Information fusion

from Wikipedia, the free encyclopedia

The information fusion comprises methods to link data from different sensors or information sources with the aim of gaining new and more precise knowledge about measured values and events.

Related disciplines are sensor fusion , data fusion and distributed measurement .

history

The theoretical origins of information fusion go back to the end of the sixties . However, these mathematical principles were not transferred to technology until later - initially in the field of artificial intelligence (AI). In this discipline, biology , especially the human brain , was often used as a model for modeling technical systems. If you take into account the performance of the brain when merging data from the different sensory organs, it is not surprising that the first approaches come from AI.

The use of information fusion is now very broad and encompasses many different disciplines - including robotics , pattern recognition , medicine , non-destructive testing , geosciences , defense and finance . Although the literature on this is very extensive, many of the methods given therein are not very systematic.

Methods

In the last few years a number of systematic merger approaches have emerged, the most important of which are briefly discussed here. For this, the fusion problem is first formulated as a parameter estimation model. A parameter is emitted from a source which represents a realization of the random variable . The target variable can be a measured variable, but it can also be latent constructs that do not have to claim physical reality. In the latter case, the size can be understood in the Platonic sense as an idealization of the sensor data , in which the desired or known properties of the target size itself are taken into account. With the help of several sensors, the data are recorded, which are also to be understood as realizations of a random process . The measurement corresponds to a figure that can be described mathematically using the conditional probability distribution (WV) of . Below it is assumed that it is in and the WV is described in terms of a probability density function is to continuously sizes.

Classic statistics

Classical statistics are based on an empirical, frequentistic interpretation of probabilities in which the sensor data are regarded as random variables, but not the measured variable itself. The estimation of using the sensor data is based on the so-called probability density function , which is understood and maximized as a function of becomes:

The associated value is called the maximum likelihood or ML estimate.

Bayesian statistics

In Bayesian statistics , the measured variable is also understood as the realization of a random variable , which is why the a priori probability density function is used to determine the a posteriori probability density function :

Maximizing this equation gives the maximum a posteriori (MAP) solution for the parameter to be estimated :

This procedure has the essential advantage that it allows the specification of the WV for the parameter to be estimated for given measurement data , whereas the classical procedure only allows specification of the WV for the sensor data with a given parameter value .

Dempster Shafer Evidence Theory

The evidence theory is often as an extension of probability theory considered or as a generalization of Bayesian statistics. It is based on two non-additive measures - the degree of belief and plausibility - and offers the possibility of expressing uncertainty in more detail. In practical tasks, however, it is not always possible to present the available knowledge about the relevant variables in such a differentiated manner and thus to fully exploit the theoretical possibilities of this approach.

Fuzzy logic

The fuzzy logic is based on the generalization of the concept of set with the aim of achieving a fuzzy knowledge representation . This is done using a so-called membership function , which assigns each element a degree of membership to a set. Due to the arbitrariness in the choice of this function, the fuzzy set theory is a very subjective method, which is therefore particularly suitable for the representation of human knowledge. In information fusion, fuzzy methods are used to handle uncertainty and vagueness related to the sensor data.

Neural Networks

Artificial neural networks (ANN) are another method of merging information . These can be based on processing units simulated by software, which are interconnected to form a network, or implemented in hardware in order to solve specific tasks. Their use is particularly advantageous when it is difficult or impossible to specify an algorithm for combining the sensor data. In such cases, the neural network is taught the desired behavior in a training phase with the aid of test data. The disadvantage of neural networks is the lack of options for integrating a priori knowledge about the variables involved in the fusion.

Web links