A novel approach for anomaly detection in dairy cow gas emission records

Mitigating the considerable contribution of ruminant livestock to global methane emissions, which accounts for 17%, is of vital importance. One significant approach to this is the genetic selection of cows for reduced methane emissions, a strategy that demands large-scale, accurate methane emission measurements from individual cows in commercial farms. The existing ‘sniffer’ technique, which obtains samples of the air near the cow’s nostrils through a tube fixed in the feed bin within an automated milking system (AMS), presents challenges due to the cow’s head movements during the milking recording, creating anomalies in the gas emission data. Moreover, correcting these anomalies directly remains a pressing issue. This study presents a novel data mining approach for detecting anomalous by approximating the cow head position using CO2 time series data collected from cows. The data was collected using sniffer machines on two dairy cattle farms and analyzed using a data-decomposition approach. The performance of the proposed method was evaluated using probabilistic methods such as the Kolmogorov–Smirnov (KS) test and confusion matrix. Results showed that the proposed method was able to accurately detect anomalous peaks in the synthetic CO2 time series, with more than 95% of anomalous peaks being detected. This study demonstrates the potential of the proposed method to improve the accuracy of methane emissions estimates from cows and the understanding of cows’ behavior. To further validate the effectiveness of the proposed method, it was tested on the gas emission dataset recorded by GreenFeed (C-LockTM), which contains various CO2 time series with annotated anomalous points. The validation results demonstrated that the method successfully detected more than 90% of the anomalous points in the time series while marking the peaks containing the anomalous data points. This approach proved to be efficient and accurate, as the peaks represent regions of high anomaly concentration and are more likely to contain true anomalies.