General information: This dataset is meant to serve as a benchmark problem for fault detection and isolation in dynamic systems. It contains preprocessed sensor data from the adaptive high-rise demonstrator building D1244, built in the scope of the CRC1244. Parts of the measurements have been artificially corrupted and labeled accordingly. Please note that although the measurements are stored in Matlab's .mat-format (Version 7.0), they can easily be processed using free software such as the SciPy library in Python.

Structure of the dataset:

train contains training data (only nominal) validation contains validation data (nominal and faulty). Faulty samples were obtained by manipulating a single signal in a random nominal sample from the validation data. test contains test data (nominal and faulty). Faulty samples were obtained by manipulating a single signal in a random nominal sample from the test data. meta contains textual labels for all signals as well as additional information on the considered fault classes

File contents: Each file contains the following data from 1200 timesteps (60 seconds sampled at 20 Hz):

t: time in seconds u: actuator forces (obtained from pressure measurements) in newtons y: relative elongations as well as bending curvatures of structural elements obtained from strain gauge measurements, and actuator displacements measured by position encoders label: categorical label of the present fault class, where 0 denotes the nominal class and faults in the different signals are encoded according to their index in the list of fault types in meta/labels.mat

Faulty samples additionally include the corresponding nominal values for reference

u_true: actuator forces without faults y_true: measured outputs without faults

Textual labels for all in- and output signals as well as all faults are given in the struct labels. Each sample's textual fault label is additionally contained in its filename (between the first and second underscore).