Signal processing for sensor arrays

Five methods have been developed to process output signals from sensor arrays to improve sensor characteristics. A multielement capacitive force sensor array with nonlinear characteristics and inter-element cross-talk was used to demonstrate and compare the performance of these methods: (1) Variable Threshold, which calculated a threshold value from a histogram of the data to determine whether the output of each element was due to an input or cross-talk; (2) Algebraic Cross-Talk Reduction, which reduced the cross-talk between array elements by solving for the input matrix from the input-output relation of the sensor; (3) 2 x 2 Edge and Shape Detection, which was adapted from image processing techniques and enhanced the resolution of the sensor output; (4) Deconvolution using FFT; and (5) Artificial Neural Network, which was trained to learn the inverse relation of the sensor transfer function. The ability of these methods to reduce the cross-talk seen between sensor elements and more closely estimate the sensor input was studied. Preprocessing of the sensor output, which consisted of calibration of each element using non-linear curve fitting, was used to compensate for the non-linear, non-uniform sensitivity characteristics of the sensor array elements. Postprocessing compensated for the dependency on the number of excited elements. A comparative study of eight force sensor arrays using a 200 element test set showed that the simpler signal processing methods, Variable Threshold and 2 x 2 Edge and Shape Detection, performed well only for a limited number of input conditions especially when forces were applied through small objects. In general, neither reduced the error in determining total force by more than 30 percent. The Artificial Neural Network method, the most complicated and computationally intensive method during initial training but simple to implement thereafter, produced superior results for the majority of the input patterns applied to the sensor array, reducing the overall error by 6 to 83 percent. All five signal processing techniques reduced the overall errors of the sensor outputs, some more consistently than others. However, when the overall error magnitude was less than 10 percent, only the Artificial Neural Network reduced the error consistently, making it the method of choice. The results of this study showed that sensors having outputs with high error values can be converted into sensors with much lower output error using appropriate signal processing methods. Incorporation of such signal processing methods into sensor arrays allows the use of sensors with less than ideal characteristics.