Submit Manuscript | http://medcraveonline.com Introduction Ultrasound systems used for object recognition and for the navigation of blind people through the environment are mostly composed of walking sticks that vibrate when approaching an object. In addition to vibrating systems, sound emission systems are also used to aid navigation for blind people, 1 these systems are composed of cameras combined with ultrasonic sensors, only cameras or only ultrasonic sensors. 2 A literature review by Almeida shows that from 2016 onwards there was a drop in the number of studies using ultrasound as the main means of object detection. The author explains that this drop is due to the fact that technologies using cameras for the recognition of objects and people are drawing the attention of the scientifc sphere and more researches in this area are being developed. 2 Ifukube 3 performed tests for detecting an object by the generation of sounds and the use of ultrasonic sensors. He showed the importance of hearing low and high frequency sounds in order to improve the detection of the object’s position. Simultaneously, studies such as Nabhani’s 4 & Manju’s 7 & Velappa’s 6 used neural networks for object recognition, having Nabhani used cameras and Manju and Velappa used ultrasonic sensors. These authors showed that the use of neural networks is a viable method for identifying objects. 4,5 Neural networks are architectures used for machine learning, their main block is the neuron, a small functional unit. The neuron has an input and an output that is calculated through an activation function and a bias. The combination of several neurons forms the neural network with its activation functions, bias and weights. Figure 1 shows a neuron in the input layer with its input variables (X1, X2, ..., Xn), weights (P1, P2, ..., Pn), correction value (Bias B1), output intermediate value (V1), activation function (F) and output value (Y1). The V1 value is calculated using Eq.1 and then the chosen activation function is applied, resulting in the Y1 output. 7 V 1 = Σ(X n P n ) + B 1 n=1, 2, 3, ... (Eq. 1) The input variables values in a neural network must be normalized for a better processing performance, 8 Eq. 2 is used to normalize the input data in a neural network, suppose that it is desired to normalize values between 2cm and 400cm (the minimum and the maximum that an ultrasound sensor model HC-SR04 can measure, for example) and transform them into 0.0 and 1.0: D N = (D – 2,0) / (400,0 – 2,0) (Eq. 2) Where D is the real distance and DN the normalized distance. Figure 1 Neuron model in the input layer of a neural network. The F activating function of a neural network depends on the problem to be solved and also on the layer in which the neuron is at. It is possible to create new activation functions, but there are already widely used functions for several artifcial intelligence problems that are effective, among them are the sigmoid function, ReLU (Rectifed Linear Unit), Leaky ReLU and Softmax. 7,8 The sigmoid function and its combinations are broadly used in classifcation networks, whereas the ReLU function is applied to the hidden layers of the network (neurons positioned between the input and the output layer). The Leaky ReLU function is generally used in cases where the ReLU function has not given good results. 8 Similar to the sigmoid function, Softmax is useful in classifcation problems of several classes, resulting in the probability that an input from the network belongs to one of these classes. 8 The mathematical model in Figure 1 is known as Perceptron, this model was developed in the 1950s and 1960s by the scientist Frank Rosenblatt and allows a clear understanding of how a neural network works in mathematical terms. 8,9 Rosenblatt proposed a simple rule for calculating the output, multiplying the input values by their respective weights, adding these results and applying a rule for the output value: 0 if the sum is less than or equal to a threshold value or 1 if the sum is above the threshold value. By varying the weights and the threshold, different decision- making can be reached. 9 The bias value defnes the neuron sensitivity Int J Biosen Bioelectron. 2020;6(3):70‒73. 70 ©2020 Almeida et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and build upon your work non-commercially. Acquisition and recognition of ultrasonic signatures using multi-layer neural network Volume 6 Issue 3 - 2020 Maicol Peterson Gandolphi de Almeida, Pedro Bertemes-Filho State University of Santa Catarina, Brazil Correspondence: Pedro Bertemes-Filho, Universidade do Estado de Santa Catarina, Rua Paulo Malschitzki 200, Zona Industrial Norte, 89219-710, Joinville, Santa Catarina, Brazil, Tel +55 47 34817848, Email Received: August 05, 2020 | Published: August 31, 2020 Abstract The echo captured by an ultrasonic sensor provides details of the geometric shape of objects, this ability common among dolphins and bats is known as echolocation. The signal detected by an ultrasound sensor contains, in addition to the distance from the object, characteristics such as the type of material and internal cracks. This article describes a method for acquiring ultrasonic signals and recognizing them as geometric shapes through the use of a neural network processed by the TensorFlow JS library. The results using basic geometries (rectangular, triangular and spherical parts) showed that TensorFlow JS recognizes their ultrasonic signatures, being an additional solution to those that use cameras for object recognition. Keywords: echolocation, neural network, tensor fow, ultrasonic signature International Journal of Biosensors & Bioelectronics Research Article Open Access