Non-linear Blind Source Separation Non-linear Blind Source Separation Applied to Ion-sensitive Field Applied to Ion-sensitive Field

Effect Transistor Sensor ArraysEffect Transistor Sensor Arrays

Guillermo Bedoya

Advanced Hardware Architectures (UPC)

Laboratoire des Images et des Signaux (INPG)

UNIVERSITAT POLITÉCNICA DE CATALUNYA

UPC INPGrenoble

OutlineIntroduction

- Overview

- What is Blind Signal Separation (briefly)

Identifiability & Algorithmic solutions

-When can it be done

-How do we do it (briefly)

Applications to semiconductor-based sensor arrays

-The ISFET/CHEMFET devices

-How does BSS apply to semiconductor-based chemical sensing

OverviewPotential advantages of integrated circuit technology applied to the field of physiological data acquisition and water monitoring systems :

• small size

• reliability

• low cost & rapid time response

• multi sensor chip

• on chip signal processing

Our objective is to design a low cost/high performance smart sensor system for Biomedical and environmental monitoring applications, using ISFET/CHEMFET sensor arrays.

IntroductionBlind Source Separation deals with the « separation » of a mixture of sources, with a little prior information about the mixing process and the sources signals

..

....

SensorsSensors

S1

S2

SN

Sources

Environment

Source Separation Algorithm

W...

..

.x2

xN

x1 Ŝ 1

Ŝ2

ŜN

Observations

x = As Ŝ = Wx

What is Blind Source Separation

x1[n]=a11s1[n]+…+a1psp[n]..

xm[n]=am1s1[n]+…+ampsp[n]

x[n]=(x1[n]…xm[n])=As[n]

y[n]=(y1[n]…ym[n])=Wx[n]=s[n]

s x yA W

Identifiability & Algorithmic solutionsPrinciples of information theory can be applied to the BSS problem.

We suppose that source signals are independents (realistic assumption).

Then, we minimize a measure of the independence (e.g., the mutual

information (MI) I(.)) of the outputs I(y), where y=Wx.

s x= As y= Wx

How do we do it?

We consider a processing function g, which operates on a scalar X using a function Y=g(X;w) in order to maximize the MI between X and Y. Parameter w is chosen to maximize I(X;Y).

I(X;Y)=H(Y) – H(Y|X)

The MI is maximized when H(Y) is maximized (for g deterministic).

H(y)=sum[H(yi)] - I(y1,…, yN)

In order to maximize H(y) (we can maximize each H(yi) or minimize I(y1,…, yN)). g must be the CDF of x.

The mutual information is minimized when all the outputs are independent !

W g(x,W)x y

How do we do it?

An adaptive scheme is to take: Δw α

Δw = ;∂ (ln |∂y/∂x|) ∂w

1w

then,Δw α + x(1-2y)

Assuming a particular functional form:

y = g(x) = 1/(1+e-(wx))

∂y/∂x = wy (1-y)

∂H(y)

∂w

And we have a weight update rule :

w[k+1]=w[k]+ stepsize Δw

Applications to semiconductor-based sensor arrays

Ion-Sensitive Field Effect Transistors (ISFETS and CHEMFETs) are basically metal oxide semiconductor field-effect devices. The construction of an ISFET differs from the conventional MOSFET devices, in that the gate metal is omitted and replaced by a

membrane sensitive to the ions of interest.

Potentiometric sensors!

ISFET/CHEMFET sensors

S D

(1) Reference Electrode (Vref)

(2) Solution (Electrolyte)

(3) Membrane (MOSFET gate

metal)

(4) Gate Insulator

VG VD

ID

Silicon Substrate

ISFET/CHEMFET sensors (Short view)

ID for the conventional MOSFET is:

ID = α[(VG –VT)-0.5VD]VD (1)

where α = μCoW VD/L

We have to establish new expressions for VG and VT to adapt equ. (1) to the ISFET.

MOSFET VG ground-to-gate metal potential

ISFET VG* ground-to-membrane potential

VG* = VG + (Electrode-electrolyte potential)+ Nernst Potential

ISFET/CHEMFET sensors (Short view)

VT = Φms + 2 ΦF - Qss+QB

Co

VT*= Φcs + 2 ΦF - +Eref – E0 Qss+QB

Co

VG* = Vref + RT ln (ai + Kij aj )

nF

zi/zj

ID* = α[(VG* –VT*)-0.5VD]VD (2)

Main characteristics

Sensitivity:

• Lowest level of chemical concentration that can be detected.

• The smallest increment of concentration that can be detected in

the sensing environment.

Selectivity:

• The sensor’s ability to detect what is of interest and to separate that from interferents

Comparison of Chemical sensors

The use of different type of sensors is recomended to obtain more robust semiconductor-based chemical sensor arrays

Current problems

Miniaturized chemical sensors have yet to achieve their full potential.

They must accomodate:

• High noise levels in chemical composition of the field environment.

• Highly variable environmental conditions (temperature, humidity)

Solutions:

1. Classical methods :

• calibration (LUTs, polygonal interpolation, progressive calibration)

• compensation (Structural, adjust, etc)

2. ISFET SOURCE SEPARATION techniques

Advantages:

• Cancel crosstalk (added noise caused by interferent ions

• Cancel cross-sensitivity (when the response varies with the temperature – drift).

• Low cost/high performance implementation

Sensor fabricationSensor fabrication Application of signal processing techniquesApplication of signal processing techniques

• New materials.

• New devices.

.

.

High cost !

How does BSS apply to semiconductor-based chemical sensing?

ID1* = α[(Vref + RT ln (ai + K1 aj ) - VT*)- 0.5VD]

nF

zi/zj

We have the observations matrix:

ID2* = α[(Vref + RT ln (ai + K2 aj ) - VT*)- 0.5VD]

nF

zi/zj

IDN* = α[(Vref + RT ln (ai + KN aj ) - VT*)- 0.5VD]

nF

zi/zj

Non-linear BSS

Spatial diversity

Adaptive algorithm for Non-Linear BSS

Linear mixing stage

Non –linear distortion

Non –linear compensation

Linear de-mixing stage

A W

f1

f2

g1

g2

ai

aj

âi

âj

ID1

ID2

gID1

gID2

ID = Φa + Φb*ln(ai + Kijaj ) ;zi/zj

gID = e(ID –Φa)/ Φb

Adaptive algorithm for Non-Linear BSS

1. Initialization W = I, gID = ID, Φa and Φb.

2. Loop

1. Compute outputs by: y =Wx;

2. Estimation of parameters

Φa[k+1]= Φa[k]+ stepsize ΔΦa

Φb[k+1]= Φb[k]+ stepsize ΔΦb

3. Normalization

4. Linear BSS algorithm: w[k+1]=w[k]+ stepsize Δw

5. Repeat until convergence

Preliminary Results• The algorithm recovers the wave forms of the main ion activity (ai) and the interferent ion activity (aj). aj is considered as noise in other approaches. Selectivity is improved.

• The algorithm works well for sensor arrays with poor selectivity coefficients. Sensor arrays with poor performance bring to the algorithm spatial diversity and more statistical information. We can use low cost sensors.

• The adaptive scheme allows the separation when environment characteristics varies (e.g., temperature). We hope to study the algorithm behaviour as a function of the device drift.

• The scheme can be adjusted to build a multi-parametric system, adding more sensors (sensitive to other ions) and adjusting the algorithm parameters.

• Hardware implementation using a DSP card is currently being developed. Previous implementations had shown good performance.

Non-linear Blind Source Separation Non-linear Blind Source Separation Applied to Ion-sensitive Field Applied to Ion-sensitive Field

Effect Transistor Sensor ArraysEffect Transistor Sensor Arrays

UNIVERSITAT POLITÉCNICA DE CATALUNYA

UPC INPGrenoble