Gas turbine abnormal state monitoring method based on frequency spectrum reconstruction errors

Abstract

The invention relates to a gas turbine abnormal state monitoring method based on frequency spectrum reconstruction errors, in particular to a multi-working-condition intelligent detection method for abnormal states generated by mechanical faults such as ground gas turbine and aviation gas turbine blades, bearings, rub-impact, non-concentric and the like. The method is characterized in that an improved depth variational auto-encoder is used for reconstructing a broadband acceleration frequency spectrum of a casing of the gas turbine, a multi-scale fused neural network structure and an improved neural network full-connection layer are introduced, and the state of the gas turbine is judged by using reconstruction errors. According to the learning method for mapping the high-dimensional data to the low-dimensional manifold, the model has high robustness and generalization performance. In practical application, the blade fracture fault sign of a certain practical industrial gas turbine is successfully found several days ahead of time under the condition that an expert cannot find the fault sign through a traditional means. The method is simple, reliable, high in flexibility, wide in application range and convenient to use in engineering practice.

Description

technical field [0001] The invention relates to a gas turbine state monitoring method, in particular to a gas turbine abnormal state monitoring method based on frequency spectrum reconstruction errors. Background technique [0002] Gas turbine fault monitoring and early warning are indispensable to maintain the healthy and stable operation of gas turbines. In the field of gas turbine fault monitoring, it is generally necessary to use abnormal state detection methods to determine whether the gas turbine is in an unhealthy and stable operating state, and then use other The method is to judge the specific fault type and carry out traceability diagnosis, and then the operator makes a judgment and takes measures such as shutting down for maintenance. In this process, it is very important to judge the health status of the gas turbine. Abnormal state monitoring mainly uses environmental parameters and parameters generated during the operation of gas turbines, such as performance p...

AI Technical Summary

Problems solved by technology

Compared with the traditional single-threshold method regardless of working conditions, this method greatly improves the accuracy of multi-working condition diagnosis, but there are still some problems: first, this method does not take more control or environmental parameters into consideration , but only use the working conditions as the classification basis without further subdivision, so there should be room for improvement in its accuracy; secondly, this method is discrete for the classification of working conditions, and the span is generally relatively large, only a few This is a common working condition, but for some extremely complex working conditions such as aeroengines, this discrete method is much weaker in constructing the distribution of operating parameters than the method that can continuously analyze the working conditions; moreover, this This method needs to model and analyze each working condition separately, which cannot be well unified, and many parameters that can be shared have not been used reasonably, so it takes up too much computing and storage resources; finally, due to the vibration spectrum Complexity, the analysis of vibration by this method is often limited to the total amplitude, and discards a large amount of effective information hidden in the frequency spectrum, so it cannot be well detected for many types of faults

First of all, the frequency range that can be analyzed by the spectrum used in the current research is generally only a few kilohertz at most, because the sampling rate of the sensors used is so high; while the passing frequency of the blades of an aeroengine can often reach tens of thousands of hertz, so it is impossible to capture relevant Valid information about blades, making it impossible to implement blade-related fault detection

In addition, the effective information in the frequency spectrum is not necessarily stored only in the amplitude of the n-order conversion frequency. In the actual frequency spectrum, sidebands often appear, and these sidebands may also carry information that is effective for inferring the state of the gas turbine. It is difficult for traditional statistical methods to take all combinations of frequencies and sidebands into account

[0005] From the above analysis, it can be seen that in the field of abnormal monitoring of gas turbines, the traditional manual design method has not yet been able to perfect the means of utilizing vibration spectrum data. The existing methods are limited by the limitations of manual selection and tracking, and it is difficult to analyze all the spectrum. The use of intelligent methods has become an important direction of technological development in this situation

At present, the analysis of the frequency spectrum by artificial intelligence algorithms is more of a supervised or semi-supervised learning. Once there is a lack of faulty samples, the learning will become difficult, and it lacks generalization performance between different working conditions. Faced with multiple working conditions The problem still requires discretization analysis

Images