Since 2019 researchers in the field of deep learning have been exploring the possibilities of Physics Informed Neural Networks (PINN). The training of regular neural networks (NNs) involved an optimization where the loss function depends exclusively on the dataset available. In PINN this loss function takes into account also the physics of the problem, if it is known and the governing equations are given. This paper explores the advantages of the use of PINNs with respect to regular NNs, in the privileged case where a multibody model is available. However, there is still uncertainty around how much weight should be associated with each of the two losses (data-driven loss and physics loss). Therefore, different weights for the two losses are considered and their effect on the performance of the model is evaluated. The research focuses on the synthesis of a four-bar mechanism for trajectory planning of a point belonging to the connecting rod. The objective is to generate a tool that synthesizes the mechanism topology given the desired trajectory. This preliminary study shows how PINN are suitable to automatize the synthesis of mechanisms, where regular NN would generally fail. Numerical analyses also demonstrate that a PINN learns relations from a physical numerical model in a more efficient way than a traditional NN.
KEYWORDS: Education and training, Vibration, Structural health monitoring, Matrices, Accelerometers, Mathematical modeling, Systems modeling, Deep learning, Data acquisition, Damage detection
In the field of structural health monitoring, the adoption of intelligent systems able to automatically detect changes in a structure are evidently attractive. A change in the baseline configuration can be an early predictor of a structural defect that has to be monitored before it reaches critical conditions. When there is no prior knowledge on the system, deep learning models such as autoencoders could effectively detect a change and enhance the capability to determine the damage location. In this paper a deep learning approach is applied to a test rig consisting of a small building model composed by four floors connected by bending springs. Modifications of the system are simulated by changing stiffness of the spring. This algorithm is compared with traditional approach based on modal parameters by carrying out experimental tests to validate the hypothesis.
KEYWORDS: Education and training, Bridges, Windows, Sensors, Deep learning, Neural networks, Distance measurement, Structural health monitoring, Signal processing, Sensor networks
The estimation of trains weight could be useful under certain circumstances. For instance, in the field of structural health monitoring, some considerations can be derived from the evaluation of the load spectrum that an infrastructure has to withstand in its lifetime. One approach to estimate the train weight is based on the use of strain gauges mounted on the rail. The procedure allows to associate the local deformations with the load on an axle. However, strain gauges present several limitations: they are regarded as delicate sensors, and their replacement is burdensome and time-consuming. Moreover, their life is usually short when subjected to weathering and numerous load cycles. For these reasons, this paper proposes a novel methodology that relies on the use of more robust sensors mounted on a bridge structure for the estimation of the train load, alongside other information, such as the number of axles, the train speed, and the train class. The idea consists in the estimation of the train load starting from a network of sensors mounted on a bridge. A deep learning model is particularly suitable to achieve this task. The sensors network must consist of robust and easy-to-replace transducers (such as velocimeters mounted on the bridge structure). In this way, when the strain gauges are removed, the system is still able to estimate the loads passing on the bridge.
In recent years, real-time monitoring of health conditions for massive structures, such as bridges and buildings, has grown in interest. Some of the key factors are the possibility to estimate continuously the health condition, as well as a reduction in the personnel involved in visual inspections and operative costs. However, while dealing with such structures, it is extremely rare to observe anomaly conditions, and when those are met is in general too late. Consequently, the structural health monitoring problem must be tackled as an unsupervised one. The idea exploited in this research is to transform the intrinsically unsupervised problem into a supervised one. Considering a structure equipped with N sensors, which measure static or quasi-static quantities (distance, inclinations, temperatures, etc.), it could be helpful to evaluate if the relations among sensors change over time. This involves the training of N models, each of them able to estimate the quantity measured by a sensor, by using the other N-1 measurements. In this way, an ensemble of models representing the system is built (iterative model). This approach allows us to compare the expected measurement of every sensor with the real one. The difference between the two can be addressed as a symptom of modifications in the structure with respect to the nominal condition. This approach is tested on a real case, i.e. the Candia bridge in Italy.
In the context of Industry 4.0, Predictive Maintenance becomes one of the main challenges for manufacturers. The monitoring of machinery health status results in huge cost savings. For this reason, the industry of automated machinery is moving in this direction, by acquiring large volumes of data, even if without full awareness of which physical variables are important to predict the status of a machine nor, consequently, which sensors to use and where to place them. This paper presents a general approach for the selection of sensors arrangement for the development of a condition monitoring system. The algorithm is based on multibody simulation tool and gives guidelines about the physical quantities to monitor and the parameters to extract. A machine learning model is then trained to demonstrate the ability of the obtained setup in identifying possible faults. The main benefit of this work regards the generality of the approach: it can be applied to different application cases (not only automated machineries), with the only constraint of developing a validated multibody model of the system.
This paper presents a deep learning approach for detecting early fault in bearings. The identification of bearings defects represents an important problem in the field of rotating machines. Sudden failures may occur, leading to breakdown of the machinery. For this reason, the prediction of possible faults has become a major issue in the study of bearing elements. Different fault diagnosis techniques have been developed during the years based on aggregated parameters (i.e. features) that are computed starting from time domain, frequency domain or time frequency domain analysis, relying on prior knowledge about signal processing. These approaches present major limitations, that can be overcome by adopting a convolutional LSTM (long short-term memory) neural network model. In this case, a more complex architecture is built, and the algorithm can identify effective features from accelerometer signal, that could not be considered by a manual computation approach. The algorithm has been applied on data obtained from a complex test rig to assess bearings failure on high speed trains. The outcome of this work indicates that the adopted approach leads to satisfactory performances.
In modern manufacturing industries, quality control systems are crucial components that are rising attention in production environments; companies are looking for new and innovative ways to identify and minimize the quantity of non-compliant products. Intelligent quality control is particularly important when evaluating the outcome of a production line is a complex task (for example when a visual inspection is not sufficient). The first step for building a smart process control system is the identification of all the process variables that are related to the final condition of a product. If key-variables are not directly accessible in real-time, their effect can be derived by means of sensor measurements, but, in this case, a learning model able to put in relation the available information to the inaccessible variables is needed. For all these reasons, in the last couples of decades the building of reliable and robust soft sensors gained a certain relevance in the academic world. In this research an automated rotating machinery is considered. The misalignment condition between two functional parts is the inaccessible process variable, whereas the signal of an accelerometer mounted on the machinery is available for a real time measurement. Changings in rotational speed, according to the production rate required, generate variations in acceleration’s amplitude and cycles’ length. A model based on neural networks is built to detect non-compliant products, while handling different operative conditions.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.