The objective was to develop a tool to measure sow postural activity during lactation while kept between fences in the farrowing crate. Several trials were carried out to select sensors and test their ability to detect 5 different positions of the sow. First, several sensors, placed in a single metal box, were attached to the upper part of the back of sows with a custom-built belt that passes underneath the belly, just behind the front legs. For validation of the positions predicted from sensor data, sow behaviour was recorded with use of a digital camera and video records were analysed by a single observer. Prediction ability was tested with machine learning applied to random forests for the three-axis data provided by the sensors in the X, Y and Z dimensions. When comparing information provided by three sensors, i.e. two accelerometers (sensors A and B) and an application developed for a smartphone (sensor C), one sensor detected five positions more accurately than the other two. The error rate of prediction from 30% of the data was 2.1% with sensor A, 3.1% with sensor B, and 4.8% with sensor C. Additional trials focused on the ability of sensor A to distinguish among the five positions, using five sows. Statistical sensitivity was 0.94 and 0.95 for the right and left lateral positions, respectively; 0.79 for the ventral position; 0.75 for the standing position; but only 0.33 for the sitting position (because it represented only 4% of sows’ time budget). One single axis was sufficient to detect the two lateral positions. Next, we concentrated our interest on the advantages and drawbacks of 2 accelerometers: we ran new trials to compare outputs from sensor A with those obtained with a sensor D. We used acquisition speeds varying from 1 to 16 obs/sec to optimize the sampling rate to detect positions accurately and to test if postural changes at risk for the piglets can be detected. Analyses are in progress. The use of accelerometers to measure finely the postural activity of lactating sows is promising.