This paper presents a study of feed-forward neural network (NN) systems developed to determine the head kinematics of subjects who are exposed to impact accelerations. The neural networks process accelerometer data collected during short-duration impact acceleration tests conducted at the National Biodynamcis Laboratory of the University of New Orleans. During an impact acceleration experiment, the subject sits on the sled chair and a piston gives impetus to the sled to travel down a track. Head data is gathered by an array of nine accelerometers. Two more accelerometers are mounted on the sled.
The neural processing systems produce the history of the rotational and translational position, velocity, and acceleration of the origin of the accelerometer array mounted on the mouth. Output produced by a least squares algorithm that uses both photographic and accelerometer raw data are used as a baseline and to provide training data for the neural networks. The main disadvantages of the NNs are their speed, and that statistical information and accurate modeling of the testing system are not required. Results show that the neural networks provide accurate information about the kinematics of the subject even when no photographic data are used.
Kaminsky, Edit J. and Anwer S. Bashi, "Determination of Head Kinematics from Impact Acceleration Test Data using Neural Networks," 1998 lASTED Internal. Conf. on Signal Processing and Communications (ICSPC'98), (Canary Islands, Spain), Feb. 11-14, 1998, pp. 224-227.