Machine Learning based Predictive Analytics for Aircraft Gearbox


Gearbox fault diagnosis is a vital tool to ensure safe operations in a mechanical transmission system. Acoustic and vibratory observations from such types of mechanical devices may be sensitive to faults. The interference transmitting through variant paths and the vibrating signals emitted from the single sensor may not be sufficient to describe the health condition of a gearbox. interference transmitting through variant paths and the vibrating signals emitted from the single sensor may not be sufficient to describe the health condition of a gearbox. To address these issues, multi sensor data fusion is primarily selected for the fault diagnosis of gearbox. In this work, the health condition of the gearbox is acquired using a vibration accelerometer and various operating conditions such as loading condition, rpm variations are considered. Vibrations signals are gathered from motors operating under different loading conditions and health states are also considered.

The three main steps in the proposed method include feature extraction, feature selection and application of bidirectional LSTM. Model based methodology performs statistical and underlying physics analysis to extract the hand crafted features from the individual sensors. However, they are not able to automatically extract spatial and temporal features. Deep learning method such as CNN extract the features automatically from multiple sensors source, while they are unable to extract temporal and spatial features without losing relevant feature information. To handle these challenges we propose a bidirectional, convolutional long short-term memory (BiConvLSTM) network to extract the temporal and spatial features from vibrational and rotational speed to find out the location, nature and direction of gearbox faults.


Gearbox is extensively used in various aircrafts, automobiles and industrial machines for transmitting power and torque. The major failures in gear systems are due to (1) wear and tear (2) tooth fracture (3) pitting (4) scoring etc. The failure can be critical for safety, may affect productivity, machine downtime, unexpected breakdown and maintenance cost may be increased.

Traditionally, fault detection and diagnosis techniques using vibration signals form the basis for monitoring the health of gearboxes. The vibration signals are affected by non-stationary load, friction, nonlinear stiffness and strong noise.

The key issues to evaluate the gearbox fault in the presence of noise are [1] to extract fault feature information effectively, [2] eliminate the uncertainty in the identification process and [3] the fault status accurately. Also the vibration signals collected from the single sensor may be unstable or unreliable. Generally, acoustic signals are mainly used for monitoring the health condition as these signals are sensitive to vibrating bodies rather than vibrating sensors and hence it is able to identify the fault in the early stage. The advantages of acoustic signals and vibrating signals may be combined using sensor fusion so that the resultant signal has less uncertainty than the signals obtained from the individual sensors. The challenges faced by the fusion of signals:(1) feature extraction from different types of sensory data (2) selection of level of fusion. For this, extensive domain expertise is essentially required.

Deep learning methods such as convolutional neural network, autoencoder, deep belief network and deep neural network have been extensively used to automatically obtain the special features for the fault diagnosis of gearbox. In each iteration, an error term is calculated after extracting the features which is then used as feedback so that the training accuracy can be improved [8]. DNN can extract spatial features due to the variations in the amplitude of vibration [9,10]and patterns extracted in frequency domain spectra [11,12]. Besides, DNN can extract the spatial relationship between two observations within a single timestep[11,13]. DNN cannot deal with temporal features since they cannot capture the time of an unusual response due to fault and time interval between normal and abnormal responses. Deep learning techniques like recurrent neural network(RNN) [14] and LSTM[15] have been widely used to automatically obtain the temporal features for health assessment of gearboxes.


This paper explains the system architecture, fusion of multiple signals and the classification of faults to monitor the health condition of the gearbox. In general, the task of assessing the health of gearbox and prognostics consists of four steps:

Data acquisition

Signal preprocessing and feature extraction

Feature selection and data fusion


To remove interference and highlight fault information, energy operator and time synchronous averaging are performed to the multiple sensor data. The statistical features from the raw and preprocessed signals are integrated to form the final feature set. As a first step, feature extraction is applied. Secondly, the feature selection algorithm combines min-redundancy, max- relevance and distance evaluation technique to provide the optimal feature set. Thirdly, a bidirectional LSTM architecture is applied to determine the fault conditions of the gearbox and classify the different types of faults. Most common gearbox faults can be categorized as, (i) root crack (ii) surface spalling and (iii) chipped tooth. The chipped tooth is the rupture of material from the working tip of gear and is extremely common in numerous industrial practices which triggers other gear faults and is very difficult to detect at its initial stage [18]. The dimensions of the seeded bearing are recorded and the gearbox is classified as healthy, minor seeded defect condition and major seeded defect condition.


Investigate different gearbox failure modes

Collect the data from multiple sensors and fuse the data

Data exploration and analysis

Experimenting different feature selection methods

Initial study of appropriate machine learning models

Leave a Reply

Your email address will not be published.