WO2023097705A1 - 一种基于多源数据融合的压气机旋转失速预测方法 - Google Patents

一种基于多源数据融合的压气机旋转失速预测方法 Download PDF

Info

Publication number
WO2023097705A1
WO2023097705A1 PCT/CN2021/135611 CN2021135611W WO2023097705A1 WO 2023097705 A1 WO2023097705 A1 WO 2023097705A1 CN 2021135611 W CN2021135611 W CN 2021135611W WO 2023097705 A1 WO2023097705 A1 WO 2023097705A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
source data
stall
prediction
time
Prior art date
Application number
PCT/CN2021/135611
Other languages
English (en)
French (fr)
Inventor
孙希明
李育卉
全福祥
Original Assignee
大连理工大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 大连理工大学 filed Critical 大连理工大学
Publication of WO2023097705A1 publication Critical patent/WO2023097705A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/15—Vehicle, aircraft or watercraft design
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/14—Force analysis or force optimisation, e.g. static or dynamic forces
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation

Definitions

  • the invention relates to a compressor rotating stall prediction method based on multi-source data fusion, and belongs to the technical field of aero-engine modeling and simulation.
  • the aeroengine is the core component that affects the overall performance and reliability of the aircraft, and the axial flow compressor has an extremely critical impact on the stability of the engine, and the aerodynamic instability of the compressor is a very destructive and rapidly developing common Therefore, the accurate identification and timely prediction of the compressor instability process under complex working conditions is the focus of research in the field of aero-engines at home and abroad.
  • compressor instability can be divided into two states: surge and rotating stall.
  • surge and rotating stall When the compressor works stably, the flow rate decreases and the pressure ratio increases. However, when the flow rate decreases to exceed the instability boundary, the compressor will produce airflow separation. , the working state is likely to enter the stall precursor, rotating stall and surge stages from the steady state.
  • the active control method that takes measures at the initial disturbance stage of the instability is more worthy of further research and application. , so it is necessary to accurately identify the working status of each stage of the compressor, detect the occurrence of instability precursors in time, and gain more time for active control.
  • the aero-engine is a complex multi-component system. Multiple independent sensors are often installed at different positions inside the compressor to observe the operating status of each component.
  • the sensor system formed by it can detect and record each gas path during normal operation and failure. The status change of the component.
  • the current discrimination methods for compressor rotating stall usually deal with the pulsating pressure signals collected by a single sensor at different time periods, or uniformly process the pulsating pressure signals collected by multiple sensors at the same time period without distinction, without comprehensive consideration of the sensor system
  • the multi-faceted instability state information contained in it does not distinguish the differences between sensors, and the discrimination results are relatively one-sided. It is also easily restricted by the working state of a single sensor, lacks flexibility, and has poor reliability.
  • the present invention provides a method for predicting compressor rotating stall based on multi-source data fusion.
  • a method for predicting compressor rotating stall based on multi-source data fusion comprising the following steps:
  • S1.1 Collect the dynamic pressure experiment data recorded by sensors at different positions as a multi-source data set, use the fast Fourier transform method to perform spectrum analysis on the experimental data, and determine the steady-state operating frequency range and stall frequency range;
  • S1.2 Use a low-pass filter to filter the pressure change data, filter the high-frequency components, and retain the steady-state operating frequency components and stall frequency components;
  • sampling point As the basic unit to generate a label column with the same time length as the data set, where each sampling point is marked as 1 or 0 according to whether it is surging;
  • S2.1 Use the sliding window technology to divide the pressure change data of each sensor separately with a fixed time step, form the sampling data covered by the sliding window into a sample, and adjust the dimensions of all data sets to (samples, seq_length, sensors ), where samples is the number of samples, seq_length is the time step, and sensors is the number of sensors;
  • each branch module is input-oriented, defines two stacked time distribution convolution modules, and uses the feature parameter sharing of the TimeDistributed layer in the keras library to construct each time Distributed convolution module, each time-distributed convolution module is specifically composed of a single-channel one-dimensional convolution layer, a normalization layer and an activation layer; the input samples are batched after passing through the convolution layer of the first time-distributed convolution module Normalization processing, introducing the correction nonlinear unit ReLU to adjust the information passed to the next time-distributed convolution module, and then repeating the processing steps of the first time-distributed convolution module, using the Flatten() function to convert the final multi-dimensional
  • the output is one-dimensional to prepare for the transition of the subsequent fully connected layer;
  • S2.3 According to the number of sensors used in the experiment, set up a multi-source data feature extraction module composed of feature extraction branch modules.
  • the feature extraction branch module corresponding to each sensor defines the convolution kernel size, filter size,
  • the parameters such as the moving step are independent of each other between the modules, and the parameters are not shared.
  • the output characteristic information of this module is independently saved and transmitted to the time pattern extraction module;
  • S3.1 connect and merge the output feature information of each branch module in the multi-source data feature extraction module, as the input of the time pattern extraction module;
  • S4.1 Build a multi-source data fusion prediction model based on the multi-source data feature extraction module and the time pattern extraction module.
  • the pressure change data measured by different sensors are processed by different feature extraction modules, and each module has an independent filter to extract correlation Features, it is convenient to follow the actual test to add and delete sensors, merge the feature information output by multiple modules and pass it to the time pattern extraction module, learn the long-term time dependence in the data, and obtain the stall prediction probability through the linearly activated fully connected layer;
  • S4.2 uses the mean square error (MSE) between the model output value f(x) and the real value y as the loss function, and the calculation method is as follows:
  • i indicates that the current training is the i-th sample
  • n indicates the number of samples calculated in each batch
  • n is the size of the entire training sample set
  • the calculation method of the root mean square error RMSE is as follows:
  • the calculation method of the Score function is as follows:
  • the Score function distinguishes early prediction and later prediction.
  • it is usually more desirable to advance earlier, so as to gain more time for active control, so this requires the evaluation index to be asymmetric , so a larger penalty is implemented in the Score function for lagged predictions.
  • step S5.1 Obtain the test set data divided after preprocessing in step S1, and adjust the data dimension to (samples, seq_length, sensors) according to the step S2.1;
  • S5.4 Select a set of dynamic pressure data, including the experimental data recorded by a sensor that is working abnormally, and use the trained model to predict the stall probability of this set of data to test the fault tolerance of the model.
  • the method provided by the present invention is used to predict the rotating stall of the compressor, which integrates multiple sensor data at different positions, improves the comprehensiveness and prediction accuracy of information features, and outputs according to the model
  • the stall prediction probability can be divided according to the actual threshold value of the project, so as to realize the hierarchical early warning.
  • the model can dynamically adjust part of the network structure following the number of sensors used in the experiment, making it easier to apply to multi-sensor environments.
  • the model can still accurately predict the compressor stall probability based on the data recorded by the remaining sensors, which has a certain degree of fault tolerance.
  • This method is based on data and has nothing to do with the engine structure. Therefore, the model can be easily transferred to different types of engines by training different data sets, which has certain universality.
  • Fig. 1 is a flowchart of a method for predicting compressor rotating stall based on multi-source data fusion
  • Fig. 2 is a flow chart of data preprocessing
  • Fig. 3 is a multi-source data feature extraction module structure diagram
  • Figure 4 is a structural diagram of the multi-source data fusion prediction model
  • Fig. 5 is a diagram of the pressure change of all sensors in one set of data in the test set, where (a) is the diagram of the change of the stator tip p in of the inlet guide vane with time, and (b) is the change of the dynamic pressure p 1 of the first-stage stator tip with time Figure (c) is the dynamic pressure p 2 at the tip of the second stage stator changing with time, (d) is the dynamic pressure p 4 at the tip of the fourth stage stator changing with time, (e) is the dynamic pressure at the outlet wall p out with time change map;
  • Fig. 6 is the prediction result diagram of the multi-source data fusion prediction model on the test data shown in Fig. 5, where (a) is the dynamic pressure p4 at the tip of the fourth-stage stator changing with time, and (b) is the multi-source data fusion prediction The stall prediction probability given by the model changes with time, (c) is the early warning signal given by the multi-source data fusion prediction model;
  • Figure 7 is the prediction result of the multi-source data fusion prediction model on the dynamic pressure data of a single sensor, where (a) is the dynamic pressure p4 at the tip of the fourth-stage stator changing with time, and (b) is the multi-source data fusion prediction model The given stall prediction probability changes with time, (c) is the early warning signal given by the multi-source data fusion prediction model;
  • Fig. 8 is the variation curve of the test data of the fault tolerance capability of the model, in which (a) is the diagram of the change of p in at the tip of the stator tip of the inlet guide vane with time, (b) is the diagram of the change of dynamic pressure p 1 at the tip of the primary stator with time, (c ) is the diagram of the dynamic pressure p 2 at the tip of the second-stage stator changing with time, (d) is the diagram of the dynamic pressure p 4 at the tip of the fourth-stage stator changing with time, (e) is the diagram of the dynamic pressure p out at the outlet wall changing with time;
  • Fig. 9 is a test diagram of the fault tolerance capability of the model, in which (a) is the diagram of the dynamic pressure p1 at the tip of the first-stage stator changing with time, (b) is the diagram of the stall prediction probability given by the multi-source data fusion prediction model with time, (c) The early warning signal given by the multi-source data fusion prediction model.
  • the background of the present invention is the stall experiment data of a certain type of aeroengine, and the process flow of the compressor rotating stall prediction method based on multi-source data fusion is shown in FIG. 1 .
  • FIG. 2 is a flow chart of data preprocessing, and the steps of data preprocessing are as follows:
  • Each set of data generates a label sequence with the sampling point as the basic unit.
  • the dimension is the same length as the data set, and each sampling point is marked as 1 or 0 according to whether it is surged;
  • FIG. 3 is a structural diagram of the multi-source data feature extraction module. The steps to construct the multi-source data feature extraction module are as follows:
  • the sliding window technology is used to divide the pressure change data of each sensor separately with a fixed time step, and the sampling data covered by the sliding window are combined into one sample, and all the data
  • the set dimensions are all adjusted to (samples, seq_length, sensors), where samples is the number of samples, seq_length is the time step, and sensors is the number of sensors;
  • each branch module defines two stacked time distribution convolution modules, and use the feature parameter sharing of the TimeDistributed layer in the keras library to construct each time distribution convolution module, each time-distributed convolution module is specifically composed of a single-channel one-dimensional convolution layer, a normalization layer, and an activation layer; the input samples are batch-normalized after passing through the convolution layer of the first time-distributed convolution module Processing, introduce the correction nonlinear unit ReLU to adjust the information passed into the next time-distributed convolution module, and then repeat the processing steps of the first time-distributed convolution module, and use the Flatten() function to output the final multi-dimensional one-dimensional to make transition preparations for the subsequent full connection;
  • a multi-source data feature extraction module composed of multiple feature extraction branch modules, use the same size sliding window to divide different sensor data, and ensure that the sample dimension is uniform; the feature extraction branch corresponding to each sensor
  • the module independently defines parameters such as the convolution kernel size, filter size, and moving step size for the input samples.
  • the parameters are not shared between modules, and the output feature information of the module is independently saved and transmitted to the time pattern extraction module.
  • Figure 4 is a structural diagram of a multi-source data fusion prediction model. The steps to build a multi-source data fusion prediction model are as follows:
  • the multi-source data feature extraction module extracts the feature information of the data recorded by each sensor respectively, and merges the feature information of the same time period in chronological order as the input of the time pattern extraction module;
  • Bi-LSTM bidirectional long-term short-term memory network
  • the goal of this method is to train the network so that the output prediction probability is as close as possible to the real probability, so the mean square error (MSE) between the model output value f(x) and the real value y is used as the loss function, and the calculation method is as follows:
  • i indicates that the current training is the i-th sample
  • n indicates the number of samples calculated in each batch
  • n is the size of the entire training sample set
  • the evaluation index is asymmetric
  • the Score function distinguishes the early For forecasting and late forecasting, a greater penalty is implemented for lagging forecasting, so in addition to RMSE, the Score function is also used as an evaluation index, and the calculation method is as follows:
  • the results of the two evaluation indicators are combined, and the optimal model is saved to obtain the final multi-source data fusion prediction model.
  • Fig. 5 is a diagram of the pressure change of all sensors in one group of test data in the test set, where (a) is the diagram of the change of p in at the stator tip of the inlet guide vane with time, and (b) is the dynamic pressure p 1 at the tip of the first-stage stator with time (c) is the dynamic pressure p 2 at the tip of the second-stage stator changing with time, (d) is the dynamic pressure p 4 at the tip of the fourth-stage stator changing with time, (e) is the dynamic pressure at the outlet wall p out with Time change graph.
  • Fig. 6 is the prediction result diagram of the multi-source data fusion prediction model on the test data shown in Fig.
  • FIG. 5 is the diagram of the dynamic pressure p4 at the tip of the four-stage stator changing with time, that is, Fig. 5(d)
  • (b ) is the graph of the stall prediction probability over time given by the multi-source data fusion prediction model
  • (c) is the early warning signal given by the multi-source data fusion prediction model.
  • Figure 7 is the prediction result of the multi-source data fusion prediction model on the dynamic pressure data of a single sensor, where (a) is the dynamic pressure p4 at the tip of the fourth-stage stator changing with time, that is, Figure 5 (d),
  • (b) The graph of the stall prediction probability over time given by the multi-source data fusion prediction model, (c) is the early warning signal given by the multi-source data fusion prediction model.
  • Figure 8 is the pressure change diagram of the test data of the fault tolerance capability of the model, in which (a) is the diagram of the variation of the stator tip p in of the inlet guide vane with time, (b) is the diagram of the dynamic pressure p 1 of the first stage stator tip with time, ( c) is the diagram of the dynamic pressure p 2 at the tip of the second-stage stator changing with time, (d) is the diagram of the dynamic pressure p 4 at the tip of the fourth-stage stator changing with time, and (e) is the diagram of the dynamic pressure p out at the outlet wall changing with time.
  • Figure 9 is the test results of the fault tolerance capability of the model, where (a) is the dynamic pressure p 1 at the tip of the first-stage stator changing with time, which is Figure 8(b), and (b) is the prediction model given by the multi-source data fusion The graph of the stall prediction probability changing with time, (c) is the early warning signal given by the multi-source data fusion prediction model.
  • the steps to perform real-time prediction on test data are as follows:
  • the five curves in the pressure change diagram of the test set in Figure 5 are the dynamic pressure data recorded by the five measuring point sensors in the same period. It can be seen from the figure that a downward-developing protrusion began to appear at 6.02s, and it was in a stall In the initial disturbance stage, with the development of the stall disturbance, it began to fluctuate violently at 6.23s, which was the precursor stage, and completely developed into a stall surge after 6.48s; the test set data was processed according to the steps in the data preprocessing, and according to The input to the source data fusion predictive model requires adjustment of data dimensions.
  • the present invention comprehensively considers the information of multiple measurement points of the compressor system, so a comparative test is carried out for single-source data and multi-source data, and (d) is selected in a group of sensor data in Figure 5, that is, the fourth-stage stator tip
  • the dynamic pressure p 4 is preprocessed according to the requirements, and the data dimension is adjusted before entering the multi-source data fusion prediction model.
  • the multi-source data feature extraction module in the model only has one feature extraction branch module.
  • Figure 7(b) and (c) are the prediction results given by the model. It can be seen that the predicted probability fluctuates slightly around 6.27s, the probability value increases from 0 to about 15%, and the probability value rises to 100% around 6.53s , indicating a complete compressor stall.
  • Figure 8 is a pressure change diagram of model fault tolerance test data, in which the sensor corresponding to Figure 8(a) is working abnormally during measurement, and the other four sensors are working normally; it can be seen from the figure that there is a sudden downward development around 6.23s Sharp, in the initial disturbance stage of the stall, and directly developed into a stall surge at about 6.35s, without obvious aura stage; process the test set data according to the steps in the data preprocessing, and adjust the data according to the input requirements of the multi-source data fusion prediction model dimension.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Geometry (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Business, Economics & Management (AREA)
  • General Engineering & Computer Science (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Computer Hardware Design (AREA)
  • Artificial Intelligence (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Tourism & Hospitality (AREA)
  • Quality & Reliability (AREA)
  • Operations Research (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Marketing (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • General Business, Economics & Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Development Economics (AREA)
  • Automation & Control Theory (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Control Of Positive-Displacement Air Blowers (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

一种基于多源数据融合的压气机旋转失速预测方法，首先，获取来自多传感器的对航空发动机动态压力数据并进行预处理，划分测试、训练和验证数据集。其次，依次构建多源数据特征提取模块、构建时间模式提取模块、构建多源数据融合预测模型、保存最优预测模型。最后，在测试数据上进行实时预测：按构建多源数据融合预测模型的输入要求调整数据维度；采用多源数据融合预测模型计算每个样本的失速预测概率；采用多源数据融合预测模型计算失速预测概率；采用多源数据融合预测模型给出包含一个工作异常的传感器所记录数据的失速预测概率。本发明综合多源数据信息，提高信息特征的全面性和预测精度，具有一定容错性；有利于提高发动机主动控制的性能，具有普适性。

Description

一种基于多源数据融合的压气机旋转失速预测方法 技术领域
本发明涉及一种基于多源数据融合的压气机旋转失速预测方法，属于航空发动机建模与仿真技术领域。
背景技术
航空发动机是影响飞机整体性能和可靠性的核心部件，其中轴流压气机对发动机的稳定性又有着极其关键的影响，而压气机气动失稳则是一种破坏性极强、发展迅速的常见故障，因此在复杂工况下对压气机失稳过程的精准判别和及时预测是国内外航空发动机领域的研究重点。一般压气机失稳可分为喘振和旋转失速两种状态，当压气机稳定工作时，流量减少，压比增大，但当流量减少到超过失稳边界时，压气机则会产生气流分离，工作状态极可能由稳态进入失速先兆、旋转失速和喘振阶段。而旋转失速和喘振一旦发生，很难自动恢复，并且发动机内部部件往往已经受到损害，因此相比于被动控制方法，在失稳初始扰动阶段就采取措施的主动控制方法更值得深入研究与应用，所以需要准确识别压气机各阶段工作状态，及时检测失稳先兆的出现，为主动控制争取更多时间。
航空发动机属于复杂多部件系统，在压气机内部的不同位置往往安装了多个相互独立的传感器来观测各部件运行状态，其构成的传感器系统可以检测并记录正常运行时和故障发生时各气路部件的状态变化情况。然而，目前压气机旋转失速的判别方法通常处理单个传感器于不同时间段采集的脉动压力信号，或不加区分地统一处理多个传感器于同一时间段采集的脉动压力信号，没有综合考虑到传感器系统所蕴含的多方面失稳状态信息，也没有区分传感器之间的差异性，判别结果相对片面，也容易受单个传感器工作状态制约，缺乏灵活性、可靠性差。
发明内容
针对现有技术中可靠性差，相对片面的问题，本发明提供一种基于多源数据融合的压气机旋转失速预测方法。
本发明的技术方案：
一种基于多源数据融合的压气机旋转失速预测方法，包括以下步骤：
S1.获取航空发动机多源失速数据，并对其进行预处理，其中多源失速数据包括实验平台中多个传感器记录的动态压力数据，包括以下步骤：
S1.1收集不同位置的传感器所记录的动态压力实验数据作为多源数据集，使用快速傅里叶变换方法对实验数据进行频谱分析，确定稳态工作频率范围以及失速频率范围；
S1.2采用低通滤波器对压力变化数据进行滤波处理，过滤高频分量，保留稳态工作频率 分量和失速频率分量；
S1.3对滤波后数据进行重采样，缩小数据维度；由于S1.1分析失速频率在60～110Hz之间，依据奈奎斯特采样定理进行5倍降采样；
S1.4对降采样后数据进行归一化处理，将数据分布通过线性变化映射到[0,1]区间；
S1.5以采样点为基本单位生成与数据集同时间长度的标签列，其中每个采样点根据是否喘振标为1或0；
S1.6为保证结果客观有效，将预处理后的全部数据集划分为训练数据集、验证数据集和测试数据集；
S2.构建多源数据特征提取模块，包括以下步骤：
S2.1采用滑动窗口技术对每个传感器的压力变化数据以固定时间步长单独进行划分，将滑动窗口所覆盖的采样数据组成一个样本，将所有数据集维度均调整为(samples,seq_length,sensors)，其中samples为样本个数，seq_length为时间步长，sensors为传感器数量；
S2.2为每个传感器构建一个独立的特征提取分支模块，每个分支模块面向输入，定义两个堆叠的时间分布卷积模块，利用keras库中TimeDistributed层的特征参数共享性来构建每个时间分布卷积模块，每个时间分布卷积模块具体由单通道一维卷积层、归一化层和激活层组成；输入样本在经过第一个时间分布卷积模块的卷积层后进行批量归一化处理，引入校正非线性单元ReLU以调整传入下一个时间分布卷积模块的信息，再重复进行第一个时间分布卷积模块的处理步骤，利用Flatten()函数将其最后的多维输出一维化，为后续全连接层做过渡准备；
S2.3根据实验中所用传感器数目，组建由特征提取分支模块构成的多源数据特征提取模块，每个传感器对应的特征提取分支模块面向输入样本定义本模块的卷积核大小、滤波器大小、移动步长等参数，模块之间相互独立，参数不共享，独立保存本模块的输出特征信息，并传至时间模式提取模块；
S3.构建时间模式提取模块，包括以下步骤：
S3.1连接合并多源数据特征提取模块中每个分支模块的输出特征信息，作为时间模式提取模块的输入；
S3.2基于双向长短期记忆网络(Bi-LSTM,Bidirectional Long short-Term Memory)构建时间模式提取模块，依次经过两层Bi-LSTM网络对输入数据进行处理，其中第一层Bi-LSTM中的记忆单元数量大于等于第二层中的记忆单元数量；
S3.3添加dropout层防止训练中过拟合情况的发生，再通过一组密集层及线性激活Linear函数，将Bi-LSTM网络的输出维度调整为(n_outputs,1)，即为旋转失速预测概率；
S4.构建多源数据融合预测模型，包括以下步骤：
S4.1基于多源数据特征提取模块和时间模式提取模块构建多源数据融合预测模型，不同传感器测量所得的压力变化数据经过不同的特征提取模块进行处理，每个模块有独立的滤波器提取相关特征，方便跟随实际试验进行传感器添加、删除等操作，合并多个模块输出的特征信息并传递给时间模式提取模块，学习数据中时间长期依赖性，通过线性激活的全连接层得到失速预测概率；
S4.2采用模型输出值f(x)和真实值y之间的均方误差(MSE)作为损失函数，计算方式如下：
Figure PCTCN2021135611-appb-000001
其中，i表示当前训练的是第i个样本，n表示每批次计算的样本数量，对于全批量学习，n为整个训练样本集大小；
S4.3保存训练后的多源数据融合预测模型并在验证集上测试，根据验证集评价指标调整模型超参数，量化模型预测性能的评价指标采用均方根误差(RMSE)和Score函数，综合两个评价指标结果，保存最优模型得到最终的多源数据融合预测模型；
所述均方根误差RMSE的计算方式如下：
Figure PCTCN2021135611-appb-000002
其中，Δy i＝y pred-y true，是模型根据第i个样本生成的预测值与真实值之差，N是每批次输入模型的总样本数量；
所述Score函数计算方式如下：
Figure PCTCN2021135611-appb-000003
其中，Score函数区分了提前预测和之后预测，在旋转失速预测的实际应用环境中，相比于滞后预测，通常更希望提前一些，从而为主动控制争取更多时间，因此这要求评价指标是非对称的，所以Score函数中为滞后预测实施更大的惩罚。
S5.在测试数据上进行实时预测
S5.1获取步骤S1中预处理后划分出的测试集数据，按照S2.1步骤将数据维度调整为(samples,seq_length,sensors)；
S5.2采用S4.3中保存的多源数据融合最优预测模型计算每个样本的失速预测概率，并按 照时间顺序排序；
S5.3随机选取测试集中一组动态压力数据，采用保存的最优预测模型仅对其中一个传感器数据进行失速预测，并同基于全部传感器数据计算的失速预测概率进行对比，以观察多源数据融合对模型预测效果的帮助性；
S5.4选取一组动态压力数据，其中包含一个工作异常的传感器所记录的实验数据，采用训练后的模型对这组数据进行失速概率预测，以测试模型的容错性。
本发明的有益效果：通过本发明所提供方法对压气机旋转失速预测，相比于以往的方法，综合了不同位置的多个传感器数据，提高信息特征的全面性和预测准确性，根据模型输出的失速预测概率，可以根据工程实际设置阈值划分概率，从而实现分级预警。模型可跟随试验中使用的传感器数量动态调整部分网络结构，更方便应用于多传感器环境。同时，模型在少部分传感器无法正常工作的情况下，仍能根据剩余传感器记录的数据准确预测压气机失速概率，具有一定的容错性。本方法基于数据，与发动机结构无关，因此通过训练不同数据集就可以将模型方便地迁移到不同型号发动机上使用，具有一定的普适性。
附图说明
图1为基于多源数据融合的压气机旋转失速预测方法流程图；
图2为数据预处理流程图；
图3为多源数据特征提取模块结构图；
图4为多源数据融合预测模型结构图；
图5为测试集其中一组数据的全部传感器压力变化图，其中(a)为进口导向叶片静子尖部p in随时间变化图，(b)为一级静子尖部动压p 1随时间变化图，(c)为二级静子尖部动压p 2随时间变化图，(d)为四级静子尖部动压p 4随时间变化图，(e)为出口壁面动压p out随时间变化图；
图6为多源数据融合预测模型在图5所示测试数据上的预测结果图，其中(a)为四级静子尖部动压p 4随时间变化图，(b)为多源数据融合预测模型给出的失速预测概率随时间的变化图，(c)为多源数据融合预测模型给出的预警信号；
图7为多源数据融合预测模型在单个传感器动压数据上的预测结果图，其中(a)为四级静子尖部动压p 4随时间变化图，(b)为多源数据融合预测模型给出的失速预测概率随时间的变化图，(c)为多源数据融合预测模型给出的预警信号；
图8为模型容错能力测试数据变化曲线，其中图(a)为进口导向叶片静子尖部p in随时间变化图，(b)为一级静子尖部动压p 1随时间变化图，(c)为二级静子尖部动压p 2随时间变化图，(d)为四级静子尖部动压p 4随时间变化图，(e)为出口壁面动压p out随时间变化图；
图9为模型容错能力测试图，其中(a)为一级静子尖部动压p 1随时间变化图，(b)为多源数 据融合预测模型给出的失速预测概率随时间的变化图，(c)为多源数据融合预测模型给出的预警信号。
具体实施方式
下面结合附图对本发明作进一步说明，本发明依托背景为某型航空发动机失速实验数据，基于多源数据融合的压气机旋转失速预测方法流程如图1所示。
图2为数据预处理流程图，数据预处理步骤如下：
S1.实验设置5个测量点，测量从正常到喘振共10s的动态压力数值，传感器测量频率为6kHz，共记录16组数据；5个测量点分别位于：进口导向叶片静子尖部、一级静子尖部、二级静子尖部、四级静子尖部、出口壁面；
S2.使用快速傅里叶变换方法对实验数据进行频谱分析，确定稳态工作频率范围以及失速频率范围；
S3.采用低通滤波器对压力变化数据进行滤波处理；为缩小数据维度、节约计算资源，对滤波后数据进行降采样，依据频谱分析中确认失速频率的范围在60～110Hz之间，根据奈奎斯特采样定理选定5倍降采样率；
S4.对降采样后数据进行归一化处理，将数据分布通过线性变化映射到[0,1]区间；
S5.每组数据以采样点为基本单位生成一列标签序列，维度与数据集同长，其中每个采样点根据是否喘振标为1或0；
S6.为保证结果客观有效，将预处理后的全部数据集划分为训练数据集、验证数据集和测试数据集。
图3为多源数据特征提取模块结构图，构建多源数据特征提取模块的步骤如下：
S1.为了捕获时间序列内每个阶段的细小特征，采用滑动窗口技术对每个传感器的压力变化数据以固定时间步长单独进行划分，将滑动窗口所覆盖的采样数据组成一个样本，将所有数据集维度均调整为(samples,seq_length,sensors)，其中samples为样本个数，seq_length为时间步长，sensors为传感器数量；
S2.针对每个传感器搭建各自独立的特征提取分支模块，每个分支模块定义了两个堆叠的时间分布卷积模块，利用keras库中TimeDistributed层的特征参数共享性来构建每个时间分布卷积模块，每个时间分布卷积模块具体由单通道一维卷积层、归一化层和激活层组成；输入样本在经过第一个时间分布卷积模块的卷积层后进行批量归一化处理，引入校正非线性单元ReLU以调整传入下一个时间分布卷积模块的信息，再重复进行第一个时间分布卷积模块的处理步骤，利用Flatten()函数将其最后的多维输出一维化，为后续的全连接做过渡准备；
S3.根据试验中所用传感器数目组建由多个特征提取分支模块构成的多源数据特征提取 模块，使用同样大小的滑动窗口划分不同传感器数据，保证样本维度大小统一；每个传感器对应的特征提取分支模块面向输入样本独立定义本模块的卷积核大小、滤波器大小、移动步长等参数，模块之间参数不共享，并且独立保存本模块的输出特征信息，传至时间模式提取模块。
图4为多源数据融合预测模型结构图，构建多源数据融合预测模型的步骤如下：
S1.多源数据特征提取模块分别提取各个传感器所记录数据的特征信息，按照时间顺序将同时间段的特征信息合并在一起，作为时间模式提取模块的输入；
S2.基于双向长短期记忆网络(Bi-LSTM)构建时间模式提取模块，依次经过两层Bi-LSTM网络层对输入数据进行处理，其中第一层Bi-LSTM中的记忆单元数量要大于或等于第二层中的记忆单元数量；
S3.通过输出密集层，即全连接层，和线性激活Linear函数处理Bi-LSTM网络的输出，同时添加dropout层防止训练中发生过拟合情况，最终得到失速预测概率值。
S4.本方法的目标是训练网络使得输出预测概率尽可能与真实概率相近，因此采用模型输出值f(x)和真实值y之间的均方误差(MSE)作为损失函数，计算方式如下：
Figure PCTCN2021135611-appb-000004
其中，i表示当前训练的是第i个样本，n表示每批次计算的样本数量，对于全批量学习，n为整个训练样本集大小；
S5.保存训练后的多源数据融合预测模型并在验证集上测试，根据验证集评价指标调整模型超参数，量化模型预测性能的评价指标采用均方根误差(RMSE)，计算方式如下：
Figure PCTCN2021135611-appb-000005
其中，Δy i＝y pred-y true，是模型根据第i个样本生成的预测值与真实值之差，N是试验中输入模型的总样本数量；
此外，在旋转失速预测的实际应用环境中，相比于滞后预测通常更希望模型做到提前预测，从而为主动控制争取更多时间，因此这要求评价指标是非对称的，而Score函数区分了早期预测和晚期预测，为滞后预测实施更大的惩罚，所以除了RMSE外，还采用Score函数作为评价指标，计算方式如下：
Figure PCTCN2021135611-appb-000006
综合两个评价指标结果，保存最优模型得到最终的多源数据融合预测模型。
图5为测试集中其中一组测试数据的全部传感器压力变化图，其中(a)为进口导向叶片静子尖部p in随时间变化图，(b)为一级静子尖部动压p 1随时间变化图，(c)为二级静子尖部动压p 2随时间变化图，(d)为四级静子尖部动压p 4随时间变化图，(e)为出口壁面动压p out随时间变化图。图6为多源数据融合预测模型在图5所示测试数据上的预测结果图，其中(a)为四级静子尖部动压p 4随时间变化图，即图5(d)，(b)为多源数据融合预测模型给出的失速预测概率随时间的变化图，(c)为多源数据融合预测模型给出的预警信号。图7为多源数据融合预测模型在单个传感器动压数据上的预测结果图，其中(a)为四级静子尖部动压p 4随时间变化图，即图5(d)，(b)为多源数据融合预测模型给出的失速预测概率随时间的变化图，(c)为多源数据融合预测模型给出的预警信号。图8为模型容错能力测试数据压力变化图，其中图(a)为进口导向叶片静子尖部p in随时间变化图，(b)为一级静子尖部动压p 1随时间变化图，(c)为二级静子尖部动压p 2随时间变化图，(d)为四级静子尖部动压p 4随时间变化图，(e)为出口壁面动压p out随时间变化图。图9为模型容错能力测试结果图，其中(a)为一级静子尖部动压p 1随时间变化图，即为图8(b)，(b)为多源数据融合预测模型给出的失速预测概率随时间的变化图，(c)为多源数据融合预测模型给出的预警信号。在测试数据上进行实时预测的步骤如下：
S1.图5测试集压力变化图中五条曲线分别为同一时段中五个测量点传感器记录的动态压力数据，从图中可以看出，6.02s开始出现了一个向下发展的突尖，处于失速初始扰动阶段，随着失速扰动的发展，在6.23s开始有剧烈的波动，为先兆阶段，在6.48s后彻底发展为失速喘振；按数据预处理中步骤对测试集数据处理，并按多源数据融合预测模型的输入要求调整数据维度。
S2.按时间顺序,用多源数据融合预测模型计算基于图5一组传感器测量数据的失速预测概率，其中图6(a)为图5(d)，用于辅助图6(b)和(c)进行预测结果分析。从图6(b)和(c)可以看到，在6.23s左右失速概率从0提升至20％左右，在6.23s到6.37s之间一直维持在20％左右，同失速先兆阶段的时间跨度相符，之后在6.49s渐渐提升至100％左右，表明彻底发展为失速喘振。根据失速喘振原理，先兆的出现大概率代表失速和喘振会在后续发生，因此在6.23s左右模型预测到失速概率有明显提升时，即预判进入先兆阶段，系统随即给出预警信号。
S3.本发明综合考虑压气机系统多个测量点的信息，因此针对单源数据和多源数据进行了对比试验，在图5的一组传感器数据中选择(d)，即四级静子尖部动压p 4，按照要求进行预处理、调整数据维度后输入多源数据融合预测模型，模型中多源数据特征提取模块仅设置一个 特征提取分支模块。图7(b)和(c)为模型给出的预测结果，可以看出在6.27s左右预测概率有小幅度波动，概率值从0提升到15％左右，6.53s左右概率值上升至100％，表明压气机彻底失速。整体来看，相较于图6，仅针对单个传感器继续预测会有滞后预测的情况，同时在先兆阶段概率提升幅度不明显，不利于系统给出失速预警信号，因此表明融合多源数据对失速预测有一定帮助。
S4.图8模型容错能力测试数据压力变化图，其中图8(a)对应的传感器在测量时工作异常，其余4个传感器正常工作；由图可知，在6.23s左右有一个向下发展的突尖，处于失速初始扰动阶段，在6.35s左右直接发展成失速喘振，没有明显的先兆阶段；按数据预处理中步骤对测试集数据处理，并按多源数据融合预测模型的输入要求调整数据维度。
S5.按时间顺序,用多源数据融合预测模型计算基于8一组传感器测量数据的失速预测概率，其中图9(a)为图8(b)，用于辅助图9(b)和(c)进行预测结果分析。观察预测结果图发现，在6.23s左右失速概率从0突然升至16％左右，说明捕捉到动压数据中向下发展的突尖信号，在6.35s后提升至100％，表明彻底进入失速喘振阶段，因此说明多源数据融合模型在有一个传感器工作异常的情况下，仍能根据其他正常工作的传感器记录信息进行准确预测，具有一定的容错能力，方便工程应用。

Claims (5)

  1. 一种基于多源数据融合的压气机旋转失速预测方法，其特征在于，包括以下步骤：
    S1.获取航空发动机多源失速数据，并对其进行预处理，其中多源失速数据包括实验平台中多个传感器记录的动态压力数据：
    S1.1收集不同位置的传感器所记录的动态压力实验数据作为多源数据集，对实验数据进行频谱分析，确定稳态工作频率范围以及失速频率范围；
    S1.2采用低通滤波器对压力变化数据进行滤波处理，过滤高频分量，保留稳态工作频率分量和失速频率分量；
    S1.3对滤波后数据进行重采样，缩小数据维度；
    S1.4对降采样后数据进行归一化处理，将数据分布通过线性变化映射到[0,1]区间；
    S1.5以采样点为基本单位生成与数据集同时间长度的标签列，其中每个采样点根据是否喘振标为1或0；
    S1.6将预处理后的全部数据集划分为训练数据集、验证数据集和测试数据集；
    S2.构建多源数据特征提取模块，包括以下步骤：
    S2.1采用滑动窗口方法对每个传感器的压力变化数据以固定时间步长单独进行划分，将滑动窗口所覆盖的采样数据组成一个样本，将所有数据集维度均调整为(samples,seq_length,sensors)，其中samples为样本个数，seq_length为时间步长，sensors为传感器数量；
    S2.2为每个传感器构建一个独立的特征提取分支模块，每个分支模块面向输入定义两个堆叠的时间分布卷积模块，并构建两个时间分布卷积模块，
    S2.3根据所用传感器数目，组建由特征提取分支模块构成的多源数据特征提取模块，每个传感器对应的特征提取分支模块面向输入样本定义本模块的参数，模块之间相互独立，参数不共享，独立保存本模块的输出特征信息，并传至时间模式提取模块；
    S3.构建时间模式提取模块，包括以下步骤：
    S3.1连接合并多源数据特征提取模块中每个分支模块的输出特征信息，作为时间模式提取模块的输入；
    S3.2基于双向长短期记忆网络Bi-LSTM构建时间模式提取模块，依次经过两层Bi-LSTM网络对输入数据进行处理，
    S3.3添加dropout层，再通过一组密集层及线性激活Linear函数，将Bi-LSTM网络的输出维度调整为(n_outputs,1)，即为旋转失速预测概率；
    S4.构建多源数据融合预测模型，包括以下步骤：
    S4.1基于多源数据特征提取模块和时间模式提取模块构建多源数据融合预测模型，不同传感器测量所得的压力变化数据经过不同的特征提取模块进行处理，每个模块有独立的滤波 器提取相关特征，合并多个模块输出的特征信息并传递给时间模式提取模块，学习数据中时间长期依赖性，通过线性激活的全连接层得到失速预测概率；
    S4.2采用模型输出值f(x)和真实值y之间的均方误差MSE作为损失函数；
    S4.3保存训练后的多源数据融合预测模型并在验证集上测试，根据验证集评价指标调整模型超参数，采用均方根误差RMSE和Score函数量化模型预测性能的评价指标，综合两个评价指标结果，保存最优模型得到最终的多源数据融合预测模型；
    S5.在测试数据上进行实时预测
    S5.1获取步骤S1中预处理后划分出的测试集数据，按照S2.1步骤将数据维度调整为(samples,seq_length,sensors)；
    S5.2采用S4.3中保存的多源数据融合最优预测模型计算每个样本的失速预测概率，并按照时间顺序排序；
    S5.3随机选取测试集中一组动态压力数据，采用保存的最优预测模型仅对其中一个传感器数据进行失速预测，并同基于全部传感器数据计算的失速预测概率进行对比，以观察多源数据融合对模型预测效果的帮助性；
    S5.4选取一组动态压力数据，其中包含一个工作异常的传感器所记录的实验数据，采用训练后的模型对这组数据进行失速概率预测。
  2. 根据权利要求1所述的一种基于多源数据融合的压气机旋转失速预测方法，其特征在于，所述的步骤S2.2中，每个时间分布卷积模块均由单通道一维卷积层、归一化层和激活层组成；输入样本在经过第一个时间分布卷积模块的卷积层后进行批量归一化处理，引入校正非线性单元ReLU以调整传入下一个时间分布卷积模块的信息，再重复进行第一个时间分布卷积模块的处理步骤，利用Flatten()函数将其最后的多维输出一维化。
  3. 根据权利要求1所述的一种基于多源数据融合的压气机旋转失速预测方法，其特征在于，所述步骤S3.2中，第一层Bi-LSTM中的记忆单元数量大于等于第二层中的记忆单元数量。
  4. 根据权利要求1所述的一种基于多源数据融合的压气机旋转失速预测方法，其特征在于，所述的步骤S4.2中，损失函数计算方式如下：
    Figure PCTCN2021135611-appb-100001
    其中，i表示当前训练的是第i个样本，n表示每批次计算的样本数量，对于全批量学习，n为整个训练样本集大小。
  5. 根据权利要求1所述的一种基于多源数据融合的压气机旋转失速预测方法，其特征在于， 所述步骤S4.3中：
    所述均方根误差RMSE的计算方式如下：
    Figure PCTCN2021135611-appb-100002
    其中，Δy i＝y pred-y true，是模型根据第i个样本生成的预测值y pred与真实值y true之差，N是每批次输入模型的总样本数量；
    所述Score函数计算方式如下：
    Figure PCTCN2021135611-appb-100003
PCT/CN2021/135611 2021-11-30 2021-12-06 一种基于多源数据融合的压气机旋转失速预测方法 WO2023097705A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111441733.5A CN114186337A (zh) 2021-11-30 2021-11-30 一种基于多源数据融合的压气机旋转失速预测方法
CN202111441733.5 2021-11-30

Publications (1)

Publication Number Publication Date
WO2023097705A1 true WO2023097705A1 (zh) 2023-06-08

Family

ID=80603059

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/135611 WO2023097705A1 (zh) 2021-11-30 2021-12-06 一种基于多源数据融合的压气机旋转失速预测方法

Country Status (2)

Country Link
CN (1) CN114186337A (zh)
WO (1) WO2023097705A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116502075A (zh) * 2023-06-28 2023-07-28 吉林大学 一种多模态水下自主航行器状态检测方法及系统
CN117454166A (zh) * 2023-10-11 2024-01-26 国网四川省电力公司电力科学研究院 基于EffNet轻量化模型的致火电弧故障识别方法
CN117910550A (zh) * 2024-03-19 2024-04-19 江苏海拓宾未来工业科技集团有限公司 一种基于深度学习的无油超高速离心压缩机自动优化系统
CN117933104A (zh) * 2024-03-25 2024-04-26 中国人民解放军国防科技大学 固体姿轨控发动机燃气调节阀压强修正方法

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118347697B (zh) * 2024-06-17 2024-08-20 西北工业大学 一种内波中水下固定平台的受力预报方法及系统

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021007084A1 (en) * 2019-07-08 2021-01-14 Purdue Research Foundation Compressor stall warning using nonlinear feature extraction algorithms
CN112613226A (zh) * 2020-12-10 2021-04-06 大连理工大学 一种用于剩余寿命预测的特征增强方法
CN113569338A (zh) * 2021-08-06 2021-10-29 大连理工大学 一种基于时间扩张卷积网络的压气机旋转失速预警方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112161784B (zh) * 2020-09-07 2022-01-18 华南理工大学 基于多传感器信息融合迁移网络的机械故障诊断方法
CN113469060A (zh) * 2021-07-02 2021-10-01 浙大城市学院 多传感器融合卷积神经网络航空发动机轴承故障诊断方法
CN113469281B (zh) * 2021-07-22 2023-11-24 西北工业大学 一种工业齿轮箱多源信息融合故障诊断方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021007084A1 (en) * 2019-07-08 2021-01-14 Purdue Research Foundation Compressor stall warning using nonlinear feature extraction algorithms
CN112613226A (zh) * 2020-12-10 2021-04-06 大连理工大学 一种用于剩余寿命预测的特征增强方法
CN113569338A (zh) * 2021-08-06 2021-10-29 大连理工大学 一种基于时间扩张卷积网络的压气机旋转失速预警方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
"Doctoral Dissertation", 20 July 2020, DALIAN UNIVERSITY OF TECHNOLOGY, CN, article SONG, TIANYU: "Research on Flash Flood Forecasting Based on Long Short-Term Memory Networks", pages: 1 - 146, XP009546074, DOI: 10.26991/d.cnki.gdllu.2020.004027 *
"Master's Thesis", 15 May 2009, LANZHOU UNIVERSITY OF TECHNOLOGY, CN, article MEN, RIXIU: "Remote Fault Diagnosis System of Rotating Machinery Based on Network", pages: 1 - 85, XP009546075 *

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116502075A (zh) * 2023-06-28 2023-07-28 吉林大学 一种多模态水下自主航行器状态检测方法及系统
CN116502075B (zh) * 2023-06-28 2023-09-12 吉林大学 一种多模态水下自主航行器状态检测方法及系统
CN117454166A (zh) * 2023-10-11 2024-01-26 国网四川省电力公司电力科学研究院 基于EffNet轻量化模型的致火电弧故障识别方法
CN117454166B (zh) * 2023-10-11 2024-05-10 国网四川省电力公司电力科学研究院 基于EffNet轻量化模型的致火电弧故障识别方法
CN117910550A (zh) * 2024-03-19 2024-04-19 江苏海拓宾未来工业科技集团有限公司 一种基于深度学习的无油超高速离心压缩机自动优化系统
CN117910550B (zh) * 2024-03-19 2024-06-07 江苏海拓宾未来工业科技集团有限公司 一种基于深度学习的无油超高速离心压缩机自动优化系统
CN117933104A (zh) * 2024-03-25 2024-04-26 中国人民解放军国防科技大学 固体姿轨控发动机燃气调节阀压强修正方法
CN117933104B (zh) * 2024-03-25 2024-06-07 中国人民解放军国防科技大学 固体姿轨控发动机燃气调节阀压强修正方法

Also Published As

Publication number Publication date
CN114186337A (zh) 2022-03-15

Similar Documents

Publication Publication Date Title
WO2023097705A1 (zh) 一种基于多源数据融合的压气机旋转失速预测方法
CN112001128B (zh) 一种基于深度学习的轴流压气机失速喘振预测方法
US20230316051A1 (en) Pre-alarming method for rotary stall of compressors based on temporal dilated convolutional neural network
CN110441065B (zh) 基于lstm的燃气轮机在线检测方法与装置
Zhou et al. Fault diagnosis of gas turbines with thermodynamic analysis restraining the interference of boundary conditions based on STN
WO2024045246A1 (zh) 一种针对航空发动机故障预测的脉冲回声状态网络模型
Xiong et al. Adaptive deep learning-based remaining useful life prediction framework for systems with multiple failure patterns
CN105930629B (zh) 一种基于海量运行数据的在线故障诊断方法
Ren et al. Aero-engine remaining useful life estimation based on multi-head networks
Togni et al. A combined technique of Kalman filter, artificial neural network and fuzzy logic for gas turbines and signal fault isolation
CN110779745A (zh) 一种基于bp神经网络的换热器早期故障诊断方法
CN113053171A (zh) 一种民机系统风险预警方法及系统
Xu et al. New RUL prediction method for rotating machinery via data feature distribution and spatial attention residual network
Wang et al. Deep-learning modeling and control optimization framework for intelligent thermal power plants: A practice on superheated steam temperature
Quan et al. Detection of Rotating Stall Inception of Axial Compressors Based on Deep Dilated Causal Convolutional Neural Networks
Wang et al. Research on anomaly detection and positioning of marine nuclear power steam turbine unit based on isolated forest
CN112052496B (zh) 一种基于var模型的高拱坝谷幅变形影响因素分析系统的操作方法
Ma et al. An interpretability research of the Xgboost algorithm in remaining useful life prediction
Eustace A real-world application of fuzzy logic and influence coefficients for gas turbine performance diagnostics
Liu et al. NPFormer: Interpretable rotating machinery fault diagnosis architecture design under heavy noise operating scenarios
Shao et al. A dimensionless immune intelligent fault diagnosis system for rotating machinery
Wang et al. A Fault Diagnosis Strategy based on Qualitative Trend Analysis Integrating Andrews Plot for Industrial Processes
Yu et al. EGT baseline model of aeroengine based on kernel principal component analysis and deep belief network
Li et al. Rotating Stall Detection of Axial Compressors Based on Multi-source Heterogeneous Data Fusion
Liu et al. A Remaining Useful Life Prediction Method of Mechanical Equipment Based on Particle Swarm Optimization-Convolutional Neural Network-Bidirectional Long Short-Term Memory

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21966131

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE