APPLICATION OF ARTIFICIAL NEURAL NETWORKS IN DIGITAL SIGNAL PROCESSORS

Summary:

In this work we propose a method for the characterization of the non - linear response of an infrared sensor used in the sensor - object distance measurement, using an artificial neural network model with supervised training. The neuronal model is developed and trained using the tool of neural networks of MATLAB ??, and then implanted via C language, in a digital signal processing applications, for its later application in embedded systems of signal acquisition. Three training algorithms are compared to verify the feasibility of a future implementation of online training. The Levenberg - Marquardt backpropagation algorithm has allowed obtaining the best results in the modeling of the sensor characteristic curve, when obtaining through its application, an error in the learning of the data of training of 6x10-5, in the smaller number of times of registered training compared with the methods Resilient backpropagation and quasi - Newton backpropagation. The results of the model and the implementation confirm a satisfactory performance of the applied method, which can be extended to the characterization of other types of sensors.

 

INTRODUCTION

A sensor is a device capable of transforming physical or chemical quantities called instrumentation variables into electrical quantities. The instrumentation variables depend on the type of sensor and can be for example, temperature, light intensity, distance, speed, acceleration, inclination, displacement, pressure, force, torsion, humidity, pH, among others. An electrical magnitude obtained can be an electrical voltage, an electric current [1]. In many cases, these electrical magnitudes and their corresponding instrumentation variables are related by a non-linear function, which is why methods, algorithms or circuits are required to deal with this problem,

 

This work proposes a method to characterize the non-linear response of the infrared sensor, by means of the implantation of an artificial neural network model with supervised off-line training on a DSP.

 

Artificial neural networks are a processing paradigm inspired by the way in which the nervous system of humans works, and characterized by its capacity for generalization, robustness and fault tolerance. Artificial neural networks are interconnected networks, which operate in parallel using simple (usually adaptive) processors and with hierarchical organization that try to interact with objects in the real world, in the same way as the central nervous system does [2]. Neural networks respond to a biological model of interconnection between processing elements called artificial neurons. Learning in living beings, particularly animals, is done by trial and error, through examples or demonstrations. Biological neural networks dynamically adjust internal parameters (weights and trends) that govern the representation of information or knowledge, adopting the ability to generalize responses to events never before raised. In an analogous way, these processes are represented, by means of algorithms or electronic circuits.

 

The neuronal model is implanted in a DSP Microchip® dsPIC 30F3013, taking advantage of the versatility and speed of processing of the device. The training of the network is done offline, so the computational cost is drastically reduced, allowing for perfectly valid implementations in architectures with lower computing resources. The network model has been validated in MATLAB, in whose environment the experimental results have been obtained that confirm the optimal performance of the application.

 

II. DEVELOPMENT

 

1.- Infrared sensor An infrared sensor is an electronic device capable of measuring the infrared electromagnetic radiation of the bodies present in its field of vision. All bodies reflect a certain amount of radiation that is invisible to the eyes, but not for these devices. This is because infrared radiation is in a range of the electromagnetic spectrum just below visible light [3]. This type of sensors can be classified as passive, in the case that they are provided only with infrared receivers, or active, in the case of being composed of transmitter-receiver pairs. This research proposes the design of a method for characterizing the non-linear response of an active sensor, model SHARP® GP2Y0A021, because the device is based on the combination of an infrared LED and a concentrator lens that forms a single beam, as emitter, and a phototransistor coupled to another lens, as a receiver. The infrared rays emitted by the infrared LED are reflected by the objects that are in line seen in front of the emitter, with a sharp reflection angle that decreases as these objects move away from the sensor and are finally captured by the receiver. The phototransistor has been constructed with a pyroelectric, artificial material, usually forming a thin sheet inside gallium nitrate (GaN), or cesium nitrate (CsNO3), derivatives of phenylpyrazine and phthalocyanine Cobalt, elements with which they are usually built These devices. Generally, they are integrated in various configurations (1,2, 4 pixels of pyroelectric material). In the case of pairs, it is customary to give opposite polarities to work with a differential amplifier, causing the self-cancellation of the infrared beam energy increases and the decoupling of the equipment. In theFigure 1 shows the infrared sensor model and the trajectory described by the infrared beam emitted in the presence of two obstacles at different distances [4]:

 

 

2.- Architecture of the neural network

 

The technology of neural networks has been widely used in problems of generalization or approximation of functions of arbitrary complexity, so the proposed design uses this feature to

 

represent the non-linear response of the infrared sensor, based on a set of distance-voltage values, for the training of a multilayer neural network, by means of a backpropagation algorithm with three different variants. The typical neuron of the proposed neural network model obeys the following relationship, which expresses the output of the i-th processing element.