The planetary gearbox is a key component in mechanical transmission systems and has been widely used in wind turbines, helicopters and other heavy machineries [1
]. The wide range of gear ratios, small room in power transmission line, and high transmission efficiency are the most significant advantages of planetary gearboxes [2
]. Planetary gearboxes generally operate in tough working environments, which makes them suffer from a variety of failures and damages [1
] causing unwanted downtime, economic losses, and even human casualties [4
]. Therefore, the fault diagnosis of planetary gearboxes is necessary to guarantee a safe and efficient operation of mechanical transmission systems.
Health conditions of planetary gearboxes can be reflected by various kinds of measurements, including vibration signal, acoustic signal, driven motor current, shaft speed, oil debris, etc. Different measurements have different drawbacks and are sensitive to different types of damage modes or operation conditions. Thus, combining and analyzing these measurements together should be an appropriate approach to detect various types of faults of complex systems. This approach, named multi-sensor data fusion, can achieve more accurate and reliable result than approaches with a single sensor [5
Nevertheless, multi-sensor data fusion for fault diagnosis suffers from two challenging problems [6
]. (1) One is the feature extraction of multi-sensory data. Generally, conventional fault diagnosis includes three steps [8
]: signal acquisition, feature extraction, and fault classification. In the feature extraction step, fault sensitive features are extracted and selected from raw data through signal processing technologies and data analysis strategies, such as Fourier spectral analysis, wavelet transformation (WT), and principal component analysis (PCA). However, the multiple types of sensory data of multi-sensor data fusion may cause a number of issues that make the feature extraction with multi-sensor much more difficult than with a single sensor. These issues [6
] include more imprecision and uncertainties in measurements, various noises, more conflicting or correlating data, higher data dimensions, etc. Extracting features from these kinds of data will be a very challenging and cruel task. At the same time, multiple types of sensory data also increases the difficulties and consumes much time to choose an optimal handcraft feature or manual feature extraction method, even though to date, the optimal handcraft feature or feature extraction method for a specific type of sensory data still remains unanswered [8
]; (2) The other challenge of the multi-sensor data fusion is the selection of different fusion levels. Similarly, different fusion levels have their own advantages and disadvantages, and the suitable one for different fault diagnosis tasks is usually different [5
]. Selecting an optimal fusion level for a specific fault diagnosis task always requires domain expertise, prior knowledge, and human labor.
Deep neural networks (DNN), also known as deep learning, have been attracting increasing attention from researchers from various fields in recent years [10
]. The key advantage of DNN is the feature learning ability [11
], which can automatically discover an intricate structure and learn useful features from raw data layer by layer. A number of studies [11
] have shown that DNN can fuse input raw data and extract basic information from it in its lower layers, fuse the basic information into higher representation information and decisions in its middle layers, and further fuse these decisions and information in its higher layers to form the final classification result. It can be seen that DNN itself is a fusion structure [11
], which fuses the feature extraction, feature selection, data-level fusion, feature-level fusion, decision-level fusion, and classification into one single learning body. For this reason, a DNN-based low level fusion, e.g., data-level fusion, can not only learn features from fused raw data automatically, but also fuse these data, features and decisions adaptively through its deep-layered structure. DNN might be an ideal model to fuse multi-sensory data and detect faults of a mechanical system. However, although some applications [14
] of DNN in feature learning and fault diagnosis with a single sensor have been found in recent years, no study, to the best of our knowledge, has investigated the effectiveness of DNN-based feature learning and the adaptive level fusion method for fault diagnosis. It is attractive and meaningful to investigate this adaptive fusion method, which can learn features from raw data automatically, and select and combine fusion levels adaptively.
Deep convolutional neural networks (DCNNs), as one of the main types of DNN models, have been successfully used in mechanical fault diagnosis with automatic feature learning from single sensory data [9
]. Benefitting from several unique structures, DCNN can achieve better results with less training time than standard neural networks. Firstly, DCNN has a large set of filter kernels in convolutional layers, which can capture representative information and patterns from raw data. Stacking these convolutional layers can further fuse information and build complex patterns; Secondly, DCNN is an unfully-connected network, where each filter shares the same weights. This structure can reduce both the training time and complication of the model. In addition, the pooling layer of DCNN further reduces the revolution of the input data as well as the training time, and improves the robustness of the extracted patterns (a detailed introduction to the DCNN model is presented in Section 2
). Thus, DCNN should have great potential in processing the multi-sensory data of a mechanical system, which usually contains rich information in the raw data and is sensitive to training time as well as model size.
Aiming to address the two problems of multi-sensor data fusion mentioned above, this paper proposes an adaptive data fusion method based on DCNN and applies it to detect the health conditions of a planetary gearbox. Different from conventional methods, the proposed method is able to (1) extract features from raw data automatically and (2) optimize a combination of different fusion levels adaptively for any specific fault diagnosis task with less dependence on expert knowledge or human labor.
The rest of the paper is organized as follows. In Section 2
, the typical architecture of DCNN and an adaptive training method are briefly described. Section 3
illustrates the procedures of the proposed method, the design of the DCNN model, and several comparative methods introduced to further analyze the performance of the proposed method. In Section 4
, an experimental system of a planetary gearbox test rig is used to validate the effectiveness of the proposed method. Finally, the conclusions are drawn in Section 5
5. Conclusions and Future Work
This paper presents an adaptive data fusion method based on DCNN to detect the health conditions of planetary gearboxes. The processes of data-level fusion, feature-level fusion, decision-level fusion, feature learning, and fault diagnosis are all fused into one DCNN model adaptively. The proposed method can learn features from raw data, and fuse data, features, and decisions adaptively through the deep-layered structure of DCNN with fewer requirements of expert knowledge or human labor for feature extraction and fusion-level selection. The performance of the proposed method is evaluated through the experiment of the planetary gearbox fault test rig. As comparisons, feature-level fusion, decision-level fusion, handcraft features, single sensory data, and two traditional intelligent models, BPNN and SVM, are also tested in the experiment. The comparative results of the experiment verify the effectiveness of the proposed method, which achieves the best testing accuracy among all the comparative methods in the experiment.
Our future work will focus on testing the DCNN model-based feature learning and data fusion approaches on more mechanical objects, fault modes, operation conditions, and sensor types, which can further confirm the effectiveness of approaches and help us to find out other useful application guidance. Moreover, due to the large number of parameters of deep learning models, manual parameter optimization often takes many trials-and-errors to find the best one, and conventional automatic searching methods are usually very time-consuming and easily converge into a local optimum. It is meaningful to investigate more effective and faster approaches to optimize the parameters automatically. Finally, combinations of different deep learning architectures should improve the effect of fault diagnosis. Adding recurrent architecture may make the model suitable to predict future fault conditions, and combining with auto-encoder architecture may improve the feature learning ability to capture more complex features.