In order to overcome the shortcomings that regular recognition methods do not extract the pivotal difference information of incipient faults with overlapped features, two optimized methods are used to enhance the learning ability of extractor and promote the recognition accuracy of classifier. Firstly, based on one-dimensional neural network (1D-CNN) and its variant temporal convolutional network (TCN), an improved network structure called temporal convolutional block-convolutional neural network is proposed. It learns historical information by causal convolution and general trend information by non-causal convolution. Secondly, a new loss item named direct discriminant is proposed to optimize the loss function based on the softmax classifier. It makes classifier inputs have better separability. Compared with existing recognition methods such as variational modal decomposition-stochastic configuration network (VMD-SCN), manifold regularized stochastic configuration network (MRSCN), kernel principal component analysis-deep belief network (KPCA-DBN) and TCN, the superior recognition of the proposed method is illustrated through a numerical example and TE chemical industrial process simulation.

Incipient faults often occur in the early stages of mature faults. Because of its small amplitude and energy, detection and diagnosis are more difficult. From the point of view of extracting frequency characteristics, to study the accurate recognition method of incipient faults can effectively improve the reliability and safety of system, which is of great significance to actual industrial process. Since the fluctuation of incipient faults and normal signals is not obvious from the amplitude features of time domain, it is difficult to recognize faults from amplitude by conventional method. But everything vibrates. So, in order to recognize faults more easily from the perspective of frequency features, a one-dimensional convolution neural network with extracting frequency features is proposed. Based on cross entropy loss function, this method adds an optimization index to guide model to extract frequency features. By constraining the minimum frequency bandwidth of node output value, the central frequency of feature extracted is closer to the central frequency of fault. Incipient faults are identified through more frequency features instead of time domain features. This method improves recognition accuracy and reduces the amount of neural network parameters. Compared with existing recognition methods such as one-dimensional convolutional neural network, temporal convolutional network and so on, the superior recognition of the proposed method is illustrated through Case Western Reserve University and Paderborn University bearing dataset. Experiments show that this method can accurately identify incipient faults by non-overlapping frequency features, and has fewer parameters and simpler structure under the same recognition accuracy.

The amplitudes of incipient fault signals are similar to health state signals, which increases the difficulty of incipient fault diagnosis. Multi-scale reverse dispersion entropy (MRDE) only considers difference information with low frequency range, which omits relatively obvious fault features with a higher frequency band. It decreases recognition accuracy. To defeat the shortcoming with MRDE and extract the obvious fault features of incipient faults simultaneously, an improved entropy named hierarchical multi-scale reverse dispersion entropy (HMRDE) is proposed to treat incipient fault data. Firstly, the signal is decomposed hierarchically by using the filter smoothing operator and average backward difference operator to obtain hierarchical nodes. The smoothing operator calculates the mean sample value and the average backward difference operator calculates the average deviation of sample values. The more layers, the higher the utilization rate of filter smoothing operator and average backward difference operator. Hierarchical nodes are obtained by these operators, and they can reflect the difference features in different frequency domains. Then, this difference feature is reflected with MRDE values of some hierarchical nodes more obviously. Finally, a variety of classifiers are selected to test the separability of incipient fault signals treated with HMRDE. Furthermore, the recognition accuracy of these classifiers illustrates that HMRDE can effectively deal with the problem that incipient fault signals cannot be easily recognized due to a similar amplitude dynamic.