Sensors aad Actuators A49 (1995) 173-180
A tactile sensor data-processing system
Hans Odeberg
Depariment of Physics and Mearuremenr Technology, Linkdping Universify, S-Sgl83 Lb&ping, Sweden
Received 20 December 1994; in revised form 10 March 1995; accepted 23 May 1995
Abstract
In this paper, a framework for distributed sensor data processing is presented, where sensors with local intelligence preprocess information.
A supervisor gathers the information from the sensor nodes and fuses the local sensor estimates into a global estimate, using fuzzy logic-
based algorithms. Since a typical system will contain a large number of sensor nodes, the local processing will probably be performed by
simple 8- or l&bit processors. Care is thus taken to make sure that the algorithms perform well on low-end hardware. The concept is then
implemented on a sensor system where local sensor processors extract information from tactile matrices, and send the data to a host over a
shared serial bus. Using the distributed processing approach yields increased data-acquisition rates and offloads calculations from the system
controller.
Keywords: Data processing; Tactile sensors
1. Introduction
Consider a measurement system such as the one depicted
in Fig. 1. A number of sensors connected to a host computer
over a shared bus. If &sensors gather large amounts of data
and transmit them over the bus, the bandwidth requirements
of this bus will be very large. Also, the sensors may be of
different types, sending very different types of data. The host
will then be burdened with interpreting all this data, regard-
less how much of it is really needed.
One way to reduce the bus bandwidth requirements is to
let the local processors preprocess data, sending relevant
high-level information to the host only when required. To
avoid making the sensor systems excessively expensive, the
Sensors --On
SI
sz
@
J
Supervisor ” .ff P
Pig. 1. A typical measurement system.
0924-4247/951$09.50 Q 1995 Etsevier Science S.A. All rights reserved
SSDIO924-4247(95)01030-5
processing power of each sensor is necessarily limited. Thus
there is a need for fast algorithms.
To get a consistent representation of the information for
all types of sensors, some issues still need to be resolved:
l How should the information from sensor to supervisor be
represented? Is there a form common to all sensors?
l How is this information extracted from the sensor signals?
l How does the supervisor combine the information from
different sensors?
The next four sections will define a theoretical framework
that solves these problems.
2. How to represent sensor information
Previous attempts at defining a consistent representation
of sensor information have centred on probabilistic methods,
using Gaussian statistics [ 1,2]. This approach has its merits:
there is a well-defined calculus for probabilities, with espe-
cially simple formulae for Gaussian distributions. While solv-
ing the consistency problem, this approach does leave other
problems unsolved. Data are not always Gaussian. While a
Gaussian distribution is often a reasonable approximation for
many continuous distributions, modelling a binary on-off
sensor with a Gaussian would be somewhat difficult.
Instead, we propose the use of a possibility measure, or
opinion [ 31, /.L The supervisor sends a hypothesis to the
sensor, which then replies with an opinion indicating to what
degree it thinks the hypothesis possible.
174 H. Odeberg /Sensors and Actuators A 49 (1995) 173-180
Fig. 2. Creating a sensor opinion.
The opinion is modelled as a number in the range [ 0, l]
with /J= 1 when data are in complete agreement with the
hypothesis, p = 0 when they completely contradict it, and
p = 1 when the sensor data have no relevance to the hypoth-
esis, indicating a zero information content. The advantage
this brings is that there is no longer a need to normalize with
respect to the entire universe of discourse. However, new
formulae for the evaluation and combination of opinions are
required.
3. Creating sensor opinions
Since the opinion is defined to be in the range [0, I], a
natural approach is to use fuzzy logic. A simple triangular
shape could be used to map a sensor input to an opinion (Fig.
2). For more complex input-to-opinion mappings, a fuzzy
IF-THEN rule base may be used (Fig. 3). The different rules
are then combined into a single opinion using one of the
available defuzzification techniques.
4. Adapting to change
When the characteristics of measured parameters change,
due to drift, wear or noise, it is desirable for the system to
note these changes and adapt to them. If the location and
width of the triangular mapping between sensor input and
opinion described above can be characterized by the mean f
and standard deviation a of the measured parameter, one
solution is to modify them as new data become available,
using the standard recursive least-squares formula [4]. In
Ref. [ 51, this method is adapted for use on simple fixed-point
processors.
5. Combining sensor opinions
Assuming that the sensors have calculated their opinions
concerning a hypothesis, the remaining problem is to find a
function that combines these into a joint estimate. When
combining the opinions of sensors, what algorithms should
be used? The standard fuzzy AND/OR connectives are cer-
tainly not suitable, since they are implemented as min and
max functions. Any algorithm that singles out the most
extreme value in a data set is not likely to be very efficient
with respect to noise suppression and outlier removal. Let us
first look at the combination of two sensors, later extending
the results to any number of sensor opinions. The require-
ments on the function used are:
( 1) If two sensors disagree, there is no knowing which one
to trust. When one sensor supports a hypothesis the other
sensor rejects, their combined opinion should be a statement
of ignorance. Thus:
f(ffx, ;-x)=f (1)
(2) When combining a sensor that holds no definite opin-
ion (p = 1) with another sensor, the result should be domi-
nated by the latter:
f(t7 CL,) =I& (2)
(3) When two sensors have reasonably similar opinions,
what should the result be? One approach is to let the opinions
reinforce each other, resulting in a synergy effect. Thus if
both sensors are reasonably sure of an opinion, the net result
should be a certainty equal to or exceeding that of both opin-
ions:
f(O,O)=O; f(f, i,=f; f(1, l)=l
f(P** CL,)
Px 1