This dataset description paper introduces a challenging sequence processing task. Specifically, the task is to recognize faults from aircraft gas turbine engine conditions where previous faults determine how the presently observed conditions should be interpreted. Several state-of-the-art deep-learning sequence processors have been tested on the task, and these preliminary results demonstrate that they cannot correctly model the phenomenon.
Time series forecasting and time series classification are
the major sequence processing tasks, and both depend on a
method’s ability to identify time-based patterns in the data.
When dealing with long sequence processing in particular,
successful methods must identify non-local patterns such
as trends, seasonality, and long-range dependencies.
Observing large-scale and high-impact repositories such as
the Monash Time Series Forecasting Archive [
The sequential integration of sub-symbolic and symbolic modules allows connectionist recognition of single events to interact with long-range dependencies between events. Such neuro-symbolic approaches are powerful, but to demonstrate their efectiveness they must be applied to datasets and tasks where the following conditions hold simultaneously: (a) the recognition of individual events in the raw data is a non-trivial pattern recognition task for which connectionist methods are more appropriate; and (b) this recognition is afected by unknown symbolic patterns that need to be discovered in conjunction with the sub-symbolic patterns themselves.
In this dataset description paper we present such a task where purely connectionist sequence processing performs substantially below the par established through the usual trend and seasonality tasks and where neuro-symbolic methods can demonstrate their efectiveness in challenging realworld applications.
Jet engines play a vital role in aviation, and ensuring their safety and reliability is paramount. The harsh environments they operate in can cause malfunctions, impacting aerothermodynamic measurements. Detecting faults in-flight from time-series data remains challenging due to limited instrumentation, similar fault efects, concurrent faults, and prior events. This paper introduces a dataset of simulated timeseries measurements from a turbofan engine, including various common faults, to aid in developing fault detection strategies. This dataset is valuable for the Machine Learning community, presenting a complex engineering problem with unique features. It includes 2,410 time series, each with 3,600 time steps, where faults are introduced, and measurements are taken at every step. Some faults have nearly indistinguishable efects, posing challenges for diferentiation, while others exhibit long-term dependencies. These attributes make the dataset suitable for research in sequential and timestepwise classification and investigating long-term dependencies. The dataset comprises time series measurements from a low bypass ratio turbofan engine, typical of modern civil aviation engines, generated using an Engine Performance Model. The dataset has been archived and is publicly accessible from https://doi.org/10.5281/zenodo.15856441
An EPM interrelates parameters that represent engine
component health and operating conditions with measurements
performed on an engine [
Y = g(u, f )
(1)
where g is a vector function representing the EPM, u is
a vector of measured quantities defining the engine’s
operating point, Y is a set of measurements for condition
monitoring, and f is a vector of engine component health
parameters. Typically, two health parameters are used for
each component: the flow factor (SW), indicating the
component’s swallowing capacity, and the eficiency factor (SE),
representing thermal eficiency. The application of such
parameters for assessing engine component health has been
discussed in [
Using the EPM, the nominal values of the available measurement set, denoted as Yo, can be computed for a specified operating point u. These values correspond to the nominal health parameters, fo, and are derived via the following relation:
Yo = g(u, fo)
Once a measurement vector Y is obtained from the engine, the relative deviation ΔY(%) from its nominal value Yo is defined as: ΔY(%) =
Y −
Yo Yo
To emulate realistic measurement conditions, random
noise is superimposed on the measurement deviations. The
noise levels considered are consistent with those typical of
this instrumentation, as documented in [
The EPM used for dataset generation was developed in
MATLAB. At each time step, the data includes a labeled vector of
measurement deltas produced by the EPM at a specified
operating point and a specific engine health condition. A total
of 10 operating points were considered, covering a broad
range of the engine’s flight envelope. Throughout a time
series, all time steps correspond to the same operating point.
Additionally, five distinct health conditions were considered,
as summarized in Table 1. The first health condition
represents the nominal, healthy operation of the engine. The
second health condition (coded as TIP) models an increased
clearance of the compressor blades. As reported in [
Each time series comprises 3,600 time steps. During this period, various health conditions may develop. A time series without any fault contains measurements indicative of healthy engine operation at all time steps. In other time series, the engine operates fault-free initially, but at a certain point, a TIP, VGV, or FOD fault occurs and persists until the end of the series. In some other time series, the engine also operates fault-free initially, but at a certain point, a brief period occurs during which a VGV fault is present. The (2) (3) (4) engine then continues to operate without faults until a later time step, when a combined VGV+FOD fault manifests and persists for the remainder of the time series. This scenario exemplifies a specific fault condition characterized by a severe VGV fault leading to FOD. The damage is attributed to the detachment of a particle from the VGV mechanism, such as a loosened bolt, which is subsequently ingested by the engine. The failure of the VGV mechanism may be preceded by a transient VGV fault event, serving as an early indicator of the impending failure. This scenario was selected because both the TIP and VGV+FOD faults exert similar impacts on engine performance. The key diferentiator between the TIP fault and the VGV+FOD fault is the long-term dependency of the latter on the initial, instantaneous VGV fault.
The dataset is used to evaluate the classification capabilities of sequential machine learning methods on time-series data with long-term dependencies. Preliminary results suggest the dataset challenges current techniques. For example, Figure 1 shows four classification methods trained on the dataset and tested on a time-series segment containing a VGV+FOD fault. The methods, include an RNN, a GRU, and an LSTM, together with an MLP, trained for point-topoint classification for comparison. All methods failed to detect the VGV+FOD fault, though they detected a precursor VGV fault at time step 740. Sequential methods have shown limited ability to use information from prior time steps, performing similarly to the point-wise MLP.
This research was co-funded by the European Union under GA no. 101135782 (MANOLO project). Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or CNECT. Neither the European Union nor CNECT can be held responsible for them.
The authors have not employed any Generative AI tools.