The use case is relative to flight test helicopters vibration measurements. The dataset has been collected and released by Airbus SAS. A main challenge in flight tests of heavily instrumented aircraft (helicopters or airplanes alike) is the validation of the generated data because of the number of signals to validate. Manual validation requires too much time and manpower. Automation of this validation is crucial. 
    
In this case, different accelerometers are placed at different positions of the helicopter, in different directions (longitudinal, vertical, lateral) to measure the vibration levels in all operating conditions of the helicopter. The data set consists of multiple 1D time series with a constant frequency of 1024 Hz taken from different flights, cut into 1 minute sequences.

We are interested in the detection of abnormal sensor behaviour. Sensors are recorded at 1024Hz and we provide sequences of one-minute length. 

Training data
The training dataset is composed of 1677 one-minute-sequences @1024Hz of accelerometer data measured on test helicopters at various locations, in various angles (X, Y, Z), on different flights. All data has been multiplied by a factor so that absolute values are meaningless, but no other normalization procedure was carried out. All sequences are considered as normal and should be used to learn normal behaviour of accelerometer data.
Validation Data
The validation dataset is composed of 594 one-minute-sequences of accelerometer data measured on test helicopters at various locations, in various angles (X, Y, Z). Locations and angles may or may not be identical to those of the training dataset. Sequences are to be tested with the normal behaviour learnt from the training data to detect abnormal behaviour. The amount of abnormal sequences in the validation dataset is a priori unknown.

Datasets are provided in a HDF5 format that can be decoded by many standard machine learning modules (like pandas for instance):
  +  In the training dataset, the dataframe is called "dftrain"
  +  In the validation dataset, the dataframe is called "dfvalid"
Each dataframe has 61440 columns corresponding all time steps contained in one minute at 1024Hz and are named from 0 to 61339.