The Research On Structural Damage Detection Methods Based On Feature Extraction And Outlier Detction With Time Domian Response

Monitoring the condition and performance of vital structures is an essential issue in mechanical, civil and aerospace structures to ensure the service quality of structures and avoid grave accident. The structural damage detection technique is the fundamental technique of the structural health monitoring. Therefore, the researches on it are of great theoretical significance and practical value.By far, vibration-based damage detection technique has been the most widely used damage detection technique. The attractiveness of the technique is that it has no requirement that the structure under performance be taken out of service for inspection and an automated continuous monitoring system can be developed because of its simplicity, minimum interaction with users. According to whether or not the feature extraction demands the structural mathematic model (generally FEM model), the vibration-based technique can be divided into two categories, namely the model-dependent method and the model-free method. The model-dependent method has to establish the accurate mathematical model of the structure, which would be quite challenging in case of large-scale structures. Instead, the model-free method has no requirement to establish the mathematical model of the structure. It solely relies on structural responses, which is easy available for practical implementation. In the whole process of the model-free method, feature extraction portion and classification portion are key parts. For the improvement and perfection of the model-free method, with the support of the project (serial number 20023) entitled'Research and Development of Health Monitoring System for Long Span Bridge'which is provided by The Department of Communication of Hunan Province, further studies on these two portions are addressed in this dissertation. The main research work of the dissertation includes two aspects: one is the damage-sensitive features extraction from time domain response, the other is the structural abnormity detection. The main work is as follows:1. The features extracted from frequency domain mostly have low sensitive to early damage. Aiming at this limitation, the singular spectrum extracted from the structural time domain response based on phase state reconstruction combined with singular value decomposition is proposed as the damage feature. In this method, the change of singular spectrum cross-entropy is introduced to indicate the structural damage. The loss of information caused by the signal transfer from time domain into frequency domain can be overcome, which could retain sensitivity of the chosen features. Applying the singular value analysis can effectively eliminate noise. For demonstration, a numerical study on the ASCE benchmark model is performed. The results show that the health condition of the structure can be effectively monitored by the proposed method.2. The outlier detection method based on statistics has to require a prior knowledge of probability distribution of the structural response. In most case, this knowledge is difficult to acquire, which would induce the method unsuitable for practical application. To solve this problem, the outlier detection method motivated by artificial immune system is put forward. First the improvement on the positive selection algorithm is implemented. Next the numerical studies on the efficiency of damage detection employing the negative and positive selection algorithms are carried out. Then the influences on the detection resulting from self match radius and detector number are discussed. Last the validity of the method for damage detection under Gauss excitation and non-Gauss excitation is investigated. The research results conclude that the proposed method hasn't to be provided with the prior knowledge. No matter which method, negative selection algorithm or positive selection algorithm, its detection rate is superior to that of the method based on statistics. In addition, it could acquire a good detect rate at low warning rate by adjusting the self match radius.3. Most outlier detection methods have to set up the learning model. Due to this problem, the outlier detection method based on density is adopted. Some appealing features of this method are: (1) the learning model is not in demand; (2) the dynamic updating of the data can be sustained. The method and the steps for structural damage detection under ambient excitation are described. Two indicators– the relative local outlier factor and the relative local sparsity coefficient for quantify the damage are defined. The ratios of the indicators pre- damage and post-damage are introduced to indicate the damage location. Numerical analysis on Benchmark model and the analysis for the monitoring data from collision between a vessel and the pier of a Bridge demonstrate that the method can reliably detect the structural abnormity. The effects of the relative parameters on the detection are studied.4. Structural ambient vibration response has the disadvantages of low energy and easy be affected by noise. In view of the fact, a method using energy ratio for identify damage in a beam subjected to moving load is developed. The idea of the method is to check the local extreme maximum value of the energy ratios curve, which is plotted with the energy ratio of the pre-damage to that of the post-damage structure at every measurement position. In the numerical simulation, the vehicle is modeled as an unchangeable moving force at invariable velocity, and the bridge is simplified as a continuous Euler-Bernoulli beam simply-supported at both ends. The feasibility of the method to detect single and multiple damages is validated. The effects of the noise, force, speed, window length, measurement position and measurement number are studied. 5. Based on the principle of edge point detection in digital image, the damaged position in a structure can be regarded as an edge point in digital image and a method using an edge operator to assess damage subjected to moving load is presented. The squares of the wavelet packet energy from the measurement positions are obtained. The ratio of the square pre- damage to the square post-damage is served as the indicator. On this condition, the edge amplitude curve is plotted with the edge operators. Then the damaged location can be judged according to the local extreme maximum value of the curve. The effectiveness of the method to detect single and multiple damages and quantify the damage is numerically validated.