Sensors have been installed on many civil infrastructures to monitor structural conditions. False alarm will be triggered, however, by the malfunctioned sensor even under normal conditions. Therefore, pre-processing is needed to identify and classify the sensor fault before structural damage detection and assessment. This paper proposes a deep learning-based method, namely, the Tsfresh Long Short-Term Memory networks (TLSTM), to address the sensor fault classification. The python package Tsfresh is used to extract features that are sensitive to sensor fault from measured signals. These features are further selected with the Benjamini–Yekutieli procedure. With the selected features, a long short-term memory (LSTM) network combining two fully-connected layers and a Softmax layer is constructed to differentiate sensor fault types. Experimental data with five types of sensor faults are obtained by mechanical and electrical simulation. The proposed method is shown able to successfully classify all these sensor fault types.

A few sensing techniques have been developed for detecting defects in civil infrastructures in structural health monitoring (SHM) instead of onsite human inspection. Many false alarms observed are due to faults in sensors, and some of them may lead to incorrect interpretation of the measured data. In this paper, the differentiation of different sources of variability, i.e. the sensor fault and structural damage, is investigated for a vibration-based SHM system. A deep learning-based method with a new architecture of Fully connected Stateful Long Short Time Memory (FS-LSTM) Neural Networks for differentiating the sensor fault and structural damage without information on details of the fault is proposed. Information from the sensor fault is assumed local, and that from structural damage is relatively global affecting several sensors. Since the LSTM are capable to learn data features automatically, the proposed method performs well without an accurate mathematical model. Stationary random acceleration responses of a three-span continuous bridge when under operational conditions are studied. A deep network model is applied to the measured data with estimation on four types of sensor fault. The response residues between the true and predicted values of the deep FS-LSTM NN is statistically analyzed to determine the threshold of sensor fault. An experimental study with a steel frame in the laboratory is also performed to validate the proposed method.

The challenge of natural language processing is from natural language to logical form (SQL). We present an F-SemtoSql neural approach that is a fuzzy decision semantic deep network query model based on demand aggregation. It aims to address the problem of the complex and cross-domain Text-to-SQL generation task. The corpus is trained as the input word vector of the model with LSTM and Word2Vec embedding technology. Combined with the dependency graph method, the problem of SQL statement generation is converted to slot filling. Complex tasks were divided into four levels via F-SemtoSql and constructed by the need of aggregation. At the same time, to avoid the order problem in the traditional model effectively, we adopted the attention mechanism and used a fuzzy decision mechanism to improve the model decision. On the challenging Text-to-SQL benchmark Spider and the other three datasets, F-SemtoSql achieves faster convergence and occupy the first position.