RICERCA ITALIANA     

Research program

Integrated Methods and Algorithms for NonDestructive Evaluations of architectural heritage

University Co-ordinator

Università degli Studi di CAGLIARI - INGEGNERIA ELETTRICA ED ELETTRONICA - CAGLIARI(CA)

Research Unit Leader

Barbara CANNAS

Description

The latest analytical methods' development opens new ways to improve Non Destructive (ND) techniques in general, allowing to overcome the main drawbacks which made this methods non-reliable in the past. This non-reliability was caused by various reasons, which are, for example:

- low accuracy of survey's results, due to not enough advanced mathematical devices;
- subjectivity of result's final interpretation, usually entrusted human operator, who needs large experience in the subject;
- impossibility to joint multisensorial data and results, except by visual comparisons, subjective and not rigorous;

In order to overtake the aforesaid problems, Unit n.4 proposes the development of Soft Computing, Modelling, Data Fusion and Decision Fusion techniques for the processing of data recorded by various ND systems for diagnosis of valuable masonry. In particular, for diagnosis of stone, brick and concrete (even reinforced) masonry structures having monumental, civil or industrial significance.
Referring to ND survey's systems, the Unit n.4 will undertake data acquisition by Sonic survey. Presently, the research Unit has a tool for sonic analysis, consisting of a digital oscilloscope, a transducer, a receiver and a PC. The tools supplies the propagation time of the sonic waves. The Unit's commercial equipment has been still used in both quarry and waste stone materials, in various monumental structures (allowed tanks to Municipality and Supervision of Cultural Goods) and in both screen-test and in-situ concretes. Hence, a large measurement database is available yet. Besides, for the chosen test case structure, data from electromagnetic, ultrasonic and thermographic surveys, available from the other research Units involved, are going to be used.
The Unit n.4 have differentiated and complementary competencies in data processing techniques, modelling, optimisation and soft computing and in Non Destructive diagnosis too. Especially, researchers belonging to ING-IND/31 Electrical Engineering area, are going to co-operate with researchers of ICAR 09 Structural Engineering area. The latter developed extensive experience on sonic and ultrasonic techniques, pull-out, pull-off, termography, flat-jack, etc, applying these methods in structural diagnosis for many years.

The use of sonic ND method is directed to the attainment of the following targets:
- Structure's homogeneity evaluation, regarding physical and elastomechanical aspects.
- Finding damaged area or structure's elements, in case of their presence, and evaluating their decay's rate.
- Estimation of materials' elastomechanical characteristics such as the elastic modulus.
- Shape, geometry and displacement of singularities, if present, such as voids or inclusions or, in general, local alteration of medium's features.

The first phase of project will be focused on identification of defects to be singled out: for example, structure's discontinuities (cracks or instabilities), higher density's areas, due to subsidences, voids or cavities in general, etc. In effect, the problems to be solved show a multitude of events, because different diagnostic operations are needed to evaluate the structure's condition change in different kinds of structures.
Once determined the kind of target to be detected, the Unit will proceed selecting a benchmark for which multisensorial data are available or easy to collect. In this part of the project, the Unit n.4 will resort to experts' estimates. Moreover particular objects having known structural defects will be studied to build a training set of examples for which the correspondence measurement-defect or measurement-object's parameter is already determined.

The second phase will concern the sonic measurements' acquisition by instrumented hammer and single piezoelectric transducer or transducers' array. Time series corresponding to different kind of measurements will be acquired, that is:

- repeated measures to study the same object with the same coupling source-transducer, necessary to evaluate the trend of environmental and instrumental errors, to verify the repeatability of measurement, purging the latest one from single test's specifics (source's impact intensity, transducer's angulation, coupling between transducers and object) and to generate synthetic measurements which constitute the supervised diagnosis systems' training set;
- more measures acquired to study the same object using a transducers' array and varying the source's location in order to obtain a set of measures to be treated with tomographic techniques or to acquire superficial waves. To realise that sub-phase, the Unit will resort to array processing techniques, developed by the Unit n.2;
- more measures pertaining the study of various adjoining sections having the same axis to obtain a 3D survey of the object.

After the acquisition's phase, primary time series processing is needed. Purposing to define an efficient system, data preprocessing is necessary. The Unit n.4 will resort to techniques, developed and implemented by Unit n.1 and Unit n.2, to identify the signals' features representing the studied phenomena, to filter the noise, to integrate missing data and to verify the outliers' presence.

The third phase of the project will involve data processing to obtain the Significant Parameters for the interpretation of the events (presence-absence of defects, classification of defects). In effect, as noted before, in survey's techniques, such as sonic, the acquired data doesn't coincide with the data significant for the structure's analysis. It is necessary an analytical processing to evaluate the significant information.
Thus, Inverse Modelling and Inverse Ill-posed problem solving techniques have to be used. In this project, optimisation techniques are going to be developed and an optimal Regularisation algorithm, chosen from innovative algorithms (Variational Regularisation, POCS Algorithms, Preconditioned Methods, algorithms to choose Tikhonov Regularisation's parameter such as Weldoz method, L-curve method and so on), will be implemented.
The Inverse Modelling problem will be solved by means of an inversion methodology based on Neural techniques, in which the Unit have developed an extensive experience.
This phase will return the evaluations of Significant Parameters needed to study considered object. In the field of sonic survey, the Significant Parameters consist in medium density and elastic parameters. These quantities could be valued or directly or via propagation speed of elastic transversal and/or longitudinal waves (in case, by way of surface waves' analysis) in the considered medium.

Moreover, the Significant Parameters have to be processed, using Feature Extraction and Data Reduction techniques, to maximise the information quality and significance and to make a first analysis. These techniques will be developed and particularised to Significant Parameters by the Unit n.1.

The fourth phase of the project will concern the development and the application of systematic and not subjective techniques for the integration of data coming from multidirectional and multisensorial acoustic surveys.
Data Fusion (DF) techniques allows one to systematically obtain multidirectional and multisensorial data interpretation increasing results' quality with respect to those obtained with a subjective-qualitative analysis. Therefore, it is necessary to implement a platform for the fusion of data acquired and processed by the Inversion techniques, to obtain a single optimised result for each section and for each sensor, which can be elaborated in the following interpretation platform. In this phase, the DF algorithms are related to information fusion at "Feature level", which is the level of measure's characteristics individuation. They return a "summary" of the information coming from various sensors or measures, and extract the significant information for the analysis. For those purposes it is useful the combination of multiple DF techniques, e.g..:
- algorithms strictly based on Bayesian approaches, which are the only ones properly definable as "Statistics". They are based on Bayes's Maximum a Posteriori Probability (MAP) modified by the Minimum Risk Theory (MRT);
- Bayesian "Naïve" algorithms. In these algorithms the MRT Discriminant function is simplified by the hypothesis of sensors' confidence level independence for each class;
- Dempster-Shafer theory, which assigns a probability mass for each decision (i.e. chosen between "intact medium" or "void's presence"). This probability mass takes into account the sensor uncertainty too;
- Fuzzy techniques. In this logic, the object membership function to a given class, varies in the range [0,1], and it is not a Boolean variable as in the statistical approaches;
- Rule-based algorithms, which consider various threshold and conditions to be verified to accept an element in a class;
- Voting fusion techniques, which consist in a special Rule-based technique case.

The fifth part of the project will involve the development of Soft Computing techniques at decision level. In the past, neural techniques to interpret ND surveys' results have been mostly used for industrial applications, such as searching defects on metallic materials. The literature is poor of SC techniques applications in valued masonry structures. The Unit n.4 will develop interpreting techniques by using neural supervised classifiers, also trained by the aforesaid data-set based on experts' evaluations for known objects. The Unit will return a first coupling data-defect, with aforesaid techniques. At the same time, the Unit n.1 will use Non-supervised and Fuzzy classification (or clustering) techniques.

Then, the implementation of a decision platform based on Decision Fusion techniques is needed, in order to obtain the final response of system.
In the last ten years, the interest on new classification approach is growing up. This approach takes into account that different classifiers present different performance in different subsets of the feature space. If these subspaces are properly differentiated, the combined information coming from different classifiers could be helpful, because of the exploitation of the advantages proper of each classifier.
In literature different criteria to combine the outputs of different classifiers are presented. These criteria differ for their complexity. The simplest ones are based on ‘fixed' combination rules, such as the majority voting or the simple mean. The complex criteria use adaptive techniques or (trainable) methodology based on knowledge, such as the weighted voting or the Behaviour Knowledge space rule.
In the sixth phase, Unit n.4 will realise a decision platform based on criteria following Weighted and Rule-based algorithms' logic.

The following diagnam reports the temporal organization of the different phases of the project, tigether with the interactions with the other units of research.