Multi-scale fault diagnosis method combining pressure-induced attention mechanism with convolutional neural network

Abstract

The present invention relates to a multi-scale fault diagnosis method combining a pressure-induced attention mechanism with a convolutional neural network, and the method comprises the following steps: step 1: collecting a vibration signal at a position of a diesel engine cylinder head; step 2: performing variational modal decomposition on the signal based on the collected vibration signal to obtain a plurality of modal components u as representations of time domain information of the vibration signal; step 3: performing fast Fourier transform on the signal based on the collected vibration signal to obtain an FFT amplitude spectrum magnitude_spectrum as representations of frequency domain information of the vibration signal; step 4: constructing a data set of a deep learning model based on step 2 and step 3; step 5: constructing a SEBlock-CNN model; and step 6: extracting important features in the time and frequency domain data by respectively inputting the time domain information u and the frequency domain information magnitude_spectrum into the SENBlock-CNN, thereby achieving the purpose of fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of diesel engine fault diagnosis, in particular to a multi-scale fault diagnosis method. Background Art

[0002] Vibration signals are one of the most commonly used data sources in mechanical fault diagnosis and can provide important information about the operating status of the equipment. Vibration signals are usually analyzed through time domain analysis and frequency domain analysis to extract features. Time domain analysis identifies fault features by directly analyzing the signal waveform, but the identification effect of complex fault modes and periodic faults is limited. Frequency domain analysis reveals the frequency components of the signal through methods such as Fourier transform, which is particularly suitable for the detection of periodic faults. However, using time domain or frequency domain analysis alone often cannot fully capture all the information of the signal. Therefore, combining time domain and frequency domain data for comprehensive analysis has become an effective way to improve the accuracy of fault diagnosis.

[0003] Deep learning, especially convolutional neural networks (CNNs), has demonstrated powerful feature extraction capabilities in mechanical fault diagnosis. CNNs automatically learn features from raw signals through convolutional layers, overcoming the limitations of manual feature extraction in traditional methods. However, although CNNs have achieved certain success in mechanical fault diagnosis, deep learning models still face challenges such as noise interference, feature redundancy, difficulty in distinguishing different fault modes, and multi-scale feature fusion.

[0004] In recent years, Squeeze-and-Excitation (SE) Block, as a channel attention mechanism, has been widely used in deep learning models. SE Block adaptively weights different channel features, improves the model's ability to focus on key features, and effectively enhances the accuracy of fault diagnosis.

[0005] In this context, a multi-scale feature extraction method combining time domain and frequency domain data has emerged. The present invention provides a multi-scale fault diagnosis method based on a combination of a pressure-induced attention mechanism and a convolutional neural network. This method can not only simultaneously consider the time domain changes and frequency domain characteristics of the signal, but also automatically learn key features through a deep learning network, significantly improving the accuracy and robustness of fault diagnosis. In particular, the multi-scale feature extraction method can capture multi-level information of the signal from different time and frequency scales, providing more comprehensive support for fault diagnosis.

[0006] Currently, using time domain or frequency domain analysis alone often cannot fully capture all the information of the signal. Convolutional neural networks (CNNs) face challenges such as noise interference, feature redundancy, difficulty in distinguishing different fault modes, and multi-scale feature fusion. Summary of the invention

[0007] This paper proposes a multi-scale fault diagnosis method that combines the pressure-induced attention mechanism with a convolutional neural network. By processing time domain and frequency domain data simultaneously, features are extracted from different time scales and frequency ranges. This multi-scale analysis can help the model capture the key features of different fault modes, improve the ability to identify complex faults, and thus improve the diagnostic accuracy.

[0008] To achieve the above purpose, the technical solution of the present invention is: a multi-scale fault diagnosis method combining a pressure-induced attention mechanism with a convolutional neural network, the steps of which are:

[0009] Step 1: Collect the vibration signal at the cylinder head of the diesel engine;

[0010] Step 2: Based on the vibration signal collected in step 1, perform variational mode decomposition (VMD) on the signal to obtain multiple modal components u as the representation of the time domain information of the vibration signal;

[0011] Step 3: Based on the vibration signal collected in step 1, perform fast Fourier transform on the signal to obtain the FFT magnitude spectrum magnitude_spectrum as the frequency domain information representation of the vibration signal;

[0012] Step 4: Based on steps 2 and 3, construct a data set for the deep learning model;

[0013] Step 5: Build the SEBlock-CNN model;

[0014] Step 6: By inputting the time domain information u and the frequency domain information magnitude_spectrum into SENBlock-CNN respectively, important features of the time and frequency domain data are extracted to achieve the purpose of fault diagnosis.

[0015] Furthermore, the method for constructing the data set is to concatenate the modal component u obtained by VMD decomposition and the magnitude spectrum magnitude_spectrum of FFT along the rows to form a new feature matrix, and save the feature data to disk for model training, verification and testing.

[0016] Furthermore, two SEBlock-CNN modules are used as a network architecture running in parallel.

[0017] Furthermore, the specific method of constructing the SEBlock-CNN model is:

[0018] (1) The input data is the time domain features and corresponding frequency domain features of a one-dimensional time series signal. These data are concatenated row by row to form a new feature matrix, which is input into the model for feature extraction.

[0019] (2) Use the Conv1D convolution layer to extract the features of the time-frequency signal, set parameters such as the convolution kernel size, input channels, and output channels; use BatchNorm for normalization after convolution, and use the ReLU activation function to introduce nonlinearity;

[0020] (3) Introducing the SEBlock module: After convolution layer 1, the SEBlock module is added to adaptively weight the extracted feature channels. SEBlock first compresses the features of each channel into a scalar through global average pooling, and then weights them through a fully connected layer to finally obtain the weight coefficient of each channel, which is normalized to a value between [0, 1] through a Sigmoid activation function. The weight is multiplied by the output feature map of the convolution layer to enhance the features of important channels and suppress irrelevant information.

[0021] (4) Set the number of convolutional layers according to actual needs to prevent overfitting, then use AdaptiveMaxPool1D for adaptive pooling to compress the spatial dimensions of features and retain key information, and finally input the fully connected layer for fault classification.

[0022] Furthermore, the SEBlock-CNN model is trained and optimized. The loss function is as follows: for the classification task, the cross entropy loss function is selected for optimization. The Adam optimizer is used for parameter update. The training data is divided into multiple batches. The gradient is calculated through the back-propagation algorithm to update the convolution kernel, fully connected layer, and SEBlock-CNN module parameters.

[0023] The beneficial effects of the present invention are:

[0024] The present invention combines the pressure-induced attention mechanism with the convolutional neural network multi-scale fault diagnosis method, introduces the SEBlock module, enhances CNN's attention to key features, and thus improves the performance of the model in diesel engine fault diagnosis. SEBlock automatically adjusts the weight of each feature channel, can more effectively capture important information in the fault signal, suppress noise, and significantly improve the accuracy and robustness of diagnosis.

[0025] The present invention extracts features from different time scales and frequency ranges by processing time domain and frequency domain data simultaneously. This multi-scale analysis can help the model capture the key features of different fault modes, improve the ability to identify complex faults, and thus improve the diagnostic accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the algorithm flow framework diagram of the present invention;

[0027] Figure 2 It is the confusion matrix diagram of CNN1d;

[0028] Figure 3 It is the Resnet confusion matrix diagram;

[0029] Figure 4 It is the GRU confusion matrix diagram;

[0030] Figure 5 This is the confusion matrix diagram of the multi-scale feature extraction model of SEBlock-CNN. DETAILED DESCRIPTION

[0031] In order to make the advantages and benefits of the technical solution provided by the present invention more clearly reflected, the technical solution provided by the present invention is now further described in detail with reference to the accompanying drawings, specifically:

[0032] Embodiment 1: Figures 1 to 5 As shown, this embodiment provides a diesel engine intake and exhaust valve fault diagnosis model based on SEBlock module and CNN. The model improves the accuracy of feature extraction by adaptively adjusting the weight of feature channels, and improves the accuracy of fault diagnosis by multi-scale feature extraction. The specific implementation steps are as follows:

[0033] Step 1: Data preprocessing. First, collect one-dimensional vibration signals from the cylinder head. Data collection can be performed through sensors; VMD decomposition is performed on the time domain signal to generate time domain features in different frequency ranges; Fast Fourier transform is performed on the time domain signal to generate corresponding frequency domain features. Time-frequency conversion can help capture the time and frequency characteristics of the signal and reveal the different frequency components in the fault signal.

[0034] Step 2: Build the SEBlock-CNN model architecture

[0035] 1. The input data is the time domain features and corresponding frequency domain features of a one-dimensional time series signal. These data are concatenated row by row to form a new feature matrix, which is then input into the model for feature extraction;

[0036] 2. Use the Conv1D convolution layer to extract the features of the time-frequency signal, set parameters such as the convolution kernel size, input channels, and output channels; use BatchNorm for normalization after convolution, and use the ReLU activation function to introduce nonlinearity.

[0037] Using convolution kernels of different sizes in the convolution layer can capture different scale features in the time domain signal. Smaller convolution kernels focus mainly on high-frequency components, while larger convolution kernels focus on low-frequency components.

[0038] 3. Introduce SEBlock module: After convolution layer 1, add SEBlock module to adaptively weight the extracted feature channels. SEBlock first compresses the features of each channel into a scalar through global average pooling, and then weights it through the fully connected layer to finally obtain the weight coefficient of each channel, and normalize it to a value between [0,1] through the Sigmoid activation function. The weight is multiplied by the output feature map of the convolution layer to enhance the features of important channels and suppress irrelevant information.

[0039] 4. Set the number of convolution layers according to actual needs to prevent overfitting. Then use AdaptiveMaxPool1D for adaptive pooling to compress the spatial dimensions of features and retain key information, and finally input the fully connected layer for fault classification.

[0040] Step 3: Model training and optimization

[0041] Loss function: For classification tasks, select the cross-entropy loss function for optimization. Use the Adam optimizer to update parameters. Adam can adaptively adjust the learning rate to meet the training needs of different models and data. Divide the training data into multiple batches, calculate the gradient through the back-propagation algorithm, and update the parameters of the convolution kernel, fully connected layer, SEBlock module, etc.

[0042] Embodiment 2:

[0043] Step 1. Preparation before training:

[0044] The data used in this embodiment are vibration signals of three cylinders on an actual diesel engine, and three different degrees of faults, namely, intake valve fault, exhaust valve fault, and intake and exhaust valve fault, are set respectively. First, the collected vibration signal is preprocessed. The preprocessing means used in the present invention is to use the vibration signal to perform a method based on VMD and FFT. This method extracts the frequency domain features and time domain features of the signal while retaining the original signal information, providing more useful input for subsequent classification tasks.

[0045] Step 2. Training parameter settings:

[0046] In order to enable the model to fully learn the important features in the time-frequency domain information of the vibration signal, the learning rate was set to 0.0001, the number of iterations was set to 20, the batch size was set to 32, and the Adam optimizer was used to optimize the cross entropy loss function.

Claims

1. A multi-scale fault diagnosis method combining pressure-induced attention mechanism and convolutional neural network, characterized in that: The steps are: Step 1: Collect the vibration signal at the cylinder head of the diesel engine; Step 2: Based on the vibration signal collected in step 1, perform variational mode decomposition (VMD) on the signal to obtain multiple modal components u as the representation of the time domain information of the vibration signal; Step 3: Based on the vibration signal collected in step 1, perform fast Fourier transform on the signal to obtain the FFT magnitude spectrum magnitude_spectrum as the frequency domain information representation of the vibration signal; Step 4: Based on steps 2 and 3, construct a data set for the deep learning model; Step 5: Build the SEBlock-CNN model; Step 6: By inputting the time domain information u and the frequency domain information magnitude_spectrum into SENBlock-CNN respectively, important features of the time and frequency domain data are extracted to achieve the purpose of fault diagnosis.

2. The method for multi-scale fault diagnosis combining pressure-induced attention mechanism and convolutional neural network according to claim 1 is characterized in that: The method for constructing the data set is to concatenate the modal component u obtained by VMD decomposition and the magnitude spectrum magnitude_spectrum of FFT along the rows to form a new feature matrix, and save the feature data to disk for model training, verification and testing.

3. The method for multi-scale fault diagnosis combining pressure-induced attention mechanism and convolutional neural network according to claim 1, characterized in that: A network architecture using two SEBlock-CNN modules running in parallel.

4. The method for multi-scale fault diagnosis combining pressure-induced attention mechanism and convolutional neural network according to claim 1, characterized in that: The specific method of building the SEBlock-CNN model is: (1) The input data is the time domain features and corresponding frequency domain features of a one-dimensional time series signal. These data are concatenated row by row to form a new feature matrix, which is input into the model for feature extraction. (2) Use the Conv1D convolution layer to extract the features of the time-frequency signal, set parameters such as the convolution kernel size, input channels, and output channels; use BatchNorm for normalization after convolution, and use the ReLU activation function to introduce nonlinearity; (3) Introducing the SEBlock module: After the convolution layer 1, the SEBlock module is added to adaptively weight the extracted feature channels. SEBlock first compresses the features of each channel into a scalar through global average pooling, and then weights them through the fully connected layer to finally obtain the weight coefficient of each channel, which is normalized to a value between [0, 1] through the Sigmoid activation function; The weights are multiplied by the output feature map of the convolutional layer to enhance the features of important channels and suppress irrelevant information; (4) Set the number of convolutional layers according to actual needs to prevent overfitting, then use AdaptiveMaxPool1D for adaptive pooling to compress the spatial dimensions of features and retain key information, and finally input the fully connected layer for fault classification.

5. The method for multi-scale fault diagnosis combining pressure-induced attention mechanism and convolutional neural network according to claim 4 is characterized in that: Train and optimize the SEBlock-CNN model. Loss function: For classification tasks, select the cross entropy loss function for optimization, use the Adam optimizer for parameter update, divide the training data into multiple batches, calculate the gradient through the back propagation algorithm, and update the convolution kernel, fully connected layer, and SEBlock-CNN module parameters.

Images