Real-Time Structural Deflection Estimation in Hydraulically Actuated Systems Using 3D Flexible Multibody Simulation and DNNs

The paper clearly states and demonstrates the two central claims on a hydraulically actuated flexible boom: (A) single-step estimation achieves test MAPE below 1.5% in multiple sensor configurations, including a two-sensor case, and (B) batch speed-up S≈10^7 over the flexible multibody simulation, with S defined as S = (tsim / tNN)·(nout / nin). These are documented via figures and text, and the paper also specifies a statistics-only procedure to compute td without equations of motion, along with the S formula used for timing comparisons. However, for replication and auditing, important experimental specifics (e.g., explicit numeric values of nin, nout, tsim, tNN per batch; the exact architecture sizes; and a clearly separated test split) are not fully enumerated in the text beyond figures and general descriptions, making the empirical validation incomplete from a strict reproducibility standpoint (e.g., the reported S uses a figure and definition but without a full tabulated breakdown; the model description gives layer patterns like ‘L’, ‘TLSLTLSLT’ but omits widths). On the candidate side, the solution provides a precise estimator class, a statistics-only method to compute td, and a rigorous training/evaluation and timing protocol that mirrors the paper’s definitions, including the same S formula and a two-sensor option; yet it deliberately does not execute experiments or report results and thus cannot verify A or B here. Therefore, the paper’s experimental proof is under-specified for independent reproduction, and the model solution is a well-defined plan that stops short of verification—both incomplete. Key supporting passages: the paper’s MAPE results and two-sensor feasibility are described in the evaluation section and summary (maximum MAPE below 1.5%; two sensors U and s suffice) , the S formula and speed-up result are given with timing considerations (GPU init 0.2 ms, S definition, factor 10^7) , and the statistics-only td computation is specified (δy-threshold method, n* persistence, averaging td over samples) .