WO2023010658A1 - 一种基于时间扩张卷积网络的压气机旋转失速预警方法 - Google Patents

一种基于时间扩张卷积网络的压气机旋转失速预警方法 Download PDF

Info

Publication number
WO2023010658A1
WO2023010658A1 PCT/CN2021/119187 CN2021119187W WO2023010658A1 WO 2023010658 A1 WO2023010658 A1 WO 2023010658A1 CN 2021119187 W CN2021119187 W CN 2021119187W WO 2023010658 A1 WO2023010658 A1 WO 2023010658A1
Authority
WO
WIPO (PCT)
Prior art keywords
time
data
convolutional network
surge
convolution
Prior art date
Application number
PCT/CN2021/119187
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
Priority to US17/795,095 priority Critical patent/US20230316051A1/en
Publication of WO2023010658A1 publication Critical patent/WO2023010658A1/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
    • F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
    • F04—POSITIVE - DISPLACEMENT MACHINES FOR LIQUIDS; PUMPS FOR LIQUIDS OR ELASTIC FLUIDS
    • F04D—NON-POSITIVE-DISPLACEMENT PUMPS
    • F04D27/00—Control, e.g. regulation, of pumps, pumping installations or pumping systems specially adapted for elastic fluids
    • F04D27/001—Testing thereof; Determination or simulation of flow characteristics; Stall or surge detection, e.g. condition monitoring
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
    • F05—INDEXING SCHEMES RELATING TO ENGINES OR PUMPS IN VARIOUS SUBCLASSES OF CLASSES F01-F04
    • F05D—INDEXING SCHEME FOR ASPECTS RELATING TO NON-POSITIVE-DISPLACEMENT MACHINES OR ENGINES, GAS-TURBINES OR JET-PROPULSION PLANTS
    • F05D2270/00—Control
    • F05D2270/01—Purpose of the control system
    • F05D2270/10—Purpose of the control system to cope with, or avoid, compressor flow instabilities
    • F05D2270/101—Compressor surge or stall
    • F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
    • F05—INDEXING SCHEMES RELATING TO ENGINES OR PUMPS IN VARIOUS SUBCLASSES OF CLASSES F01-F04
    • F05D—INDEXING SCHEME FOR ASPECTS RELATING TO NON-POSITIVE-DISPLACEMENT MACHINES OR ENGINES, GAS-TURBINES OR JET-PROPULSION PLANTS
    • F05D2270/00—Control
    • F05D2270/70—Type of control algorithm
    • F05D2270/709—Type of control algorithm with neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/08—Probabilistic or stochastic CAD
    • 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 rotation stall warning method based on a time-expanded convolutional network, and belongs to the technical field of aeroengine modeling and simulation.
  • the performance stability of an aero-engine is directly related to the flight safety of the whole machine, and the stability of the air circuit components in terms of overall flow, pressure and energy affects the overall working state of the engine.
  • the compressor rotates Stall is one of the most destructive and fast-changing faults, so its accurate identification and timely early warning is the focus of research in the field of aero-engines at home and abroad.
  • the development process of compressor instability mainly includes four stages: steady state, stall precursor, rotating stall, and surge.
  • the flow rate of the compressor decreases and the pressure ratio increases.
  • the compressor will experience flow instability, resulting in rotating stall or surge, and the flow fluctuation is extremely severe.
  • the discrimination of rotating stall usually extracts the characteristics of the precursor signal from the pulsating pressure signal of the compressor to detect the stall.
  • the detection algorithm mainly includes time-domain analysis method, frequency-domain analysis method and time-frequency analysis method.
  • the time-domain analysis method is based on the time-domain characteristic changes of the pressure signal, using variance analysis, correlation analysis and other methods to distinguish, the calculation speed is fast, and it is convenient for engineering applications, but it is highly dependent on the auxiliary value of the signal, has poor stability, and is easily affected by noise.
  • the frequency domain analysis method detects by analyzing the characteristic changes of the signal spectrogram, but there is a prerequisite for signal stationarity, so its application is limited.
  • the time-frequency analysis method combines time-domain information and spectrum features, increases the dimension of the analysis information, and can better analyze non-stationary signals, but it is not very versatile for stall signals with large differences in appearance.
  • the present invention provides a compressor rotation stall early warning method based on time-expanded convolutional network.
  • a compressor rotating stall early warning method based on time-expanded convolutional network comprising the following steps:
  • the aero-engine surge data is preprocessed, including the following steps:
  • S1.4 Sliding window technology is used to construct data set samples, and the time-domain data is divided in units of steps-sized time steps.
  • the sampling points covered by each data window form a sample, and each sample is marked with a surge or not. label 1 or 0;
  • S2.2 Build an expanded convolution module based on causal convolution and expanded convolution.
  • the basic module of each layer in the temporal convolutional network consists of two expanded convolution modules with the same kernel size and expansion factor; Batch normalization processing is performed after convolution, and the correction nonlinear unit ReLU is introduced to adjust the information passed to the next layer. After the second expansion convolution, batch normalization processing is performed, and the obtained features are the same as those extracted by the previous layer. Summing, the output features of this layer are obtained by calculating the ReLU activation function;
  • the designed Resnet-v network allows two parts of data input, one part is historical data characteristics, and the other part is data covariates; calculate the time-domain statistical characteristics of each sample, including variance, mean , the most value and other data features, plus the measurement point number corresponding to the sample, constitute a set of covariate features, which are used as one of the inputs of the Resnet-v network module;
  • S4.1 uses a similar Seq2Seq architecture to build a time-expanded convolutional network prediction model, which is divided into two parts: Encoder and Decoder, where the Encoder part is a time convolutional network module, and the Decoder part is composed of a Resnet-v network module and an output dense layer;
  • the output dense layer of the Decoder module receives the fusion output of the previous step, and through the processing of the dense layer, batch normalization layer, ReLU activation function, etc., the predicted value of the probability of surge is obtained;
  • MHL Modified HuberLoss
  • HuberLoss is a more robust loss function, which can effectively combine the advantages of MSE and MAE loss functions , to avoid the problem that MAE is not derivable when the value is 0 and MSE is greatly affected by outliers, the formula is:
  • y represents the real data value
  • f(x) represents the current predicted value
  • ⁇ is a hyperparameter that determines how to calculate the error
  • L ⁇ (y,f(x)) is the currently calculated loss value
  • the influence factor ⁇ is introduced, ⁇ [0,1], which is used to represent the impact of the samples in the surge state (category 1) being misclassified as non-surge state (category 0), and the weight coefficient ⁇ is defined as 0- 1 for:
  • the F_score indicator is as follows:
  • P is the precision rate (Precision), which indicates the proportion of the samples that are classified as positive classes that are actually positive classes: TP is the number of true cases, FP is the number of false positive cases; R is the recall rate (Recall), which indicates the proportion of the positive class in the sample that is correctly predicted: TP is the number of true cases and FN is the number of false negatives. Since the positive class is predicted to be the negative class, that is, the loss caused by the surge sample is predicted to be non-surge, so ⁇ is taken as 2 to increase the importance of the recall rate in the evaluation index.
  • step S5.1 Obtain the test set data divided after preprocessing in step S1, and adjust the data dimension according to the input requirements of the time-dilated convolutional network prediction model;
  • the method provided by the present invention integrates the pressure data change trend and time-domain statistical feature covariates to improve the prediction accuracy. Since the output of the model is the predicted probability of surge, different thresholds can be set to divide the probability to achieve hierarchical alarms, and the engine operating state can be adjusted according to the level of surge probability. Therefore, this method is conducive to improving the performance of active engine control.
  • the invention is based on data and has nothing to do with the engine structure, so the model can be easily transferred to different types of engines by training different data sets, and has certain universality.
  • Fig. 1 is a flow chart of a compressor rotating stall warning method based on a time-expanded convolutional network
  • Fig. 2 is a flow chart of data preprocessing
  • Figure 3 is a time convolutional network structure diagram
  • Figure 4 is a Resnet-v network structure diagram
  • Figure 5 is a structural diagram of the time-expanded convolutional network prediction model
  • Figure 6 is the prediction result of the time-expanded convolutional network prediction model on the test data, where (a) is the graph of dynamic pressure at the tip of the inlet zero-stage stator changing with time, and (b) is the time-expanded convolutional network prediction model.
  • the graph of the surge prediction probability changing with time, (c) is the early warning signal given by the time-expanded convolutional network prediction model;
  • Figure 7 is a comparison of the influence of covariates on the prediction effect of the time-expanded convolutional network prediction model, where (a) is the dynamic pressure at the tip of the imported second-stage stator changing with time, and (b) is the surge when the input does not have covariates The change graph of the predicted probability over time, (c) is the change graph of the predicted probability of surge with time when the input has covariates.
  • the background of the present invention is the surge experimental data of a certain type of aeroengine, and the flow chart of the compressor rotation stall warning method based on time-dilated convolutional network is shown in FIG. 1 .
  • Figure 2 is the flow chart of data preprocessing.
  • the experiment sets 10 measurement points to measure the dynamic pressure value from normal to surge for t seconds.
  • the sensor measurement frequency is 6kHz, and a total of 16 sets of data are recorded; the 10 measurement points are respectively located at the entrance Guide vane stator tips, zero-stage stator tips, first-stage stator tips (three in the circumferential direction), second-stage stator tips, third-stage stator tips, fourth-stage stator tips, fifth-stage stator tips, and outlet wall surface.
  • the data preprocessing steps are as follows:
  • the filtered data is down-sampled, and the down-sampling rate is determined according to the Nyquist sampling theorem according to the numerical distribution interval of the surge frequency;
  • FIG. 3 is a structural diagram of the temporal convolutional network.
  • the steps for constructing the temporal convolutional network module are as follows:
  • the basic module of each layer in the temporal convolution network consists of two expansion convolution modules with the same kernel size and expansion factor value.
  • the width k of the convolution filter is set to 2, and each layer of convolution uses 11 filters, and the expansion factors of each expanded convolution layer are [1, 2, 4, 8, 16, 32], so that the receptive field reaches 128,;
  • Batch normalization processing is performed after the first expansion convolution, and the correction nonlinear unit ReLU is introduced to adjust the information passed to the next layer, and batch normalization processing is performed after the second expansion convolution.
  • the residual connection between the convolution modules is used to directly obtain the feature information of the previous step while avoiding the disappearance of the gradient.
  • FIG. 4 is a diagram of the Resnet-v network structure. The steps to build a Resnet-v network are as follows:
  • the designed Resnet-v network allows two parts of data input, one part is historical data characteristics, and the other part is data covariates; calculate the time-domain statistical characteristics of each sample, including variance, mean, The data features such as the most value, plus the measurement point number corresponding to the sample, constitute a set of covariate features, which are used as one of the inputs of the Resnet-v network module;
  • Figure 5 is a structural diagram of a time-expanded convolutional network prediction model. The steps for constructing a time-expanded convolutional network prediction model are as follows:
  • Encoder is a time convolutional network module
  • Decoder is composed of a Resnet-v network module and an output dense layer
  • the output of the Encoder module is:
  • x is the current input time series
  • t represents the current moment
  • w is the convolution kernel
  • d represents the expansion factor of the time convolution network module
  • K represents the size of the kernel
  • h t is the input data of the time convolution network module at time t
  • the extracted features represent the output of the Encoder module.
  • h t is the output of the encoder
  • X t is the covariate of the input data at time t
  • R( ) is the residual function applied to the covariate X t
  • ⁇ t represents the output of the Resnet-v network module.
  • the output dense layer of the Decoder module receives the fusion output ⁇ t of the previous step, and designs the operation mode of the output dense layer according to the requirements of the model prediction task, and obtains the probability of surge through the dense layer, batch normalization layer, activation function, etc.
  • the predicted value as shown in the following formula:
  • Dense refers to the output dense layer operation set according to the prediction task.
  • the current surge probability prediction task it is specifically set to the "dense-BN-ReLU-Dropout-Soft ReLU" structure, where the Soft ReLU activation function is used for Satisfy the positive definiteness condition of the parameter; Z is the predicted probability estimate of the output.
  • Compressor surge data has the following problems, which will affect the model training effect to a certain extent:
  • the stall precursor is generally in the shape of a sharp wave. Before the arrival of the surge, the data is very stable, but after the surge occurs, the data jitters very violently, and the value changes greatly compared with the previous period. These samples are easier to classify , belonging to simple samples; samples in the process of developing from sharp wave to surge have small fluctuations or basically no fluctuations, and it is relatively difficult to identify the characteristics of these samples; the time interval from the appearance of stall precursor to the occurrence of surge is relatively short, Swift, so the easy/hard sample ratio is unbalanced;
  • the second is that different types of misclassification have different impacts. Compared with the non-surge state being judged as the surge state, the actual cost of misclassifying the surge state as a non-surge state is much higher than the former.
  • MHL Modified HuberLoss
  • HuberLoss effectively combines the advantages of both MSE and MAE loss functions, avoiding the disadvantages of MAE being non-conductive when the value is 0 and MSE being greatly affected by outliers, helping The model learns from difficult samples, and its formula is:
  • influence factor ⁇ , ⁇ [0,1] is introduced to indicate that the samples in the surge state (category 1) are misclassified as non-surge state (category 0).
  • Influence size define the weight coefficient ⁇ 0-1 as:
  • P is the precision rate (Precision), which indicates the proportion of real positive samples in the samples predicted as positive class, namely R is the recall rate (Recall), which indicates the proportion of the true positive samples that are correctly predicted, that is TP is the number of true cases, FP is the number of false positives, and FN is the number of false negatives.
  • Precision the precision rate
  • Recall the recall rate
  • the positive class is predicted to be the negative class, that is, the loss caused by the surge sample is predicted to be non-surge, so ⁇ is taken as 2 to increase the importance of the recall rate in the evaluation index. Save the model that optimizes the evaluation index to obtain the final time-expanded convolutional network prediction model.
  • Figure 6 is the prediction result of the time-expanded convolutional network prediction model on the test data, where (a) is the real-time change diagram of the dynamic pressure at the tip of the imported zero-stage stator, and (b) is the result given by the time-expanded convolutional network prediction model The real-time change diagram of surge prediction probability, (c) is the early warning signal given by the time-expanded convolutional network prediction model.
  • Figure 7 is a comparison of the influence of covariates on the prediction effect of the time-expanded convolutional network prediction model, where (a) is the dynamic pressure at the tip of the imported second-stage stator changing with time, and (b) is the surge when the input does not have covariates
  • the change graph of the predicted probability over time (c) is the change graph of the predicted probability of surge with time when the input has covariates.
  • the present invention comprehensively considers the historical data and the time-domain feature information contained in the covariates of the data, so a comparative experiment is conducted on the degree of improvement of the covariates on the prediction effect. It can be seen from Figure (a) of the prediction result diagram of the dynamic pressure data of the tip of the second-stage stator that an upwardly developing protrusion began to appear at 7.43s, which was in the initial disturbance stage of the stall. With the development of the stall disturbance, at 7.79s It began to fluctuate violently and completely developed into a stall surge. Also follow the steps in data preprocessing to process the test set data. One set of data contains covariate information, and one set of data does not contain covariate information.
  • Figure 7(b) shows that without the assistance of covariate information , the surge probability rises after about 7.49s, which is 0.05s later than the start point of the initial stage of the stall.
  • Figure 7(c) shows that the surge probability rises after 7.44s, which is basically the same as the start time of the initial stage of the stall, thus indicating that the covariate The information is helpful to predict the probability of surge.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Geometry (AREA)
  • Mechanical Engineering (AREA)
  • Computer Hardware Design (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Medical Informatics (AREA)
  • Evolutionary Biology (AREA)
  • Automation & Control Theory (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Control Of Positive-Displacement Air Blowers (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种基于时间扩张卷积网络的压气机旋转失速预警方法，首先，对航空发动机动态压力数据进行预处理，在实验数据中划分出测试数据集和训练数据集；其次，依次构建时间卷积网络模块、构建Resnet-v网络模块、构建时间扩张卷积网络预测模型，保存最优预测模型；最后，在测试数据上进行实时预测：按照时间卷积网络预测模型的输入要求调整测试集数据维度；按时间顺序，通过时间扩张卷积网络预测模型计算每个样本的喘振预测概率；通过时间扩张卷积网络预测模型计算含有协变量与不含协变量的一对样本的实时喘振概率，观察协变量对模型预测效果的提升作用。该方法综合了时域统计特征和变化趋势，提高了预测精度；有利于提高发动机主动控制的性能，具有一定的普适性。

Description

一种基于时间扩张卷积网络的压气机旋转失速预警方法 技术领域
本发明涉及一种基于时间扩张卷积网络的压气机旋转失速预警方法，属于航空发动机建模与仿真技术领域。
背景技术
航空发动机的性能稳定性直接关系到整机的飞行安全，而气路部件在整体流量、压力和能量方面维稳作用影响着发动机的整体工作状态，在各种常见的气路故障中，压气机旋转失速是其中一种破坏性极强、变化极快的故障，因此对其进行精准识别和及时预警是国内外航空发动机领域的研究重点。一般来说，压气机失稳发展过程主要有稳态、失速先兆、旋转失速和喘振四个阶段，每个阶段表现特征各异，发生机理较为复杂，传播非常迅速。稳定工作时压气机流量减少，压比增大，当流量减少到超过失稳边界时，压气机则会发生流动失稳，导致旋转失速或喘振的发生，流量波动极为剧烈。喘振发生时，发动机内的机械部件往往已经受到实质性损坏，因此，迫切需要将发动机失稳过程终止在旋转失速的初期，在未对部件造成破坏时就可识别早期微小故障征兆，从而给主动控制留出更多时间。
旋转失速的判别通常从压气机的脉动压力信号提取先兆信号的特征来检测失速，检测算法主要包括时域分析法、频域分析法和时频分析法。时域分析法依据压力信号的时域特征变化，利用方差分析、相关性分析等方法判别，计算速度快，方便工程应用，但对信号辅值依赖大，稳定性差，易受噪声影响。频域分析法通过分析信号频谱图的特征变化来检测，但有信号平稳性的前提要求，应用受限。时频分析法将时域信息和频谱特征结合，分析信息的维度增加，对非平稳信号可以更好地分析，但对于表现形态差异大的失速信号通用性不强。
发明内容
针对现有技术中准确性低，可靠性差的问题，本发明提供一种基于时间扩张卷积网络的压气机旋转失速预警方法。
本发明采用的技术方案如下：
一种基于时间扩张卷积网络的压气机旋转失速预警方法，包括以下步骤：
S1.对航空发动机喘振数据进行预处理，包括以下步骤：
S1.1导入测量点的实验数据作为数据集，采用低通滤波器，对压力变化数据进行滤波处理；
S1.2对滤波后数据进行降采样；根据喘振频率的数值分布区间，基于奈奎斯特采样定理，选定降采样率；
S1.3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；
S1.4采用滑动窗口技术构建数据集样本，以steps大小的时间步长为单位切分时域数据，每个数据窗口覆盖的采样点组成一个样本，并为每个样本标注喘振与否的标签1或0；
S1.5将整体数据集划分为训练数据集和测试数据集，再将训练数据集按3:1的比例划分为训练集和验证集；
S2.构建时间卷积网络模块，包括以下步骤：
S2.1将每个样本维度调整为(steps,1)，作为时间卷积网络模块的输入，其中steps为时间步；
S2.2搭建基于因果卷积和扩张卷积的扩张卷积模块，时间卷积网络内部每层的基本模块由两个内核大小和扩张因子数值相同的扩张卷积模块组成；在第一个扩张卷积后进行批量归一化处理，引入校正非线性单元ReLU以调整传入下一层的信息，第二个扩张卷积后再进行批量归一化处理，所得特征与上一层提取的特征加和，通过ReLU激活函数计算得到本层的输出特征；
S2.3通过堆叠多个扩张卷积模块组建时间扩张卷积网络，扩大网络接受域，采用跳跃连接保留每个卷积层的输出信息，并采用ReLU激活函数激活得到时间卷积网络模块的输出；
S3.构建Resnet-v网络模块，包括以下步骤：
S3.1考虑到喘振数据特性，设计的Resnet-v网络允许两部分数据输入，其中一部分是历史数据特征，另一部分是数据协变量；计算每个样本的时域统计特征，包括方差、均值、最值等数据特征，外加样本所对应的测量点编号，构成一组协变量特征，将其作为Resnet-v网络模块的输入之一；
S3.2通过一组密集层和批量归一化层处理输入的协变量，再应用ReLU激活函数传入下一组密集层和批量归一化层；将其输出与时间卷积网络获取的数据特征加和，通过ReLU激活函数得到Resnet-v网络的输出；
S4.构建时间扩张卷积网络预测模型，包括以下步骤：
S4.1采用类似Seq2Seq架构来构建时间扩张卷积网络预测模型，分为Encoder和Decoder两部分，其中Encoder部分为时间卷积网络模块，Decoder部分由Resnet-v网络模块和输出密集层组成；
S4.2将Encoder模块的输出特征h t输入Resnet-v网络模块，如步骤S3所述进行运算得到融合输出；
S4.3 Decoder模块的输出密集层接收上步的融合输出，通过密集层、批量归一化层、ReLU激活函数等处理，得到喘振的概率预测值；
S4.4针对喘振数据训练中存在的问题，选用MHL(Modified HuberLoss)作为损失函数；所 述HuberLoss作为一种鲁棒性更强的损失函数，能够有效结合MSE和MAE两种损失函数的优点，避免MAE在值为0时不可导问题与MSE受离群点影响大的缺点，其公式为：
Figure PCTCN2021119187-appb-000001
其中，y代表真实数据值，f(x)代表当前预测值，δ是决定如何计算误差的超参数，L δ(y,f(x))为当前计算的损失值。
另外引入影响因子β，β∈[0,1]，用来表示喘振状态的样本(类别为1)被错误分类为非喘振状态(类别为0)的影响大小，定义权重系数β 0-1为：
Figure PCTCN2021119187-appb-000002
最终损失函数形式为MHL＝β 0-1L δ(y,f(x))。
S4.5保存训练后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，评价指标采用F_score指标，保存使评价指标最优的模型得到最终的时间扩张卷积网络预测模型；
所述F_score指标如下：
Figure PCTCN2021119187-appb-000003
其中，P为精确率(Precision)，表示被分为正类的样本中实际为正类的比例：
Figure PCTCN2021119187-appb-000004
TP为真正例数，FP为假正例数；R为召回率(Recall)，表示样本中正类被正确预测的比例：
Figure PCTCN2021119187-appb-000005
TP为真正例数，FN为假负例数。由于正类被预测为负类的，即喘振样本预测为非喘振造成的损失更大，所以β取为2，增大召回率在评估指标中的重要性。
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，按照时间扩张卷积网络预测模型的输入要求调整数据维度；
S5.2采用训练后的时间扩张卷积网络预测模型计算每个样本的喘振预测概率，并按照时间顺序排序；
S5.3在测试数据中随机选取一组动压变化数据，将输入模型的数据分别整理为带有协变量参数的和不带有协变量参数的两组对比样本，采用训练后的时间扩张卷积网络预测模型，分别给出两组对比数据实时喘振概率，以观察协变量对模型预测效果的帮助性。
本发明的有益效果为：
通过本发明所提供方法对压气机旋转失速预警，相比于以往时域分析方法，综合了压力 数据变化趋势和时域统计特征协变量，提高了预测精度。由于模型的输出为喘振预测概率，可以设置不同阈值划分概率以实现分级报警，根据喘振概率高低调整发动机运行状态，因此该方法有利于提高发动机主动控制的性能。本发明基于数据，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为基于时间扩张卷积网络的压气机旋转失速预警方法流程图；
图2为数据预处理流程图；
图3为时间卷积网络结构图；
图4为Resnet-v网络结构图；
图5为时间扩张卷积网络预测模型结构图
图6为时间扩张卷积网络预测模型在测试数据上的预测结果图，其中(a)为进口零级静子尖部动压随时间变化图，(b)为时间扩张卷积网络预测模型给出的喘振预测概率随时间的变化图，(c)为时间扩张卷积网络预测模型给出的预警信号；
图7为协变量对时间扩张卷积网络预测模型预测效果的影响对比图，其中(a)为进口二级静子尖部动压随时间变化图，(b)为输入不带协变量时喘振预测概率随时间的变化图，(c)为输入带有协变量时喘振预测概率随时间的变化图。
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喘振实验数据，基于时间扩张卷积网络的压气机旋转失速预警方法流程如图1所示。
图2为数据预处理流程图，实验设置10个测量点，测量从正常到喘振共t秒的动态压力数值，传感器测量频率为6kHz，共记录16组数据；10个测量点分别位于：进口导向叶片静子尖部、零级静子尖部、一级静子尖部(周向三个)、二级静子尖部、三级静子尖部、四级静子尖部、五级静子尖部、出口壁面。数据预处理步骤如下：
S1.采用低通滤波器，对训练数据集中所有测量点测量的压力变化数据滤波处理；
S2.为节约计算资源，对滤波后数据进行降采样，根据喘振频率的数值分布区间，依据奈奎斯特采样定理确定降采样率；
S3.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；
S4.采用滑动窗口技术构建数据集样本，按时间步长大小steps切分时域数据，由数据窗口覆盖的采样时刻过程变量组成一个样本，并为每个样本标注喘振与否的标签1或0；
S5.将整体数据集划分为训练数据集和测试数据集，再将训练数据集照按3:1的比例划分为训练集和验证集。
图3为时间卷积网络结构图，构建时间卷积网络模块的步骤如下：
S1.将每个样本维度调整为(steps,1)，作为时间卷积网络模块的输入，其中steps为时间步长大小；
S2.搭建基于因果卷积和扩张卷积的扩张卷积模块，时间卷积网络内部每层的基本模块由两个内核大小和扩张因子数值相同的扩张卷积模块组成。卷积滤波器的宽度k设为2，每层卷积采用11个滤波器，每个扩张卷积层的扩张因子分别为[1、2、4、8、16、32]，使感受野达到128，；在第一个扩张卷积后进行批量归一化处理，引入校正非线性单元ReLU以调整传入下一层的信息，第二个扩张卷积后再进行批量归一化处理。卷积模块之间采取残差连接，直接前步特征信息，同时避免梯度消失。残差连接将本层所得特征与上一层的输出特征加和，即进行x l+1＝f(x l,w l)+x l运算，所得输出再通过ReLU激活函数，最终得到本层的输出特征；
S3.通过堆叠多个扩张卷积模块组建时间卷积网络，扩大网络接受域，通过信息叠加保留每个卷积层的输出特征信息，并采用ReLU激活函数激活得到时间卷积网络模块的输出。
图4为Resnet-v网络结构图，构建Resnet-v网络的步骤如下：
S1.考虑到喘振数据特性，设计的Resnet-v网络允许两部分数据输入，其中一部分是历史数据特征，另一部分是数据协变量；计算每个样本的时域统计特征，包括方差、均值、最值等数据特征，外加样本所对应的测量点编号，构成一组协变量特征，将其作为Resnet-v网络模块的输入之一；
S2.通过一组密集层和批量归一化层处理输入的协变量，再应用ReLU激活函数传入下一组密集层和批量归一化层；将其输出与时间卷积网络获取的数据特征加和，通过ReLU激活函数得到Resnet-v网络的输出。
图5为时间扩张卷积网络预测模型结构图，构建时间扩张卷积网络预测模型的步骤如下：
S1.采用类似Seq2Seq架构来构建时间扩张卷积网络预测模型，分为Encoder和Decoder两部分，其中Encoder部分为时间卷积网络模块，Decoder部分由Resnet-v网络模块和输出密集层组成；
Encoder模块的输出为：
Figure PCTCN2021119187-appb-000006
其中，x为当前输入时间序列，t代表当前时刻，w是卷积核；d代表时间卷积网络模块的扩张因子，K代表内核的大小；h t是时间卷积网络模块对t时刻输入数据所提取的特征，代表Encoder模块的输出。
S2.计算数据协变量X t，同Encoder模块的输出特征h t一起输入到Resnet-v网络模块，得到融合输出为：
δ t＝R(X t)+h t
其中，h t是编码器的输出，X t是t时刻所输入数据的协变量，R(·)是应用在协变量X t的残差函数，δ t代表Resnet-v网络模块的输出。
S3.Decoder模块的输出密集层接收上步的融合输出δ t，根据模型预测任务的要求设计输出密集层运算模式，通过密集层、批量归一化层、激活函数等处理，得到喘振的概率预测值，如下式所示：
Z＝Dense(δ t)
其中，Dense是指的根据预测任务设定的输出密集层运算，在当前喘振概率预测任务中具体设置为“dense-BN-ReLU-Dropout-Soft ReLU”结构，其中使用Soft ReLU激活函数是为了满足参数的正定性条件；Z是输出的预测概率估计值。
S4.压气机喘振数据具有以下问题，在一定程度上会影响模型训练效果：
一是简单/困难样本不平衡，失速先兆一般为突尖波形态，在其到来之前数据非常平稳，而喘振发生后数据抖动非常剧烈，数值与前期相比变化较大，这些样本比较容易分类，属于简单样本；从突尖波发展到喘振过程中的样本有小幅波动或基本没有波动，这些样本在识别特征时相对困难；从失速先兆出现到喘振发生之间时间间隔较短，发展迅速，因此简单/困难样本比例不平衡；
二是不同类别的错误分类所造成的影响不同，相比于非喘振状态被判别为喘振状态，将喘振状态错误分为非喘振状态所造成的实际代价远高于前者。
为解决以上问题，选用MHL(Modified HuberLoss)作为损失函数。
HuberLoss作为一种鲁棒性更强的损失函数，有效的结合了MSE和MAE两种损失函数的优点，避免了MAE在值为0时不可导情况与MSE受离群点影响大的缺点，帮助模型对困难样本的学习，其公式为：
Figure PCTCN2021119187-appb-000007
另外，针对错误判别所造成的问题，引入影响因子β，β∈[0,1]，用来表示喘振状态的样本(类别为1)被错误分类为非喘振状态(类别为0)的影响大小，定义权重系数β 0-1为：
Figure PCTCN2021119187-appb-000008
最终损失函数形式为MHL＝β 0-1L δ(y,f(x))。
S5.保存训练后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，评价指标采用F_score指标：
Figure PCTCN2021119187-appb-000009
其中，P为精确率(Precision)，表示被预测为正类的样本中真实正样本所占比例，即
Figure PCTCN2021119187-appb-000010
R为召回率(Recall)，表示真实正样本中被正确预测的比例，即
Figure PCTCN2021119187-appb-000011
TP为真正例数，FP为假正例数，FN为假负例数。
由于正类被预测为负类的，即喘振样本预测为非喘振造成的损失更大，所以β取为2，增大召回率在评估指标中的重要性。保存使评价指标最优的模型得到最终时间扩张卷积网络预测模型。
图6为时间扩张卷积网络预测模型在测试数据上的预测结果图，其中(a)为进口零级静子尖部动压实时变化图，(b)为时间扩张卷积网络预测模型给出的喘振预测概率实时变化图，(c)为时间扩张卷积网络预测模型给出的预警信号。图7为协变量对时间扩张卷积网络预测模型预测效果的影响对比图，其中(a)为进口二级静子尖部动压随时间变化图，(b)为输入不带协变量时喘振预测概率随时间的变化图，(c)为输入带有协变量时喘振预测概率随时间的变化图。在测试数据上进行实时预测的步骤如下：
S1.从零级静子尖部动压数据的预测结果图中(a)可以看出，7.95s开始出现了一个向上的尖波，处于失速初始扰动阶段，随着失速扰动的发展，在8.65s开始有剧烈的波动，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，调整数据维度后输入训练后的时间扩张卷积网络预测模型进行预测。结合图6(b)和图6(c)可以看出，在7.95s左右喘振概率从0提升至80％的预测值，随后喘振概率在7.95s到8.25s的区间一直维持较高的预测值，8.3s后由于原始动压的数据处于较为平稳阶段，喘振概率回落，之后在8.62s处喘振概率跟随原始数据的波动而再次升高。根据喘振发生原理，失速先兆的发生代表旋转失速和喘振的大概率出现，因此在7.95s处预判出失速先兆时，系统便在初始扰动阶段给出预警信号，以防对部件造成破坏。
S2.本发明综合考虑历史数据与数据协变量含有的时域特征信息，因此针对协变量对预测效果的改进程度做了对比实验。从二级静子尖部动压数据的预测结果图的图(a)可以看出，7.43s开始出现了一个向上发展的突尖，处于失速初始扰动阶段，随着失速扰动的发展，在7.79s开始有剧烈的波动，彻底发展为失速喘振。同样按数据预处理中步骤对测试集数据处理，一组数据含有协变量信息，一组数据不含协变量信息，调整数据维度后进行预测；图7(b)显示在没有协变量信息辅助下，喘振概率约在7.49s之后上升，比失速初始阶段起始点晚了0.05s， 图7(c)显示喘振概率在7.44s之后上升，与失速初始阶段起步时间基本持平，因此表明协变量信息对于喘振概率预测有一定帮助。
以上所述实施例仅表达本发明的实施方式，但并不能因此而理解为对本发明专利的范围的限制，应当指出，对于本领域的技术人员来说，在不脱离本发明构思的前提下，还可以做出若干变形和改进，这些均属于本发明的保护范围。

Claims (4)

  1. 一种基于时间扩张卷积网络的压气机旋转失速预警方法，其特征在于，包括以下步骤：
    S1.对航空发动机喘振数据进行预处理，包括以下步骤：
    S1.1导入测量点的实验数据作为数据集，对压力变化数据进行滤波处理；
    S1.2对滤波后数据进行降采样；
    S1.3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；
    S1.4构建数据集样本，以steps大小的时间步长为单位切分时域数据，每个数据窗口覆盖的采样点组成一个样本，并为每个样本标注喘振与否的标签1或0；
    S1.5将整体数据集划分为训练数据集和测试数据集，再将训练数据集按比例划分为训练集和验证集；
    S2.构建时间卷积网络模块，包括以下步骤：
    S2.1将每个样本维度调整为(steps,1)，作为时间卷积网络模块的输入，其中steps为时间步；
    S2.2搭建基于因果卷积和扩张卷积的扩张卷积模块，时间卷积网络内部每层的基本模块由两个内核大小和扩张因子数值相同的扩张卷积模块组成；在第一个扩张卷积后进行批量归一化处理，引入校正非线性单元ReLU以调整传入下一层的信息，第二个扩张卷积后再进行批量归一化处理，所得特征与上一层提取的特征加和，通过ReLU激活函数计算得到本层的输出特征；
    S2.3通过堆叠多个扩张卷积模块组建时间扩张卷积网络，扩大网络接受域，采用跳跃连接保留每个卷积层的输出信息，并采用ReLU激活函数激活得到时间卷积网络模块的输出；
    S3.构建Resnet-v网络模块，包括以下步骤：
    S3.1考虑到喘振数据特性，Resnet-v网络允许两部分数据输入，其中一部分是历史数据特征，另一部分是数据协变量；计算每个样本的时域统计特征，和样本所对应的测量点编号，构成一组协变量特征，将其作为Resnet-v网络模块的输入之一；
    S3.2通过一组密集层和批量归一化层处理输入的协变量，再应用ReLU激活函数传入下一组密集层和批量归一化层；将其输出与时间卷积网络获取的数据特征加和，通过ReLU激活函数得到Resnet-v网络的输出；
    S4.构建时间扩张卷积网络预测模型，包括以下步骤：
    S4.1构建时间扩张卷积网络预测模型，分为Encoder和Decoder两部分，其中Encoder部分为时间卷积网络模块，Decoder部分由Resnet-v网络模块和输出密集层组成；
    S4.2将Encoder模块的输出特征h t输入Resnet-v网络模块，如步骤S3所述进行运算得到融合输出；
    S4.3 Decoder模块的输出密集层接收上步的融合输出，依次通过密集层、批量归一化层、ReLU激活函数处理，得到喘振的概率预测值；
    S4.4针对喘振数据训练中存在的问题，选用MHL作为损失函数，其公式为：
    Figure PCTCN2021119187-appb-100001
    其中，y代表真实数据值，f(x)代表当前预测值，δ是决定如何计算误差的超参数，L δ(y,f(x))为当前计算的损失值；
    另外引入影响因子β，β∈[0,1]，用来表示喘振状态的样本被错误分类为非喘振状态的影响大小，定义权重系数β 0-1为：
    Figure PCTCN2021119187-appb-100002
    最终损失函数形式为MHL＝β 0-1L δ(y,f(x))；
    S4.5保存训练后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，评价指标采用F_score指标，保存使评价指标最优的模型得到最终的时间扩张卷积网络预测模型；
    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，按照时间扩张卷积网络预测模型的输入要求调整数据维度；
    S5.2采用训练后的时间扩张卷积网络预测模型计算每个样本的喘振预测概率，并按照时间顺序排序；
    S5.3在测试数据中随机选取一组动压变化数据，将输入模型的数据分别整理为带有协变量参数的和不带有协变量参数的两组对比样本，采用训练后的时间扩张卷积网络预测模型，分别给出两组对比数据实时喘振概率。
  2. 根据权利要求1所述的一种基于时间扩张卷积网络的压气机旋转失速预警方法，其特征在于，所述步骤1.2中，根据喘振频率的数值分布区间，基于奈奎斯特采样定理，选定降采样率。
  3. 根据权利要求1所述的一种基于时间扩张卷积网络的压气机旋转失速预警方法，其特征在于，所述步骤1.5中，训练集和验证集的比例为3:1。
  4. 根据权利要求1所述的一种基于时间扩张卷积网络的压气机旋转失速预警方法，其特征在于，所述步骤S4.5中F_score指标如下：
    Figure PCTCN2021119187-appb-100003
    其中，P为精确率，表示被分为正类的样本中实际为正类的比例：
    Figure PCTCN2021119187-appb-100004
    TP为真正例数，FP为假正例数；R为召回率，表示样本中正类被正确预测的比例：
    Figure PCTCN2021119187-appb-100005
    TP为真正例数，FN为假负例数；β取为2。
PCT/CN2021/119187 2021-08-06 2021-09-18 一种基于时间扩张卷积网络的压气机旋转失速预警方法 WO2023010658A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/795,095 US20230316051A1 (en) 2021-08-06 2021-09-18 Pre-alarming method for rotary stall of compressors based on temporal dilated convolutional neural network

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110899542.7 2021-08-06
CN202110899542.7A CN113569338B (zh) 2021-08-06 2021-08-06 一种基于时间扩张卷积网络的压气机旋转失速预警方法

Publications (1)

Publication Number Publication Date
WO2023010658A1 true WO2023010658A1 (zh) 2023-02-09

Family

ID=78170633

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/119187 WO2023010658A1 (zh) 2021-08-06 2021-09-18 一种基于时间扩张卷积网络的压气机旋转失速预警方法

Country Status (3)

Country Link
US (1) US20230316051A1 (zh)
CN (1) CN113569338B (zh)
WO (1) WO2023010658A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116504005A (zh) * 2023-05-09 2023-07-28 齐鲁工业大学(山东省科学院) 一种基于改进CDIL-Bi-LSTM的周界安防入侵信号识别方法
CN116579230A (zh) * 2023-03-31 2023-08-11 中国长江电力股份有限公司 基于tcn神经网络的闸门启闭状态实时预警方法及装置
CN117474925A (zh) * 2023-12-28 2024-01-30 山东润通齿轮集团有限公司 一种基于机器视觉的齿轮点蚀检测方法及系统

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114186337A (zh) * 2021-11-30 2022-03-15 大连理工大学 一种基于多源数据融合的压气机旋转失速预测方法
CN114548555B (zh) * 2022-02-22 2022-09-13 大连理工大学 一种深度自回归网络的轴流压气机失速喘振预测方法
CN114857062B (zh) * 2022-04-13 2023-03-24 大连理工大学 一种基于tcn的航空发动机失稳状态检测系统
CN117609704B (zh) * 2024-01-23 2024-04-19 青岛高科通信股份有限公司 一种基于神经网络模型的智能电表的电计量误差分析装置
CN117892099B (zh) * 2024-03-18 2024-06-18 湖南科技大学 基于概率神经网络的风电机组服役效能评估方法和系统
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

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111737910A (zh) * 2020-06-10 2020-10-02 大连理工大学 一种基于深度学习的轴流压气机失速喘振预测方法
CN112580267A (zh) * 2021-01-13 2021-03-30 南京航空航天大学 基于多分支特征融合网络的航空发动机喘振预测方法
CN112668775A (zh) * 2020-12-25 2021-04-16 西安翔迅科技有限责任公司 一种基于时序卷积网络算法的空气质量预测方法

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3486816A1 (en) * 2017-11-16 2019-05-22 Institut Pasteur Method, device, and computer program for generating protein sequences with autoregressive neural networks
US11315570B2 (en) * 2018-05-02 2022-04-26 Facebook Technologies, Llc Machine learning-based speech-to-text transcription cloud intermediary
EP3608844A1 (en) * 2018-08-10 2020-02-12 Naver Corporation Methods for training a crnn and for semantic segmentation of an inputted video using said crnn
CN111192245B (zh) * 2019-12-26 2023-04-07 河南工业大学 一种基于U-Net网络的脑肿瘤分割网络及分割方法
CN112365040B (zh) * 2020-11-03 2022-03-01 哈尔滨工业大学 一种基于多通道卷积神经网络和时间卷积网络的短期风电功率预测方法
CN112990536B (zh) * 2021-01-27 2022-06-03 广东工业大学 一种基于分位数卷积网络的公交到站时间预测方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111737910A (zh) * 2020-06-10 2020-10-02 大连理工大学 一种基于深度学习的轴流压气机失速喘振预测方法
CN112001128A (zh) * 2020-06-10 2020-11-27 大连理工大学 一种基于深度学习的轴流压气机失速喘振预测方法
CN112668775A (zh) * 2020-12-25 2021-04-16 西安翔迅科技有限责任公司 一种基于时序卷积网络算法的空气质量预测方法
CN112580267A (zh) * 2021-01-13 2021-03-30 南京航空航天大学 基于多分支特征融合网络的航空发动机喘振预测方法

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
LIU ZHUO: "Research on the Auxiliary Discriminant Method for Stall and Surge Conditions of Aero-Engine", JILIANG-YU-CESHI-JISHU = METROLOGY AND MEASUREMENT TECHNIQUE, JILIANG YU CESHI JISHU ZASHISHE, CN, vol. 46, no. 3, 31 March 2019 (2019-03-31), CN , pages 70 - 72, XP093031499, ISSN: 1004-6941, DOI: 10.15988/j.cnki.1004-6941.2019.3.021 *
SHAOJIE BAI; J. ZICO KOLTER; VLADLEN KOLTUN: "An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 4 March 2018 (2018-03-04), 201 Olin Library Cornell University Ithaca, NY 14853 , XP081228078 *
ZHAO HONGYANG, QUAN FUXIANG, XIA WEIGUO, SUN XIMING: "Detection of Spike-type Stall of Axial Compressors Based on Dilated Causal Convolutional Neural Networks ", CISAI 2020 JOURNAL OF PHYSICS: CONFERENCE SERIES, IOP PUBLISHING, 27 September 2020 (2020-09-27), pages 12028, XP093031370, Retrieved from the Internet <URL:https://iopscience.iop.org/article/10.1088/1742-6596/1693/1/012028/pdf> [retrieved on 20230314] *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116579230A (zh) * 2023-03-31 2023-08-11 中国长江电力股份有限公司 基于tcn神经网络的闸门启闭状态实时预警方法及装置
CN116504005A (zh) * 2023-05-09 2023-07-28 齐鲁工业大学(山东省科学院) 一种基于改进CDIL-Bi-LSTM的周界安防入侵信号识别方法
CN116504005B (zh) * 2023-05-09 2024-02-20 齐鲁工业大学(山东省科学院) 一种基于改进CDIL-Bi-LSTM的周界安防入侵信号识别方法
CN117474925A (zh) * 2023-12-28 2024-01-30 山东润通齿轮集团有限公司 一种基于机器视觉的齿轮点蚀检测方法及系统
CN117474925B (zh) * 2023-12-28 2024-03-15 山东润通齿轮集团有限公司 一种基于机器视觉的齿轮点蚀检测方法及系统

Also Published As

Publication number Publication date
US20230316051A1 (en) 2023-10-05
CN113569338B (zh) 2022-10-14
CN113569338A (zh) 2021-10-29

Similar Documents

Publication Publication Date Title
WO2023010658A1 (zh) 一种基于时间扩张卷积网络的压气机旋转失速预警方法
WO2021248746A1 (zh) 一种基于深度学习的轴流压气机失速喘振预测方法
WO2023056614A1 (zh) 一种基于堆叠长短期记忆网络的轴流压气机旋转失速预测方法
US9818242B2 (en) Gas turbine engine anomaly detections and fault identifications
US11669080B2 (en) Abnormality detection device, abnormality detection method, and program
WO2023097705A1 (zh) 一种基于多源数据融合的压气机旋转失速预测方法
WO2023159336A1 (zh) 一种深度自回归网络的轴流压气机失速喘振预测方法
CN113485302B (zh) 基于多元时序数据的车辆运行过程故障诊断方法及系统
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
WO2023103268A1 (zh) 泵阀故障检测方法
CN115098962A (zh) 一种基于隐半马尔夫模型的机械设备退化态下剩余寿命的预测方法
CN114708885A (zh) 一种基于声音信号的风机故障预警方法
Oliinyk et al. Synthesis of neuro-fuzzy networks on the basis of association rules
Viale et al. Least squares smoothed k-nearest neighbors online prediction of the remaining useful life of a NASA turbofan
CN109894495B (zh) 一种基于能耗数据和贝叶斯网络的挤压机异常检测方法及系统
Mi et al. A nonparametric cumulative sum-based fault detection method for rolling bearings using high-level extended isolated forest
Quan et al. Detection of Rotating Stall Inception of Axial Compressors Based on Deep Dilated Causal Convolutional Neural Networks
CN113836787A (zh) 基于区分性特征学习的超声速进气道流动状态监测方法
Ma et al. An interpretability research of the Xgboost algorithm in remaining useful life prediction
CN110826600B (zh) 基于自适应共振网络在线增量学习的发动机喘振预测方法
CN114548555B (zh) 一种深度自回归网络的轴流压气机失速喘振预测方法
Chen et al. Prediction and classification of faults in electric submersible pumps
CN116227172A (zh) 一种基于卷积神经网络的滚动轴承性能退化评估方法
Liu et al. Incipient fault detection enhancement based on spatial-temporal multi-mode siamese feature contrast learning for industrial dynamic process
CN111913463B (zh) 一种核电厂化学容积控制系统状态监测方法

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: 21952527

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21952527

Country of ref document: EP

Kind code of ref document: A1