A novel convolutional neural network based fault recognition method via image fusion of multi-vibration-signals

Highlights

• A novel feature extraction method is proposed by converting multi-vibration-signals to images.

• A novel recognition method is proposed with multi-vibration-signals fusion and improved deep CNN.

• A new CNN model is optimized by bottleneck layer with efficient convolution kernel to fuse data at same time.

Abstract

This paper proposed a novel fault recognition method for rotating machinery on the basis of multi-sensor data fusion and bottleneck layer optimized convolutional neural network (MB-CNN). A conversion method converting vibration signals from multiple sensors to images is proposed that can integrate information to get richer features than vibration signals from single sensor. By this method feature maps of different fault types can be obtained without tedious parameter adjustments. Based on the feature maps from multi-sensor data, a corresponding novel convolutional neural network is also constructed. The constructed network performs the bottleneck layers with an increased number of input features to avoid information lost. The data at the same time node can be fused by the convolutional kernels of which the size matches the number of sensors. Practical examples of diagnosis for the wind power test rig and the centrifugal pump test rig are given in order to verify the effectiveness of the proposed approaches, and prediction accuracy of 99.47% and 97.32% is obtained respectively. Otherwise, the performances of other conventional methods such as deep belief network (DBN), support vector machine (SVM) and artificial neural network (ANN) are evaluated for contrast with the proposed method. As shown in the results, the novel convolutional neural network obtains higher recognition accuracy and faster convergence speed.

Introduction

Bearings and gears have been widely used as vital parts in intelligent manufacturing. Once the transmission components in the rotating machine fail, the resulting vibrations will affect the overall process and safety in industrial manufacturing, which makes fault diagnosis essential in the industry. Many experts and scholars have made efforts to the failure mechanisms and fault diagnosis methods of bearing and gear based on vibration signals, which leads to the development of equipment health monitoring [[1], [2], [3]]. In general, the fault diagnosis system consists of two key steps, data processing (feature extraction) [[4], [5], [6], [7]] and fault recognition [[8], [9], [10], [11]]. Essentially, fault diagnosis is pattern recognition of the rotating machine states. Some conventional diagnostic methods are performed by calculating parameters in the time domain, frequency domain or time-frequency domain of the signals.

With the development of industrial Internet of Things and data-driven technology, machine health monitoring has become inclined to use big data for fault diagnosis. In modern industry, the application of data-driven methods in controlling and monitoring has become a hot topic of research [12,13]. Traditional data-driven methods usually include the following key components: manually designed features, feature extraction and model training [14]. Liu et al. proposed a short-time matching method and used oscillatory characters as inputs of support vector machine (SVM) for bearing fault diagnosis [15]. Hong et al. proposed a probability-based classification model and used SVM classifiers for real-time fault detection of power transformers [16]. Partially-linearized neural network can sequentially identify fault types for a rolling bearing based on symptom parameters with high accuracy [17]. Zhang et al. proposed a naive Bayes and selective support vector machine bearing fault diagnosis method based on enhanced independence of data [18]. Chine et al. proposed fault diagnostic technique for photovoltaic systems based on Artificial Neural Networks (ANN) [19]. Liao proposed a variable neural network based on regrouping particle swarm optimization, which can effectively optimize the network structure and diagnosis the gearbox faults [20]. Pandya et al. applied logistic regression technique for the energy and kurtosis parameters extracted by wavelet packet in fault diagnosis of rolling element bearings and achieved better results than ANN and SVM [21]. However, papers have shown that traditional machine learning above is difficult to extract and recognize raw data [22]. Complex feature extraction processes require specialized knowledge strongly and influence the final diagnosis result significantly.

Recently, deep learning has been widely used in computer vision [[23], [24], [25]], acoustic recognition [26] and medical images [27,28]. Hinton et al. proposed hierarchical neural network model possessing strong pattern learning ability, for example deep belief network (DBN), which accelerated the development of deep learning [29,30]. Deep learning can fully explore the feature information in big data, and the huge amount of data is able to offset the complexity increase behind deep learning and improve its generalization capability [31]. Fault diagnosis based on deep learning has gradually become a hot topic. Various researchers have demonstrated the success of deep learning models such as CNN, Auto-encoder, Deep Boltzmann Machines (DBM) and Recurrent Neural Networks (RNN) in machine health monitoring applications [[32], [33], [34]]. The above papers show that deep learning can effectively extract the hidden features in the raw data. Deep Convolutional Neural Network (CNN) is a representative graphic recognition network in deep learning. Ince et al. proposed a fast and accurate motor condition monitoring and early fault-detection system based on one-dimensional CNN [35]. Liu et al. proposed feature extraction and fault diagnosis method for planetary gears based on variational mode decomposition and CNN [36]. Ding et al. proposed a multi-scale feature mining method for energy fluctuation of spindle bearings based on wavelet packet energy images and CNN [37]. Chen proposed a time-domain and frequency-domain of multi-sensor data feature fusion technique through multiple two-layer sparse self-encoder neural network, and identify machine operating by deep confidence network [38]. However, the above papers all rely on expert experience to construct features, and some of which are limited to analyzing data from one sensor.

This paper proposed a novel fault recognition method for rotating machinery based on multi-vibration-signals fusion and bottleneck layer optimized convolutional neural network (MB-CNN). Firstly, a method for converting multi-vibration-signals into two-dimensional images is proposed, which can obtain the feature maps without tedious parameter adjustments efficiently. The influence from the experience of the experts on the images is eliminated. Secondly, an improved CNN is proposed by using bottleneck layer and special convolutional kernel to further extract features and fuse the multi-sensor data. Thirdly, as is shown in the results of diagnosis of experimental data from wind power test rig, MB-CNN based fault diagnosis method has higher prediction accuracy than the single-channel data diagnosis methods and other traditional methods mentioned above.

The rest of this paper is organized as following. Section 2 introduces deep convolutional neural network. Section 3 presents the diagnosis methodologies, including the conversion method converting multi-vibration-signals to images and the improved CNN structure for data fusion. Section 4 presents two types of experimental verification and discussion of the diagnostic methods presented in this paper. The conclusions will be introduced in Section 5.