2021 CareerHack [TSMC X Microsoft]February 22, 2021
Smart Factory, smart manufacturing
Using Azure IoT and Azure Sphere Board
By the research in Manufacturing statistics in Great Britain, 2020, we know that up to 15% of fatal injuries were caused by contact with machinery.
When an emergency happens, when will the machine stop? As we know, the machine stops when detecting it involves the abnormal object. When a machine involves abnormal objects, involvement might cause injuries that are already too late to stop the machine.
So, we decide to stop the machine immediately when emergence happens by using sensors and machine-learning techniques.
To stop the machine, we acknowledge that putting sensors on the machine makes difficulty high caused by different volumes of machines. So we change our target to human because the relation of human is an obvious signal to tell whether this person is in an emergency or not.
The fight-or-flight response is a physiological reaction that occurs in response to a perceived harmful event, attack, or threat to survival. 
The most known physical effects are
- Acceleration of heart and lung action
- Paling or flushing, or alternating between both
- Inhibition of stomach and upper-intestinal action to the point where digestion slows down or stops
- Constriction of blood vessels in many parts of the body
- Dilation of blood vessels for muscles
- Inhibition of the lacrimal gland (responsible for tear production) and salivation
- Dilation of the pupil
Besides detecting the signal of the physical effects caused by the fight-or-flight response, to stop the emergency timely, we need the reaction that generates fast enough. In this case, we select the effect from the lung, which is breath.
Breath can be easily detected by the microphone, we design a mask that contains a microphone and build a Deep-Learning model to detect the emergence by sound inside the mask.
Before building our model, we would have to make sure that breath is detectable from our microphone, by using Fourier transform, we could draw a spectrogram to visualize the composition of sound.
It could be seen that the feature is obvious, which means that detecting emergence is quite possible.
Input Data: Time series of 22050 values per second
Output Data: 4 categories (no breath, breath, cough, speak)
In the end, we use 11,128 params and achieve 95.778% accuracy.