FastICA-Based Localization of Dynamic Loads in Continuous Structures

Dynamic load identification is a demanding inverse problem in structural vibration analysis because applied forces often cannot be measured directly and must be determined from the structural responses they produce. In continuous systems, accelerations recorded at selected points may be the only available evidence of loads acting at unknown positions and changing rapidly with time. The problem becomes more demanding when several loads act at the same time. For a single excitation, the problem is often more manageable, especially when the load position is known or the event is a short impact. Under multi-point excitation, however, the responses produced by different loads are superposed through the structure’s modal behavior. When every possible loading-point combination must be tested, the computational burden can become a practical barrier, especially in continuous systems where engineers need to recover both the force position and the force history. In a recently research paper published in International Journal of Mechanical Sciences Dr. Zhengshu Wang, Professor Jinhui Jiang, Dr.  Wen Jing, and Dr. Rutong Chen from the Nanjing University of Aeronautics and Astronautics, developed a rapid dynamic load identification method that combines numerical integration, FastICA, and modal shape comparison. The technically distinct element is the formulation of modal load histories as mixed observations of statistically independent physical load signals, allowing multiple load contributions to be separated before localization. The new method then uses proportional relationships among modal shape values to determine load locations and recover corresponding magnitudes. This differs from prior search-based localization strategies by decoupling the multi-load problem and reducing the growth of candidate-location screening from combinational to load-wise.

The research team begins by establishing a mapping from measured structural response to modal loads. The researchers formulate the continuous system in modal coordinates and use a numerical integration procedure, illustrated through the Newmark explicit method, to express the measured response as a convolution-like relation involving modal load histories. Solving this inverse relation yields modal loads of different orders. This step is important because it moves the problem from physical response space, where several load effects are entangled, into a modal-load representation where the mixing relationship can be treated more directly.  The key design choice is the treatment of the modal load vector as a set of observed mixed signals. Since the contribution of each physical load to each modal load is proportional to the modal shape evaluated at the load location, the transformation between physical loads and modal loads has the same mathematical form as the standard Independent Component Analysis model. FastICA is then used to separate statistically independent load components from the identified modal loads. The scientific consequence is clear: multi-load localization no longer requires the simultaneous testing of all possible loading-point combinations. Each separated component can be associated with its own modal-shape signature and localized individually. After separation, modal shape comparison supplies the spatial interpretation. Because ICA cannot determine the absolute order or scale of the separated sources, the method does not rely on direct equality between the separated mixing matrix and the physical modal transformation matrix. Instead, it uses proportional relationships among modal shape values. For each separated load component, the method constructs a modal-shape deviation measure and searches for the location where that deviation is minimized. Once the position is identified, the load magnitude can be recovered by accounting for the modal shape relationship.

Afterwards, the researchers examined pairs of sinusoidal loads with different frequencies, combinations of sinusoidal and impact loading, broadband random loading combined with impact loading, and broadband random loading combined with sinusoidal excitation. Across these cases, the method localizes loads with either exact agreement or small errors relative to the beam length, even when response signals include 5% measurement noise. The reconstructed load histories also remain close to the applied histories, with signal-to-noise ratios reported for continuous loads and peak relative error metrics used for impact loads. One interesting point is that the identified load signals sometimes have better signal-to-noise measures than the original response signals, which the authors attribute to regularization in the inverse step and the noise-reduction behavior of signal separation. Computational efficiency is part of the method’s design logic. The authors compared the method against a modal load reconstruction approach using optimization. Even when the proposed method searches all possible loading positions, they found it requires less time than the comparison method, while approaching the computational time needed for magnitude reconstruction when locations are already known. The reason is structural: FastICA decouples the multi-load localization problem, reducing the required search from a combinational form to a load-by-load search, and modal shape comparison keeps each screening operation simple. The complex wing model extends the evaluation beyond the one-dimensional beam. With two sinusoidal loads applied to a finite element wing structure, the method identifies both load coordinates accurately and reconstructs the corresponding load histories with high reported signal-to-noise ratios. A separate model-error analysis introduces a 10% thickness error into the beam model. The team finds that localization remains relatively stable, while magnitude reconstruction is more affected. This distinction is physically reasonable within the proposed method because spatial localization depends mainly on the modal shape distribution, whereas magnitude reconstruction depends more directly on quantitative model parameters. Experimental verification on a simply supported beam adds a practical test: a sinusoidal excitation and a double-impact load are applied, accelerations are measured at five points, and the method identifies locations within the scale of the force sensor while reconstructing both load types with reported accuracy and faster computation than the comparison approach.

The findings of the Nanjing University of Aeronautics and Astronautics scientists are most directly relevant to engineering situations where structures experience unknown, time-varying loads and where placing force sensors at the actual loading positions is difficult or impossible. The new method uses measured vibration responses to infer both the location and time history of multiple dynamic loads, which makes it useful for structural vibration analysis, load monitoring, fault diagnosis, and structural health assessment in continuous systems. A major application is in aerospace structures, especially because the method was evaluated on a complex wing model as well as a simply supported beam. This extension matters because wing structures, including skins, spars, and ribs, may experience several simultaneous excitations during service. A rapid method that can separate and locate multiple loads from response signals could help engineers identify where dynamic excitation is entering the structure and how those loads evolve over time. This is especially useful for vibration control, fatigue assessment, and post-event interpretation after impacts or abnormal loading events.

The new method also has value for mechanical systems exposed to mixed loading conditions. The simulations include sinusoidal loads, random broadband loads, impact loads, and combined load cases. These conditions resemble practical situations where rotating machinery, aerodynamic excitation, operational vibration, and accidental impacts may overlap. Because the method can separate statistically independent load contributions, it could help distinguish different sources of excitation rather than treating the measured response as one inseparable vibration record. Another application is rapid localization for monitoring systems with limited sensor access. The investigators emphasized that conventional multi-load localization may require searching many possible loading-point combinations, while the FastICA-based strategy decouples the problem and improves computational efficiency. This makes the approach relevant to online or near-real-time diagnostic workflows, where localization must be completed within practical computational limits. The experimental beam validation gives the method practical weight beyond numerical simulation. Using measured responses, together with the reported noise and model-error tests, the paper shows that the method can support response-based identification of multiple dynamic loads when direct force measurement is difficult. Its most natural role is as a response-based tool for rapid localization and load-history reconstruction in continuous systems.