The remaining useful life (RUL) prediction of turbofan engines is beneficial to safe operation and maintenance. Turbofan engines generally require multi-sensors to monitor the operational state, but current mainstream RUL prediction models tend to ignore the spatial correlations between sensors. Although few research has used graph neural networks to capture such correlations, there are some limitations in the way the graph structure is determined. To this end, we introduce sensor embeddings and propose a novel RUL prediction model based on a Convolution-Graph Attention Network (ConvGAT). Firstly, the time-series of every sensor is segmented using a sliding time window and the data within that window is taken as the model input. Next, every sensor is treated as a node in the graph. The features of sensor data, which serve as the initial features of the corresponding sensor node, are extracted using a convolutional layer. Then, a learnable embedding vector is specially introduced for every sensor and the spatial correlations between sensors are captured using a graph attention network. Particularly, the sensor embeddings play an important role in the graph structure learning as well as the graph attention mechanism. Finally, the sensor node features output by the GAT are fused with the corresponding sensor embeddings and passed through a fully connected layer to predict the RUL of turbofan engines. The ConvGAT-based RUL prediction model’s performance is evaluated using a benchmark dataset regarding turbofan engines, i.e., the C-MAPSS dataset. The experimental results indicate the superior performance of the proposed model.

Vibration signals acquired by accelerometers are commonly used in the field of bearing fault diagnosis. The measured vibration signals contain a series of impulse responses when localized bearing faults occur. These impulses are often viewed as fault features and exhibit nonlocal self-similarity in terms of morphological characteristics. However, the impulses are generally contaminated by background noise. In this paper, a group-based K-SVD denoising algorithm, which exploits the nonlocal self-similarity property, is presented to extract bearing fault features. In contrast to regular K-SVD denoising, this denoising algorithm forms groups by clustering similar signal segments and regards each group rather than each segment as the basic unit of sparse representation. To group similar segments, we further propose a particular grouping method that exploits bearing prior knowledge (i.e., possible impulse spacings). The new grouping method could form groups in a targeted manner, thereby achieving more accurate grouping. Group-based K-SVD denoising is able to extract fault features better as similar segments collaboratively provide a large amount of shared feature information, especially in the presence of strong noise (such as signal-to-noise ratios lower than -10 dB). The numerical and experimental results demonstrate the superior performance of group-based K-SVD denoising over regular K-SVD denoising and other two benchmark algorithms (i.e., wavelet shrinkage and spectral kurtosis).

Remaining useful life (RUL) prediction is a key technique for the condition maintenance of mechanical equipment. Deep learning networks can fully exploit the degradation information embedded in multi-sensor data and have achieved great success in RUL prediction tasks. However, most of such methods only consider the temporal dependencies in sensor time-series, ignoring the spatial dependencies between sensors as well as the importance of sensors and time-steps simultaneously. To this end, an RUL prediction method (referred to as AGATT), integrating attention mechanisms, graph attention networks, and Transformer, is proposed. The AGATT uses two parallel channels (i.e., a spatial channel and a temporal channel), each of which comprises an attention mechanism and a graph attention network. The spatial channel identifies important sensors and learns spatial dependencies, while the temporal one identifies important time-steps and learns temporal dependencies. The temporal and spatial features extracted by the two channels are fused to form a sequence that is fed into the Transformer to predict RUL. The Transformer’s prediction performance is enhanced as its input contains both temporal and spatial information at each time-step. A well-known turbofan engine dataset (i.e., the C-MAPSS dataset), which includes four prediction tasks, is used to verify the RUL prediction performance of the AGATT. The experimental results show that the AGATT outperforms state-of-the-art RUL prediction methods in the three prediction tasks and achieves comparable prediction results in another one. In addition, we visualize the degradation feature vectors extracted by the AGATT as well as the importance of different sensors and time-steps.

Identifying the health state of machinery using sensor information is one of the key tasks for intelligent operation and maintenance of machinery. In recent years, deep learning models have been widely used in this field. However, these models usually lack interpretability. To this end, we use the algorithm unrolling technique to unroll the convolutional version of Iterative Shrinkage-Thresholding Algorithm (CISTA) into a neural network called CISTA-Net. The CISTA-Net’ s structure is determined by the iterative algorithm and thus has a clear theoretical basis. By adding a classification module and a reconstruction module in parallel to its output end, the CISTA-Net is interpretable in terms of both its structure and output. Interpretability analysis through the well-known bearing dataset from the Case Western Reserve University (CWRU) show that the reconstruction module can enhance the interpretability. Besides, a simple residual block can be optionally added to the input end of the CISTA-Net for feature pre-extraction, resulting in an alternative CISTA-Net called ResCISTA-Net. The ResCISTA-Net is compared with some conventional and interpretable fault diagnosis models on another two mechanical fault diagnosis tasks. Experimental results suggest that: (1) The residual block can not only improve diagnostic accuracy but also prevent performance degradation for some tasks; (2) The reconstruction module can also improve diagnostic accuracy; (3) The ResCISTA-Net performs excellently in terms of diagnostic accuracy, training sample size requirement, and noise robustness.