One-class Learning Based Novelty Detection And Its Application In The Heavy Equipment Condition Recognition

Condition monitoring and fault diagnosis are very important for the maintenance and safe operation of equipment. Currently, the most intelligent fault diagnosis system must need historical data labeled by existing categories to establish classification model. However, in the actual industry, it is very difficult to get complete monitoring data under different condition, especially for expensive heavy equipment which have high reliability and less fault data. There often exists the problems of lack of performance data and a priori knowledge, such as insufficient normal samples, scarce fault samples and incomplete failure modes. To solve these problems, a novelty detection method based on one-class learning is proposed. This method attempts to complete the novelty detection using the single class of normal samples and gradually establish a complete diagnosis system with the data obtained in the process. There are three key issues faced during the novelty detection, including lack of feature selection techniques for specific subjects, lack of accurate and robust detection algorithms and lack of online diagnosis system to cope with unknown status. To solve these issues, the corresponding solutions are deeply studied, and three key technologies including the novelty detection algorithm based on feature evaluation, the novelty detection model based on hybrid ensemble and the novelty detection system based on evolutionary learning are developed. Then the overall processes and implementation strategies of the novelty detection integrated system based on one-class evolutionary learning are proposed. And the practical verification is carried out in the condition monitoring for the key components of the heavy roll grinder as study object. The main research work includes:(1) In order to carry out the effective feature selection under the situation of sample class missing, a feature evaluation technology based on support vector data description is proposed, then the novelty detection algorithm is designed based on this feature evaluation technology. The results of bearing condition test verification show that the feature evaluation technology will select sensitive features which can effectively distinguish between different states, remove those insensitive features which are less effective or even affect accuracy, and ultimately improve the novelty detection performance.(2) To improve the accuracy, stability and usability of one single detection algorithm, the hybrid novelty detection model is proposed. The detection model with varieties of novelty detection algorithms integrated is built based on ensemble learning. The corresponding hybrid algorithms and integration guidelines are designed for the purpose of novelty detection and the basic implementation process of this model is introduced. The results of bearing condition test verification show that the detection performance of the hybrid detector based on weighted integration is greatly improved when the low performance single detector uses poor sensitivity features.(3) To solve the problem of unknown status detection with the acquired data and status categories increased, the novelty detection system is proposed based on evolutionary learning. The evolutionary intelligent diagnosis system with novelty detection and supervised learning classifier is designed. The novelty detection method for one-class classification problem is extended for multi-class classification problem, so that even there are only normal samples, this system can still evolve through self-learning and gradually establish the diagnostic capabilities of the unknown status. The results of bearing condition novelty detection and multi-class classification show that this system with the one-class detector constructed can accomplish the novelty detection for unknown fault and its own upgrade, and complete the multi-class fault recognition using the auxiliary arbitration function of the supervised learning classifier if necessary.(4) The condition monitoring and diagnosis system is designed for the key components of a heavy roll grinder. This system is composed by three levels including filed monitoring level, remote service level and multi-application level. The implementation strategy of the novelty detection integrated diagnosis system is introduced and the specific data acquisition scheme is designed for the key component monitoring. During monitoring application, a misalignment between the apexes of the header stock and the tail stock is successfully detected using only normal samples data.