Free-stall cubicles are designed so that cows do not defecate in the bedding material, while still providing comfortable lying places. However, fixed cubicle elements, such as the partition and neck rail, restrict available movement space and may hinder cows from performing natural lying down and standing up movement patterns. Although there are various types of cubicle partitions that differ in shape, dimensions, or materials, there is no method other than visual observations to assess their effects on cow welfare. An automated detection system could improve the efficiency and promote objectivity of such assessments. Therefore, the aim of this research was to explore which atypical lying down and standing up behaviors could be detected using body-mounted accelerometers and machine learning. Three leg- and one head-mounted accelerometer set to record at 20 Hz were fitted to 48 lactating dairy cows (Brown Swiss and Holstein × Swiss Fleckvieh). Lying down and standing up events were simultaneously assessed through video observations, by assigning binary presence/absence labels for atypical behaviors, such as lunging the head sideways when standing up and pawing the bedding material before lying down. Different time series classification algorithms were employed for model development using a nested cross-validation strategy. The best performing classifiers were MiniRocket and the deep learning algorithm InceptionTime. Atypical behaviors performed during standing up events, namely Hesitant head lunge and Crawling backwards, were identified as most promising candidates for accelerometer-based detection. These behaviors were detected with balanced accuracies of 0.67 and 0.74, respectively, and their learning curves indicated that more training data might further improve model performance. Overall, achieved performances were not yet satisfactory for application in the evaluation of new dairy cow housing installations. Potentially, ethograms designed for human observers are not optimal for machine learning and adjustments with machine learnability in mind might be necessary. The behaviors identified as promising are good candidates for further development into an efficient and objective method that could complement human observations in the assessment of dairy cow housing installations.
