Equipment fault prediction method based on DCFM model

Abstract

The invention discloses an equipment fault prediction method based on a DCFM model, and the method achieves the deeper and more comprehensive extraction of equipment information through building a DCFM model structure which can fully explore the high-order and low-order feature information of equipment, improves the classification effect of an equipment fault prediction model, and meanwhile, the model performs automatic cross multiplication on the characteristics of the equipment, so that the disadvantage that a traditional machine learning method depends on characteristic engineering can be made up to a certain extent. In addition, an FM module of the DCFM model can realize fine learning of second-order characteristic parameters of sparse data based on implicit vectors; the learning efficiency of the FM module and the CrossNetwork module is linear level complexity, the training speed of the model is improved to a certain extent, rapid construction of the model is facilitated, and compared with other complex machine learning and deep learning models, the DCFM model can meet the requirements of online prediction of equipment components for time response and accuracy.

Description

technical field [0001] The invention relates to the fields of data mining and electromechanical failure prediction, in particular to a DCFM model-based equipment failure prediction method. Background technique [0002] Most of the core ideas of existing equipment fault state prediction methods are based on signal processing or analytical models, which cannot meet the real-time and accuracy required for fault prediction under large-scale data conditions; the prediction schemes disclosed in some documents Although machine learning is involved, these solutions are relatively rough in mining the original features of the equipment, and it is difficult to achieve satisfactory prediction results only by using the original features for training, and most algorithms rely on artificially designed features. [0003] LR (Logistic Regression), as a commonly used classifier in the industry, has the advantages of simple form, strong interpretability, and easy implementation of parallelism,...

AI Technical Summary

Problems solved by technology

FFM (Field-aware Factorization Machines) introduces a feature domain based on the FM model to make the model learning more refined, but the problem is that the computational complexity of the model increases, which affects the learning efficiency of the model.

The DNN model has the advantage of exploring high-order features, but DNN ignores the important information carried by low-order features, which will also affect the prediction effect

Images