Induction motors are electromechanical devices used in most industrial applications. Due to their simple design and well-developed manufacturing technologies, induction motors are considered relatively reliable and robust [1
]. However, the motors fall into failure mode and seriously affect industrial operations. Eventually, this leads to failure of the entire operating system if the failure condition is not identified or if it is neglected. Several types of faults related to winding, stator, rotor, and bearing can be observed in an induction motor [2
]. There are mainly four types of fault diagnosis methods such as signal-based, model-based, knowledge-based, and active/hybrid methods. Signal-based methods use the measured signal to extract the features and make a diagnostic decision based on the prior knowledge of the diagnostic process. Signal-based methods can be classified into a time-domain signal-based method, frequency-domain signal-based method, and time-frequency signal-based method. Model-based methods can be categorized into deterministic fault diagnosis methods, fault diagnosis methods for discrete-events and hybrid systems; stochastic fault diagnosis; and fault diagnosis methods for distributed and network systems, which are categorized by the model type used [4
]. Hybrid methods are studied as a combination of two or more fault diagnosis methods. For example, He et al. [5
] diagnosed plastic bearing faults by combining the signal-based and data-driven methods. By combining signal-based and knowledge-based techniques, a fault diagnosis method was studied to detect the inter-turn faults in induction motors, in which wavelet transform is applied to extract the features from the collected vibration signals, and the principal component analysis (PCA) and neural network (NN) were used as classifiers to classify healthy from faulty motors [6
]. Active methods are studied as a system in which a suitably designed input signal is injected into a dynamic process during a test period to distinguish accurately and quickly the faulty modes from the normal modes. Stochastic active fault diagnosis and deterministic active fault diagnosis are the two active methods studied [7
]. For example, Campbell et al. [10
], used two candidate models one as a normal system and another as a fault system in a multimodel system, and an auxiliary signal was designed to detect the correct model under a given interval of test time. As an extended study, an active fault detection method for multiple faults generated simultaneously or sequentially [11
Knowledge-based methods, also considered as data-driven methods, are the most commonly used methods for analyzing signals such as vibration, temperature, electrical tension, and current. These methods require a huge amount of historical data to find the patterns in the given signal. However, data signals can be captured using sensors [12
]. Signal-based features are extracted, and feature selection methods are applied to reduce feature dimensions and also to avoid the repeated information, which in turn improves the performance by holding the significant features [14
]. The extracted features are used for fault diagnosis by various traditional machine learning methods [15
]. Traditional machine learning methods have achieved prominent results. However, feature extraction depends significantly on diagnostic knowledge and signal processing expertise. Furthermore, traditional methods are incapable of extracting discriminative features from raw data and always require a process to extract the feature from the signal [16
In spite of the advanced development in machine learning, deep learning (DP) has become the most effective study that can significantly overcome the drawbacks of traditional machine learning methods for fault diagnosis. DP can automatically extract and learn abstract features from raw data, and avoid manual feature extraction [18
]. Many deep learning models, such as deep belief network (DBN) [22
], stacked sparse auto-encoder [19
], sparse auto-encoder [21
], denoising auto-encoder [23
] and sparse filtering [24
] have been studied to diagnose the faults, and very significant results have been achieved. One of the most effective used deep learning model, called convolutional neural network (CNN), has been used to learn hierarchical feature representation from raw data and has delivered promising results [20
Most pattern recognition tasks deal with time-series data signals. Financial data (stock and currency exchange rates), video processing, music mining, weather and forecasting, biometric data, and biomedical signal processing are few examples, in which time-series data have been studied [26
]. Likewise, electrical industrial devices also often work with time-series data such as measurements of voltage, current, temperature, and vibration signals. One-dimensional CNN has been studied and applied to time-domain machinery data signals to diagnose the faults in induction motors [29
]. In a few cases, machinery data can also be represented in two-dimensional (2D) model, such as the time-frequency domain using the wavelet transform technique [30
]. In addition, time-series data can be represented in 2D texture images using the concept called recurrence plot (RP) [31
]. Image representation of time-series data provides a different set of features that are not available for 1D signals. Therefore, 2D texture images can be used for classification [32
In this study we use a time-series data signals converted as recurrence plots (RP) to feed the proposed deep CNN for fault diagnosis. Current signals with a phase difference were collected from 3-phase induction motors, and each phase current was used as an input data sample. Our results show that RP provides an effective method for viewing trajectory periodicity over a phase space, allowing us to discover specific elements of the m-dimensional phase space trajectory using a 2D figure. The main contributions of this paper are illustrated as follows. First, the raw current signals are represented as RP images. Second, an efficient deep CNN model is studied and applied on RP images to extract the multi-level features for fault classification. Lastly, the proposed deep CNN based framework achieves significant results compared to other deep learning methods.
The rest of the paper is arranged in the following way: Section 2
discusses the related works. Section 3
explains the proposed framework—time-series data to 2D texture image conversion—and proposed deep CNN models are discussed. Section 4
presents the experimental results followed by the conclusions and future work presented in Section 5
2. Related Works
This section briefly reviews recent deep learning contributions on induction motors’ fault diagnosis. Several types of signal processing methodologies have been studied in the time-domain, frequency-domain, and time-frequency domain to extract and learn the features in order to classify the working condition of the motor. Lee et al. [33
] studied the convolutional deep belief network (CDBN) to classify audio signals. They converted time-domain into frequency-domain data to learn the features form audio signals. A multi-channel CNN has been studied to handle multi-variate time-series data [34
]. A separate CNN is used to learn the features from individual time-series data, and result from all the CNNs are combined and classified using a fully connected multilayer perceptron (MLP) classifier. Audio signals are transformed into a time-frequency domain to feed into CNN for classification [35
]. The Gramian Angular Field (GAF) and Markov Transition Field (MTF) are used to convert time-series signals as images. A tiled CNN is used to classify time-series images [36
Ngaopitakkul et al. [15
] explain a decision algorithm based on ANN for fault diagnosis. Pandya et al. [17
] propose an efficient KNN (k-Nearest Neighbours) classifier using the asymmetric proximity function for fault diagnosis. Yang et al. [16
] constructed an SVM (Support Vector Machine)-based method to diagnose the fault patterns of the roller bearings. Jia et al. [38
] propose a fault diagnosis method based on deep neural networks using an auto-encoder. Deep learning models such as deep autoencoder (DAE), deep belief network (DBN), and CNN have been discussed for fault diagnosis [19
]. Ince et al. [26
] propose a one-dimensional (1D) CNN to diagnose faults using real-time motor data. Abdeljaber et al. [39
] studied 1D CNN to detect real-time structural damages. A deep CNN was used to analyze multichannel time-series data signals for human activities [40
]. However, these models only used a small amount of low-level features in hidden layers. However, in this paper, we study a deep CNN method to automatically learn the useful texture features in order to classify faults. 1D raw current signals were converted to 2D images, and the proposed CNN model was able to successfully capture the temporal and spatial dependencies in the images by applying relevant filters. Furthermore, the proposed model was able to extract and learn high-level features from these images along with the low-level features. The performance of fault classification improved by the combined implementation of feature extraction and the CNN classifier.
3. Proposed Study and Framework
This section explains the proposed framework based on RP images and deep CNN for fault diagnosis. It consists of two subsections: (1) the time-series data signals are converted into 2D texture images, and (2) a deep convolutional neural network model is discussed to learn features from texture images for fault classification.
3.1. Time-Series Data to 2D Texture Images
The time-series data can be categorized using a unique recurring behavior such as periodic and irregular cyclic aspects. Moreover, time-series data are generated as the repetition of states, which is a normal phenomenon for ever-changing irregular systems or random processes. RP [31
] is a tool for visualizing and investigating the m-dimensional phase space trajectory using a 2D representation of its repetitive occurrences. The primary idea of RP is to disclose trajectory movements from the current state to the previous state and it can be formulated as:
is the number of states
is a threshold value of distance, ‖. ‖ is the norm and θ
(.) is the Heaviside function. The recurrence matrix (R
) comprises two sets of values called texture and typology. The texture information belongs to individual dots, sloping lines, perpendicular lines, and horizontal lines, whereas the typology information categorized by uniform, regular, shift, and interrupted. Obviously, in RP, there are patterns and information that are not easily visually seen and interpreted. The detailed explanation can be found in [32
Raw current signals are collected from 3-phase induction motors for the fault analysis every 5 s with a sampling rate of 10,000 samples per second. Data samples taken for different periods of time from 1 s to 5 s were investigated with recurrence plots. Five seconds of data samples gave the most distinguishable patterns in the recurrence plots. The collected raw current signals from the operating induction motor are represented as recurrence plot shown in Figure 1
As shown in the Figure 2
, nondistinguishable recurrence plots were generated for two different motors operating with different modes of failure, when the raw current signal values were used to generate recurrence plots. Even though, the motors working with different modes such as faulty or healthy, it can be clearly seen that it is almost impossible to find a distinguishable pattern in these recurrence plots—they look exactly the same with no difference in any color or pattern. However, to find distinguishable patterns in RP images, an effective preprocessing technique called Max–Min difference was used in this study, and it is implemented as follows:
Step 1: maximum and minimum peaks of the current signal are collected for each one full cycle.
Step 2: difference between the maximum and minimum peak value is then used to generate the recurrence plot for the whole signal.
Step 3: the above two steps are repeated for all types of faults, and healthy motor signals generate distinguishable recurrence plot.
As shown in the Figure 3
, after applying the preprocessing technique to raw current signal values, the generated recurrence plots are well distinguishable and can be considered for classification of different faults and healthy conditions of the motors by CNN. Initially, the 2D recurrence texture images were generated by raw one dimensional (1D) current signals and then classifier automatically learned the features from texture images to classify the motors’ fault condition.
3.2. The Methodology’s Architecture
The methodology’s architecture consists of two parts. Part one explains the architecture used to train the CNN model and part two belongs to motor testing using the trained model.
The relevant data were collected for a total of five conditions of the induction motor. A setup of four induction motors for the following four faults and one motor for the healthy condition were used. The four faults were: (1) bearing axis deviation, (2) stator and rotor friction, (3) rotor end ring break, and (4) poor insulation.
As shown in Figure 4
, the training setup had two stages. The first stage setup was done in one of the lab servers (lab server). Data for all conditions of the motor were collected in this server in CSV format. The dataset comprised of 3-phase current signals. Data preprocessing was applied to the raw current signals to generate the recurrence plots. Recurrence plots’ 2D texture images were stored into the S3 database hosted on the cloud platform (EI-PaaS).
The second stage setup is in the cloud. The Edge Intelligent Platform as a Service (EI-PaaS) has the analytical framework service, where the CNN model was implemented along with the S3 database. While training the CNN model, the images were maintained in the temporary directory structure. Figure 5
illustrates the setup used for the deployment/testing phase of the application. The architecture has a 3-phase induction motor connected to a data acquisition system (DAQ), which sends motor-related data such as current signal values in binary format to an Edge device. The Edge device reads the binary data and converts the data into decimal format and stores them as a CSV file. The data stored in the CSV file are used to generate the relevant RP image. The generated RP texture image is fed to the well-trained CNN model to diagnose and classify the motor condition as one of the four faults or healthy condition.
3.3. Architecture of the Proposed CNN Model
The proposed deep CNN model has a three-stage structure. Each stage representing a feature learning stage with different feature-levels and it includes convolution, activation, and pooling layers.
As shown in Figure 6
, the proposed deep CNN model has three convolutional layers with 32–3 × 3 filter, 64–3 × 3 filter, and 128–3 × 3 filter, respectively. In addition, three max-pooling layers of pooling size 2 × 2 were used. Type of layers, output shape of each layer, along with the number of trainable parameters are listed in Table 1
The activation function Leaky ReLU (Rectified Linear Units) was applied to introduce nonlinearity into each stage, allowing CNN to learn complex models. A specific reason for adding Leaky ReLU was to avoid and attempt to fix the problem of dying ReLUs. It has proven to be more effective than the logistic sigmoid function. However, during the training, ReLU units can die and this could occur when large gradient flows through a ReLU neuron. It causes the weights to update such that the neuron will never activate again on any data point. Leaky ReLU makes an attempt to solve this problem [41
]. Pooling layers were introduced to reduce the resolution of the input image by the process of subsampling and the max-pooling was applied in the proposed model.
At the end of the three stages, the feature maps were flattened into a column vector. The flatted output vector supplied to a feed-forward neural network and backpropagation was employed to every iteration of training. During training, the proposed model was able to distinguish among the dominating and also low-level features in texture images and classify by a fully connected layer for five types of faults. To estimate the parameters of the proposed model, one of the gradient-based optimization (backpropagation algorithm) methods was used. Adam optimizer was used to update the parameters to achieve faster convergence [43
5. Conclusions and Future Work
In this study, we investigate and discuss a novel framework to diagnose the faults in 3-phase induction motors based on recurrence plots and the deep CNN model. The important contributions of this paper are: proposing a method to transform a time-series data signal to 2D texture images (recurrence plots) and applying the proposed deep CNN model to learn the features from the recurrence plots to classify the 2D texture images for the fault diagnosis. The proposed framework is implemented for four types of faults including bearing axis deviation, stator and rotor friction, rotor end ring break, and poor insulation, and achieved a very prominent accuracy of 99.81%. The proposed framework outperforms other traditional and deep learning models due to its ability to learn both high-level and low-level features. The proposed framework demonstrates promising results by considering a single variable as an input feature compared to the rule-based diagnosis methods that require multiple features for fault diagnosis.
The limitations of the proposed methodology are discussed as follows. First, the dataset collected for the experiment is comparatively small and a huge amount of data samples is needed for different load conditions such as no load, half load, or full load. Second, data from the motors with different specifications are needed to extract and learn more generalized features. Therefore, important future work should focus on motors working with different load conditions to collect more data samples, and investigating data to generate more generalized features for CNN model training. Furthermore, future work includes the review of transfer learning to avoid any unnecessary time required to train the model and utilize the model to learn other feature types.