Improving the operational safety of systems is essentially based on fault detection and isolation algorithms, these algorithms mainly consist in deriving the different kinds of faults while minimizing false alarms, nondetections and delays in fault detection. The choice of robust diagnosis by the bond graph approach (BG) is based on the use of a robust analytical relations redundant (ARR) generation algorithm from linear fractional transformation models (LFT), where the uncertain part of the ARR is used to generate the adaptive residue thresholds. These relationships not only allow the detection and isolation of defects on the various elements of the system, but also the location since the industrial system is modeled element by element. The BG-LFT modeling and the robust diagnosis of an accelerometer using micro-electromechanical are presented in this paper. The interaction of the various phenomena is taken into account, thanks to the energy properties of the bond graph tool. The residuals and the adaptive thresholds for normal operation and in abnormal operation are determined. The results suggest that the use of the Bond Graph model for normal operation, the evolutions of the residuals converge towards zero, whereas a fault caused by an element (for example, the stiffness springs [Formula: see text] or [Formula: see text] are modeled, respectively, by capacitive elements [Formula: see text]and [Formula: see text] shows that the residues [Formula: see text], [Formula: see text] and [Formula: see text] become different from zero. These variations explain that these residues are sensitive to these elements, which are confirmed by results presented in Table 2. The analysis of the sensitivity of the residuals to parametric and structural faults is carried out to determine the detectable values of the faults and therefore to monitor the performance of the diagnosis.