Actors: Inria / MODAL, Inria / STATIFY, Cerema / GIPI, Cerema / GITEX, Cerema / ENDSUM, Cerema / DTerSO, Cerema / DTerMed
The data available is potentially very changeable because depending on the technological evolution of the sensors used and the instrumentation campaigns of the moment (e.g .: the sampling frequency changes with the new sensor or a sensor measuring a new physical quantity is added. on the site). In addition, one type of data that will be systematically encountered corresponds to temporally and spatially dependent data (typical example of this double dependence: multivariate time series). Difficulties specific to these data will also have to be taken into account, such as the alignment of curves (a fairly open question still today). It will also be necessary to take into account the fact that the data are often scattered (many works not / little instrumented) which leads us to problems of missing or partially missing data with variable uncertainties and errors. All this contributes to the great versatility of the data available on road infrastructure and OA. This axis 2 will be fed by the data collected under axis 1 and conversely can feed the task T1.1 to enrich the study of the places and methods of data capture.
Task 3.1: Understanding / visualizing sensor outputs
Actors: Inria / MODAL, Inria / STATIFY, Cerema / ENDSUM, Cerema / DTerSO, Cerema / DTerMed, Cerema / GITEX
The measurements of the sensors over time give rise to numerous curves. Analysis of these curves usually requires a time-consuming visual inspection. The objective of this WP is to provide aids for understanding and visualizing sensor outputs, through the following points:
View curve data from dimension reduction methods;
Classify the curves in order to determine homogeneous subsets of curves, and thus summarize the information of the curve by the homogeneous group from which it comes;
Detection of breaks in the signal and confidence intervals for functional parameters (mean);
Automatic detection of anomalies in curves.
The problem will be addressed as part of the functional data analysis, and will take into account the dependence between the different sensors.
Task 3.2: Detection of anomalies in images
Actors: Inria / STATIFY, Cerema / ENDSUM
The objective is to detect defects or anomalies from image data (eg: cracks in tunnels). Three types of difficulties are identified:
The combination of heterogeneous information coming from multimodal measurements: the objective is to multiply the sources of information to move towards detection methods requiring less supervision, because the production of ground truths (manual identification of faults) is generally expensive.
Taking into account a priori knowledge: in addition to measurement information, it is a question of taking into account in the search for defects their spatial or geometric aspect (eg: the cracks have regular linear shapes).
Decrease in the number of false alarms (wrongly detected faults): classic detection approaches focus on classification performance but do not for the most part provide a reliable measure of the uncertainty on this classification. A better quantification of this uncertainty should make it possible to better identify the “false alarm” candidates.
Task 3.3: Forecasting the evolution of key quantities
Actors: Inria / MODAL, Inria / STATIFY, Cerema / DTerSO, Cerema / DTerMed, Cerema / GIPI, Cerema / GITEX.
We have measurements of a key quantity (for example the size of cracks or the evaluation score of a part of a structure) measured at several points of a structure over time. These data can be modeled by multivariate time series. There is therefore a double dependence, in time and in space. We are interested in two difficult problems from a practical point of view:
Estimate the probability that the size of a crack exceeds a critical threshold (not yet observed in the data);
Conversely, given a fixed risk (in the form of a probability), what is the size of the crack associated with this probability?
In both cases, the problem can be approached by extreme value statistics which provide tools for extrapolation beyond the field of observation of the data. However, these tools cannot be used blindly and need to be adapted to the complex data available. It is necessary to take into account the double dependence already mentioned and the non-stationarity (seasonality of the phenomena due to meteorological conditions and the overall tendency for structures to deteriorate).
Task 3.4: Help with predictive maintenance
Actors: Inria / MODAL, Inria / STATIFY, Cerema / GIPI, Cerema / GITEX.
This involves early identification of the drifts of the parameters observed in the monitoring process and therefore to move towards an on-line prediction (the data arrive in continuous flow). This WP will use some outputs from WP 3.1 and 3.3, with an objective of reusing the associated probabilistic tools (degradation law, law of evolution) in a maintenance context. One stake will then be to mix the various laws already estimated and to allow their use on-line.
An appropriate formalism should be proposed for:
- take into account the various uncertainties associated with the repair action itself, with the various associated costs;
- carry out optimization that takes into account these different uncertainties;
- consider several criteria, sometimes conflicting, when doing optimization (notion of Pareto front).
