Hello friends!
I present to your attention another interesting project of Vadim Zhukovsky-an electronic nose based on a neural network.
-------------------------------------------------------
Currently, in the design of gas analyzers, more attention is paid to two main areas: the search for new sorbent-sorbate systems and the development of an effective Converter of the response of the recognition element to the measured signal (transducer). This is reflected, for example, in the vast number of publications on the topic of biosensors.
In this paper, we ask a slightly different question: how does our nose actually work? It is unlikely that we have our own receptor for each line of Chanel perfumes. And, according to modern knowledge, where the brain is represented as a neural network, in order to recognize olfactory images, we do not necessarily need to have a narrowly selective system of receptors. To do this, it is enough to have a dozen types of receptors with different degrees of sensitivity to stimuli. For rice. 1 shows a neural network diagram where the signal, represented as a one-dimensional matrix of values for each type of receptor (1), is multiplied by a two-dimensional matrix of weight coefficients (2), which serves to compress data, and multiplied by a two-dimensional matrix (3), which classifies elements.
Rice. 1. Diagram of a neural network for pattern recognition of olfactory images:
1) vector of values of different receptors, 2) hidden layer, 3) classifier.
The result of matrix multiplication will be a one-dimensional matrix, whose values will characterize the probability of detecting an element that was previously trained by the neural network.
In this paper, we propose the concept of an optical transducer (Fig. 2), which can be used to implement this scheme.
Figure 2. Scheme of the optoelectronic "nose": 1) laser, 2) two plane-parallel transparent plates (waveguides), 3) array of sensors, 4) CCD matrix.
The optoelectronic "nose" is a Mach-Zener interferometer with an array of sensors (3) applied to the shoulder of one of the two waveguides (2). The state of sensor activity in this design will be characterized by an interference pattern. In this way, we get a graphical representation of olfactory images that can be trained by the neural network.
---------------------------------------------------
Until next time, @r3benok
2641.724 SRY