No description, website, or topics provided.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

Intel Hacks 2017 - Hackathon


Intel & Curie

Bluetooth LE

Central: A Bluetooth device such as smart phone, tablet, or PC that initiates an outgoing connection request to an advertising peripheral device. Once connected to the peripheral, the central device can exchange data, read values from the peripheral device, and execute commands on the peripheral devices.

Peripheral: A BLE device that accepts an incoming connection request after advertising. It gathers and publishes data for other devices to consume.

The central device communicates with peripherals through advertising packages. Peripheral devices send out the advertisements, and the central device scans for advertisements.

Central and Peripheral Devices


Communication Diagrams

Node-Red Flows

Node-Red Flow Main Pipeline

Node-Red Flow Dashboard

Node-Red Dashboard

Node-Red Dashboard


Demo CountPages alpha

Data Processing

We'll be using an MEMS (MicroElectoMechanical System) IMU (Inertial Measurement Unit) in the form of MPU9250 which includes an accelerometer, gyroscope and magnetometer. For our purposes, the magnetometer, which gives "global" orientation referenced to the earths magnetic field, is useless.

The gyroscope returns the pitch yaw and roll of the unit itself, allowing us to orientate the ragdoll model on the frontend to better represent the orientation of the person when the seizures started. While gyroscopes over time do drift by a nontrivial degree, realistically for the display purposes required, if we're within 15° of the true value, then this is good enough.

The accelerometer returns angular accelerations in the x y z directions, and is primarily used to show the intensity of the seizures.

The gyroscope will give pitch, yaw and roll of the unit itself, with the acceleration allowing us to measure the magnitude of the seizures. Previous experience with the MPU9250 shows that there will be a reasonable amount of noise at all frequencies. It is expected that there will be more noise in the kHz range due the the buck converter used.

To improve the accuracy of the accelerometer, outputs of the gyroscope and accelerometer will be combined using a form of sensor fusion called a complementary filter. The drift from the gyroscope won't be a large issue due to the small timescale used for the differentiation.

The formula for this complementary filter is $acceleration = accelerometer values 0.85 + 0.15 (current gyro value - previous gyro value)/(current millis() - previous millis()) .$

Note: to keep it simple to read i'm not using proper mathematical notation

This acceleration vector, which contains elements x,y,z, will be turned into a magnitude, since information about x y z accelerations will be useless, considering its relative to the sensor itself. This is transformed into a scalar magnitude using a 3D version of 3D Pythagorean theorem.

$Magnitude = sqrt(x^2+y^2+z^2)$

Once we've got this information there are two main ways that it can be interpreted

Use the Arduino 101s built in 128 node neural network hardware

Creating a neural network to analyze this data is less than desirable

  • There are no existing pre-trained neural networks for this problem, we would need to train our own network which would involve manually trimming seizure events (time consuming on both fronts).
  • These neural networks would likely be specific to the region that the sensor is applied (an arm seizure would have different characteristics to a leg seizures).
  • In order for this model to be generalized, data would be required from multiple patients, a time consuming effort.

Characterize high changes in magnitude as seizure events

This would also be undesirable for three main reasons

  • The accelerometer is a noisy signal. With the accelerometer being heavily weighted by the complementary filter, this noise will also be heavily weighted. Seizures can be low magnitude events, making it possible for noise which is either difficult or impossible to eliminate to be characterized as a seizure.
  • The change in magnitude is highly dependent on the sampling frequency of the accelerometer, too high of a sampling frequency and we may not have a high enough resolution to properly characterise the event. Too low and we may miss the event.

Use Fourier transforms

Not to be confused with the four year transform


This is where we use some observed properties of a seizure combined with some mathematical knowledge to make our life easier. The key observed property, is that a seizure is likely to occur in the 1-10Hz range. Therefore, we are looking for changes for these specific frequencies. To do this, we use the Fourier transform to go from the time (sample) domain into the frequency domain. The clearest example of this is provided at Mathworks, Where we take a noisy signal from the time domain.


And apply a Fast Fourier Transform (FFT) which takes this discrete signal and transforms it from the time time domain into the frequency domain.


Which shows that there are two signals in there: one at 50Hz and one at ~125Hz.

We will be applying this principle for our acceleration data, but will be looking for spikes in the 1-10Hz range, which we know is approximately the frequency that a seizure will occur at (try to shake your hand faster than 10 times per second).

It's important to note that we're starting at 1 instead of 0Hz, because we expect that our accelerometer data will oscillate around a semi constant non zero value, with the Fourier transform of a constant being represented by a spike at 0Hz.


We will be using the Arduino FFT library provided by Kosme.

Immediately we know that we must be sampling our data at least as fast as the Nyquist rate, which is 2x the highest frequency that we want to sample. In this case as we want to measure up to 10Hz, the sample rate (Fs) will be 20Hz.

Once this sample rate is determined, we need to figure out the number and size of the FFT bins.The FFT function as implemented by Kosme returns the frequencies in "bins" eg [0 to binwidth, binwidth to 2xbinwidth] Knowing on the lower bound that we want to cut out at least the 0-1hz region, we will want a bin width that allows us to cut out approximatly 1hz. Having a preset sampling frequency (Fs), and knowing that $bin width = Fs/(N Bins)$, where number of bins has to be a power of 2. We incur a non trivial speed penalty on all FFT functions when increasing the number of bins (which is expected).

Speed characteristics

Function run reorder window lin lin8 log
N (ms) (us) (us) (us) (us)* (us)
256 6.32 412 608 588 470 608
128 2.59 193 304 286 234 290
64 1.02 97 152 145 114 144
32 0.37 41 76 80 59 74
16 0.12 21 37 46 30 39

Data from openmusiclabs

We therefore pick 32 bins, giving us a bin width of 0.625Hz, meaning we will be ignoring the first two bins, giving us a range of 1.25Hz to 10Hz.

Each of these bins will return a scalar intensity value. The units themselves are reasonably arbitrary

Improving our FFT model

There are two ways that our FFT model can be improved

  • remove the noise floor (see the single sided amplitude spectrum)
  • Consider that the higher the frequency the event, the worse it is, and apply a multiplier based on that.

##Model //System model #Subsystem1 #Subsystem2 #Subsystem3 #Datalogger This is the component that measures acceleration values It consists of:

  • HC-10 Bluetooth module to allow for BLE communication with the Arduino101 router
  • A MPU-9250 IMU to get accelerometer and gyroscope values.
  • An arduino mini to do read and do the math for the accelerometer and gyroscope values, as well as stream the data over bluetooth.
  • A Lipo charger
  • A 18650 rechargable lithium battery
  • A boost converter to step up the voltage from the 3-4.2V from the battery to 5V required by the arduino and IMU.
  • A case to fit it all in. //Wiring Diagram img