Development of structural health monitoring system using data fusion and deep learning method

Damages in structures may cause strength decreases in system elements due to aging and environmental factors, changes in the structure geometry that cause the system to weaken, and additional stresses on the structural elements, and unforeseen situations may arise in the design. Structural health monitoring (SPI) is a process in which certain strategies are applied to detect damage to a structure, determine its location, evaluate its magnitude and determine the remaining life of the structure. Destructive and non-destructive testing methods are used to detect damage in aircraft. While destructive testing techniques refer to methods that eliminate or partially destroy the possibility of reuse of the inspected materials or products, non-destructive testing (NDT) refers to a type of inspection performed without compromising the integrity of the material or part. NDT procedures today have more mature technologies than structural health monitoring (SPI) techniques, but require human intervention and require more labor. It is also more costly due to the use of external probes or equipment. NDT is not a viable option for the condition-based maintenance concept. Therefore, the installation of sensor-based smart structures is gaining importance as a way to monitor damages during use in the field. The use of composite materials allows instantaneous and continuous monitoring of structural health with sensors in design and production processes from nano level to macro level. In this way, it may be possible to turn these structures into smart structures and monitor them effectively during use. Automating damage detection methods can increase analysis sensitivity and reduce dependence on human-based methods by providing technical innovations and convenience to institutions. In this study, real structures produced for use in aircraft were used. In the originally created data set, images taken over these structures and vibration data obtained from vibration sensors placed on the structures were used. A total of 133 images were taken from these structures, 101 of which were damaged and 32 of which were undamaged, taken from different angles, and were divided into two groups as damaged and undamaged. A total of 37 measurements were made, 25 measurements on damaged structural parts and 12 measurements on undamaged parts. A total of 185 sensor data were obtained from all sensors, 5 sensor measurements from each part, and were recorded in the computer environment to be used in the data set. Sensor data is divided into two groups: damaged and undamaged. Image processing algorithms trained with deep learning techniques have been used to examine sensor data along with structural images. Considering that a large amount of data and a long time is required to train artificial neural networks, the SqueezeNet network structure and a problem-specific CNN structure network model were used using the transfer learning method to analyze the originally obtained structural part images. By applying the sensor data alone to the input of the 1 D-CNN Structure, the network was trained and the data was classified. In addition, in the study, the success of the network was evaluated by creating a Multi-Input CNN structure specific to the problem. In this network model, both damaged/undamaged part images and images of damaged/undamaged sensor data changing over time were defined as input, and training and test classification was made as a result of using the two data together in the network structure. As a result of the studies, damaged and undamaged part classification was performed in all models and the most successful result was obtained when using the Multi-Input CNN network architecture. It is envisaged that the models created will be improved and provide better results by using a dataset containing more data. Within the framework of the results obtained, it proposes a cost-effective and rapid damage detection method that can contribute to existing structural health monitoring procedures. Thanks to the deep learning methods used, it offers a solution that is not dependent on auditor experience and eliminates the rate of human error. Damage detection of aircraft structures can be made more practical and precise with computer-based solutions, due to situations such as subjectivity arising from human factors, in addition to the need for expert personnel training and experience.