Research On Industrial Fault Diagnosis And Forecasting Methods Based On Limited Supervised Learning

With the rapid development of the modern smart manufacturing industry,industrial equipment has become increasingly complex and overloaded.Once the equipment failure occurs,it will cause enormous economic losses and threaten the safety of operators.To reduce the risk of equipment failure,intelligent fault diagnosis and forecasting technology has been widely used and plays an important role in monitoring the health status of equipment and maintaining the safe operation of the system.In recent years,with the help of deep learning which owns strong feature representation capability,researchers have made a significant technological breakthrough in the performance improvement of intelligent diagnosis and forecasting models.However,most of the existing methods depend heavily on a large number of labeled fault data,which is usually expensive to obtain,thus bringing a new challenge to the practical application of the algorithms.This paper focuses on this challenge and conducts in-depth research on the problems of fault diagnosis and remaining useful life forecasting under the scenario of insufficient labeled data,using domain adaptation and self-supervised learning techniques in limited supervised learning.The main works of this paper are as follows:(1)A new method based on improved convolution neural network and domain adversarial learning is presented to solve the under-adapting problem caused by insufficient feature extraction capability of cross-domain fault diagnosis models.This method uses a convolution neural network to extract features for sub-sequences of signals and uses feature fusion technologies to explicitly capture the correlation information among sub-sequences,thereby enhancing the feature extraction ability of the model.In addition,the method combines domain adversarial training and maximum mean discrepancy to align the distribution differences between domains.The experimental results show that the proposed method can capture more domain-invariant features for two domains and has obvious advantages in cross-domain diagnostic performance.(2)A new method based on self-supervised domain adaptation is presented to solve the under-adapting problem caused by the insufficient domain alignment capability of cross-domain fault diagnosis models.This method mixes source and target domain samples for temporal relation sampling to generate a relation dataset and performs a self-supervised relation classification auxiliary task to enable the feature extractor to capture the consistent semantic representation of the two domains,thereby enhancing the domain alignment capability of the model.The experimental results show that the proposed method can be directly applied to the existing cross-domain fault diagnosis methods,and further improves the diagnostic performance.(3)To improve the forecasting accuracy of remaining useful life in the case of insufficient supervisory signals,a semi-supervised remaining useful life forecasting method based on autoencoder and deep metric learning is proposed.Two label-independent ancillary tasks are designed to capture common knowledge on unlabeled data.Specifically,the feature distance measurement assistant task enables samples with similar residual life in the feature space to obtain closer feature representations.The signal reconstruction assistant task enables the information of the original signal in the feature space to be preserved as much as possible.The experimental results show that the proposed method can show excellent forecasting performance in unlabeled datasets.