WO2023044979A1 - 类不平衡数据集下的机械故障智能诊断方法 - Google Patents

类不平衡数据集下的机械故障智能诊断方法 Download PDF

Info

Publication number
WO2023044979A1
WO2023044979A1 PCT/CN2021/123198 CN2021123198W WO2023044979A1 WO 2023044979 A1 WO2023044979 A1 WO 2023044979A1 CN 2021123198 W CN2021123198 W CN 2021123198W WO 2023044979 A1 WO2023044979 A1 WO 2023044979A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
fault
model
diagnosis
mechanical
Prior art date
Application number
PCT/CN2021/123198
Other languages
English (en)
French (fr)
Inventor
王俊
戴俊
石娟娟
江星星
姚林泉
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 WO2023044979A1 publication Critical patent/WO2023044979A1/zh

Links

Images

Classifications

    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G06F18/24137—Distances to cluster centroïds
    • G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/24323—Tree-organised classifiers
    • 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
    • G06N3/047—Probabilistic or stochastic 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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 the field of fault intelligent diagnosis, in particular to an intelligent fault diagnosis method for mechanical faults under a type unbalanced data set.
  • CNN convolutional neural network
  • DNN deep belief network
  • ResNet residual network
  • the class imbalance data set will easily lead to a decline in the performance of the diagnostic model, that is, the model is easy to overfit the normal signal with a large number of samples, and the Underfitting to a small number of faulty signals.
  • the model is easy to learn some redundant or even irrelevant features in the process of extracting fault data features, which reduce the generalization ability of the model.
  • the dynamic weight method gives more attention to a small number of fault samples by adjusting the weight parameters in the network, so as to improve the underfitting problem of fault samples.
  • the data generation method is to generate new samples of the same category by using a small amount of fault data to expand the fault samples, balance the fault data and normal data, and use the balanced data set to train the intelligent diagnosis model.
  • Traditional data generation methods include Synthetic Minority Upsampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN) and so on.
  • the dynamic weight method needs to dynamically adjust the weight according to the unbalance rate between normal and fault samples, so it is suitable for applications where the unbalance rate is known. And when the data is extremely class unbalanced, the dynamic weight method is easily disturbed by redundant features in a small number of fault samples, resulting in over-fitting of the model and reducing the accuracy of fault diagnosis.
  • the data generation method makes the classes in the data set reach balance by upsampling a small number of fault signals, and fundamentally solves the class imbalance phenomenon.
  • the mechanical structure is complex and has nonlinear characteristics. Its vibration signal often has strong background noise under actual working conditions, and it shows obvious non-stationary characteristics under fault conditions.
  • the traditional data generation method does not learn the distribution characteristics of the data, and directly generates the signal through interpolation technology in the time domain signal, which is easily disturbed by the measurement noise component, and the quality of the generated data is not high, which is also easy to cause the performance of the intelligent diagnosis model to decline.
  • the technical problem to be solved by the present invention is to provide an intelligent diagnosis method for mechanical faults under the unbalanced data set, aiming at the limited application scenarios of the dynamic weight method, the traditional data generation method is susceptible to noise interference, and the generated data quality is not high.
  • the invention proposes a new data generation method, which is based on deep neural network, through the combination of autoencoder and generative confrontation network, learns the low-dimensional distribution characteristics of fault signals, and generates fault data according to low-dimensional features, so as to avoid Noise interference, obtain high-quality generated data, and use balanced data sets to achieve high-performance intelligent diagnosis of mechanical faults.
  • the present invention provides a method for intelligent diagnosis of mechanical faults under a class-unbalanced data set, including:
  • Step (1) data preprocessing: convert the mechanical vibration signal to the frequency domain, and normalize the amplitude to the [0,1] range;
  • Step (2) model building: combine the autoencoder and the generative confrontation network to build a data generation model
  • Step (3) model training: using the fault data to train the data generation model according to the preset loss function and optimization algorithm;
  • Step (4) data generation: Utilize the low-dimensional features of the fault data learned by the data generation model in training, generate corresponding fault data after multiple interpolation and noise addition, and realize various data balances;
  • Step (5) fault diagnosis: use the class balance data set to train the preset fault diagnosis model, and use the trained fault diagnosis model to perform intelligent diagnosis on mechanical faults.
  • the autoencoder is composed of an encoder and a decoder
  • the generated confrontation network is composed of a generator and a discriminator
  • the decoder is exactly a generator
  • the autoencoder Learn the low-dimensional features of the input data, that is, the real data, through the encoder, and then output the generated data that is consistent with the distribution characteristics of the input data through the decoder through the low-dimensional features and its category labels, that is, fake data
  • the discriminators in the generated confrontation network are respectively Perform authenticity discrimination and category classification on the input data and the generated data.
  • the encoder, the decoder, and the discriminator are constructed by one of a deep convolutional network, a deep belief network, and a residual network.
  • the preset loss function includes the mean square error loss function between the data generated by the generator and the input data of the encoder, and the cross-entropy classification loss of the discriminator for true and false data
  • the Wasserstein distance or binary cross-entropy loss function used by the discriminator to identify true and false data and the mean square error loss function between the output features of the encoder and the implicit features of the discriminator.
  • the preset optimization algorithm includes but not limited to stochastic gradient descent (SGD), stochastic gradient descent with momentum (Momentum), Nesterov momentum method, Adagrad algorithm, automatic One of the adaptive moment estimation methods (Adam).
  • SGD stochastic gradient descent
  • Momentum stochastic gradient descent with momentum
  • Nesterov momentum method Nesterov momentum method
  • Adagrad algorithm automatic One of the adaptive moment estimation methods (Adam).
  • step (4) the interpolation is carried out in different low-dimensional features of the same category of fault samples, and the label of this category needs to be embedded before generating the fault data, and the added noise is low-amplitude of random noise.
  • the preset fault diagnosis model includes one of support vector machine, k-nearest neighbor algorithm, random forest, fuzzy system or deep neural network.
  • the present application also provides a computer device, including a memory, a processor, and a computer program stored on the memory and operable on the processor.
  • a computer device including a memory, a processor, and a computer program stored on the memory and operable on the processor.
  • the processor executes the program, any one of the steps of the method described above.
  • the present application also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the steps of any one of the methods described above are implemented.
  • the present application also provides a processor, the processor is used to run a program, wherein the program executes any one of the methods when running.
  • the invention discloses an intelligent diagnosis method for mechanical faults under a class unbalanced data set.
  • This method aims at the problem of the decline of diagnostic accuracy caused by the imbalance of data sets in mechanical fault diagnosis, and proposes a new data generation method, which uses the feature mining ability of deep learning and the confrontation training mechanism to learn the data distribution characteristics of a small number of fault samples.
  • the low-dimensional feature space of the data uses interpolation and noise to generate new features, and after embedding labels, new fault samples are obtained through the generator. Interpolation in low-dimensional space can eliminate the influence of measurement noise in the signal, adding random noise can increase the diversity of generated samples, and embedding labels can ensure the consistency of the data distribution between generated samples and the same type of fault samples.
  • this method has at least the following advantages: (1) It can learn the low-dimensional distribution characteristics of the data and eliminate the interference of measurement noise; (2) The generated data is consistent with the fault data of the same category and has a certain diversity at the same time. The quality of generated data is high; (3) The accuracy rate of intelligent identification of mechanical faults is high.
  • Fig. 1 is a flow chart of the method for intelligent diagnosis of mechanical faults under the class unbalanced data set of the present invention.
  • Fig. 2 is the comparison figure of the generated data and the real data of four kinds of fault types obtained in the mechanical fault intelligent diagnosis method under the class unbalanced data set of the present invention
  • the left column is the real data under the four kinds of fault states
  • the right column Generate data corresponding to it.
  • Fig. 3 is the variation curve of the classification accuracy of the mechanical fault intelligent diagnosis method and the traditional method under the unbalanced data set of the present invention under five kinds of unbalanced rates.
  • a flow chart of a mechanical fault intelligent diagnosis method under a class unbalanced data set the technology specifically includes:
  • Step 101 data preprocessing. Perform Fourier transform on the vibration signal, convert the mechanical vibration signal to the frequency domain, and normalize the amplitude to the [0,1] range.
  • Step 102 Model building. Combine autoencoders and generative adversarial networks to build data generation models.
  • the autoencoder is composed of an encoder and a decoder
  • the generative confrontation network is composed of a generator and a discriminator
  • the decoder is the generator.
  • the autoencoder learns the low-dimensional features of the input data (true data) through the encoder, and then outputs the generated data (false data) that is consistent with the distribution characteristics of the input data through the decoder through the low-dimensional features and their category labels.
  • the discriminator in the Generative Adversarial Network performs authenticity discrimination and category classification on the input data and the generated data respectively.
  • the encoder, decoder, and discriminator include but are not limited to one of deep convolutional networks, deep belief networks, and residual networks.
  • Step 103 model training. Use the fault data to train the data generation model according to the preset loss function and optimization algorithm.
  • the loss function of the data generation model during training includes:
  • the generated data is finally close to the data distribution of the input data of the same category, but it is difficult for the discriminator to distinguish the authenticity of the generated data and the input data of the same category, and the balance between the generator and the discriminator is reached, and the data is completed.
  • the preset optimization algorithm includes, but is not limited to, one of stochastic gradient descent (SGD), stochastic gradient descent with momentum (Momentum), Nesterov momentum method, Adagrad algorithm, and adaptive moment estimation method (Adam).
  • SGD stochastic gradient descent
  • Momentum stochastic gradient descent with momentum
  • Nesterov momentum method Nesterov momentum method
  • Adagrad algorithm Adagrad algorithm
  • Adam adaptive moment estimation method
  • Step 104 Data generation. Using the low-dimensional features of the fault data learned by the data generation model during training, the fault data of the corresponding class is generated through multiple interpolation and noise addition to achieve various data balances.
  • Interpolation is carried out in different low-dimensional features of the same category of fault samples.
  • the label of this category needs to be embedded, and the noise added is low-amplitude random noise.
  • Step 105 fault diagnosis.
  • the preset fault diagnosis model includes but not limited to one of support vector machine, k-nearest neighbor algorithm, random forest, fuzzy system, and deep neural network.
  • a planetary gearbox fault simulation test platform was built, and four fault states were manually set: broken teeth, missing teeth, root cracks, tooth surface wear, and a total of five health states including normal states.
  • An acceleration sensor is installed on the planetary gearbox to collect the vibration signal of the gearbox, and the sampling frequency is 5kHz.
  • Each health state contains 2000 sets of signals, of which 1000 sets of signals are used as test data and do not participate in training, and the length of each set of signals is 2048 data points.
  • five kinds of imbalance rates are set in the example, that is, the ratio of the number of healthy samples of the gearbox to the number of samples of each type of fault, respectively are 5:1, 10:1, 20:1, 50:1, and 100:1, and the data volume of healthy samples under each imbalance ratio is 1000.
  • the technology disclosed in the present invention is used to process the 5 groups of unbalanced data sets.
  • the steps are shown in FIG. 1 , and the detailed information is as follows.
  • Step (1) data preprocessing. Perform Fourier transform on the vibration signal, convert the mechanical vibration signal to the frequency domain, and normalize the amplitude to the [0,1] range.
  • the length of the original time domain signal is 2048 data points, after Fourier transform, the frequency domain signal of length 1024 is taken as the input data of the model.
  • Step (2) model building. Combine the autoencoder and the generation confrontation network to build a data generation model.
  • the specific implementation is as follows:
  • 1Autoencoder It includes an encoder and a decoder, and its main function is to encode and decode input data.
  • the encoder adopts a four-layer one-dimensional convolutional neural network structure. The dimensions of each layer are 8, 16, 32, and 64 respectively. A convolution kernel with a length of 15 is used.
  • the LeakyReLU activation function layer is connected between the convolution layers. After the samples pass through the encoder Outputs a 64-dimensional latent feature vector.
  • the decoder adopts a four-layer one-dimensional deconvolution neural network structure, and the dimensions of each layer are 64, 32, 16 and 8 respectively.
  • the deconvolution kernel with a length of 15 is used, and the ReLU activation function layer is connected between the deconvolution layers.
  • the decoder The last layer of deconvolution is connected to a sigmoid activation function, which limits the magnitude of the generated data to the [0,1] range.
  • 2 Generative confrontation network including generator and discriminator.
  • the generator is the decoder in the autoencoder.
  • the discriminator designs four one-dimensional convolutional layers and two fully connected layers. The dimensions of each layer are 8, 16, 32, and 64 respectively.
  • a convolution kernel with a length of 15 is used.
  • Each convolutional layer is connected to a LeakyRelu activation.
  • the function layer and the convolutional layer finally output a feature vector with a length of 64 dimensions.
  • the feature is then input into two fully connected layers, the first fully connected layer reduces the 64-dimensional feature vector to 1-dimensional to calculate the Wasserstein distance between the generated data and the real data.
  • the second fully connected layer reduces the 64-dimensional feature vector to 4 dimensions (that is, the number of fault categories that need to be upsampled) and connects the Softmax activation layer to judge the signal category.
  • Step (3) model training. Use the fault data to train the data generation model according to the preset loss function and optimization algorithm.
  • the loss function there are 4 parts of the loss function:
  • Step (4) data generation.
  • the fault data of the corresponding class is generated through multiple interpolation and noise addition to achieve a balance of various data;
  • Step (5) fault diagnosis.
  • the support vector machine is selected as the fault diagnosis model, and its input data are 6 main features of each data sample extracted by principal component analysis method. Firstly, the class balance data set is used to train the support vector machine, and then the classification accuracy of the trained support vector machine is tested by using the test set data (the data volume of each category is 1000).
  • Fig. 2 has provided the classification accuracy rate obtained after using the method proposed by the present invention and the class balance data set training support vector machine obtained by synthetic minority class upsampling technology respectively, and also provided the classification accuracy rate obtained without using the data generation method Rate.
  • the method proposed by the present invention and the synthetic minority class upsampling technique can both improve the classification accuracy of the classifier, and the method proposed by the present invention can obtain the highest classification accuracy, which proves that the data proposed by the present invention
  • the generated data obtained by the generative method is of high quality, which is beneficial to improve the performance of the classifier.

Abstract

一种类不平衡数据集下的机械故障智能诊断方法，包括：步骤(1)、数据预处理：把机械振动信号转换到频域，并把幅值归一化到[0,1]范围；步骤(2)、模型搭建：把自动编码器和生成对抗网络进行组合，搭建数据生成模型；步骤(3)、模型训练：利用故障数据按照预设的损失函数和优化算法训练所述数据生成模型；步骤(4)、数据生成：利用所述数据生成模型在训练中学习到的故障数据低维特征，通过多次插值、加噪后生成对应类的故障数据，实现各类数据平衡；步骤(5)、故障诊断：利用类平衡数据集训练预设的故障诊断模型，利用训练好的故障诊断模型对机械故障进行智能诊断。利用自动编码器、生成对抗网络的结合，实现机械故障诊断。

Description

类不平衡数据集下的机械故障智能诊断方法 技术领域
本发明涉及故障智能诊断领域，具体涉及一种类不平衡数据集下的机械故障智能诊断方法。
背景技术
随着旋转机械设备不断朝着智能化、精密化、复杂化方向发展，机械设备的结构日趋复杂和紧凑。机械设备在服役过程中，一旦某个零部件出现故障，将会影响整个机械设备的运行，甚至引发安全事故。为了确保机械设备的健康运行，深度学习理论作为模式识别和机器学习领域最新的研究成果开始逐步运用在机械故障智能诊断中。相较于传统故障诊断方法，基于深度学习的智能诊断模型利用深度网络模型自适应地从信号中提取有效故障特征，其诊断效率高、不依赖操作者信号处理经验，受到了广泛的关注。
目前在机械故障智能诊断中常用的模型包括卷积神经网络(CNN)、深度置信网络(DBN)、残差网络(ResNet)等。这些模型在训练过程中，往往需要输入大量历史数据集作为训练样本，从而建立数据与健康状态类别的对应关系。机械设备发生故障虽然会给设备运行带来很大的安全隐患，但是故障的发生是一个偶发事件，设备不会在故障状态下长期运行，所以正常状态数据多、故障状态数据少，从而导致数据集的类不平衡问题。正常类与故障类的不平衡给机械健康状态识别带来了很大困难和挑战，类不平衡数据集容易导致诊断模型性能下降，即模型容易对样本数量较多的正常信号过拟合，而对数量较少的故障信号欠拟合。此外，由于故障样本稀少，模型在提取故障数据特征的过程中很容易 学习到其中的一些冗余甚至是不相关的特征,这些特征降低了模型的泛化能力。
为了解决类不平衡带来的机械故障智能诊断模型性能下降问题，常用的方法有动态权重法和数据生成法。动态权重法通过调整网络中的权重参数给予数量较少的故障样本更多的关注，从而改善对于故障样本的欠拟合问题。数据生成法则是通过利用少量故障数据来生成同类别新的样本，用以扩充故障样本，使故障数据与正常数据达到平衡，用平衡的数据集训练智能诊断模型。传统的数据生成方法有合成少数类增采样技术(SMOTE)、自适应合成抽样(ADASYN)等。
传统技术存在以下技术问题：
在实际类不平衡数据集下的机械故障智能诊断中，动态权重法需要根据正常和故障样本间的不平衡率来动态调整权重，所以适合应用在不平衡率已知的情况。且当数据出现极端类不平衡时，动态权重法容易受少量故障样本中的冗余特征干扰，造成模型过拟合，降低了故障诊断的准确率。数据生成法通过对少量故障信号进行增采样，使数据集中的各类达到平衡，从根本上解决了类不平衡现象。然而机械结构复杂，呈非线性特性，其振动信号在实际工况下往往具有较强的背景噪声，在故障状态下表现为明显的非平稳特性。传统的数据生成方法没有学习数据的分布特性，直接在时域信号中通过插值技术进行信号生成，容易受测量噪声成分的干扰，且生成数据的质量不高，同样容易造成智能诊断模型性能下降。
发明内容
本发明要解决的技术问题是提供一种类不平衡数据集下的机械故障智能诊断方法，针对动态权重法应用场景受限、传统数据生成方法易受噪声干扰和生成数据质量不高的问题，本发明提出一种新的数据生成方法，该方法以深度神 经网络为基础，通过自动编码器与生成对抗网络的结合，学习故障信号的低维分布特性，根据低维特征生成故障数据，从而免受噪声的干扰，获得高质量的生成数据，利用平衡数据集实现高性能的机械故障智能诊断。
为了解决上述技术问题，本发明提供了一种类不平衡数据集下的机械故障智能诊断方法，包括：
步骤(1)、数据预处理：把机械振动信号转换到频域，并把幅值归一化到[0,1]范围；
步骤(2)、模型搭建：把自动编码器和生成对抗网络进行组合，搭建数据生成模型；
步骤(3)、模型训练：利用故障数据按照预设的损失函数和优化算法训练所述数据生成模型；
步骤(4)、数据生成：利用所述数据生成模型在训练中学习到的故障数据低维特征，通过多次插值、加噪后生成对应类的故障数据，实现各类数据平衡；
步骤(5)、故障诊断：利用类平衡数据集训练预设的故障诊断模型，利用训练好的故障诊断模型对机械故障进行智能诊断。
在其中一个实施例中，步骤(2)中，所述自动编码器由编码器和解码器组成，所述生成对抗网络由生成器和判别器组成，解码器就是生成器；所述自动编码器通过编码器学习输入数据即真数据的低维特征，再通过低维特征及其类别标签经过解码器输出与输入数据分布特性一致的生成数据即假数据；所述生成对抗网络中的判别器分别对所述输入数据和所述生成数据进行真假判别和类别分类。
在其中一个实施例中，所述编码器、所述解码器和所述判别器都包括通过深度卷积网络、深度置信网络、残差网络中的一种来构建。
在其中一个实施例中，步骤(3)中，所述预设的损失函数包括生成器生成 数据与编码器输入数据之间的均方误差损失函数、判别器对真假数据的交叉熵分类损失函数、判别器对数据真假鉴别的Wasserstein距离或者二值交叉熵损失函数以及编码器输出特征和判别器中间隐含特征之间的均方误差损失函数。
在其中一个实施例中，步骤(3)中，所述预设的优化算法包括但不限于随机梯度下降法(SGD)、带动量的随机梯度下降(Momentum)、Nesterov动量法、Adagrad算法、自适应矩估计法(Adam)中的一种。
在其中一个实施例中，步骤(4)中，所述插值是在同类别故障样本的不同低维特征中进行的，在生成故障数据之前需嵌入该类别的标签，加入的噪声为低幅值的随机噪声。
在其中一个实施例中，步骤(5)中，所述预设的故障诊断模型包括支持向量机、k最近邻算法、随机森林、模糊系统或者深度神经网络中的一种。
基于同样的发明构思，本申请还提供一种计算机设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述程序时实现任一项所述方法的步骤。
基于同样的发明构思，本申请还提供一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现任一项所述方法的步骤。
基于同样的发明构思，本申请还提供一种处理器，所述处理器用于运行程序，其中，所述程序运行时执行任一项所述的方法。
本发明的有益效果：
与现有技术相比，本发明公开了一种类不平衡数据集下的机械故障智能诊断方法。本方法针对机械故障诊断中数据集类不平衡导致的诊断精度下降问题，提出一种新的数据生成方法，利用深度学习的特征挖掘能力和对抗训练机制，学习少量故障样本的数据分布特性，在数据的低维特征空间利用插值和加噪产生新的特征，嵌入标签后通过生成器获得新的故障样本。在低维空间进行插值可以排除信号中测量噪声的影响，加入随机噪声可以提高生成样本的多样性， 嵌入标签可以保证生成样本与同类别故障样本数据分布的一致性。因此，该方法至少具有以下优点：(1)能够学习数据的低维分布特性，排除测量噪声的干扰；(2)生成数据具有与同类别故障数据的一致性，同时兼具一定的多样性，生成数据的质量高；(3)机械故障智能识别的准确率高。
附图说明
图1是本发明类不平衡数据集下的机械故障智能诊断方法的流程图。
图2是本发明类不平衡数据集下的机械故障智能诊断方法中得到的四种故障类型的生成数据与真实数据的对比图，左侧一列为四种故障状态下的真实数据，右侧一列为与之相对应的生成数据。
图3是本发明类不平衡数据集下的机械故障智能诊断方法及传统方法在五种不平衡率下的分类精度变化曲线。
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。
如图1所示的一种类不平衡数据集下的机械故障智能诊断方法流程图，该技术具体包括：
步骤101：数据预处理。对振动信号进行傅立叶变换，将机械振动信号转换到频域，并把幅值归一化到[0,1]范围。
步骤102：模型搭建。把自动编码器和生成对抗网络进行组合，搭建数据生成模型。
自动编码器由编码器和解码器组成，生成对抗网络由生成器和判别器组成， 解码器就是生成器。自动编码器通过编码器学习输入数据(真数据)的低维特征，再通过低维特征及其类别标签经过解码器输出与输入数据分布特性一致的生成数据(假数据)。生成对抗网络中的判别器分别对输入数据和生成数据进行真假判别和类别分类。
编码器、解码器、判别器包括但不限于通过深度卷积网络、深度置信网络、残差网络中的一种来构建。
步骤103：模型训练。利用故障数据按照预设的损失函数和优化算法训练数据生成模型。
数据生成模型在训练过程中的损失函数包括：
1)生成器生成数据与编码器输入数据之间的均方误差损失函数。优化该损失函数可以保证生成数据与输入数据分布特性的一致性。
2)判别器对真假数据的交叉熵分类损失函数。优化真数据的交叉熵分类损失函数可以提高判别器对于真实数据的分类能力；优化假数据的交叉熵分类损失函数可以提高判别器对于生成数据的类别鉴别能力和生成器对于类别特征的学习能力，保证同类别生成数据之间的特征一致性和不同类别生成数据之间的特征差异性。
3)判别器对数据真假鉴别的Wasserstein距离或者二值交叉熵损失函数。优化该损失函数可以进一步提高生成器生成数据的质量和判别器的鉴别能力。
4)编码器输出特征和判别器中间隐含特征之间的均方误差损失函数。优化该损失函数可以提高生成器和判别器对同类别数据所提取特征的一致性。
通过对以上损失函数的优化，最终使生成数据接近于同类别输入数据的数据分布，而判别器难以鉴别同类别生成数据与输入数据的真假，生成器与鉴别器之间达到平衡，完成数据生成模型的训练。
预设的优化算法包括但不限于随机梯度下降法(SGD)、带动量的随机梯度 下降(Momentum)、Nesterov动量法、Adagrad算法、自适应矩估计法(Adam)中的一种。
步骤104：数据生成。利用数据生成模型在训练中学习到的故障数据低维特征，通过多次插值、加噪后生成对应类的故障数据，实现各类数据平衡。
插值是在同类别故障样本的不同低维特征中进行的，在生成故障数据之前需嵌入该类别的标签，加入的噪声为低幅值的随机噪声。
步骤105、故障诊断。利用类平衡数据集训练预设的故障诊断模型，利用训练好的故障诊断模型对机械故障进行智能诊断。
预设的故障诊断模型包括但不限于支持向量机、k最近邻算法、随机森林、模糊系统、深度神经网络中的一种。
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。
以齿轮箱故障智能诊断为例，搭建行星齿轮箱故障仿真试验平台，分别人工设置四种故障状态：断齿、缺齿、齿根裂纹、齿面磨损，加上正常状态共五种健康状态。在行星齿轮箱上安装加速度传感器来采集齿轮箱的振动信号，采样频率为5kHz。每一种健康状态都包含了2000组信号，其中1000组信号作为测试数据不参与训练，每组信号长度为2048个数据点。为了验证本发明中提出的类不平衡数据集下的机械故障智能诊断方法的有效性，实例中设置了5种不平衡率，即齿轮箱健康样本数量与每一类故障样本数量的比值，分别为5:1、10:1、20:1、50:1和100:1，每种不平衡率下健康样本的数据量都是1000个。
采用本发明公开的技术对所述5组类不平衡数据集进行处理，步骤如图1所示，详细信息如下。
步骤(1)、数据预处理。对振动信号进行傅立叶变换，将机械振动信号转换到频域，并把幅值归一化到[0,1]范围。原始时域信号长度为2048个数据点，经过傅里叶变换后取1024长度的频域信号作为模型的输入数据。
步骤(2)、模型搭建。把自动编码器和生成对抗网络进行组合，搭建数据生成模型，具体实施例如下：
①自动编码器：包含编码器和解码器，其主要作用是对输入数据进行编码与解码。编码器采用四层一维卷积神经网络结构，每层维度分别是8、16、32和64，采用长度为15的卷积核，卷积层间连接LeakyReLU激活函数层，样本经过编码器后输出64维的潜在特征向量。解码器采用四层一维反卷积神经网络结构，每层维度分别是64、32、16和8，采用长度为15的反卷积核，反卷积层间连接ReLU激活函数层，解码器最后一层反卷积连接到一个Sigmoid激活函数，将生成数据的幅值限制在[0,1]范围。
②生成对抗网络：包含生成器和判别器。生成器就是自动编码器中的解码器。判别器设计了四个一维卷积层和两个全连接层，每层维度分别是8、16、32和64，采用长度为15的卷积核，每个卷积层间连接了LeakyRelu激活函数层，卷积层最后输出长度为64维的特征向量。随后将该特征分别输入到两个全连接层中，第一个全连接层将64维的特征向量降至1维，用以计算生成数据和真实数据之间的Wasserstein距离。第二个全连接层将64维的特征向量降至4维(即需要增采样的故障类别数目)并连接Softmax激活层，判断信号的类别。
步骤(3)、模型训练。利用故障数据按照预设的损失函数和优化算法训练数据生成模型。在该实施例中有4部分损失函数：
①生成器生成数据与编码器输入数据之间的均方误差损失函数；
②判别器对真假数据的交叉熵分类损失函数；
③判别器对数据真假鉴别的Wasserstein距离函数；
④编码器输出特征和判别器中间隐含特征之间的均方误差损失函数。
对各部分损失函数相加后，通过均方根传递算法(RmsPorp)进行反向传播，依次优化判别器和自动编码器。重复执行模型训练，迭代2000次后模型损失趋于平衡，结束网络训练。
步骤(4)、数据生成。利用数据生成模型在训练中学习到的故障数据低维特征，通过多次插值、加噪后生成对应类的故障数据，实现各类数据平衡；
将同类别的训练样本输入到编码器中，获取输入数据的潜在特征向量。随后选取同类的特征向量进行插值，本实施例中采用K最近邻方法选取特征向量，从低维特征向量任取一个特征向量，并找出其最近邻的3个向量，再从近邻向量中任取一个，进行向量插值。插值扩增后，对新获取的向量加入0.02倍的标准高斯白噪声，并将样本的标签嵌入到加噪后的向量中，实现潜在特征向量的扩增。最后将处理完成的特征向量输入到解码器中，生成新的样本。图2给出了四种故障类型的生成信号与真实信号的对比图，可以看出生成信号服从真实信号的分布规律，同时具有一定的差异性。
步骤(5)、故障诊断。利用类平衡数据集训练预设的故障诊断模型，利用训练好的故障诊断模型对机械故障进行智能诊断。
故障诊断模型选用支持向量机，其输入数据是利用主成分分析方法提取的每个数据样本的6个主特征。首先采用类平衡数据集训练支持向量机，然后采用测试集数据(各类别数据量均为1000)测试训练后的支持向量机的分类准确率。图2给出了分别利用本发明提出的方法和合成少数类增采样技术获得的类平衡数据集训练支持向量机后得到的分类准确率，同时也给出了未采用数据生成方法得到的分类准确率。在不同的不平衡率下，本发明提出的方法和合成少数类增采样技术都可以提高分类器的分类准确率，而本发明提出的方法可以获得最高的分类准确率，证明本发明提出的数据生成方法获得的生成数据的质量高，有利于提高分类器的性能。
综上所述，本发明通过将自动编码器和生成对抗网络结合，利用深度学习的特征挖掘能力和对抗训练机制，可以学习少量故障样本的数据分布特性。此外，通过在低维空间利用插值和加噪来生成潜在特征，再通过解码器生成数据，可以提高对测量噪声的抗干扰能力和数据质量，提升机械故障智能诊断的性能。
以上所述实施例仅是为充分说明本发明而所举的较佳的实施例，本发明的保护范围不限于此。本技术领域的技术人员在本发明基础上所作的等同替代或变换，均在本发明的保护范围之内。本发明的保护范围以权利要求书为准。

Claims (10)

  1. 一种类不平衡数据集下的机械故障智能诊断方法，其特征在于，包括：
    步骤(1)、数据预处理：把机械振动信号转换到频域，并把幅值归一化到[0,1]范围；
    步骤(2)、模型搭建：把自动编码器和生成对抗网络进行组合，搭建数据生成模型；
    步骤(3)、模型训练：利用故障数据按照预设的损失函数和优化算法训练所述数据生成模型；
    步骤(4)、数据生成：利用所述数据生成模型在训练中学习到的故障数据低维特征，通过多次插值、加噪后生成对应类的故障数据，实现各类数据平衡；
    步骤(5)、故障诊断：利用类平衡数据集训练预设的故障诊断模型，利用训练好的故障诊断模型对机械故障进行智能诊断。
  2. 如权利要求1所述的类不平衡数据集下的机械故障智能诊断方法，其特征在于，步骤(2)中，所述自动编码器由编码器和解码器组成，所述生成对抗网络由生成器和判别器组成，解码器就是生成器；所述自动编码器通过编码器学习输入数据即真数据的低维特征，再通过低维特征及其类别标签经过解码器输出与输入数据分布特性一致的生成数据即假数据；所述生成对抗网络中的判别器分别对所述输入数据和所述生成数据进行真假判别和类别分类。
  3. 如权利要求2所述的类不平衡数据集下的机械故障智能诊断方法，其特征在于，所述编码器、所述解码器和所述判别器都包括通过深度卷积网络、深度置信网络、残差网络中的一种来构建。
  4. 如权利要求1所述的类不平衡数据集下的机械故障智能诊断方法，其特征在于，步骤(3)中，所述预设的损失函数包括生成器生成数据与编码器输入数据之间的均方误差损失函数、判别器对真假数据的交叉熵分类损失函数、判 别器对数据真假鉴别的Wasserstein距离或者二值交叉熵损失函数以及编码器输出特征和判别器中间隐含特征之间的均方误差损失函数。
  5. 如权利要求1所述的类不平衡数据集下的机械故障智能诊断方法，其特征在于，步骤(3)中，所述预设的优化算法包括随机梯度下降法、带动量的随机梯度下降、Nesterov动量法、Adagrad算法、自适应矩估计法中的一种。
  6. 如权利要求1所述的类不平衡数据集下的机械故障智能诊断方法，其特征在于，步骤(4)中，所述插值是在同类别故障样本的不同低维特征中进行的，在生成故障数据之前需嵌入该类别的标签，加入的噪声为低幅值的随机噪声。
  7. 如权利要求1所述的类不平衡数据集下的机械故障智能诊断方法，其特征在于，步骤(5)中，所述预设的故障诊断模型包括支持向量机、k最近邻算法、随机森林、模糊系统或者深度神经网络中的一种。
  8. 一种计算机设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，其特征在于，所述处理器执行所述程序时实现权利要求1到7任一项所述方法的步骤。
  9. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，该程序被处理器执行时实现权利要求1到7任一项所述方法的步骤。
  10. 一种处理器，其特征在于，所述处理器用于运行程序，其中，所述程序运行时执行权利要求1到7任一项所述的方法。
PCT/CN2021/123198 2021-09-27 2021-10-12 类不平衡数据集下的机械故障智能诊断方法 WO2023044979A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111136682.5 2021-09-27
CN202111136682.5A CN113935460B (zh) 2021-09-27 2021-09-27 类不平衡数据集下的机械故障智能诊断方法

Publications (1)

Publication Number Publication Date
WO2023044979A1 true WO2023044979A1 (zh) 2023-03-30

Family

ID=79276976

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/123198 WO2023044979A1 (zh) 2021-09-27 2021-10-12 类不平衡数据集下的机械故障智能诊断方法

Country Status (2)

Country Link
CN (1) CN113935460B (zh)
WO (1) WO2023044979A1 (zh)

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116401596A (zh) * 2023-06-08 2023-07-07 哈尔滨工业大学（威海） 基于深度指数激励网络的早期故障诊断方法
CN116432091A (zh) * 2023-06-15 2023-07-14 山东能源数智云科技有限公司 基于小样本的设备故障诊断方法、模型的构建方法及装置
CN116701948A (zh) * 2023-08-03 2023-09-05 东北石油大学三亚海洋油气研究院 管道故障诊断方法及系统、存储介质和管道故障诊断设备
CN116993319A (zh) * 2023-07-14 2023-11-03 南京先维信息技术有限公司 一种基于物联网的远程设备健康监测方法及装置
CN117056734A (zh) * 2023-10-12 2023-11-14 山东能源数智云科技有限公司 基于数据驱动的设备故障诊断模型的构建方法及装置
CN117056814A (zh) * 2023-10-11 2023-11-14 国网山东省电力公司日照供电公司 一种变压器声纹振动故障诊断方法
CN117076935A (zh) * 2023-10-16 2023-11-17 武汉理工大学 数字孪生辅助的机械故障数据轻量级生成方法及系统
CN117332342A (zh) * 2023-11-29 2024-01-02 北京宝隆泓瑞科技有限公司 一种基于半监督学习的机泵设备运行故障分类方法及装置
CN117593783A (zh) * 2023-11-20 2024-02-23 广州视景医疗软件有限公司 基于自适应smote的视觉训练方案生成方法及装置
CN117610614A (zh) * 2024-01-11 2024-02-27 四川大学 基于注意力引导的生成对抗网络零样本核电密封检测方法

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114057053B (zh) * 2022-01-18 2022-04-26 杭州浅水数字技术有限公司 用于特种机械的部件疲劳程度监测方法
CN114611233B (zh) * 2022-03-08 2022-11-11 湖南第一师范学院 一种旋转机械故障不平衡数据生成方法及计算机设备
CN114993677B (zh) * 2022-05-11 2023-05-02 山东大学 不平衡小样本数据的滚动轴承故障诊断方法及系统
CN116204786B (zh) * 2023-01-18 2023-09-15 北京控制工程研究所 生成指定故障趋势数据的方法和装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110428004A (zh) * 2019-07-31 2019-11-08 中南大学 数据失衡下基于深度学习的机械零部件故障诊断方法
US20200202221A1 (en) * 2018-12-20 2020-06-25 Shandong University Of Science And Technology Fault detection method and system based on generative adversarial network and computer program
CN112396088A (zh) * 2020-10-19 2021-02-23 西安交通大学 一种小样本下隐式激励对抗训练的机械故障智能诊断方法
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
CN113298230A (zh) * 2021-05-14 2021-08-24 西安理工大学 一种基于生成对抗网络的不平衡数据集的预测方法

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113239991B (zh) * 2021-04-28 2024-02-27 浙江工业大学 基于回归生成对抗网络的火焰图像氧浓度预测方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200202221A1 (en) * 2018-12-20 2020-06-25 Shandong University Of Science And Technology Fault detection method and system based on generative adversarial network and computer program
CN110428004A (zh) * 2019-07-31 2019-11-08 中南大学 数据失衡下基于深度学习的机械零部件故障诊断方法
CN112396088A (zh) * 2020-10-19 2021-02-23 西安交通大学 一种小样本下隐式激励对抗训练的机械故障智能诊断方法
CN113298230A (zh) * 2021-05-14 2021-08-24 西安理工大学 一种基于生成对抗网络的不平衡数据集的预测方法
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

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116401596A (zh) * 2023-06-08 2023-07-07 哈尔滨工业大学（威海） 基于深度指数激励网络的早期故障诊断方法
CN116401596B (zh) * 2023-06-08 2023-08-22 哈尔滨工业大学（威海） 基于深度指数激励网络的早期故障诊断方法
CN116432091A (zh) * 2023-06-15 2023-07-14 山东能源数智云科技有限公司 基于小样本的设备故障诊断方法、模型的构建方法及装置
CN116432091B (zh) * 2023-06-15 2023-09-26 山东能源数智云科技有限公司 基于小样本的设备故障诊断方法、模型的构建方法及装置
CN116993319A (zh) * 2023-07-14 2023-11-03 南京先维信息技术有限公司 一种基于物联网的远程设备健康监测方法及装置
CN116993319B (zh) * 2023-07-14 2024-01-26 南京先维信息技术有限公司 一种基于物联网的远程设备健康监测方法及装置
CN116701948A (zh) * 2023-08-03 2023-09-05 东北石油大学三亚海洋油气研究院 管道故障诊断方法及系统、存储介质和管道故障诊断设备
CN116701948B (zh) * 2023-08-03 2024-01-23 东北石油大学三亚海洋油气研究院 管道故障诊断方法及系统、存储介质和管道故障诊断设备
CN117056814B (zh) * 2023-10-11 2024-01-05 国网山东省电力公司日照供电公司 一种变压器声纹振动故障诊断方法
CN117056814A (zh) * 2023-10-11 2023-11-14 国网山东省电力公司日照供电公司 一种变压器声纹振动故障诊断方法
CN117056734A (zh) * 2023-10-12 2023-11-14 山东能源数智云科技有限公司 基于数据驱动的设备故障诊断模型的构建方法及装置
CN117056734B (zh) * 2023-10-12 2024-02-06 山东能源数智云科技有限公司 基于数据驱动的设备故障诊断模型的构建方法及装置
CN117076935A (zh) * 2023-10-16 2023-11-17 武汉理工大学 数字孪生辅助的机械故障数据轻量级生成方法及系统
CN117076935B (zh) * 2023-10-16 2024-02-06 武汉理工大学 数字孪生辅助的机械故障数据轻量级生成方法及系统
CN117593783A (zh) * 2023-11-20 2024-02-23 广州视景医疗软件有限公司 基于自适应smote的视觉训练方案生成方法及装置
CN117593783B (zh) * 2023-11-20 2024-04-05 广州视景医疗软件有限公司 基于自适应smote的视觉训练方案生成方法及装置
CN117332342A (zh) * 2023-11-29 2024-01-02 北京宝隆泓瑞科技有限公司 一种基于半监督学习的机泵设备运行故障分类方法及装置
CN117332342B (zh) * 2023-11-29 2024-02-27 北京宝隆泓瑞科技有限公司 一种基于半监督学习的机泵设备运行故障分类方法及装置
CN117610614A (zh) * 2024-01-11 2024-02-27 四川大学 基于注意力引导的生成对抗网络零样本核电密封检测方法
CN117610614B (zh) * 2024-01-11 2024-03-22 四川大学 基于注意力引导的生成对抗网络零样本核电密封检测方法

Also Published As

Publication number Publication date
CN113935460B (zh) 2023-08-11
CN113935460A (zh) 2022-01-14

Similar Documents

Publication Publication Date Title
WO2023044979A1 (zh) 类不平衡数据集下的机械故障智能诊断方法
CN109580215B (zh) 一种基于深度生成对抗网络的风电传动系统故障诊断方法
Liang et al. Rolling bearing fault diagnosis based on one-dimensional dilated convolution network with residual connection
WO2023044978A1 (zh) 基于对抗流模型的机械设备无监督故障诊断方法
CN110212528B (zh) 一种配电网量测数据缺失重构方法
CN105738109A (zh) 基于稀疏表示与集成学习的轴承故障分类诊断方法
CN106443447A (zh) 一种基于iSDAE的航空发电机故障特征提取方法
Liu et al. Imbalanced fault diagnosis of rotating machinery using autoencoder-based SuperGraph feature learning
CN108106844A (zh) 一种自适应参数调节的自动编码机的轴承故障诊断方法
CN105678343A (zh) 基于自适应加权组稀疏表达的水电机组噪声异常诊断方法
CN114778112A (zh) 风电发电机组机械故障音频识别和故障诊断方法
CN105424366A (zh) 基于eemd自适应消噪的轴承故障诊断方法
CN111783531A (zh) 一种基于sdae-ielm的水轮机组故障诊断方法
CN105335698A (zh) 一种基于自适应遗传算法和som网络的齿轮故障诊断方法
CN111275108A (zh) 基于生成对抗网络对局部放电数据进行样本扩展的方法
CN109060347A (zh) 基于堆叠消噪自动编码器和门控循环单元神经网络的行星齿轮故障识别方法
Lu et al. Unbalanced bearing fault diagnosis under various speeds based on spectrum alignment and deep transfer convolution neural network
CN116628592A (zh) 一种基于改进型生成式对抗网络的动设备故障诊断方法
Liang et al. Multibranch and multiscale dynamic convolutional network for small sample fault diagnosis of rotating machinery
CN114897138A (zh) 基于注意力机制和深度残差网络的系统故障诊断方法
CN115290326A (zh) 一种滚动轴承故障智能诊断方法
CN115409052A (zh) 一种数据失衡下风力发电机组轴承的故障诊断方法及系统
Huo et al. A Generative Adversarial Network Based a Rolling Bearing Data Generation Method Towards Fault Diagnosis
CN113505817A (zh) 不均衡数据下轴承故障诊断模型样本自适应加权训练方法
CN114235409A (zh) 一种轻量化通信的滚动轴承多用户协同智能故障诊断方法

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: 21958094

Country of ref document: EP

Kind code of ref document: A1