CAPACITIVELY COUPLED EMG ELECTRODES WITH FINGER GESTURE RECOGNITION

Figure 6: Capacitive EMG electrode placement and testing
2. Post Filtering circuit testing :
The filters of the post processing circuit were required to be tested. A sinusoidal signal of 20 mV was given from the function generator to the filter and the frequency was swept from 0 Hz to 1000 Hz. The amplitude started increasing from about 5 Hz and remained stable and constant until 200 Hz and started decreasing and completely attenuated at 600 Hz, showing the proper functioning of the analog filter. 60 Hz signal could be seen superimposed on the desired signal, but we chose not to have an analog notch filter, and chose to filter out the 60 Hz using digital filter. We did not connect the actively driven ground yet. Reading a few research papers on non contact electrodes, we realized the importance of an actively driven ground [5],[6]. Thus, we connected all the inverting terminals from the filters of all electrodes to an inverting amplifier. The inverting amplifier inverts the signal and feeds back to the body via gel contact electrode. The negative feedback drives the common mode voltage to a low voltage. Having connected the actively driven ground, we saw a significant reduction in the noise level. After testing on the breadboard, we chose to move on to the solder board, being the cheapest option if compared to the printed circuit boards.

To significantly be able to differentiate the signals from every finger, we chose to have multiple EMG electrodes. As the biopotential signals are best understood as differential signal, we chose to give the Nidaq differential input. Initially, we made one common negative input to the Nidaq, and had 3 positive input. This configuration was not suitable as we found two signals tracking each other. So, we chose to have two differential negative inputs with their two respective positive inputs. We had to compromise on the number of signals, as now we just had 2 signals as compared to 3 initially. We could get decent signals on the MATLAB GUI now, with two signals significantly changing with every finger moving. We encountered an unexpected issue with this configuration. Even if one electrode was disconnected from the arm, tapping it would significantly change the output of the other connected electrode. It meant all the four electrodes were coupled to it each other, and any movement of one would show a change in the output of the other. The reason of such coupling was not very clear, but one of the reason was the long wires connecting all across the circuit from sensors to the board, and to the Nidaq. We eventually chose to work with single differential signal, i.e. with two electrodes only being one positive and second negative.

SOFTWARE DESIGN
There are two MATLAB programs. The first program (signals.m) basically displays the EMG on a GUI. The second program displays a hand on a GUI which can be controlled by different finger motions. Both programs are similar in the setup of the Nidaq. Each program starts by setting up the nidaq for data acquisition. The command delete(daqfind) is used to remove any existing device objects in the workspace. Then, ‘nidaq’ is defined as the input, and an analog input object is created and assigned one channel input. The analog input is then configured to trigger manually, with the sampling frequency set to 1000 samples per second and samples per trigger set to 1000. The input type is also set to differential mode:

delete(daqfind);
adaptor='nidaq';
adaptorData = daqhwinfo(adaptor);
id = adaptorData.InstalledBoardIds{1} ;
ai = analoginput(adaptor, id);
set(ai,'InputType','Differential');
ch0 = addchannel(ai, 0); %differential input 1
Fs=1000; %sampling frequency
set(ai, 'SampleRate', Fs);
set(ai, 'SamplesPerTrigger', 1000);
set(ai, 'TriggerRepeat', 1);
set(ai, 'TriggerType', 'manual'); %trigger mode

In Signals.m, after setting up the Nidaq, the GUI is designed. A quit button was added to exit the program, and a capture button was added to be used to acquire data for each finger motion to train a neural network. A notch filter was designed using the signal processing toolbox:

function Hd = notch
Fs = 1000; % Sampling Frequency
N = 20; % Order
Fc1 = 55; % First Cutoff Frequency
Fc2 = 65; % Second Cutoff Frequency

% Construct an FDESIGN object and call its BUTTER method.
h = fdesign.bandstop('N,F3dB1,F3dB2', N, Fc1, Fc2, Fs);
Hd = design(h, 'butter');
end

This was done by designing a 20-order bandstop filter with stopband from 55 Hz to 65 Hz. Next a low pass butterworth filter with order 7 was designed. The cutoff frequency was set to 80 Hz. It could be as high as 300 Hz but we decided to limit the frequency to 80 Hz because there was some high frequency noise observed in the testing we decided would corrupt the classification. A high pass filter was designed with cutoff frequency of 10 Hz. The program keeps running until the quit button is hit. The program keeps reading the data from the Nidaq at a sampling frequency of 1000 Hz. The data is read using the following code:

start(ai); %start the analog input object
trigger(ai); %trigger the ai
[data,time]= getdata(ai, ai.SamplesPerTrigger);%read the nidaq
stop(ai);

The last 400 ms of the previous 1 second buffer is appended to the current buffer before filtering in order to get rid of transients and make the display more fluid. The data is first passed through a notch filter to get rid of the 60 Hz noise. Then it is passed through a low pass filter and high pass filter to restrict the signal to the EMG range that we want (10-80 Hz). Next the data is displayed on the GUI by setting the xdata and ydata of the current plot: “set(ch0plot,'xdata',time,'ydata', ch0data);”. The ‘drawnow’ command is used to ensure that the plot happens immediately so it looks real time. The program keeps running until the quit button is hit.

In ‘Processing.m’, the data is processed the same way as in ‘Signals.m’ except this time, we try to classify the signals and match it to a particular finger. The GUI in this case is a picture of a right hand with five pushbuttons which change color depending on which finger is perceived to have moved. The hand is displayed on the GUI by using the ‘imshow’ function on the hand.png picture which is in the same folder as the program. The window is then resized, and the pushbuttons are added. The classification is done using a probabilistic neural network (pnn) designed in Matlab. According to the Matlab help page, the pnn is a kind of radial basis network which has two layers. The first layer has biases which are set by the function ‘0.8326/spread’. We experimented with different values of the spread but none of them worked. The pnn requires a training set data which are the files that were recorded for each finger. There were five signals for each finger except the ring finger. We didn’t record signals for the ring finger because it was extremely difficult to move it without moving other fingers. Thus there were 20 training set vectors which form the 20 columns of the training matrix. A 4x20 target matrix is created. Each column is a particular column of the 4x4 identity matrix. Each column also corresponds to a particular column in the training set matrix. Each column has a 1 in the row which corresponds to the finger the training set vector corresponds to (1st row- thumb, 2nd row- index finger, 3rd row- middle finger, 4th row- little finger). The neural network is created using the code: “net = newpnn(P,T,spread);” where P is the training set matrix, T is the target matrix and ‘spread’ is the spread as described above. When each signal is sampled and processed, it is then passed through this neural network using the code: “y= sim(net, ch0data);”. The result is a vector with a 1 in the row the signal was classified as belonging to and a 0 elsewhere. The program checks to see if the signal in the buffer could be a real signal (its maximum should be greater than 8 mV and its minimum less than -2mV. If so, a flag is set depending on which finger was deduced to have been moved. The pushbutton above that finger changes color for two seconds. This program keeps running until the quit button is hit to exit.

SOFTWARE TESTING
For software testing, we plotted the signal in frequency domain with no hand motion to see which frequency bands the noise came from. As expected, there was a significant 60 Hz component. What was surprising is that, there was a lot of noise around 80 Hz, 200 Hz and 320 Hz. We weren’t sure where these noise signals came from but what was interesting was that each frequency from 80 Hz differed from the next by about 120 Hz. We designed notch filters to get rid of these frequencies. The 60 Hz noise was sufficiently attenuated. However, the other frequencies kept changing location (even though they differed by 120 Hz). After changing the bandwidth of the stop band for different notch filters, and not being able to get rid of these noise signals, we decided to limit the EMG signal to a band from 10-80 Hz. The high pass cutoff was set to 10 Hz to limit the interference that could arise from ECG signals. The next test that was performed was the GUI that displays the hand. A test was made to make sure that the push buttons change color when a finger move is detected and that the right push button changes color. The next tests that were performed have to do with the process of obtaining signals for training the neural network. A program was written to obtain five training set signals for each finger apart from the ring finger of the test subject. The program detected when the signal crossed a certain threshold after the ‘capture’ button on the GUI had been pressed. The final test that was performed was changing the bias of the second layer of the probabilistic neural network designed in MATLAB. This was done by changing the ‘spread’ argument in the call “net = newpnn(P,T,spread);” This function is explained more in the documentation section.

RESULTS
We were successful in getting noise free EMG for every finger. We could see the change in amplitude significantly with the application of force on the complete arm. We saw consistency in the output of EMG for thumb and the little finger (pinky).