Automatic feature extraction of time-series applied to fault severity assessment of helical gearbox in stationary and non-stationary speed operation Diego Cabrera a,b,∗ , Fernando Sancho b , Chuan Li c , Mariela Cerrada d , René-Vinicio Sánchez a , Fannia Pacheco a , José Valente de Oliveira e a Department of Mechanical Engineering, Universidad Politécnica Salesiana sede Cuenca, Ecuador b Department of Computer Science and Artificial Intelligence, Universidad de Sevilla, Spain c Chongqing Key Laboratory of Manufacturing Equipment Mechanism Design and Control, Chongqing Technology and Business University, Chongqing, China d CEMISID, Universidad de Los Andes, Mérida, Venezuela e CEOT, Universidade do Algarve, Portugal Keywords: Deep learning Convolution Auto-encoder Wavelet packets Helical gearbox abstract Signals captured in rotating machines to obtain the status of their components can be considered as a source of massive information. In current methods based on artificial intelligence to fault severity assess- ment, features are first generated by advanced signal processing techniques. Then feature selection takes place, often requiring human expertise. This approach, besides time-consuming, is highly dependent on the machinery configuration as in general the results obtained for a mechanical system cannot be reused by other systems. Moreover, the information about time events is often lost along the process, preven- ting the discovery of faulty state patterns in machines operating under time-varying conditions. In this paper a novel method for automatic feature extraction and estimation of fault severity is proposed to overcome the drawbacks of classical techniques. The proposed method employs a Deep Convolutional Neural Network pre-trained by a Stacked Convolutional Autoencoder. The robustness and accuracy of this new method are validated using a dataset with different severity conditions on failure mode in a helical gearbox, working in both constant and variable speed of operation. The results show that the proposed unsupervised feature extraction method is effective for the estimation of fault severity in helical gear- box, and it has a consistently better performance in comparison with other reported feature extraction methods. 1. Introduction Gearboxes are fundamental components in rotating machines mainly composed by gears, bearings and shafts. These parts interact in a lubricated environment to minimize the friction effects [1]. In this context, the most common failure modes in gearboxes can hap- pen either by mechanical components or by lubrication conditions [2]. Gearbox failures can produce undesired machinery stops, caus- ing huge economic losses and even fatal accidents [3]. Hence, it is important to be able to recognize the condition of each component in an easy way and at a reasonable cost. ∗ Corresponding author at: Department of Mechanical Engineering, Universidad Politécnica Salesiana sede Cuenca, Ecuador. E-mail address: dcabrera@ups.edu.ec (D. Cabrera). Gear wear is a specific failure mode that can appear in all the stages of the device useful life. This fault mode might start from the beginning of the machine operation and increases over time. The gear wear identification allows detecting incipient faults and facilitates the synchronization with planning process, inventory management and it is close to “on-time” maintenance. The work in Jardine et al. [4] shows that historically the main approaches, to diagnose the device conditions in rotating machin- ery, are: (i) waveform data analysis, (ii) value type data analysis, and (iii) data analysis combining event data and condition moni- toring data. In the first case, time-frequency based techniques have out-stood in the representation of information; e.g., Fan and Zuo [5] have shown that Hilbert transform combined with Wavelet Packet Decomposition are suitable to obtain the fault characteris- tic features. In the second one, condition indicators are designed to predict the status of machinery devices. Finally, the third approach has emerged from advances in machine learning techniques over