Data-driven Complex Equipment Anomaly Detection Methods

For the past few years, with the wide spread use of big data and the internet of things technologies in the field of manufacture, industrial big data based modern condition monitoring systems are widely applied in the condition monitoring of complex equipment. With the help of a large amount of condition monitoring data collected from complex equipment by condition monitoring systems, effective and efficient analysis of complex equipment's working condition, help to optimize the scheduling of maintenance task, improve operational efficiency and feedback to the design and manufacturing of complex equipment. Therefore, condition monitor data driven complex equipment anomaly detection, are more and more widely focused in the field of manufacture.Taking condition monitoring data collected from complex equipment as the research object, this thesis focuses on one main topic: complex equipment anomaly detection. The main contributions of this thesis are summarized as follows.According to a large amount of complex equipment condition monitoring data,this thesis proposes a condition monitoring space oriented to complex equipment anomaly detection. Based on this logical data space, this thesis applies Cassandra system to design a free table based condition monitoring data physical storage structure, which reflects the hierarchy of condition monitoring data from monitoring equipment to collected values.? For the single equipment's single condition monitoring data based complex equipment anomaly detection problem, this thesis proposes a condition monitoring metrics space, including each equipment condition parameter corresponding condition monitoring series' first metrics and second metrics. According to second metric's thresholds of each equipment condition parameter, it can detect each equipment condition parameter corresponding anomalous condition monitoring series.For the single equipment's multiple condition monitoring data based complex equipment anomaly detection problem, this thesis proposes a condition monitoring series group's latent correlation based anomaly detection method. It first extracts a latent correlation vector from a condition monitoring series group in each equipment work cycle, and then build a latent correlation probabilistic model. According to the probabilistic model's maximum likelihood classifier, it can detect each anomalous condition monitoring series group corresponding anomalous equipment work cycle.? For the multiple equipment's multiple condition monitoring data based complex equipment anomaly detection problem, this thesis proposes a multiple equipment cooperative anomaly detection method. It extracts equipment's working condition matrix from time dimension and region dimension, combining equipment historical anomaly information in the framework of collaborative filtering, intended to use matrix factorization technology to detect the quantity of anomaly equipment for each region.