技术领域Technical field
本发明涉及神经网络技术领域，尤其涉及一种包含先验知识神经网络的辊压机多传感器诊断方法。The present invention relates to the field of neural network technology, and in particular to a multi-sensor diagnosis method for a roller press that includes a priori knowledge neural network.
背景技术Background technique
传统的机械故障诊断主要依赖于信号处理方法，如傅里叶变换、小波变换等对振动信号进行时频转换，以提取故障相关特征，但这种方法需要大量专业知识，难以推广。近年来，基于深度学习的端到端数据驱动方法因其高诊断准确率受到广泛关注，其中，卷积神经网络(CNN)能够自动从原始样本中提取高维特征，在机械故障诊断中应用广泛。Traditional mechanical fault diagnosis mainly relies on signal processing methods, such as Fourier transform, wavelet transform, etc., to perform time-frequency conversion of vibration signals to extract fault-related features. However, this method requires a lot of professional knowledge and is difficult to promote. In recent years, end-to-end data-driven methods based on deep learning have received widespread attention due to their high diagnostic accuracy. Among them, convolutional neural networks (CNN) can automatically extract high-dimensional features from original samples and are widely used in mechanical fault diagnosis. .
但是，卷积神经网络(CNN)是一个典型的黑箱模型，其决策机制不够清晰，降低了结果的可信度，限制了模型在高可靠性要求的故障诊断场景中的应用，当前CNN可解释性研究主要面向2D图像，不适用于1D机械振动信号，少数工作试图提高CNN在机械故障诊断中的可解释性，但解释通常不明确，需要主观理解，并可能降低诊断性能或泛化能力。因此，研发一种可解释的CNN模型对机械系统故障诊断具有重要意义。However, the convolutional neural network (CNN) is a typical black box model, and its decision-making mechanism is not clear enough, which reduces the credibility of the results and limits the application of the model in fault diagnosis scenarios with high reliability requirements. The current CNN can explain Sexual research is mainly oriented to 2D images and is not applicable to 1D mechanical vibration signals. A few works try to improve the interpretability of CNN in mechanical fault diagnosis, but the explanation is usually unclear, requires subjective understanding, and may reduce diagnostic performance or generalization ability. Therefore, developing an interpretable CNN model is of great significance for fault diagnosis of mechanical systems.
发明内容Contents of the invention
本发明的目的是为了解决现有技术中的问题，而提出的一种包含先验知识神经网络的辊压机多传感器诊断方法。The purpose of the present invention is to propose a multi-sensor diagnosis method for roller presses that includes a priori knowledge neural network in order to solve the problems in the prior art.
一种包含先验知识神经网络的辊压机多传感器诊断方法，包括以下步骤：A multi-sensor diagnosis method for roller presses containing prior knowledge neural networks includes the following steps:
S1、构建时频网络，所述时频网络包含时频卷积层和CNN主干网络，所述时频卷积层对原始振动信号进行时频转换，提取故障相关特征，所述CNN主干网络对特征进行进一步提取和分类；S1. Construct a time-frequency network. The time-frequency network includes a time-frequency convolution layer and a CNN backbone network. The time-frequency convolution layer performs time-frequency conversion on the original vibration signal and extracts fault-related features. The CNN backbone network Features are further extracted and classified;
S2、在辊压机上安装多个传感器进行状态检测；S2. Install multiple sensors on the roller press for status detection;
S3、将多个测点传感器所获得的振动数据输入时频网络，获得诊断结果。S3. Input the vibration data obtained by multiple measuring point sensors into the time-frequency network to obtain the diagnosis results.
在上述的一种包含先验知识神经网络的辊压机多传感器诊断方法中，所述时频卷积层的核函数为：In the above-mentioned multi-sensor diagnosis method for roller presses containing a priori knowledge neural network, the kernel function of the time-frequency convolution layer is:
其中，s为缩放因子，n为采样点索引；未缩放的Morlet母小波函数为：Among them, s is the scaling factor, n is the sampling point index; the unscaled Morlet mother wavelet function is:
其中，Ψ(n)为母小波函数；Nc为时频卷积层的通道数；σ为窗口参数；f0为中心频率。Among them, Ψ(n) is the mother wavelet function; Nc is the number of channels in the time-frequency convolution layer; σ is the window parameter; f0 is the center frequency.
在上述的一种包含先验知识神经网络的辊压机多传感器诊断方法中，在时频卷积层中，振动数据卷积得到实部时频特征图和虚部时频特征图，所述实部时频特征图和虚部时频特征图融合得到经过时频变换的融合特征。In the above-mentioned multi-sensor diagnosis method for roller presses containing a priori knowledge neural network, in the time-frequency convolution layer, the vibration data is convolved to obtain the real part time-frequency feature map and the imaginary part time-frequency feature map, as described The real part time-frequency feature map and the imaginary part time-frequency feature map are fused to obtain the fusion feature after time-frequency transformation.
在上述的一种包含先验知识神经网络的辊压机多传感器诊断方法中，所述时频网络中，振动数据经过时频卷积层后得到融合特征，再经过CNN主干网络对特征进行进一步提取得到单传感器分支输出。In the above-mentioned multi-sensor diagnosis method for roller presses that includes a priori knowledge neural network, in the time-frequency network, the vibration data is passed through the time-frequency convolution layer to obtain fusion features, and then the features are further processed through the CNN backbone network. Extract the single sensor branch output.
在上述的一种包含先验知识神经网络的辊压机多传感器诊断方法中，所述时频网络中，将多个单传感器分支输出进行特征融合后依次经过展平层、全连接层、激活函数、全连接层后输出分类结果。In the above-mentioned multi-sensor diagnosis method for a roller press that includes a priori knowledge neural network, in the time-frequency network, the outputs of multiple single sensor branches are feature fused and then go through the flattening layer, the fully connected layer, and the activation layer in sequence. function and fully connected layer to output the classification result.
与现有的技术相比，本发明优点在于：Compared with existing technology, the advantages of the present invention are:
1、本发明中的时频卷积层，将时频转换方法参数化并嵌入卷积层中，既提升性能，又具解释性，通过小波时频转换分析对时频卷积层进行可解释，揭示CNN模型的决策时频依据。1. The time-frequency convolution layer in the present invention parameterizes the time-frequency conversion method and embeds it in the convolution layer, which not only improves performance but is also interpretable. The time-frequency convolution layer is interpretable through wavelet time-frequency conversion analysis. , revealing the decision-making time-frequency basis of the CNN model.
2、本发明将多传感器的时频信息进行融合，使得诊断更加准确全面，同时时频卷积层相比CNN，减少了参数数量，具有更快收敛速度和更强的诊断能力。2. This invention fuses time-frequency information from multiple sensors to make diagnosis more accurate and comprehensive. At the same time, compared with CNN, the time-frequency convolution layer reduces the number of parameters, has faster convergence speed and stronger diagnostic ability.
3、构建端到端、高性能且可解释的时频网络，实现机械故障的智能诊断，并且时频卷积层核函数的设计增加了模型的物理约束，融合了人工的专家知识。3. Construct an end-to-end, high-performance and interpretable time-frequency network to achieve intelligent diagnosis of mechanical faults. The design of the time-frequency convolution layer kernel function increases the physical constraints of the model and integrates artificial expert knowledge.
附图说明Description of the drawings
图1为本发明中时频卷积层结构示意图。Figure 1 is a schematic diagram of the structure of the time-frequency convolution layer in the present invention.
图2为本发明中单传感器分支网络结构示意图。Figure 2 is a schematic structural diagram of a single sensor branch network in the present invention.
图3为本发明中时频网络结构示意图。Figure 3 is a schematic diagram of the time-frequency network structure of the present invention.
图4为本发明实施例诊断结果混淆矩阵图。Figure 4 is a confusion matrix diagram of diagnosis results according to the embodiment of the present invention.
图5为本发明中PBCNN与其他主流神经网络在不同信噪比下的测试准确率柱状图。Figure 5 is a histogram of the test accuracy of PBCNN and other mainstream neural networks under different signal-to-noise ratios in the present invention.
具体实施方式Detailed ways
为使本发明实现的技术手段、创作特征、达成目的与功效易于明白了解，下面结合具体实施方式，进一步阐述本发明。In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention will be further elaborated below in conjunction with specific implementation modes.
参照图1-3所示，一种包含先验知识神经网络的辊压机多传感器诊断方法，包括以下步骤：Referring to Figure 1-3, a multi-sensor diagnosis method for roller presses including a priori knowledge neural network includes the following steps:
S1、构建时频网络，所述时频网络包含时频卷积层和CNN主干网络，所述时频卷积层对原始振动信号进行时频转换，提取故障相关特征，所述CNN主干网络对特征进行进一步提取和分类；S1. Construct a time-frequency network. The time-frequency network includes a time-frequency convolution layer and a CNN backbone network. The time-frequency convolution layer performs time-frequency conversion on the original vibration signal and extracts fault-related features. The CNN backbone network Features are further extracted and classified;
S2、在辊压机上安装多个传感器进行状态检测；S2. Install multiple sensors on the roller press for status detection;
S3、将多个测点传感器所获得的振动数据输入时频网络并进行特征融合，获得诊断结果，相对于单测点数据包含的信息更多，判断更加准确。S3. Input the vibration data obtained by multiple measuring point sensors into the time-frequency network and perform feature fusion to obtain diagnostic results. Compared with the data of a single measuring point, it contains more information and the judgment is more accurate.
时频卷积层将小波时频转换方法嵌入其中作为核函数，通过小波时频转换分析对时频卷积层进行解释，揭示CNN模型的决策时频依据，所述时频卷积层的核函数为：The time-frequency convolution layer embeds the wavelet time-frequency conversion method as a kernel function. The time-frequency convolution layer is explained through wavelet time-frequency conversion analysis to reveal the time-frequency basis for decision-making of the CNN model. The kernel of the time-frequency convolution layer The function is:
其中，s为缩放因子(可训练参数，在训练过程中，s会根据目标函数不断更新，以调整Morlet小波核的时域和频域特性，从而对信号进行自适应的时频转换)，n为采样点索引；未缩放的Morlet母小波函数为：Among them, s is the scaling factor (a trainable parameter. During the training process, s will be continuously updated according to the objective function to adjust the time domain and frequency domain characteristics of the Morlet wavelet kernel, thereby performing adaptive time-frequency conversion of the signal), n is the sampling point index; the unscaled Morlet mother wavelet function is:
其中，Ψ(n)为母小波函数；Nc为时频卷积层的通道数；σ为窗口参数；f0为中心频率。Among them, Ψ(n) is the mother wavelet function; Nc is the number of channels in the time-frequency convolution layer; σ is the window parameter; f0 is the center frequency.
在时频卷积层中，振动数据卷积得到实部时频特征图和虚部时频特征图，所述实部时频特征图和虚部时频特征图融合得到经过时频变换的融合特征，时频卷积层的可训练参数为核函数的控制参数，而不是随机初始化的卷积核权重，不同的时频转换方法对应不同的核函数。In the time-frequency convolution layer, the vibration data is convolved to obtain a real part time-frequency feature map and an imaginary part time-frequency feature map. The real part time-frequency feature map and the imaginary part time-frequency feature map are fused to obtain a fusion after time-frequency transformation. Features: The trainable parameters of the time-frequency convolution layer are the control parameters of the kernel function, rather than the randomly initialized convolution kernel weights. Different time-frequency conversion methods correspond to different kernel functions.
所述时频网络中，振动数据经过时频卷积层后得到融合特征，再经过CNN主干网络对特征进行进一步提取得到单传感器分支输出，所述时频网络中，将多个单传感器分支输出进行特征融合后依次经过展平层、全连接层、激活函数、全连接层后输出分类结果，时频卷积层可以泛化到不同的CNN模型中，显著提升其诊断性能。In the time-frequency network, the vibration data obtains fusion features after passing through the time-frequency convolution layer, and then further extracts the features through the CNN backbone network to obtain the output of a single sensor branch. In the time-frequency network, multiple single sensor branches are output After feature fusion, the classification results are output after passing through the flattening layer, fully connected layer, activation function, and fully connected layer. The time-frequency convolution layer can be generalized to different CNN models, significantly improving its diagnostic performance.
实施例Example
实验使用水泥厂辊压机设备，分别以辊压机轴承B3G11B作为实验研究对象，系统采样频率为51200HZ，取一次故障数据进行验证。The experiment used cement plant roller press equipment, and the roller press bearing B3G11B was used as the experimental research object. The system sampling frequency was 51200HZ, and one fault data was taken for verification.
在辊压机设备上设置16个检测点，取16个检测点的振动数据，将数据按照4096的长度进行切分，16个测点的一段长度为4096的数据为一个样本，正常和故障状态分别取2000个样本组成数据集，将数据集样本按8:1:1的比例划分为训练集、测试集和验证集，处理完成后的细节如表1所示：Set 16 detection points on the roller press equipment, take the vibration data of the 16 detection points, and divide the data according to the length of 4096. A section of data with a length of 4096 from the 16 measurement points is a sample, normal and fault status Take 2000 samples to form a data set, and divide the data set samples into a training set, a test set and a verification set in a ratio of 8:1:1. The details after the processing are completed are shown in Table 1:
表1：训练集、测试集和验证集训练数量分配表Table 1: Allocation table of training set, test set and validation set training numbers
一种包含先验知识神经网络的辊压机多传感器网络，简称PBCNN，对数据集进行诊断，诊断结果混淆矩阵如图4所示，为了验证准确性，将数据集添加噪声后，与其他主流神经网络进行准确率对比。A roller press multi-sensor network containing a priori knowledge neural network, referred to as PBCNN, diagnoses the data set. The diagnosis result confusion matrix is shown in Figure 4. In order to verify the accuracy, after adding noise to the data set, it is compared with other mainstream Neural network accuracy comparison.
表2：PBCNN与其他主流神经网络的噪声测试准确率表Table 2: Noise test accuracy table of PBCNN and other mainstream neural networks
PBCNN与其他主流神经网络在不同信噪比下的测试准确率采用柱状图显示(图5)。The test accuracy of PBCNN and other mainstream neural networks under different signal-to-noise ratios is displayed in a histogram (Figure 5).
综上所述，PBCNN中的时频卷积层，将时频转换方法参数化并嵌入卷积层中，既提升性能，又具解释性，PBCNN融合了多传感器的时频信息，诊断更加准确全面，时频卷积层相比CNN，减少了参数数量，具有更快收敛速度和更强的诊断能力。In summary, the time-frequency convolution layer in PBCNN parameterizes the time-frequency conversion method and embeds it in the convolution layer, which not only improves performance but is also interpretable. PBCNN integrates time-frequency information from multiple sensors and makes diagnosis more accurate. Comprehensive, the time-frequency convolution layer reduces the number of parameters compared to CNN, has faster convergence speed and stronger diagnostic ability.
由技术常识可知，本发明可以通过其它的不脱离其精神实质或必要特征的实施方案来实现。因此，上述公开的实施方案，就各方面而言，都只是举例说明，并不是仅有的。所有在本发明范围内或在等同于本发明的范围内的改变均被本发明包含。It is known from common technical knowledge that the present invention can be implemented by other embodiments without departing from its spirit or essential characteristics. Therefore, the above-disclosed embodiments are in all respects illustrative and not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311448674.3ACN117473441A (en) | 2023-11-02 | 2023-11-02 | A multi-sensor diagnosis method for roller presses containing prior knowledge neural network |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311448674.3ACN117473441A (en) | 2023-11-02 | 2023-11-02 | A multi-sensor diagnosis method for roller presses containing prior knowledge neural network |
| Publication Number | Publication Date |
|---|---|
| CN117473441Atrue CN117473441A (en) | 2024-01-30 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311448674.3APendingCN117473441A (en) | 2023-11-02 | 2023-11-02 | A multi-sensor diagnosis method for roller presses containing prior knowledge neural network |
| Country | Link |
|---|---|
| CN (1) | CN117473441A (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100106458A1 (en)* | 2008-10-28 | 2010-04-29 | Leu Ming C | Computer program and method for detecting and predicting valve failure in a reciprocating compressor |
| CN113807190A (en)* | 2021-08-20 | 2021-12-17 | 西安理工大学 | A fault diagnosis method for train bearings based on STFT and CNN |
| CN114705427A (en)* | 2022-03-04 | 2022-07-05 | 重庆邮电大学 | Rolling bearing intelligent fault diagnosis method under strong noise environment |
| CN116593160A (en)* | 2023-06-27 | 2023-08-15 | 嘉兴南湖学院 | Rolling bearing fault diagnosis method under variable working condition and heavy load |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100106458A1 (en)* | 2008-10-28 | 2010-04-29 | Leu Ming C | Computer program and method for detecting and predicting valve failure in a reciprocating compressor |
| CN113807190A (en)* | 2021-08-20 | 2021-12-17 | 西安理工大学 | A fault diagnosis method for train bearings based on STFT and CNN |
| CN114705427A (en)* | 2022-03-04 | 2022-07-05 | 重庆邮电大学 | Rolling bearing intelligent fault diagnosis method under strong noise environment |
| CN116593160A (en)* | 2023-06-27 | 2023-08-15 | 嘉兴南湖学院 | Rolling bearing fault diagnosis method under variable working condition and heavy load |
| Title |
|---|
| 陈仁祥;黄鑫;杨黎霞;汤宝平;余腾伟;周君;: "基于卷积神经网络和离散小波变换的滚动轴承故障诊断", 振动工程学报, no. 05, 15 October 2018 (2018-10-15)* |
| Publication | Publication Date | Title |
|---|---|---|
| CN110059601B (en) | An Intelligent Fault Diagnosis Method Based on Multi-feature Extraction and Fusion | |
| CN109299705B (en) | Fault diagnosis method for rotating machinery based on one-dimensional deep residual convolutional neural network | |
| CN115979643B (en) | Rolling bearing fault diagnosis method and system | |
| Zhang et al. | Efficient federated convolutional neural network with information fusion for rolling bearing fault diagnosis | |
| CN110516305B (en) | Intelligent fault diagnosis method under small sample based on attention mechanism meta-learning model | |
| CN110702411B (en) | Residual error network rolling bearing fault diagnosis method based on time-frequency analysis | |
| CN111964908A (en) | A bearing fault diagnosis method under variable working conditions based on MWDCNN | |
| CN111256965B (en) | A Stacked Sparse Autoencoder Fault Diagnosis Method for Rotating Machinery Based on Multiscale Information Fusion | |
| CN110595775A (en) | Rolling bearing fault diagnosis method based on multi-branch and multi-scale convolutional neural network | |
| CN111986699B (en) | Sound event detection method based on full convolution network | |
| CN115392333B (en) | A device fault diagnosis method based on an improved end-to-end ResNet-BiLSTM dual-channel model | |
| CN112729835B (en) | A multi-domain semi-supervised fault diagnosis method and device for axial piston pump bearing | |
| CN113158984B (en) | Bearing fault diagnosis method based on complex Morlet wavelet and lightweight convolutional network | |
| CN111914703A (en) | A fault diagnosis method for mechanical rotating parts based on wavelet transform and transfer learning GoogLeNet | |
| CN111751133A (en) | An Intelligent Fault Diagnosis Method Based on Non-local Mean Embedding Deep Convolutional Neural Network Model | |
| CN114331214A (en) | Domain-adaptive bearing voiceprint fault diagnosis method and system based on reinforcement learning | |
| CN112146882A (en) | Bearing fault diagnosis method based on transfer learning vibration signal image recognition | |
| CN115687973A (en) | A Fault Diagnosis Method for Mechanical Equipment Based on Multi-signal Fusion | |
| CN116106016A (en) | A Rolling Bearing Fault Diagnosis Method Based on Joint Learning Convolutional Neural Network | |
| WO2023035757A1 (en) | Biopolymer characterization method, apparatus, and device, and computer storage medium | |
| CN118776881A (en) | A bearing fault diagnosis method based on CA-Transformer multi-source signal fusion | |
| CN115908142B (en) | Visual identification-based damage inspection method for tiny contact net parts | |
| CN112305379A (en) | Mode identification method and system for GIS insulation defect | |
| CN116662872A (en) | Fault Diagnosis Method of Bearing Machinery in Automobile Panel Finishing Line | |
| CN114997210B (en) | A machine abnormal noise recognition and detection method based on deep learning |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |