential risk of injury.

Learn More

if you would like to learn more about this project contact me here