Norhisham Bakhary

Structure damage detection using neural network with multi-stage substructuring

Artificial neural network (ANN) method has been proven feasible by many researchers in detecting damage based on vibration parameters. However, the main drawback of ANN method is the requirement of enormous computational effort especially when complex structures with large degrees of freedom are involved. Consequently, almost all the previous works described in the literature limited the structural members to a small number of large elements in the ANN model which resulted ANN model being insensitive to local damage. This study presents an approach to detect small structural damage using ANN method with progressive substructure zooming. It uses the substructure technique together with a multi-stage ANN models to detect the location and extent of the damage. Modal parameters such as frequencies and mode shapes are used as input to ANN. To demonstrate the effectiveness of this approach, a two-span continuous concrete slab structure and a three-storey portal frame are used as examples. Different damage scenarios have been introduced by reducing the local stiffness of the selected elements at different locations in the structures. The results show that this technique successfully detects all the simulated damages in the structure.

A comparison of artificial neural network learning algorithms for vibration-based damage detection

This paper investigates the performance of Artificial Neural Network (ANN) learning algorithms for vibration-based damage detection. The capabilities of six different learning algorithms in detecting damage are studied and their performances are compared. The algorithms are Levenberg-Marquardt (LM), Resilient Backpropagation (RP), Scaled Conjugate Gradient (SCG), Conjugate Gradient with Powell-Beale Restarts (CGB), Polak-Ribiere Conjugate Gradient (CGP) and Fletcher-Reeves Conjugate Gradient (CGF) algorithms. The performances of these algorithms are assessed based on their generalisation capability in relating the vibration parameters (frequencies and mode shapes) with damage locations and severities under various numbers of input and output variables. The results show that Levenberg-Marquardt algorithm provides the best generalisation performance.

Application of Neural Network for Prediction of Unmeasured Mode Shape in Damage Detection

The major problem in the vibration-based damage detection field is still a limited number of sensors and the existence of uncertainties. In this paper, a new approach combines a multi-stage ANN model and statistical method to detect damage based on the limited number of sensors with consideration of uncertainties. The first stage of the ANN is used to predict the unmeasured mode shapes data based on limited measured modal data. The second stage ANN is devoted to predicting the damage location and severity using the complete modal data from the first-stage ANN. To incorporate the uncertainties in modal data, Gaussian noise is applied to the input variables and the probability of damage existence is calculated using Rosenblueths point estimate method. The feasibility of the proposed method is demonstrated using an analytical model of a continuous two-span reinforced concrete slab. The application of a multi-stage ANN showed results having a high potential of overcoming the issue of using a limited number of sensors in structural health monitoring.

Substructuring Technique for Damage Detection Using Statistical Multi-Stage Artificial Neural Network

Artificial Neural networks (ANN) have been proven in many studies to be able to efficiently detect damage from vibration measurements. Their capability to recognize patterns and to handle non-linear and non-unique problems provides an advantage over traditional mathematical methods in correlating the vibration data to damage location and severity. However, one shortcoming of ANN is they require enormous computational effort and sometimes prohibitive time and computer memory for training a reliable ANN model, especially when structures with many degrees of freedom are involved. Therefore, in most cases, rather large elements are used in the structure model to reduce the degrees of freedom. This results in the structural vibration properties not being sensitive to small damage in a large element. As a result, direct application of ANN to detecting damage in a large civil engineering structures is not feasible. In this study, a multi-stage ANN incorporating a probability method is proposed to tackle this problem. Through this method, a structure is divided into several substructures, and each substructure is assessed independently. In each subsequent stage, only the damaged substructures are analyzed, and eventually the location and severity of small structural damage can be detected. This approach greatly reduces the computational time and the required computer memory. Moreover, a probabilistic method is also used to include the uncertainties in vibration frequencies and mode shapes in damage detection analysis. It is found that this method reduces the uncertainty effect in frequencies due to duplication error in the multi-stage ANN model and reduces the uncertainty effect in mode shapes due to the damage in other substructures. The developed approach is applied to detect damage in numerically simulated and laboratory tested concrete slab. The results demonstrate that the proposed method can detect small damage with a higher level of confidence, and the undamaged elements are less likely to be falsely detected.

Comparative Study on Design of Experiment in Frequency-Based Response Surface Methodology for Damage Detection

This paper investigates the performance of design of experiment (DOE) in response surface methodology (RSM) for vibration-based damage detection. The ability of three major types of DOE, namely central composite design (CCD), Box-Behnken (BBD) and D-optimal (Dopt) for damage detection based on modal frequency are investigated and compared. A procedure comprising three main stages—sampling, response surface (RS) modelling and model updating—are employed for damage localisation and quantification. By considering Young’s modulus and modal frequency as respective input and output, a set of samples is generated from each DOE. Full quadratic functions are considered in RS modelling while model updating is performed for damage detection. The performances of DOE are compared based on damage detectability. A numerical simply supported beam is used as case study by considering several single damage cases. The results show that CCD provides better prediction compared to other DOEs.