A key challenge in many data-analysis problems is removal of unwanted noise and nuisance signals from your data, so you can access the buried information it contains relating to the physical process you are actually interested in. For example, this is necessary when measuring natural frequencies using accelerometer sensors mounted on a structure, in order to deliver condition monitoring and predictive maintenance. As well as the natural frequencies associated with a structure itself (that will change if there is an issue with the structure), there are also daily and monthly variations (due to changes in temperature and other factors). On top of this, there is a noise element due to variations in the loading of a structure over time.
PolyChord have developed robust techniques that apply Bayesian machine learning to fit out nuisance signals and noise, to leave residuals containing the key information about the structure. This allows us to evaluate and monitor the condition of the structure, and predict its likely behaviour into the future. Maintenance teams are then provided with a prioritised list of which structures are most in need of attention, with alarms raised if any serious problems occur.
Customer Case Study - PolyStructure
PolyChord (in collaboration with Full Scale Dynamics) have fitted non-intrusive low-cost accelerometers to various bridges owned by National Highways and TfL. These accelerometers measure the vibrational response of the bridge as vehicles pass over them on a daily basis.
By applying advanced statistical inference and machine learning techniques, we provide evaluation of the current condition of the structure and a probabilistic prediction of its remaining operational lifetime. This provides the operators of the infrastructure with evidence-based, quantitative, outputs which informs their maintenance strategy and the optimisation of resources.
PolyChord applies multiple modal-analysis approaches to extract the natural frequency of the structure. We then apply Bayesian machine learning to fit out the seasonal and daily variations, as well as noise due to varying loading conditions, i.e. different volumes and types of vehicles crossing the bridge. This provides residuals, as well as knowledge of the inherent uncertainty in this quantity. Our understanding of these uncertainties guarantees robust answers, and evaluation of best and worst-case scenarios. Methods that ignore this unavoidable uncertainty have a large potential of drawing incorrect and misleading conclusions by ignoring the full picture of the data
With the cleaned residuals in hand, PolyChord applies a cutting-edge statistical inference approach to state (with confidence) whether a structure is detected to be degrading or not. In other words, rather than a yes or no answer, which is prone to producing false alarms, we can make statements such as "We are 99.99% certain that the structure's condition is degrading, but it will remain within safe operational limits for at least 5 years, and potentially up to 12 years." This process is repeated on a continuous basis, so that its future predictions are constantly refined as more data is acquired.
PolyStructure is not just another AI/Machine-Learning solution. Our solutions are underpinned by a strong physics and engineering understanding of the problem. We work closely with engineers to understand the nature of the issues they face, in order to gather the correct data, and design analysis pipelines that exploit well-known physical processes.
We put robust statistical analysis at the core of our development, rather than blindly applying machine learning. As a result our solution has fewer issues with false positives than pure machine-learning approaches and is explicable.