An Online Fault Diagnosis Method Based on Massive Operating Data

Abstract

The invention provides an online fault diagnosis method based on massive operating data. First, determine the sample parameters of the fault diagnosis model, obtain steady-state sample data, and eliminate redundant sample data; then use the optimal sample to train the fault diagnosis model; finally, in the real-time diagnosis process In the process, the monitoring data is standardized, and the unsteady data is eliminated from the processed data stream, and the trained fault diagnosis model is used to diagnose the sensor fault on the steady data, and the unsteady data is not used as the basis for judging the sensor fault , if there is fault data in the steady-state data sample, the fault diagnosis model will send out an alarm and handle the fault. The present invention performs sample optimization in the process of training the diagnosis model, performs data filtering in the process of real-time diagnosis, reduces the misdiagnosis rate of the fault diagnosis model, improves the reliability, diagnosis rate and fault tolerance of the fault diagnosis model, and further improves the Online monitoring accuracy.

Description

technical field [0001] The invention relates to the technical field of online fault diagnosis, in particular to an online fault diagnosis method based on massive operating data. Background technique [0002] Machine learning (Machine Learning) is the means and mechanism of acquiring knowledge from known sample data or information through mining, induction, deduction, analogy, etc. It is another important research field of artificial intelligence application after the expert system, and has caused extensive attention. The purpose of machine learning is to learn the training samples given in advance according to a certain method or algorithm designed, and then obtain an estimate of the dependence between the input and output of a certain system, and make the estimate better for the unknown output. To make predictions or judgments about their nature as accurately as possible. [0003] Fault diagnosis is not only an important research content in the field of massive data minin...

AI Technical Summary

Problems solved by technology

If a parameter in the sample set is poorly correlated with other parameter data, the established model is limited to fitting the relationship between limited training samples, has poor generalization ability, and cannot meet the requirements of real-time diagnosis. The environment is poor or the current technology cannot measure accurately, resulting in large fluctuations in measured values ​​or large deviations from the true value. Even if there is a certain mechanism relationship with other parameters, it cannot be identified from the process measurement data. If the historical measurement of these parameters Participation of data in model building may reduce the accuracy of diagnostic models, and diagnostic models based on process data generally determine model parameters through mechanism relationships. However, there are differences among various sensors. Satisfy the measurement requirements, so that there is a large error between the measurement data and the real value, resulting in the process measurement data of these parameters losing their inherent mechanism relationship, so it is necessary to conduct data inspection on the relationship between parameters; Unsteady data and redundant data will also destroy the accuracy of the diagnostic model and reduce the learning ability and generalization ability of the model. There is a large amount of unsteady data in the original historical data set. These data cannot accurately reflect the mechanism relationship between parameters, and there is a large amount of redundant data in the steady-state sample data. On the one hand, it increases the amount of calculation in the model training process and reduces the calculation speed of the network model. On the other hand, it may cause sample loss. data imbalance

In addition, there are many uncertain factors in the non-stationary data in the real-time diagnosis process, it is difficult to ensure that the relationship between variables conforms to a strict mathematical model, and a large number of misdiagnoses will occur, which seriously affects the effectiveness of the model. Selecting appropriate samples from data is of great significance to improve the practicability of machine learning models

Images