WO2023137807A1 - 滚动轴承类不平衡故障诊断方法及系统 - Google Patents

滚动轴承类不平衡故障诊断方法及系统 Download PDF

Info

Publication number
WO2023137807A1
WO2023137807A1 PCT/CN2022/076031 CN2022076031W WO2023137807A1 WO 2023137807 A1 WO2023137807 A1 WO 2023137807A1 CN 2022076031 W CN2022076031 W CN 2022076031W WO 2023137807 A1 WO2023137807 A1 WO 2023137807A1
Authority
WO
WIPO (PCT)
Prior art keywords
time
frequency
fault diagnosis
fault
frequency analysis
Prior art date
Application number
PCT/CN2022/076031
Other languages
English (en)
French (fr)
Inventor
王俊
许雪梅
丁传仓
沈长青
朱忠奎
Original Assignee
苏州大学
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 苏州大学 filed Critical 苏州大学
Publication of WO2023137807A1 publication Critical patent/WO2023137807A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/04—Bearings
    • G01M13/045—Acoustic or vibration analysis
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield 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
    • 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
    • G06N3/08—Learning methods

Definitions

  • the invention relates to the technical field of bearing fault diagnosis, in particular to a method and system for unbalanced fault diagnosis of rolling bearings.
  • Rolling bearings are the most commonly used general-purpose parts in various types of rotating machinery, and they are also one of the most vulnerable parts. Due to the complex operating conditions of rotating machinery and the long-term work of bearings under heavy load and high speed, failures are prone to occur. Therefore, monitoring the health status of bearings, timely and accurately diagnosing faults, and handling them as soon as possible can avoid major economic losses and safety accidents, which has important practical significance.
  • deep learning is widely used in rolling bearing fault diagnosis, but there are still some problems that need to be solved urgently in the application of traditional deep learning methods in actual bearing fault diagnosis.
  • the deep learning model requires a large amount of balanced historical data of the bearing condition for training.
  • the amount of bearing normal state data is often far greater than the amount of fault state data, which leads to the problem of class imbalance, which brings challenges to the training of fault diagnosis models. If a small number of samples contain too few features, it will be difficult for the model to learn the distribution of its data, and it will over-rely on limited data samples, which will lead to over-fitting problems, making the model's recognition of the fault state of bearings with few samples less accurate.
  • deep learning models are sensitive to clear features and are easily disturbed by noise.
  • data augmentation methods are often used, which generate sufficient generated samples from limited minority class data samples to balance the data set.
  • Commonly used data enhancement methods are divided into single-sample data enhancement and multi-sample data enhancement.
  • the former includes image geometric transformation, image color transformation, etc.
  • the latter includes Synthetic Minority Oversampling Technology (SMOTE), Generative Adversarial Network (GAN), Auxiliary Classification Generative Adversarial Network (ACGAN), etc.
  • SMOTE Synthetic Minority Oversampling Technology
  • GAN Generative Adversarial Network
  • ACGAN Auxiliary Classification Generative Adversarial Network
  • the generated samples obtained by the single-sample data augmentation method are quite different from the original samples, and forcing the "label invariance" constraint between the enhanced samples and the original samples may damage the model performance.
  • image geometric transformation is not suitable for classification with orientation information
  • image color transformation is not suitable for classification where color information is important.
  • the SMOTE algorithm has two limitations. One is that there is a certain blindness in the selection of neighbors, and the other is that it cannot learn the data distribution of the minority class, which is prone to distribution marginalization. GAN and ACGAN generate minority class samples by constructing a model, and the stability of the model training is poor, requiring a high amount of calculation and calculation time.
  • the above data enhancement methods are likely to cause a large difference in the data distribution between the generated sample and the original sample, and cannot guarantee the consistency between the generated sample and the original sample.
  • the parameter optimization process cannot be carried out simultaneously with the training of the deep learning model, and end-to-end bearing fault diagnosis cannot be realized.
  • the time-frequency model under the optimal parameters only reflects a kind of bearing fault state information, ignoring other forms of state information.
  • this type of method can effectively filter out the noise outside the fault frequency band, the noise within the fault frequency band is not removed.
  • the technical problem to be solved by the present invention is to overcome the problems existing in the prior art, and propose a rolling bearing unbalanced fault diagnosis method and system, whose expanded data samples have high consistency and diversity, without optimizing the parameters of time-frequency analysis, and can effectively filter out the noise in the time-frequency feature, and the time-frequency feature extraction can be integrated with the fault diagnosis model training, greatly improving the accuracy of bearing unbalanced fault diagnosis.
  • the present invention provides a method for diagnosing unbalanced faults of rolling bearings, which includes the following steps:
  • S10 Perform variable-parameter time-frequency analysis on the minority class of fault samples obtained in the fault state, and perform single-parameter time-frequency analysis on the majority class of normal samples obtained in the normal state to obtain a class-balanced time-frequency feature data set;
  • S20 Design a time-frequency attention mechanism network model, use the time-frequency attention mechanism network model to perform feature enhancement processing on the time-frequency feature dataset, and obtain a time-frequency feature enhancement dataset;
  • S50 Perform time-frequency analysis of a single parameter on the vibration signal of a bearing in an unknown health state to obtain a time-frequency feature, and input the time-frequency feature into a trained fault diagnosis model to obtain the health state of the bearing.
  • carry out variable parameter time-frequency analysis to the minority class fault sample that obtains under fault state comprise:
  • Different parameters are used for time-frequency analysis, and the number of parameters is to make the number of time-frequency characteristics of fault samples the same as the number of normal samples.
  • a single parameter time-frequency analysis is performed on the majority of normal samples obtained in the normal state, including:
  • the time-frequency analysis is performed with a fixed parameter, which is one of multiple parameters in the variable-parameter time-frequency analysis.
  • using the time-frequency attention mechanism network model to perform feature enhancement processing on the time-frequency feature data set including:
  • time-frequency attention mechanism network model to assign different weights to each frequency information and time information in the time-frequency feature data set for feature enhancement.
  • the method of time-frequency analysis of a single parameter in S10 is the same as the method of time-frequency analysis of a single parameter in S50.
  • the model training method includes an adaptive moment estimation algorithm, a stochastic gradient descent method, and a root mean square transfer algorithm.
  • the present invention also provides a rolling bearing unbalance fault diagnosis system, including:
  • a data enhancement module the data enhancement module is used to perform variable parameter time-frequency analysis on the minority class fault samples obtained under the fault state, and perform single parameter time-frequency analysis on the majority class normal samples obtained under the normal state, so as to obtain a class-balanced time-frequency feature data set;
  • a feature enhancement module the feature enhancement module is used to design a time-frequency attention mechanism network model, and utilizes the time-frequency attention mechanism network model to perform feature enhancement processing on the time-frequency feature data set to obtain a time-frequency feature enhancement data set;
  • a fault diagnosis model construction module the fault diagnosis model construction module is used to build a deep learning network model, the deep learning network model is embedded in the back end of the time-frequency attention mechanism network model, and a fault diagnosis model is constructed;
  • a model training module the model training module is used to use the time-frequency feature enhancement data set to train the constructed fault diagnosis model to obtain a trained fault diagnosis model;
  • a fault diagnosis module the fault diagnosis module is used to perform a single parameter time-frequency analysis on the vibration signal of a bearing in an unknown health state to obtain a time-frequency feature, and input the time-frequency feature into a trained fault diagnosis model to obtain the health state of the bearing.
  • the data enhancement module includes:
  • the fault sample data enhancement sub-module the fault sample data enhancement sub-module is used to perform variable parameter time-frequency analysis on the minority class fault samples obtained under the fault state, adopting different parameters for time-frequency analysis, and the number of parameters is to make the number of fault sample time-frequency characteristics the same as the number of normal samples.
  • the data enhancement module includes:
  • a normal sample time-frequency analysis sub-module the normal sample time-frequency analysis sub-module is used to perform a single parameter time-frequency analysis on most types of normal samples obtained in a normal state, using a fixed parameter for time-frequency analysis, which is one of multiple parameters in the variable parameter time-frequency analysis.
  • the feature enhancement module includes:
  • the time-frequency feature enhancement sub-module is used to use the time-frequency attention mechanism network model to assign different weights to each frequency information and time information in the time-frequency feature data set for feature enhancement.
  • the present invention uses the time-frequency analysis method under different parameters to expand the data volume of minority types of fault samples, and considers the state information of different time-frequency modes of bearing faults.
  • the time-frequency features of the same fault type have the consistency of the fault occurrence time center and the frequency band center, and the diversity of time-frequency state features; the time-frequency attention mechanism is used to enhance the time-frequency state features on the time axis and frequency axis respectively, effectively filtering the noise in the time-frequency features.
  • the expanded data samples have high consistency and diversity, without optimizing the parameters of the time-frequency analysis, and can effectively filter out the noise in the time-frequency features, and the time-frequency feature extraction can be integrated with the fault diagnosis model training, which greatly improves the accuracy of bearing unbalanced fault diagnosis.
  • Fig. 1 is a flow chart of a method for diagnosing unbalanced faults of rolling bearings disclosed in an embodiment of the present invention.
  • Fig. 2 is a schematic diagram of data enhancement based on variable parameter time-frequency analysis in an embodiment of the present invention.
  • Fig. 3 is a schematic diagram of a time-frequency attention mechanism network model in an embodiment of the present invention.
  • Fig. 4 is a schematic diagram of a deep learning network model in an embodiment of the present invention.
  • FIG. 5 is a time-frequency characteristic diagram of each fault state before and after the time-frequency attention mechanism network model in the trained fault diagnosis model in the embodiment of the present invention.
  • this embodiment provides a method for diagnosing rolling bearing imbalance faults, including the following steps:
  • S10 Perform variable-parameter time-frequency analysis on the minority class of fault samples obtained in the fault state, and perform single-parameter time-frequency analysis on the majority class of normal samples obtained in the normal state to obtain a class-balanced time-frequency feature data set;
  • S20 Design a time-frequency attention mechanism network model, use the time-frequency attention mechanism network model to perform feature enhancement processing on the time-frequency feature dataset, and obtain a time-frequency feature enhancement dataset;
  • S50 Perform time-frequency analysis of a single parameter on the vibration signal of a bearing in an unknown health state to obtain a time-frequency feature, and input the time-frequency feature into a trained fault diagnosis model to obtain the health state of the bearing.
  • time-frequency analysis is a method that can transform time series signals into the time-frequency domain, including but not limited to wavelet transform (WT), short-time Fourier transform (STFT), and Wigner-Willi distribution (WVD).
  • WT wavelet transform
  • STFT short-time Fourier transform
  • WVD Wigner-Willi distribution
  • performing variable parameter time-frequency analysis on a minority of fault samples obtained under fault conditions includes using different parameters for time-frequency analysis, and the number of parameters is to make the number of time-frequency characteristics of fault samples the same as the number of normal samples; performing single-parameter time-frequency analysis on most types of normal samples obtained under normal conditions includes using a fixed parameter to perform time-frequency analysis, and this parameter is one of the multiple parameters in variable parameter time-frequency analysis.
  • the time range and frequency range of the above time-frequency features are consistent, and the size of the time-frequency features is consistent.
  • using the time-frequency attention mechanism network model to perform feature enhancement processing on the time-frequency feature data set includes using the time-frequency attention mechanism network model to assign different weights to each frequency information and time information in the time-frequency feature data set to perform feature enhancement.
  • the above-mentioned time-frequency attention mechanism network model is constructed by artificial neural networks, including but not limited to fully connected neural networks, convolutional neural networks, deep belief networks, deep residual networks, autoencoders, recurrent neural networks, recurrent neural networks, restricted Boltzmann machines, and generative adversarial networks.
  • the deep learning network model is constructed by an artificial neural network, including but not limited to a fully connected neural network, a convolutional neural network, a deep belief network, a deep residual network, an autoencoder, a recurrent neural network, a recurrent neural network, a restricted Boltzmann machine, and a generated confrontation network.
  • an artificial neural network including but not limited to a fully connected neural network, a convolutional neural network, a deep belief network, a deep residual network, an autoencoder, a recurrent neural network, a recurrent neural network, a restricted Boltzmann machine, and a generated confrontation network.
  • the model training methods include but not limited to Adaptive Moment Estimation Algorithm (Adam), Stochastic Gradient Descent (SGD), Root Mean Square Transfer Algorithm (RmsPorp).
  • Adam Adaptive Moment Estimation Algorithm
  • SGD Stochastic Gradient Descent
  • RmsPorp Root Mean Square Transfer Algorithm
  • the single parameter time-frequency analysis method in S10 is the same as the single parameter time-frequency analysis method in S50.
  • the present invention uses the time-frequency analysis method under different parameters to expand the data volume of minority types of fault samples, and considers the state information of different time-frequency modes of bearing faults.
  • the time-frequency features of the same fault type have the consistency of the fault occurrence time center and the frequency band center, and the diversity of time-frequency state features; the time-frequency attention mechanism is used to enhance the time-frequency state features on the time axis and frequency axis respectively, effectively filtering the noise in the time-frequency features.
  • the expanded data samples have high consistency and diversity, without optimizing the parameters of the time-frequency analysis, and can effectively filter out the noise in the time-frequency features, and the time-frequency feature extraction can be integrated with the fault diagnosis model training, which greatly improves the accuracy of bearing unbalanced fault diagnosis.
  • the vibration data of wheel-set bearings under different health states are collected by using the wheel-set bearing test bench.
  • the test table uses a sensor motor to drive a small wheel pair through the motor belt.
  • the two hydraulic cylinders add a radial load on both sides of the small wheel pair through the two wheel pairs.
  • the lotus transmits to the parallel wheel pair, and the two wheels of the big theory are in contact with the two wheel pairs, and the bearing seats of the bearing of the bearing are fixed on the base and cannot be rotated.
  • the wheel set bearing on one side of the small wheel set is the bearing to be tested.
  • the bearing model is NF210EM, the rotational speed is 1739.13 ⁇ 1757.43RPM, and the sampling frequency is 32kHz.
  • Its health status includes normal state (NO), inner ring fault state (IF), rolling element fault state (BF) and outer ring fault state (OF).
  • Table 1 Wheelset Bearing Vibration Dataset Information
  • the method disclosed in the present invention is used to diagnose the faults of the above-mentioned wheel set bearings in different health states, and the specific steps are as follows.
  • Step (1) data enhancement.
  • the wavelet transform is used to analyze the time-frequency of each data sample.
  • the mother wavelet is complex Morlet wavelet, which has two parameters: center frequency and bandwidth.
  • the center frequency parameter is fixed at 1.
  • the bandwidth parameter uniformly selects 100 values from the interval [1, 3], so that each fault type can obtain 600 time-frequency features, as shown in Figure 2; for the normal samples in the training set, the bandwidth parameter is fixed at 2, so that the normal state type obtains 600 time-frequency features; for the samples in the test set, the bandwidth parameter is fixed at 2, and each sample generates a time-frequency feature. All time-frequency features are resized to 64 ⁇ 64.
  • the 600 time-frequency features of each type in the training set constitute a balanced time-frequency feature dataset for training the fault diagnosis model.
  • Fig. 3 is a schematic diagram of a time-frequency attention mechanism network module in an embodiment of the present invention, which consists of a frequency attention module and a time attention module.
  • the frequency attention module includes a convolution layer and an average pooling layer.
  • the average pooling layer reduces the time dimension to 1, and then uses the activation function sigmoid to map the excited features to the [0,1] interval as the weight of each frequency band information in the time-frequency feature; the time attention module also includes a convolution layer and an average pooling layer. .
  • Step (3) construction of fault diagnosis model.
  • the model consists of two convolutional layers, two pooling layers, and a fully connected layer, and the feature sizes of each layer are listed in Figure 4.
  • the time-frequency attention mechanism network module shown in Figure 3 is embedded in the network model shown in Figure 4, that is, the output of the model in Figure 3 is used as the input of the model in Figure 4 to construct a fault diagnosis model.
  • Step (4) model training.
  • a balanced time-frequency feature dataset is used for supervised training of the built fault diagnosis model.
  • the training method is the adaptive moment estimation algorithm (Adam), the loss function is the cross-entropy loss function, the learning rate is 0.001, and the iteration is 50 times.
  • Adam adaptive moment estimation algorithm
  • the loss function is the cross-entropy loss function
  • the learning rate is 0.001
  • the iteration is 50 times.
  • Step (5) fault diagnosis.
  • the wavelet transform is performed on each sample in the test set, and the time-frequency feature size is adjusted to 64 ⁇ 64, and then input into the trained fault diagnosis model.
  • Figure 5 shows the time-frequency characteristic diagrams of each fault state before and after the time-frequency attention mechanism network module in the trained fault diagnosis model. It can be seen that the time-frequency attention mechanism proposed by the method of the present invention can accurately enhance the time-frequency characteristics of the fault state and effectively remove the noise inside and outside the fault frequency band. Further input the enhanced time-frequency features into the deep learning network model, the health status category of each sample can be obtained.
  • the accuracy of fault diagnosis of each sample in the test set by the method of the present invention can be known, and the accuracy rate of bearing fault diagnosis can be calculated.
  • the final results are listed in Table 2.
  • the method of the invention can obtain a fault diagnosis accuracy rate as high as 98.96%, which shows that the class imbalance fault diagnosis model proposed by the method of the invention has very high performance.
  • Table 2 also provides the results of three comparison methods.
  • comparison method 1 does not have data enhancement and feature enhancement, but directly inputs the sample into the network model shown in Figure 4 after wavelet transform of a single parameter; comparison method 2 only performs data enhancement, but does not perform feature enhancement; comparison method 3 does not perform data enhancement, but only performs feature enhancement. All the results in Table 2 are the average of the test set accuracy obtained by training ten times.
  • Comparison method 1 Comparison Method 2 Comparison method 3
  • the method of the invention Accuracy 93.89% 96.92% 97.10% 98.96%
  • the time-frequency attention mechanism network module can enhance the fault state information, and at the same time filter out the noise inside and outside the fault frequency band, and enhance the time-frequency characteristics of the fault state.
  • Both the data enhancement and feature enhancement methods proposed by the method of the present invention are beneficial to the improvement of the accuracy rate of rolling bearing fault diagnosis.
  • the feature enhancement method can be embedded in the deep learning network model to achieve end-to-end bearing fault diagnosis, which has important theoretical and practical value for rolling bearing fault diagnosis.
  • Embodiment 2 of the present invention The rolling bearing unbalance fault diagnosis system disclosed in Embodiment 2 of the present invention is introduced below.
  • the rolling bearing unbalance fault diagnosis system described below and the rolling bearing unbalance fault diagnosis method described above can be referred to each other.
  • Embodiment 2 of the present invention discloses a rolling bearing unbalance fault diagnosis system, including:
  • a data enhancement module the data enhancement module is used to perform variable parameter time-frequency analysis on the minority class fault samples obtained under the fault state, and perform single parameter time-frequency analysis on the majority class normal samples obtained under the normal state, so as to obtain a class-balanced time-frequency feature data set;
  • a feature enhancement module the feature enhancement module is used to design a time-frequency attention mechanism network model, and utilizes the time-frequency attention mechanism network model to perform feature enhancement processing on the time-frequency feature data set to obtain a time-frequency feature enhancement data set;
  • a fault diagnosis model construction module the fault diagnosis model construction module is used to build a deep learning network model, the deep learning network model is embedded in the back end of the time-frequency attention mechanism network model, and a fault diagnosis model is constructed;
  • a model training module the model training module is used to use the time-frequency feature enhancement data set to train the constructed fault diagnosis model to obtain a trained fault diagnosis model;
  • a fault diagnosis module the fault diagnosis module is used to perform a single parameter time-frequency analysis on the bearing vibration signal of an unknown health state to obtain a time-frequency feature, and input the time-frequency feature into a trained fault diagnosis model to obtain the health state of the bearing.
  • the data enhancement module includes:
  • the fault sample data enhancement sub-module the fault sample data enhancement sub-module is used to perform variable parameter time-frequency analysis on the minority class fault samples obtained under the fault state, adopting different parameters for time-frequency analysis, and the number of parameters is to make the number of fault sample time-frequency characteristics the same as the number of normal samples.
  • the data enhancement module includes:
  • a normal sample time-frequency analysis sub-module the normal sample time-frequency analysis sub-module is used to perform a single parameter time-frequency analysis on most types of normal samples obtained in a normal state, using a fixed parameter for time-frequency analysis, which is one of multiple parameters in the variable parameter time-frequency analysis.
  • the feature enhancement module includes:
  • the time-frequency feature enhancement sub-module is used to use the time-frequency attention mechanism network model to assign different weights to each frequency information and time information in the time-frequency feature data set for feature enhancement.
  • the rolling bearing unbalanced fault diagnosis system of this embodiment is used to realize the aforementioned rolling bearing unbalanced fault diagnosis method, so the specific implementation of the system can be seen in the embodiment part of the rolling bearing unbalanced fault diagnosis method mentioned above, so its specific implementation can refer to the description of the corresponding embodiments of each part, and will not be introduced here.
  • the rolling bearing unbalance fault diagnosis system of this embodiment is used to implement the above-mentioned rolling bearing unbalance fault diagnosis method, its function corresponds to that of the above method, and will not be repeated here.
  • the embodiments of the present application may be provided as methods, systems, or computer program products. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.
  • computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • These computer program instructions can also be stored in a computer-readable memory capable of directing a computer or other programmable data processing equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means that implement the functions specified in one or more flows of the flowchart and/or one or more blocks of the block diagram.
  • These computer program instructions can also be loaded on a computer or other programmable data processing equipment, so that a series of operation steps are executed on the computer or other programmable equipment to generate computer-implemented processing, so that the instructions executed on the computer or other programmable equipment provide steps for realizing the functions specified in one flow or multiple flows of the flow chart and/or one or more square blocks of the block diagram.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Acoustics & Sound (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

一种滚动轴承类不平衡故障诊断方法及系统，诊断方法包括：对少数类故障样本进行变参数时频分析，对多数类正常样本进行单一参数时频分析，得到类平衡的时频特征数据集（S10）；利用时频注意力机制网络模型进行特征增强处理，得到时频特征增强数据集（S20）；构建故障诊断模型（S30）；利用时频特征增强数据集对构建的故障诊断模型进行训练，得到训练好的故障诊断模型（S40）；对未知健康状态的轴承振动信号进行单一参数时频分析，利用故障诊断模型得到轴承的健康状态（S50）。通过诊断方法及系统，扩充的数据样本具有较高的一致性和多样性，无需优化时频分析的参数，能够有效滤除时频特征中的噪声，且大大提高轴承类不平衡故障诊断的准确率。

Description

滚动轴承类不平衡故障诊断方法及系统 技术领域
本发明涉及轴承故障诊断技术领域，尤其是指一种滚动轴承类不平衡故障诊断方法及系统。
背景技术
滚动轴承是各类旋转机械中最常用的通用零部件，也是最易损耗的零部件之一。由于旋转机械运行工况复杂，且轴承长期在重载荷和高速下工作，极易发生故障。因此对轴承的健康状态进行监测，及时准确地诊断故障，并尽早处理，可以避免重大经济损失及安全事故的发生，其有着重要的现实意义。当前，深度学习在滚动轴承故障诊断中具有广泛的应用，但是传统的深度学习方法在实际的轴承故障诊断应用时还存在一些亟需解决的问题。
首先，深度学习模型需要大量平衡的轴承状态历史数据来训练。在工程实践中，轴承正常状态数据量常远远大于故障状态数据量，这就出现了类不平衡问题，给故障诊断模型的训练带来了挑战。数量较少的样本所包含的特征过少，模型就会很难学习其数据分布规律，并且会过度依赖有限的数据样本，从而导致过拟合问题，使得模型对轴承少样本的故障状态的识别准确率不高。其次，深度学习模型对清晰的特征敏感，易受噪声的干扰。在工程实践中，机械设备运行时会产生大量振动噪声，采集的轴承振动数据会受到噪声干扰，导致轴承故障冲击信息不够明显，从而使深度学习模型不能准确学习到轴承故障状态特征。因此，如何在类不平衡和噪声干扰下提升故障诊断模型的性能，是深度学习应用到轴承故障诊断工程实践中的关键。
在类不平衡故障诊断中，常采用数据增强方法，其通过有限的少数类数据样本产生充足的生成样本使数据集达到平衡。常用的数据增强方法分为单样本数据增强和多样本数据增强，前者包括图像几何变换、图像颜色变换等，后者包括合成少数类过采样技术(SMOTE)、生成对抗网络(GAN)、辅助分类生成对抗网络(ACGAN)等。但是在数据增强方法中，单样本数据增强方法得到的生成样本与原样本的差异较大，强制对增强样本和原样本之间施加“标签不变性”约束可能会损害模型性能。此外，图像几何变换对于有方向信息的分类不适用，图像颜色变换对于色彩信息很重要的分类不适用。在多样本数据增强方法中，SMOTE算法具有两方面的局限性，一是近邻选择存在一定的盲目性，二是无法学习少数类的数据分布，容易产生分布边缘化问题。GAN和ACGAN通过构建模型来生成少数类样本，其模型训练的稳定性较差，需要 较高的计算量和计算时间。总的来说，以上数据增强方法均容易导致生成样本与原样本的数据分布差异较大，不能保证生成样本与原样本的一致性。另外，这些方法根据少量的轴承故障数据样本生成新的样本，而已有的少量样本包含的轴承故障状态信息有限，导致生成样本也缺乏多样性的状态信息，不利于故障诊断模型的训练。在噪声干扰下的故障特征提取中，常采用时频分析方法滤除噪声，比如小波变换(WT)、经验模态分解(EMD)、变分模态分解(VMD)等。这些方法通过先进的信号处理技术提取轴承故障的显著时频特征，滤除故障特征频带以外的噪声。但是在基于时频分析的轴承故障特征提取方法中，需要优化相应方法的参数来提取最优的时频状态特征，然而参数的优化过程不能与深度学习模型的训练同时进行，无法实现端到端的轴承故障诊断，而且最优参数下的时频模式只体现了一种轴承故障状态信息，忽略了其它形式的状态信息。此外，该类方法虽然能够有效地滤除故障频带以外的噪声，但是故障频带以内的噪声没有去除。
发明内容
为此，本发明所要解决的技术问题在于克服现有技术存在的问题，提出一种滚动轴承类不平衡故障诊断方法及系统，其扩充的数据样本具有较高的一致性和多样性，无需优化时频分析的参数，能够有效滤除时频特征中的噪声，而且时频特征提取能够与故障诊断模型训练融为一体，大大提高了轴承类不平衡故障诊断的准确率。
为解决上述技术问题，本发明提供一种滚动轴承类不平衡故障诊断方法，包括以下步骤：
S10：对在故障状态下获取的少数类故障样本进行变参数时频分析，对在正常状态下获取的多数类正常样本进行单个参数时频分析，获得类平衡的时频特征数据集；
S20：设计时频注意力机制网络模型，利用所述时频注意力机制网络模型对所述时频特征数据集进行特征增强处理，获得时频特征增强数据集；
S30：搭建深度学习网络模型，将所述深度学习网络模型嵌入所述时频注意力机制网络模型的后端，构建故障诊断模型；
S40：利用所述时频特征增强数据集对构建的故障诊断模型进行训练，获得训练好的故障诊断模型；
S50：对未知健康状态的轴承振动信号进行单个参数时频分析，获得时频特征，将所述时频特征输入到训练好的故障诊断模型中，得到所述轴承的健康状态。
在本发明的一个实施例中，在S10中，对在故障状态下获取的少数类故 障样本进行变参数时频分析，包括：
采用不同的参数进行时频分析，参数的个数为使故障样本时频特征的个数与正常样本的个数相同。
在本发明的一个实施例中，在S10中，对在正常状态下获取的多数类正常样本进行单个参数时频分析，包括：
采用一个固定的参数进行时频分析，该参数是变参数时频分析中多个参数中的一个。
在本发明的一个实施例中，在S20中，利用所述时频注意力机制网络模型对所述时频特征数据集进行特征增强处理，包括：
利用所述时频注意力机制网络模型给时频特征数据集中的各频率信息和时间信息赋予不同的权重进行特征增强。
在本发明的一个实施例中，S10中的单个参数时频分析的方法和S50中的单个参数时频分析的方法相同。
在本发明的一个实施例中，在S40中，模型训练方法包括自适应矩估计算法、随机梯度下降法以及均方根传递算法。
此外，本发明还提供一种滚动轴承类不平衡故障诊断系统，包括：
数据增强模块，所述数据增强模块用于对在故障状态下获取的少数类故障样本进行变参数时频分析，对在正常状态下获取的多数类正常样本进行单个参数时频分析，获得类平衡的时频特征数据集；
特征增强模块，所述特征增强模块用于设计时频注意力机制网络模型，利用所述时频注意力机制网络模型对所述时频特征数据集进行特征增强处理，获得时频特征增强数据集；
故障诊断模型构建模块，所述故障诊断模型构建模块用于搭建深度学习网络模型，将所述深度学习网络模型嵌入所述时频注意力机制网络模型的后端，构建故障诊断模型；
模型训练模块，所述模型训练模块用于利用所述时频特征增强数据集对构建的故障诊断模型进行训练，获得训练好的故障诊断模型；
故障诊断模块，所述故障诊断模块用于对未知健康状态的轴承振动信号进行单个参数时频分析，获得时频特征，将所述时频特征输入到训练好的故障诊断模型中，得到所述轴承的健康状态。
在本发明的一个实施例中，所述数据增强模块包括：
故障样本数据增强子模块，所述故障样本数据增强子模块用于对在故障状态下获取的少数类故障样本进行变参数时频分析，采用不同的参数进行时频分析，参数的个数为使故障样本时频特征的个数与正常样本的个数相同。
在本发明的一个实施例中，所述数据增强模块包括：
正常样本时频分析子模块，所述正常样本时频分析子模块用于对在正常状态下获取的多数类正常样本进行单个参数时频分析，采用一个固定的参数进行时频分析，该参数是变参数时频分析中多个参数中的一个。
在本发明的一个实施例中，所述特征增强模块包括：
时频特征增强子模块，所述时频特征增强子模块用于利用所述时频注意力机制网络模型给时频特征数据集中的各频率信息和时间信息赋予不同的权重进行特征增强。
本发明的上述技术方案相比现有技术具有以下优点：
本发明利用不同参数下的时频分析方法来扩充少数类故障样本的数据量，考虑了轴承故障不同时频模式的状态信息，同一故障类型的各时频特征之间具有故障发生时间中心和所处频带中心的一致性，以及时频状态特征的多样性；利用时频注意力机制分别增强时间轴和频率轴上的时频状态特征，有效滤除了时频特征中的噪声，该时频特征提取方法可以嵌入在故障诊断模型的训练中，实现端到端的轴承故障诊断。并且扩充的数据样本具有较高的一致性和多样性，无需优化时频分析的参数，能够有效滤除时频特征中的噪声，而且时频特征提取能够与故障诊断模型训练融为一体，大大提高了轴承类不平衡故障诊断的准确率。
附图说明
为了使本发明的内容更容易被清楚的理解，下面根据本发明的具体实施例并结合附图，对本发明作进一步详细的说明。
图1为本发明实施例公开的滚动轴承类不平衡故障诊断方法的流程图。
图2为本发明实施例中基于变参时频分析的数据增强示意图。
图3为本发明实施例中时频注意力机制网络模型示意图。
图4为本发明实施例中深度学习网络模型示意图。
图5为本发明实施例中训练好的故障诊断模型里时频注意力机制网络模型前后的各故障状态时频特征图。
具体实施方式
下面结合附图和具体实施例对本发明作进一步说明，以使本领域的技术人员可以更好地理解本发明并能予以实施，但所举实施例不作为对本发明的限定。
实施例一
请参阅图1所示，本实施例提供一种滚动轴承类不平衡故障诊断方法，包括以下步骤：
S10：对在故障状态下获取的少数类故障样本进行变参数时频分析，对在正常状态下获取的多数类正常样本进行单个参数时频分析，获得类平衡的时频特征数据集；
S20：设计时频注意力机制网络模型，利用所述时频注意力机制网络模型对所述时频特征数据集进行特征增强处理，获得时频特征增强数据集；
S30：搭建深度学习网络模型，将所述深度学习网络模型嵌入所述时频注意力机制网络模型的后端，构建故障诊断模型；
S40：利用所述时频特征增强数据集对构建的故障诊断模型进行训练，获得训练好的故障诊断模型；
S50：对未知健康状态的轴承振动信号进行单个参数时频分析，获得时频特征，将所述时频特征输入到训练好的故障诊断模型中，得到所述轴承的健康状态。
其中，在S10中，时频分析是可以把时间序列信号变换到时频域的方法，其包括但不限于小波变换(WT)、短时傅里叶变换(STFT)、维格纳-威利分布(WVD)。
具体地，对在故障状态下获取的少数类故障样本进行变参数时频分析包括采用不同的参数进行时频分析，参数的个数为使故障样本时频特征的个数与正常样本的个数相同；对在正常状态下获取的多数类正常样本进行单个参数时频分析包括采用一个固定的参数进行时频分析，该参数是变参数时频分析中多个参数中的一个。上述时频特征的时间范围和频率范围均保持一致，时频特征的尺寸保持一致。
其中，在S20中，利用所述时频注意力机制网络模型对所述时频特征数据集进行特征增强处理包括利用所述时频注意力机制网络模型给时频特征数据集中的各频率信息和时间信息赋予不同的权重进行特征增强。
上述时频注意力机制网络模型由人工神经网络构造，包括但不限于全连接神经网络、卷积神经网络、深度置信网络、深度残差网络、自编码器、循 环神经网络、回归神经网络、受限玻尔兹曼机以及生成对抗网络。
其中，在S30中，深度学习网络模型由人工神经网络构造，包括但不限于全连接神经网络、卷积神经网络、深度置信网络、深度残差网络、自编码器、循环神经网络、回归神经网络、受限玻尔兹曼机、生成对抗网络。
其中，在S40中，模型训练方法包括但不限于自适应矩估计算法(Adam)、随机梯度下降法(SGD)、均方根传递算法(RmsPorp)。
其中，S10中的单个参数时频分析的方法和S50中的单个参数时频分析的方法相同。
本发明利用不同参数下的时频分析方法来扩充少数类故障样本的数据量，考虑了轴承故障不同时频模式的状态信息，同一故障类型的各时频特征之间具有故障发生时间中心和所处频带中心的一致性，以及时频状态特征的多样性；利用时频注意力机制分别增强时间轴和频率轴上的时频状态特征，有效滤除了时频特征中的噪声，该时频特征提取方法可以嵌入在故障诊断模型的训练中，实现端到端的轴承故障诊断。并且扩充的数据样本具有较高的一致性和多样性，无需优化时频分析的参数，能够有效滤除时频特征中的噪声，而且时频特征提取能够与故障诊断模型训练融为一体，大大提高了轴承类不平衡故障诊断的准确率。
为了更加清楚地了解本发明的技术方案及其效果，下面结合一个具体的实施例进行详细说明。
采用轮对轴承试验台采集不同健康状态下的轮对轴承振动数据。试验台用一个感应电机通过电机皮带驱动一个小轮对转动，两个液压缸通过两个轮对轴承分别在小轮对的两边添加径向载荷，轮对轴承的轴承座底部固定在试验台的基座上，它的上部可以围绕底部基座固定点自由地做圆周运动，从而使小轮对的径向载荷传递到与之平行的大轮对上，大论对的两个轮子与小轮对的两个轮子接触转动，其两端轮对轴承的轴承座固定在基座上且不可转动。小轮对一侧的轮对轴承是被测试的轴承，该轴承型号为NF210EM，转速为1739.13～1757.43RPM，采样频率为32kHz，其健康状况包括正常状态(NO)、内圈故障状态(IF)、滚动体故障状态(BF)和外圈故障状态(OF)。
将振动数据集分为训练集和测试集，设置训练集中正常样本和每类故障样本的数据量比例为100：1，数据集信息如表1所示。
表1：轮对轴承振动数据集信息
Figure PCTCN2022076031-appb-000001
Figure PCTCN2022076031-appb-000002
采用本发明公开的方法对上述不同健康状态的轮对轴承进行故障诊断，具体步骤如下。
步骤(1)、数据增强。采用小波变换对各数据样本进行时频分析，母小波为复Morlet小波，它有两个参数：中心频率和带宽。在本实施例中，中心频率参数固定为1。对于训练集中的每个故障样本，带宽参数从区间[1,3]中均匀选择100个值，使每个故障类型都可以获得600个时频特征，如图2所示；对于训练集中的正常样本，带宽参数固定为2，使正常状态类型获得600个时频特征；对于测试集中的样本，带宽参数固定为2，每个样本生成一个时频特征。所有时频特征的大小均调整为64×64。训练集中每种类型的600个时频特征构成平衡的时频特征数据集，用于故障诊断模型的训练。
步骤(2)、特征增强。图3是本发明实施例中时频注意力机制网络模块示意图，它由频率注意力模块和时间注意力模块组成。频率注意力模块包括一个卷积层和一个平均池化层，平均池化层将时间维度减少到1，再通过激活函数sigmoid，将激发后的特征映射到[0,1]区间，作为时频特征中各频带信息的权重；时间注意力模块也包括一个卷积层和一个平均池化层，平均池化层将频率维度减少到1，再通过激活函数sigmoid，将激发后的特征映射到[0,1]区间，作为时频特征中各时间信息的权重。
步骤(3)、故障诊断模型的构建。首先搭建深度学习网络模型，如图4所示。该模型由两个卷积层、两个池化层和一个全连接层组成，每层的特征尺寸列于图4中。将图3所示的时频注意力机制网络模块嵌入在图4所示的网络模型中，即把图3模型的输出作为图4模型的输入，从而构建故障诊断模型。
步骤(4)、模型训练。采用平衡的时频特征数据集对构建的故障诊断模型进行有监督式训练，训练方法为自适应矩估计算法(Adam)，损失函数为交叉熵损失函数，学习率为0.001，迭代为50次。
步骤(5)、故障诊断。对测试集中的各样本进行小波变换，并调整时频特征尺寸为64×64，然后输入到训练好的故障诊断模型中。图5给出了训练好的故障诊断模型里时频注意力机制网络模块前后的各故障状态时频特征图，可以看出本发明方法提出的时频注意力机制可以准确增强故障状态的时频特征，有效去除故障频带内、外噪声。进一步将增强的时频特征输入到深度学习网络模型中，可以得到各样本的健康状态类别。通过与各样本的真实健康状态类别进行对比可以知道本发明方法对测试集各样本的故障诊断准确性，计算轴承故障诊断的准确率，最终结果列于表2中。本发明方法可以获得高达98.96％的故障诊断准确率，说明本发明方法提出的类不平衡故障诊断模型具有很高的性能。
为了证明本发明方法的优越性，表2也给出了三种对比方法的结果。其中，对比方法1没有数据增强和特征增强，而是对样本进行单个参数的小波变换后直接输入到图4所示的网络模型中；对比方法2只进行数据增强，不 进行特征增强；对比方法3不进行数据增强，只进行特征增强。表2中所有结果均为训练十次得到的测试集准确率的平均值。
表2：本发明方法与其它方法的故障诊断准确率结果对比
方法 对比方法1 对比方法2 对比方法3 本发明方法
准确率 93.89％ 96.92％ 97.10％ 98.96％
从表2可以看出，没有数据增强和特征增强的对比方法1的故障诊断准确率最低，单独采用数据增强和特征增强均可以提升故障诊断准确率，而同时采用数据增强和特征增强可以获得最高的故障诊断准确率。这证明本发明提出的数据增强和特征增强方法对故障诊断模型性能具有提升效果。
综上所述，通过对少数类故障样本进行变参时频分析，可以获得充足的故障状态时频特征，从而使训练数据集达到平衡，且同一故障类型的变参时频特征之间具有一致性和多样性；时频注意力机制网络模块可以增强故障状态信息，同时滤除故障频带内、外的噪声，增强故障状态时频特征。本发明方法提出的数据增强和特征增强方法均有利于滚动轴承故障诊断准确率的提升。此外，特征增强方法可以嵌入在深度学习网络模型中，实现端到端的轴承故障诊断，对滚动轴承故障诊断具有重要的理论和实用价值。
实施例二
下面对本发明实施例二公开的一种滚动轴承类不平衡故障诊断系统进行介绍，下文描述的一种滚动轴承类不平衡故障诊断系统与上文描述的一种滚动轴承类不平衡故障诊断方法可相互对应参照。
本发明实施例二公开了一种滚动轴承类不平衡故障诊断系统，包括：
数据增强模块，所述数据增强模块用于对在故障状态下获取的少数类故障样本进行变参数时频分析，对在正常状态下获取的多数类正常样本进行单个参数时频分析，获得类平衡的时频特征数据集；
特征增强模块，所述特征增强模块用于设计时频注意力机制网络模型，利用所述时频注意力机制网络模型对所述时频特征数据集进行特征增强处理，获得时频特征增强数据集；
故障诊断模型构建模块，所述故障诊断模型构建模块用于搭建深度学习网络模型，将所述深度学习网络模型嵌入所述时频注意力机制网络模型的后端，构建故障诊断模型；
模型训练模块，所述模型训练模块用于利用所述时频特征增强数据集对构建的故障诊断模型进行训练，获得训练好的故障诊断模型；
故障诊断模块，所述故障诊断模块用于对未知健康状态的轴承振动信号进行单个参数时频分析，获得时频特征，将所述时频特征输入到训练好的故 障诊断模型中，得到所述轴承的健康状态。
其中，所述数据增强模块包括：
故障样本数据增强子模块，所述故障样本数据增强子模块用于对在故障状态下获取的少数类故障样本进行变参数时频分析，采用不同的参数进行时频分析，参数的个数为使故障样本时频特征的个数与正常样本的个数相同。
其中，所述数据增强模块包括：
正常样本时频分析子模块，所述正常样本时频分析子模块用于对在正常状态下获取的多数类正常样本进行单个参数时频分析，采用一个固定的参数进行时频分析，该参数是变参数时频分析中多个参数中的一个。
其中，所述特征增强模块包括：
时频特征增强子模块，所述时频特征增强子模块用于利用所述时频注意力机制网络模型给时频特征数据集中的各频率信息和时间信息赋予不同的权重进行特征增强。
本实施例的滚动轴承类不平衡故障诊断系统用于实现前述的滚动轴承类不平衡故障诊断方法，因此该系统的具体实施方式可见前文中的滚动轴承类不平衡故障诊断方法的实施例部分，所以，其具体实施方式可以参照相应的各个部分实施例的描述，在此不再展开介绍。
另外，由于本实施例的滚动轴承类不平衡故障诊断系统用于实现前述的滚动轴承类不平衡故障诊断方法，因此其作用与上述方法的作用相对应，这里不再赘述。
本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本申请实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功 能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
显然，上述实施例仅仅是为清楚地说明所作的举例，并非对实施方式的限定。对于所属领域的普通技术人员来说，在上述说明的基础上还可以做出其它不同形式变化或变动。这里无需也无法对所有的实施方式予以穷举。而由此所引申出的显而易见的变化或变动仍处于本发明创造的保护范围之中。

Claims (10)

  1. 一种滚动轴承类不平衡故障诊断方法，其特征在于，包括以下步骤：
    S10：对在故障状态下获取的少数类故障样本进行变参数时频分析，对在正常状态下获取的多数类正常样本进行单个参数时频分析，获得类平衡的时频特征数据集；
    S20：设计时频注意力机制网络模型，利用所述时频注意力机制网络模型对所述时频特征数据集进行特征增强处理，获得时频特征增强数据集；
    S30：搭建深度学习网络模型，将所述深度学习网络模型嵌入所述时频注意力机制网络模型的后端，构建故障诊断模型；
    S40：利用所述时频特征增强数据集对构建的故障诊断模型进行训练，获得训练好的故障诊断模型；
    S50：对未知健康状态的轴承振动信号进行单个参数时频分析，获得时频特征，将所述时频特征输入到训练好的故障诊断模型中，得到所述轴承的健康状态。
  2. 根据权利要求1所述的滚动轴承类不平衡故障诊断方法，其特征在于，在S10中，对在故障状态下获取的少数类故障样本进行变参数时频分析，包括：
    采用不同的参数进行时频分析，参数的个数为使故障样本时频特征的个数与正常样本的个数相同。
  3. 根据权利要求2所述的滚动轴承类不平衡故障诊断方法，其特征在于，在S10中，对在正常状态下获取的多数类正常样本进行单个参数时频分析，包括：
    采用一个固定的参数进行时频分析，该参数是变参数时频分析中多个参数中的一个。
  4. 根据权利要求1所述的滚动轴承类不平衡故障诊断方法，其特征在于，在S20中，利用所述时频注意力机制网络模型对所述时频特征数据集进行特征增强处理，包括：
    利用所述时频注意力机制网络模型给时频特征数据集中的各频率信息和时间信息赋予不同的权重进行特征增强。
  5. 根据权利要求1所述的滚动轴承类不平衡故障诊断方法，其特征在于：S10中的单个参数时频分析的方法和S50中的单个参数时频分析的方法相同。
  6. 根据权利要求1所述的滚动轴承类不平衡故障诊断方法，其特征在于：在S40中，模型训练方法包括自适应矩估计算法、随机梯度下降法以及均方根传递算法。
  7. 一种滚动轴承类不平衡故障诊断系统，其特征在于，包括：
    数据增强模块，所述数据增强模块用于对在故障状态下获取的少数类故障样本进行变参数时频分析，对在正常状态下获取的多数类正常样本进行单个参数时频分析，获得类平衡的时频特征数据集；
    特征增强模块，所述特征增强模块用于设计时频注意力机制网络模型，利用所述时频注意力机制网络模型对所述时频特征数据集进行特征增强处理，获得时频特征增强数据集；
    故障诊断模型构建模块，所述故障诊断模型构建模块用于搭建深度学习网络模型，将所述深度学习网络模型嵌入所述时频注意力机制网络模型的后端，构建故障诊断模型；
    模型训练模块，所述模型训练模块用于利用所述时频特征增强数据集对构建的故障诊断模型进行训练，获得训练好的故障诊断模型；
    故障诊断模块，所述故障诊断模块用于对未知健康状态的轴承振动信号进行单个参数时频分析，获得时频特征，将所述时频特征输入到训练好的故障诊断模型中，得到所述轴承的健康状态。
  8. 根据权利要求7所述的滚动轴承类不平衡故障诊断系统，其特征在于，所述数据增强模块包括：
    故障样本数据增强子模块，所述故障样本数据增强子模块用于对在故障状态下获取的少数类故障样本进行变参数时频分析，采用不同的参数进行时频分析，参数的个数为使故障样本时频特征的个数与正常样本的个数相同。
  9. 根据权利要求8所述的滚动轴承类不平衡故障诊断系统，其特征在于，所述数据增强模块包括：
    正常样本时频分析子模块，所述正常样本时频分析子模块用于对在正常状态下获取的多数类正常样本进行单个参数时频分析，采用一个固定的参数进行时频分析，该参数是变参数时频分析中多个参数中的一个。
  10. 根据权利要求7所述的滚动轴承类不平衡故障诊断系统，其特征在于， 所述特征增强模块包括：
    时频特征增强子模块，所述时频特征增强子模块用于利用所述时频注意力机制网络模型给时频特征数据集中的各频率信息和时间信息赋予不同的权重进行特征增强。
PCT/CN2022/076031 2022-01-21 2022-02-11 滚动轴承类不平衡故障诊断方法及系统 WO2023137807A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210074063.6 2022-01-21
CN202210074063.6A CN114563186B (zh) 2022-01-21 2022-01-21 滚动轴承类不平衡故障诊断方法及系统

Publications (1)

Publication Number Publication Date
WO2023137807A1 true WO2023137807A1 (zh) 2023-07-27

Family

ID=81713498

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/076031 WO2023137807A1 (zh) 2022-01-21 2022-02-11 滚动轴承类不平衡故障诊断方法及系统

Country Status (2)

Country Link
CN (1) CN114563186B (zh)
WO (1) WO2023137807A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117669388A (zh) * 2024-01-30 2024-03-08 武汉理工大学 故障样本生成方法、装置、计算机介质
CN117760731A (zh) * 2023-12-19 2024-03-26 江南大学 基于Soft-IntroVAE和1D-2DA-FCNN的不平衡样本滚动轴承故障诊断方法及系统
CN117892188A (zh) * 2023-12-19 2024-04-16 济南大学 一种基于轴承振动信号关系及图神经网络的故障分类方法
CN118094368A (zh) * 2024-04-28 2024-05-28 湘江实验室 一种基于扩散模型和注意力机制的轴承故障诊断方法及装置
CN118376409A (zh) * 2024-06-25 2024-07-23 中国特种设备检测研究院 一种轴承故障诊断方法、产品、介质及设备
CN118395364A (zh) * 2024-07-01 2024-07-26 山东大学 基于改进eemd和生成对抗网络的旋转机械故障诊断方法及系统

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115481692B (zh) * 2022-09-23 2023-10-10 江苏安控智汇科技股份有限公司 一种基于sgan的水泵机组故障诊断方法
CN116128882B (zh) * 2023-04-19 2023-06-23 中汽数据(天津)有限公司 基于不平衡数据集的电机轴承故障诊断方法、设备及介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH07218334A (ja) * 1994-02-07 1995-08-18 Nippon Steel Corp 軸受の異常診断方法及び装置
CN109799090A (zh) * 2019-01-08 2019-05-24 长安大学 采用频带3分区的经验小波变换的轴承特征频率提取方法
CN109902399A (zh) * 2019-03-01 2019-06-18 哈尔滨理工大学 一种基于att-cnn的变工况下滚动轴承故障识别方法
CN112304614A (zh) * 2020-11-27 2021-02-02 中国人民解放军海军工程大学 采用多注意力机制的端到端滚动轴承智能故障诊断方法
CN112649198A (zh) * 2021-01-05 2021-04-13 西交思创智能科技研究院(西安)有限公司 类不平衡滚动轴承智能故障诊断方法、系统、设备及应用

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102589884A (zh) * 2012-03-01 2012-07-18 西北工业大学 一种基于GentleBoost的飞机发电机轴承故障诊断方法
CN103645249B (zh) * 2013-11-27 2017-05-10 国网黑龙江省电力有限公司 基于精简集下采样不均衡svm变压器在线故障检测方法
CN104408302A (zh) * 2014-11-19 2015-03-11 北京航空航天大学 一种基于lmd-svd和极限学习机的轴承变工况故障诊断方法
CN105841961A (zh) * 2016-03-29 2016-08-10 中国石油大学(华东) 一种基于Morlet小波变换和卷积神经网络的轴承故障诊断方法
CN108195587B (zh) * 2018-02-12 2023-08-25 西安交通大学 一种电机滚动轴承故障诊断方法及其诊断系统
CN208125377U (zh) * 2018-03-05 2018-11-20 西安交通大学 一种电机滚动轴承故障诊断系统
CN111665051A (zh) * 2020-07-01 2020-09-15 天津大学 基于能量权重法的强噪声变转速条件下轴承故障诊断方法
CN113204738B (zh) * 2021-04-30 2022-04-08 山东大学 一种基于小波变换与拓扑数据分析的机械故障特征提取方法
CN113432874A (zh) * 2021-05-31 2021-09-24 西安理工大学 一种基于2dpca-cwt与cnn的滚动轴承故障诊断方法
CN113255078A (zh) * 2021-05-31 2021-08-13 南京信息工程大学 一种样本不均衡条件下的轴承故障检测方法及装置
CN113933054B (zh) * 2021-09-23 2024-04-09 吴江万工机电设备有限公司 一种滚动轴承故障监测诊断方法、设备、介质及处理器

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH07218334A (ja) * 1994-02-07 1995-08-18 Nippon Steel Corp 軸受の異常診断方法及び装置
CN109799090A (zh) * 2019-01-08 2019-05-24 长安大学 采用频带3分区的经验小波变换的轴承特征频率提取方法
CN109902399A (zh) * 2019-03-01 2019-06-18 哈尔滨理工大学 一种基于att-cnn的变工况下滚动轴承故障识别方法
CN112304614A (zh) * 2020-11-27 2021-02-02 中国人民解放军海军工程大学 采用多注意力机制的端到端滚动轴承智能故障诊断方法
CN112649198A (zh) * 2021-01-05 2021-04-13 西交思创智能科技研究院(西安)有限公司 类不平衡滚动轴承智能故障诊断方法、系统、设备及应用

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
BAO PING, LIU YUNJIE: "Research on Fault Identification Based on Improved Deep Model in Combination of Generative Adversarial Networks under Unbalanced Data Sets", JOURNAL OF ELECTRONIC MEASUREMENT AND INSTRUMENT, vol. 33, no. 3, 15 March 2019 (2019-03-15), pages 176 - 183, XP093080012, ISSN: 1000-7105, DOI: 10.13382/j.jemi.B1801557 *
WU JINGRAN, DING ENJIE; CUI RAN; LIU JIANHUA: "A Diagnostic Approach for Rotating Machinery Using Multi-Scale Feature Attention Mechanism", JOURNAL OF XI'AN JIATONGUNIVERSITY, XI'AN JIATONG DAXUE CHUBANSHE, XIAN, CN, vol. 54, no. 2, 29 February 2020 (2020-02-29), CN , pages 51 - 58, XP093080013, ISSN: 0253-987X, DOI: 10.7652/xjtuxb202002007 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117760731A (zh) * 2023-12-19 2024-03-26 江南大学 基于Soft-IntroVAE和1D-2DA-FCNN的不平衡样本滚动轴承故障诊断方法及系统
CN117892188A (zh) * 2023-12-19 2024-04-16 济南大学 一种基于轴承振动信号关系及图神经网络的故障分类方法
CN117669388A (zh) * 2024-01-30 2024-03-08 武汉理工大学 故障样本生成方法、装置、计算机介质
CN117669388B (zh) * 2024-01-30 2024-05-31 武汉理工大学 故障样本生成方法、装置、计算机介质
CN118094368A (zh) * 2024-04-28 2024-05-28 湘江实验室 一种基于扩散模型和注意力机制的轴承故障诊断方法及装置
CN118376409A (zh) * 2024-06-25 2024-07-23 中国特种设备检测研究院 一种轴承故障诊断方法、产品、介质及设备
CN118395364A (zh) * 2024-07-01 2024-07-26 山东大学 基于改进eemd和生成对抗网络的旋转机械故障诊断方法及系统

Also Published As

Publication number Publication date
CN114563186A (zh) 2022-05-31
CN114563186B (zh) 2023-01-06

Similar Documents

Publication Publication Date Title
WO2023137807A1 (zh) 滚动轴承类不平衡故障诊断方法及系统
CN111323220B (zh) 风力发电机齿轮箱故障诊断方法及系统
He et al. A new hybrid deep signal processing approach for bearing fault diagnosis using vibration signals
CN106124212B (zh) 基于稀疏编码器和支持向量机的滚动轴承故障诊断方法
CN112418013B (zh) 一种小样本下基于元学习的复杂工况轴承故障诊断方法
CN110595780B (zh) 基于振动灰度图像和卷积神经网络的轴承故障识别方法
CN110516305B (zh) 基于注意机制元学习模型的小样本下故障智能诊断方法
CN115688018B (zh) 一种多工况下轴承的状态监测与故障诊断方法
CN110795843B (zh) 一种识别滚动轴承故障的方法及装置
CN108318249A (zh) 一种旋转机械轴承的故障诊断方法
CN109211546B (zh) 基于降噪自动编码器及增量学习的旋转机械故障诊断方法
CN111046916A (zh) 一种基于空洞卷积胶囊网络的电机故障诊断方法及系统
CN111351665B (zh) 基于emd和残差神经网络的滚动轴承故障诊断方法
Liu et al. Imbalanced fault diagnosis of rotating machinery using autoencoder-based SuperGraph feature learning
CN110487547B (zh) 基于振动图和迁移学习的变工况下滚动轴承故障诊断方法
CN116226646B (zh) 轴承健康状态及剩余寿命的预测方法、系统、设备及介质
CN110243603B (zh) 基于Welch转换-径向基神经网的滚动轴承故障诊断方法
CN106441896A (zh) 滚动轴承故障模式识别及状态监测的特征向量提取方法
CN107506710A (zh) 一种滚动轴承复合故障提取方法
CN111122161B (zh) 一种基于fast kurtogram和深度残差学习的变工况轴承故障诊断方法
CN111753891A (zh) 一种无监督特征学习的滚动轴承故障诊断方法
CN112364706A (zh) 一种基于类不平衡的小样本轴承故障诊断方法
Han et al. An intelligent fault diagnosis method of variable condition gearbox based on improved DBN combined with WPEE and MPE
CN114048787B (zh) 一种基于Attention CNN模型的轴承故障实时智能诊断方法与系统
CN116429428A (zh) 一种基于多特征融合和igwo-lssvm的滚动轴承故障诊断方法

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: 22921243

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE