In recent years, transfer learning has been widely applied in fault diagnosis for solving the problem of inconsistent distribution of the original training dataset and the online-collecting testing dataset. In particular, the domain adaptation method can solve the problem of the unlabeled testing dataset in transfer learning. Moreover, Convolutional Neural Network (CNN) is the most widely used network among existing domain adaptation approaches due to its powerful feature extraction capability. However, network designing is too empirical, and there is no network designing principle from the frequency domain. In this paper, we propose a unified convolutional neural network architecture from a frequency domain perspective for a domain adaptation named Frequency-domain Fusing Convolutional Neural Network (FFCNN). The method of FFCNN contains two parts, frequency-domain fusing layer and feature extractor. The frequency-domain fusing layer uses convolution operations to filter signals at different frequency bands and combines them into new input signals. These signals are input to the feature extractor to extract features and make domain adaptation. We apply FFCNN for three domain adaptation methods, and the diagnosis accuracy is improved compared to the typical CNN.

Aiming at predicting the purity of the extract and raffinate components in the simulated moving bed (SMB) chromatographic separation process, a soft-sensor modeling method was proposed by adoptig the hybrid learning algorithm based on an improved particle swarm optimization (PSO) algorithm and the least means squares (LMS) method to optimize the adaptive neural fuzzy inference system (ANFIS) parameters. The hybrid learning algorithm includes a premise parameter learning phase and a conclusion parameter learning phase. In the premise parameter learning stage, the input data space division of the SMB chromatographic separation process and the initialization of the premise parameters are realized based on the fuzzy C-means (FCM) clustering algorithm. Then, the improved PSO algorithm is used to calculate the excitation intensity and normalized excitation intensity of all the rules for each individual in the population. In the conclusion parameter learning phase, these linear parameters are identified by the LMS method. In order to improve population diversity and convergence accuracy, the population evolution rate function was defined. According to the relationship between population diversity, population fitness function and particle position change, a new adaptive population evolution particle swarm optimization (NAPEPSO) algorithm was proposed. The inertia weight is adaptively adjusted according to the evolution of the population and the change of the particle position, thereby improving the diversity of the particle swarm and the ability of the algorithm to jump out of the local optimal solution. The simulation results show that the proposed soft-sensor model can effectively predict the key economic and technical indicators of the SMB chromatographic separation process so as to meet the real-time and efficient operation of the SMB chromatographic separation process.

Blending multiple convolutional kernels is proved advantageous in neural architecture design. However, current two-stage neural architecture search methods are mainly limited to single-path search spaces. How to efficiently search models of multi-path structures remains a difficult problem. In this paper, we are motivated to train a one-shot multi-path supernet to accurately evaluate the candidate architectures. Specifically, we discover that in the studied search spaces, feature vectors summed from multiple paths are nearly multiples of those from a single path. Such disparity perturbs the supernet training and its ranking ability. Therefore, we propose a novel mechanism called Shadow Batch Normalization (SBN) to regularize the disparate feature statistics. Extensive experiments prove that SBNs are capable of stabilizing the optimization and improving ranking performance. We call our unified multi-path one-shot approach as MixPath, which generates a series of models that achieve state-of-the-art results on ImageNet.

To discover powerful yet compact models is an important goal of neural architecture search. Previous two-stage one-shot approaches are limited by search space with a fixed depth. It seems handy to include an additional skip connection in the search space to make depths variable. However, it creates a large range of perturbation during supernet training and it has difficulty giving a confident ranking for subnetworks. In this paper, we discover that skip connections bring about significant feature inconsistency compared with other operations, which potentially degrades the supernet performance. Based on this observation, we tackle the problem by imposing an equivariant learnable stabilizer to homogenize such disparities. Experiments show that our proposed stabilizer helps to improve the supernet's convergence as well as ranking performance. With an evolutionary search backend that incorporates the stabilized supernet as an evaluator, we derive a family of state-of-the-art architectures, the SCARLET series of several depths, especially SCARLET-A obtains 76.9% top-1 accuracy on ImageNet. Code is available at https://github.com/xiaomi-automl/ScarletNAS.