Study On The Method Of Damage Detection For Railway Bridges Based On Vibration Response Caused By Vehicles

Since the railway bridge structures in the complex environment have a large size, possess a strong stiffness, and take a heavy live load, some bridge damage states correlate with the live load. However, this correlation isn't considered by existing methods for damage detection, so when these methods are applied to the real railway bridge structures, many limitations are found. Therefore, by using statistical learning theory, which studies the law of machine learning while the number of samples is small, damage detection of railway bridge structure is studied under train, and the following works have been done:Firstly, aiming to the bridge structural damage states which correlate with train load, the time subdomain method for bridge structural damage detection is put forward. In this method, the structural damage vulnerability is analyzed to find out the vulnerable sections firstly. Then owing to the damage state of the vulnerable sections, right subdomains are selected from the time domain when the train is running on railway bridge. Based on the suppose that the damge state of the studied sections remains the same in a subdomain, damage detection is studied.Secondly, according to the correlation between damage degree with load value, five-step strategy is adopted to study the damage detection in every subdomain:damage alarming, division of damage grade, identification of damage location, recognition of damage degree, and precise damage detection.Thirdly, damage alarming is viewed as a problem of pattern recognition since there are a lot of differences between damage state and healthy state. Acceleration time history is used to construct damage index, sample set is optimized, and the classified algorithm, which is carried out by support vector machine in statistical learning theory, is applied to establish a model of damage alarming. At last, the proposed method for damage alarming is applied to a three-span continuous beam, and the result shows that when the train model of test data is close to the model of the sample, the damage alarming model has a preferable capability of identification and noise resistance.Fourthly, as structural states in different damage grades are distinct, grade division is viewed as a problem of pattern recognition in statistics, and the damage grade of the test structure, which is one of the three grades:inferior, poor and dangerous grades, can be judged by applying support vector machine to classification. A model of grade division is erected according to a three-span continuous bearn, and some different data are used to examine the model. At last, the result shows that the test structure can be classified rightly by this method.Fifthly, based on the characteristics of damage location identification, the hierarchical method has been introduced:the zone where the damage is located is estimated firstly, and then the subzone is identified. In every step, location identification is viewed as a problem of pattern recognition in statistics, and support vector classification is used to establish a model of damage location identification, and then proper index of location identification is constructed, and sample set is optimized to match the model. At last, the example of continuous beam is used to test the proposed method for damage location identification, and the result shows that the model can identify the damage location exactly, and has an appropriate capability of noise resistance.Sixthly, because damage degrees vary continuously, damage degree recognition is regarded as a problem of regression estimation. The index, which is constructed by acceleration time history, is taken as independent variable, and damage degree is regarded as dependent variable. Then support vector regression is applied to solve the problem. The result of the three-span continuous beam test shows that this method can recognize the gobal damage degree of the subzone where the damage is located, but the result has some discreteness.Lastly, based on the characteristics of precise damage detection, a method for precise detection is proposed:after damage degree recognition, the damage subzone is divided into two sections, and one of the sections is viewed as a solving section, of which damage degree is the detection target. By regarding damage degree recognition as a problem of regression estimation, support vector regression is used to recognize the damage degree of the solving section. Then the damage degree of another section is calculated according to the consistency of the subzone global damage degree. The proposed method is applied to the continuous beam, and the result shows that the damage degree of the solving section can be recognized by this method, but the calculated result has some discreteness.