A Method of Tactile Resistive Sensor Array Calibration

Abstract

Resistive sensor arrays (RSA) comprising of pressure-sensitive elements find broad use in pressure-sensing applications. However, sensor manufacturing inevitably yields a number of non-ideal properties. The gain, offset or non-linearities of an RSA’s elements may vary between individual sensors (taxels). Hence a simple calibration procedure to suppress these effects is desirable. The paper presents a method of and apparatus for the calibration of a pressure-sensing RSA. A forward mathematical model for each sensor is obtained during the calibration phase. In subsequent online measurement, the inverse model is utilised to compensated for the non-uniformity in sensor gain and offset. This approach leads to an appreciable improvement in picture quality, which can be mathematically quantified using classification accuracy in our tactile anti-decubitus platform for human monitoring. Namely, in a four-class classification experiment involving a support vector machine, the classification error decreased from 11 % before calibration to 3.5 % after calibration. Owing to the calibration procedure, a classifier trained using a calibrated RSA can be deployed to another calibrated RSA, without further data collection.

Resistive sensor arrays (RSA) comprising of pressure-sensitive elements find broad use in pressure-sensing applications. However, sensor manufacturing inevitably yields a number of non-ideal properties. The gain, offset or non-linearities of an RSA’s elements may vary between individual sensors (taxels). Hence a simple calibration procedure to suppress these effects is desirable. The paper presents a method of and apparatus for the calibration of a pressure-sensing RSA. A forward mathematical model for each sensor is obtained during the calibration phase. In subsequent online measurement, the inverse model is utilised to compensated for the non-uniformity in sensor gain and offset. This approach leads to an appreciable improvement in picture quality, which can be mathematically quantified using classification accuracy in our tactile anti-decubitus platform for human monitoring. Namely, in a four-class classification experiment involving a support vector machine, the classification error decreased from 11 % before calibration to 3.5 % after calibration. Owing to the calibration procedure, a classifier trained using a calibrated RSA can be deployed to another calibrated RSA, without further data collection.