A Caltech Library Service

Neuromorphic VLSI circuits for an electronic nose chip


Tang, Kea-Tiong (2001) Neuromorphic VLSI circuits for an electronic nose chip. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/sdb6-3c88.


Human olfaction is still the primary instrument used in many industries to classify the smell or flavor of products. This is a costly process since trained experts are required who can only work for relatively short periods of time. Therefore using machine olfaction to perform the task would be a significant advance. Many researchers have investigated electronic noses, but currently only relatively large "instrument" electronic noses have been built. We have designed an electronic nose on a single silicon chip, including adaption, signal processing, and classification. The electronic nose chip is composed of four different stages: Sensor stage, Signal Processing stage, Database stage, and Classifier stage. The sensor stage, as its name suggests, deals with the sensor directly. The sensor we use is a carbon black-organic polymer whose relative resistance change is proportional to the given odor concentration. The function of the sensor stage is to adapt this resistive sensor to a preset baseline value, take the AC signal voltage, then output a current proportional to the signal voltage. So the sensor stage outputs a signal current which contains information about the odor concentration. An adaptive electronics stage, a peak detector, and a transconductance amplifier are designed to complete the sensor stage. Adaptive electronics are implemented to adapt the sensor to be within a proper working range of the circuit while tuning out the environment background. Adaption is done by constructing an adjustable current source. After adaption is done, the bias current value is held, so the sensor voltage at this time contains two different types of information: baseline value and signal value. The peak detector traces the input signal to its maximum value, then holds the value for further signal processing. This is needed because the signal voltage is defined as the difference between the maximum sensor voltage and the baseline sensor voltage. The transconductance amplifier converts voltage to current linearly while functioning as high pass filter. The output current is equal to the difference between the two input voltages multiplied by some gain (called transconductance). The output of the peak detector, the maximum sensor voltage, is used as the noninverting input, while the baseline sensor voltage is used as the inverting input. By the differential input characteristic of the transconductance amplifier, the baseline information is cancelled, and only the signal information remains. Thus, the output current from the transconductance amplifier contains the signal information, i.e., odor concentration. The signal processing stage performs two important tasks for further signal processing. First, normalization throughout the signal array is realized. Then the Euclidean distance between the signal vector and the data vector is calculated. A normalizer using city-blocks distance is designed and an Euclidean distance calculation circuit is built. A normalization circuit using city-blocks distance is implemented to generate a normalized signal vector. This normalized signal vector is stored in a SRAM through an A/D in the LEARNING State. On the other hand, in the CLASSIFYING state, Euclidean distance between the normalized signal vector and the data vector is calculated. Euclidean distance circuit is implemented to calculate the Euclidean distance between signal vector and data vector. The Euclidean distance is output in the form of a current. This distance measure is utilized for classification. The database stage stores the signal vector in the data storage device (LEARNING state) or outputs the data vector from the data storage device (CLASSIFYING state). This stage also takes care of the interface between the electronic nose chip and the outside world. A central control unit is designed to generate all the control signals and arrange the time sequence. Eight-bit Static Random-Access Memory (SRAM) is used for data storage. A/D converter is used to convert the signal vector into a digital word. This happens during the LEARNING process. A D/A converter is used to convert the data from SRAM into the data vector. This happens during the CLASSIFYING process. A current copier cell is designed to maintain the value of the data current. Several current copier cells are used to form a data vector. The central control unit is designed to generate all the control signals needed and their time sequence. The classifier stage receives all the Euclidean distances between signal vector and data vectors, and generates the output corresponding to the shortest Euclidean distance, while inhibiting all the other outputs. The generated output is denoted as the answer to the pattern recognition problem. The current copier cell in the database stage is used to maintain the value of the Euclidean distance current. Several current copier cells are used to generate inputs for the LTA circuit. The Loser-Take-All (LTA) is used for parallel classification. Global inhibition can be done by using an LTA circuit.

Item Type:Thesis (Dissertation (Ph.D.))
Subject Keywords:Electrical Engineering
Degree Grantor:California Institute of Technology
Division:Engineering and Applied Science
Major Option:Electrical Engineering
Thesis Availability:Public (worldwide access)
Research Advisor(s):
  • Goodman, Rodney M.
Thesis Committee:
  • Unknown, Unknown
Defense Date:11 May 2001
Record Number:CaltechTHESIS:10122010-085150166
Persistent URL:
Default Usage Policy:No commercial reproduction, distribution, display or performance rights in this work are provided.
ID Code:6130
Deposited By: Benjamin Perez
Deposited On:12 Oct 2010 16:12
Last Modified:16 Apr 2021 23:18

Thesis Files

PDF - Final Version
See Usage Policy.


Repository Staff Only: item control page