Intelligent diagnosis of petroleum equipment faults using a deep hybrid model

Research Article Intelligent diagnosis of petroleum equipment faults using a deep hybrid model Rasim Alguliyev1 · Yadigar Imamverdiyev1 · Lyudmila Sukhostat1 Received: 19 December 2019 / Accepted: 10 April 2020 © Springer Nature Switzerland AG 2020 Abstract Performance assessment and timely failure detection of the electric submersible pump can reduce operation costs and maintenance in the oil and gas field. Features of equipment malfunction are changes in vibration signals. Evaluation of vibrations based on accelerometer sensors can detect failures and allows assessment of system failures. This paper proposes a reliable deep learning-based method for electric submersible pump faults detection. The frequency, time and spectral information of the vibrational signal are considered as input to the deep hybrid model. The spectral information includes the spectrogram obtained using the short-time Fourier transform and the scalogram as a result of the continuous wavelet transform and provides a detailed study of the vibration signal. The proposed approach is compared with k-nearest neighbors, support vector machines, logistic regression, and random forest. The experimental evaluation shows that the proposed deep hybrid model is superior to these machine learning methods, and can automatically and simultaneously detect failures of the electric submersible pump according to the vibration signal that is generated during system operation. The proposed approach gives good results and can help an expert in automatic diagnostics of equipment and several complex technical systems. Keywords Vibration signal · Fault diagnostics · Electrical submersible pump · Classification · Deep neural network · Convolutional neural network 1 Introduction One of the most effective ways to artificially lift oil to the surface is to use the electric submersible pump (ESP) systems. ESPs are complex subsystems that support the lift of oil and gas to the surface on the shelf. Installation and possible disposal of ESP due to maintenance are expensive operations. The system must reliably work after it is deployed. Removing faulty equipment should be avoided. Thus, a thorough assessment of the reliability is important [1]. Moreover, deep-sea work makes real-time monitoring of the system virtually impossible. This need motivates a thorough inspection of the equipment in a special test environment [2, 3]. Before installation, the ESP system is tested in the laboratory on large datasets. An intelligent diagnostic system helps professionals detect faults in equipment. The expert should be provided with supporting information about the quality of the system. Therefore, the decision of the intelligent diagnostic system should consider the expert’s opinion. The goal of this paper is to develop a reliable method for assessing the state of ESP using a deep hybrid model. The model combines the advantages of a deep neural network (DNN) and a convolutional neural network (CNN). The frequency- and time-domain features of vibration signals are considered as input features for the DNN model. * Lyudmila Sukhostat, ; Rasim Alguliyev, ; Yadigar Imamverdiyev, | 1 Institute of Information Technology, Azerbaijan National Academy of Sciences, 9A, B. Vahabzade Street, AZ1141 Baku, Azerbaijan. SN Applied Sciences (2020) 2:924 | https://doi.org/10.1007/s42452-020-2741-0 Vol.:(0123456789) Research Article SN Applied Sciences (2020) 2:924 CNN can process two-dimensional (2D) images according to the principle of the human brain, effectively extracts features from images, and also requires fewer training parameters, unlike DNN. The short-time Fourier transform (STFT) spectrogram and continuous wavelet transform (CWT) scalogram of the vibration signal are considered as inputs to CNN in this study. The spectrogram carries information about a fixed time–frequency representation of the signal, which does not allow obtaining a full understanding of what is happening. At the same time, the scalogram provides a more detailed view of the vibration signal. Therefore, the spectrogram and scalogram are sent to CNNs and fused for a more informative study of the signal. However, CNN gives unsatisfactory results on highdimensional images. In this regard, the size of the spectrogram and scalogram is reduced to 128 × 128 × 3 pixels. The proposed approach is compared with k-nearest neighbors (KNN) [4], support vector machines (SVM) [5], logistic regression (LR) [6] and random forest (RF) [7] as the classifier of the ESP state for the implementation of an automatic diagnostic system. The results of this study show that the proposed deep hybrid model can automatically and simultaneously extract features of vibration signals from accelerometers that are sensitive to failures in the time-, frequency- and time–frequency domains. Thus, the proposed hybrid model using deep neural networks can be applied in the diagnosis of ESP failures based on vibrational signals obtained from accelerometers. Testing of the proposed approach is carried out on an ESP system faults dataset that includes various types of failures. The rest of this paper is organized as follows: a literature review is given in Sect. 2. Section 3 describes the features extracted from vibration signals. The proposed deep hybrid model is presented in Sect. 4. The experimental results on evaluating the effectiveness of the proposed approach for ESP faults detection are presented and analysed in Sect. 5. Conclusions are given in Sect. 6. 2 Related work Finding deviations from the normal operation of the ESP that could cause it to malfunction is an important research field. Researchers offer new methods and extend existing fault detection algorithms (Table 1). The currently proposed approaches include vibration analysis and fault diagnosis to solve this problem [8–11]. The vibration signal carries the most important information about the state of mechanical devices, including ESP. Fault-sensitive signs are extracted to intelligently analyse raw signals and improve diagnostic accuracy. A method of centrifugal pump fault (incorrect alignment, Vol:.(1234567890) | https://doi.org/10.1007/s42452-020-2741-0 unbalance, and looseness) diagnosis based on empirical mode decomposition (EMD) method was proposed in [12]. A spectral regression-based approach for fault feature extraction of bearing accelerometer sensor signals was proposed [13]. K-Means method was considered to evaluate the performance of spectral regression (SR), principal component analysis (PCA), factor analysis (FA), locality preserving projections (LPP), Laplacian eigenmaps (LE) and linear discriminant analysis (LDA). The stacked denoising autoencoder (SDA) based fault diagnosis method, where sparsity representation and data compression are used to obtain high-order features was proposed [14]. The SDA model was compared to PCA, SAE (stacked autoencoder) and AE methods and showed relatively better results, because of the abil (...truncated)