Fault detection in the sniffer-based gas emission measurement systems

Abstract

To ensure the demanded sustainability level in relation to methane emissions during the animal-based production, the continuous emissions monitoring and the diverse emissions mitigation strategies are actively demanded. In this context, a proper measurement technique plays a crucial role. The “sniffers” is relatively novel measurement technique which became favored during the last several years for large-scale methane measurements in commercial farms and to these days is among the most used techniques in cattle. This study addressed the problem of fault detection in such measurement systems. The problem was formulated for the first time and was considered as the model-free detection for stochastic signals with limited prior information. The novel detection approach was developed, verified and analyzed. Specifically, the data model for sniffers measurements was formulated in terms of indexed stochastic processes. To account the effect of non-trivial complex noise in the data, the approach to data transformation based on the Karhunen-Loeve expansion was proposed. Upon this, the fault detection was formulated as the statistical hypothesis-testing problem and the sufficient test statistic was derived alongside with the related threshold. Among the novel results, a formalized notion of an unreliable data is provided in the context of fault detection. The general detection procedure requires calculation of signals’ covariance matrices (constructed from the related trajectory matrices), their diagonalization to allow signals approximation by the Karhunen-Loeve expansion; and calculation of the derived test statistic using the approximated signals. The proposed approach was verified using simulated and real data. Validation tests showed that the use of Karhunen-Loeve transformed signals demonstrate better detection rate than the non-transformed signals. Overall, the proposed approach found to be robust and suitable to automated off-line and on-line data processing.