Self-adaptive collection method, system and device for equipment health stage detection and medium

Abstract

The invention discloses a self-adaptive collection method, system and device for equipment health stage detection and a medium, and the method comprises the following steps: obtaining a data stream of equipment, carrying out the state extraction of the data stream based on a sliding window and reservoir sampling, and obtaining a concept representation; performing adaptive health stage detection on the data stream according to the concept representation to obtain health stage data; and performing fusion processing on the health stage data to increase training data of each health stage. According to the method, health stage division and multi-data-stream stage fusion processing are realized based on state extraction of sliding window and reservoir sampling and adaptive health stage detection based on concept drift detection, and when multiple groups of health data streams exist, for example, data acquired by multiple devices are subjected to stage division respectively, multiple groups of stage data are fused into a single group of stage data, the data volume of each health stage is increased, the training effect is improved, and the method can be widely applied to the fields of intelligent manufacturing and data mining.

Description

technical field [0001] The invention relates to the fields of intelligent manufacturing and data mining, in particular to an adaptive collection method, system, device and medium for equipment health stage detection. Background technique [0002] In the context of intelligent manufacturing, a large amount of equipment data is collected by sensors at all times and presented in the form of data streams. Health data in the form of data streams has the characteristics of fast speed, large capacity, difficult feature analysis, high timing correlation, and fuzzy distribution change points. When using traditional machine learning frameworks, it is impossible to deal with the phenomenon of concept drift in the data flow, that is, the distribution of data is not stable, but changes over time. When concept drift occurs, the data distribution changes, which makes the performance of the model decline. In the equipment health prediction task, there is an obvious concept drift phenomenon....

AI Technical Summary

Problems solved by technology

[0003] The current device health detection method does not consider the stage division of health data, but directly processes the entire data set. Although incremental learning, pruning and other methods are adopted, there are still high training pressure, low model accuracy, and low prediction speed. Slow and other limitations, did not fully consider the equipment health as the operating time increases and the working environment is different, the equipment health presents stages, and there are differences in different equipment health stages

Current methods use the entire dataset for training without partitioning, which suffers from difficult feature extraction and poor interpretability

[0004] Active detection and adaptation is a relatively popular method of concept drift processing. Among them, the method of monitoring model performance is model-independent and has better general-purpose capabilities. Typical methods include DDM, EDDM, ADWIN, HDDM, McDDM and other algorithms. However, the above algorithm only considers the sensitivity of drift detection, and does not fully consider the guiding role of drift detection to drift adaptation. In the alarm mechanism, the coefficient of the detection mechanism is simply adjusted, and it is impossible to balance the delay and the size of the retraining set to achieve a better performance. Good performance, which makes the division of health stages offset, resulting in too many stages of division, and increased pressure on model training.

Images