Fault diagnosis method based on convolutional neural network and improved cascade forest

Abstract

The invention belongs to the technical field of fault diagnosis, and particularly relates to a fault diagnosis method based on a convolutional neural network and an improved cascade forest, and the method comprises the steps: obtaining one-dimensional vibration data, carrying out the preprocessing of the one-dimensional vibration data, obtaining a preprocessing data set, and dividing the preprocessing data set into a training set, a test set and a verification set according to a proportion; constructing a double-branch convolutional neural network, and training the double-branch convolutional neural network by adopting the training set; constructing an improved cascade forest, and forming the machine-changed cascade forest and the trained double-branch convolutional neural network to obtain a fault diagnosis model; inputting the to-be-diagnosed data into the fault diagnosis model to obtain a diagnosis result; according to the method, a better effect can be achieved on a fault diagnosis data set.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular relates to a fault diagnosis method based on convolutional neural networks and improved cascade forests. Background Art

[0002] Rotating machinery is widely used in industrial production, where bearings play a crucial role in reducing friction and providing support. Bearing failure can lead to economic losses and even threaten personnel safety. Therefore, detecting and repairing bearing failures is essential to ensure the smooth operation of industrial equipment.

[0003] Traditional methods generally use sensors to collect one-dimensional vibration signals from rotating machinery. Signal processing methods such as Fourier transform (FFT), singular spectrum analysis (SSA), and wavelet transform (WT) are then used to extract data features. These features are then input into classifiers such as support vector machines (SVM) and K-nearest neighbor (KNN) for fault diagnosis. With the development of deep neural networks, intelligent fault diagnosis methods based on convolutional neural networks (CNNs) have been proposed and widely recognized. Among them, researchers have proposed deep forests and demonstrated their strong competitiveness compared to deep neural networks. Replacing classifiers such as SVMs and KNNs with deep forests has achieved promising results.

[0004] However, deep forests consist of multi-granularity scans and cascaded forests. Directly using a CNN for feature extraction and feeding it into a deep forest can significantly increase computational cost and complexity without significantly improving performance, even when the CNN and multi-granularity scans are functionally equivalent. Furthermore, multi-granularity scans require domain expertise for design and tuning to achieve optimal results across diverse datasets. Therefore, a new approach is needed to replace multi-granularity scans. Summary of the Invention

[0005] To solve the above problems, the present invention provides a fault diagnosis method based on convolutional neural network and improved cascade forest, which is characterized by comprising the following steps:

[0006] S1. Acquire one-dimensional vibration data and preprocess it to obtain a preprocessed dataset, and divide the preprocessed dataset into a training set, a test set, and a validation set in proportion;

[0007] S2. Construct a two-branch convolutional neural network, train the two-branch convolutional neural network using the training set and obtain feature data;

[0008] S3. Build an improved cascade forest and train it using feature data. Combine the trained improved cascade forest with a two-branch convolutional neural network to obtain a fault diagnosis model.

[0009] S4. Input the data to be diagnosed into the fault diagnosis model to obtain the diagnosis result.

[0010] Furthermore, the preprocessing of step S1 includes:

[0011] S11. Cut the one-dimensional vibration data every 1024 data points to obtain multiple cutting data;

[0012] S12. Normalizing each cutting data to obtain normalized data;

[0013] S13. Reshape each normalized data into an integer grayscale value in the range [0, 255];

[0014] S14. Normalize each integer grayscale image value and reshape it into a sample that meets the input of the convolutional neural network;

[0015] S15. All samples form a preprocessed dataset, which is divided into a training set, a test set, and a validation set in proportion.

[0016] Furthermore, the dual-branch convolutional neural network includes a first branch, a second branch and a fully connected layer;

[0017] The first branch includes convolutional layer C1, the first BN layer, pooling layer S1, convolutional layer C2, pooling layer S2, and the first global average pooling layer;

[0018] The second branch includes convolution layer C3, the second BN layer, pooling layer S3, convolution layer C4, pooling layer S4, and the second global average pooling layer.

[0019] Furthermore, the convolution layer C1 includes 64 5×5 convolution kernels, the convolution layer C2 includes 128 3×3 convolution kernels, the convolution layer C3 includes 64 7×7 convolution kernels, and the convolution layer C4 includes 128 4×4 convolution kernels; the pooling layer S1, pooling layer S2, pooling layer S3 and pooling layer S4 all use a 2×2 kernel size.

[0020] Furthermore, the PReLU function is used as the activation function in the two-branch convolutional neural network.

[0021] Furthermore, the Adam algorithm is used to train and optimize the two-branch convolutional neural network.

[0022] Furthermore, each layer in the improved cascade forest includes a random forest, two extreme forests and an extreme gradient boosting tree.

[0023] Beneficial effects of the present invention:

[0024] The present invention addresses the problem that multi-granularity scanning in deep forests requires the experience of domain experts to design and adjust feature extractors, which limits its versatility on different data sets. A fault diagnosis method based on convolutional neural networks and improved cascade forests is proposed. Complex features are automatically extracted from input data through a two-branch convolutional neural network without the need for manual design and adjustment of feature extractors, which reduces dependence on the experience of domain experts and improves the versatility and adaptability of the model. At the same time, the multi-branch structure allows different branches to focus on extracting different features of the data, thereby capturing more diverse and rich feature information. Compared with other multi-branch models, the two-branch design reduces the model complexity and the number of training parameters while maintaining the characteristics of a lightweight network, improves computational efficiency, and ensures a more stable and convergent training process. Therefore, in order to address the shortcomings of multi-granularity scanning, feature data is extracted through a two-branch convolutional neural network to obtain more representative features. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flow chart of a method for fault diagnosis using a hybrid model of a convolutional neural network and an improved cascade forest according to the present invention;

[0026] Figure 2 A structural diagram of a hybrid model of a convolutional neural network and an improved cascade forest according to the present invention;

[0027] Figure 3 FIG. 4 is a schematic diagram of converting some signal samples into grayscale images according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0029] Deep learning provides a method for fault detection and has been used in intelligent fault diagnosis since the 2010s. Compared to traditional intelligent fault diagnosis methods that require additional feature extractors or selectors, deep learning can automatically learn valuable features. In other words, deep learning methods avoid various signal processing techniques and can achieve end-to-end learning. Models such as convolutional neural networks (CNNs) and deep forests have attracted much attention due to their excellent performance in fault diagnosis. However, the multi-granularity scanning in deep forests requires redesign and parameter adjustment for different datasets, and directly combining CNNs with deep forests increases complexity. This paper proposes a hybrid deep learning model that combines a multi-branch CNN with an improved cascade forest. First, the use of a multi-branch CNN enables better feature extraction, and removing the multi-granularity scanning reduces model complexity. Second, one of the random forests in the cascade forest is replaced with XGBoost. Because XGBoost is more capable of handling nonlinear data than random forests, it achieves better results on fault diagnosis datasets. This optimizes the overall performance of the model.

[0030] The present invention provides a fault diagnosis method based on convolutional neural network and improved cascade forest, such as Figure 1 As shown, the following steps are included:

[0031] S1. Obtain one-dimensional vibration data and preprocess it to obtain a preprocessed dataset, and divide the preprocessed dataset into a training set, a test set, and a validation set in proportion.

[0032] Specifically, if Figure 3 As shown, the preprocessing of step S1 includes:

[0033] S11. The one-dimensional vibration data is cut every 1024 data points to obtain multiple cutting data; wherein the actual fault data may have various factors such as data entry errors, sensor failures, communication errors, etc. In order to simulate the actual data characteristics, the present invention adds noise to part of the cutting data.

[0034] S12. Normalizing each cutting data to obtain normalized data;

[0035] S13. Reshape each normalized data into an integer grayscale value in the range [0, 255];

[0036] S14. Normalize each integer grayscale image value and reshape it into a sample that meets the input of the convolutional neural network; specifically, reshape it into a size of 1×n or reshape it into an m×n size according to row arrangement;

[0037] S15. All samples form a preprocessed dataset, which is then divided into a training set, a test set, and a validation set in proportion.

[0038] S2. Construct a two-branch convolutional neural network and use the training set to train the two-branch convolutional neural network; input the training set into the trained two-branch convolutional neural network to obtain feature data.

[0039] Specifically, if Figure 2 As shown, the dual-branch convolutional neural network includes a first branch, a second branch and a fully connected layer;

[0040] The first branch includes convolutional layer C1, the first BN layer, pooling layer S1, convolutional layer C2, pooling layer S2, and the first global average pooling layer;

[0041] The second branch includes convolution layer C3, the second BN layer, pooling layer S3, convolution layer C4, pooling layer S4, and the second global average pooling layer.

[0042] Specifically, the convolution layer C1 includes 64 5×5 convolution kernels, the convolution layer C2 includes 128 3×3 convolution kernels, the convolution layer C3 includes 64 7×7 convolution kernels, and the convolution layer C4 includes 128 4×4 convolution kernels; the pooling layer S1, pooling layer S2, pooling layer S3 and pooling layer S4 all use a 2×2 kernel size.

[0043] During training, the parameters of the two-branch convolutional neural network, including batch_size and epochs values, are first initialized. Based on the training set, the optimal batch_size and epochs values ​​of the two-branch convolutional neural network are obtained through grid search. Specifically, the specific process of the two-branch convolutional neural network processing input data includes:

[0044] S21. The input data is input into convolution layer C1 and convolution layer C3 respectively to perform convolution operations to extract the first feature and the second feature.

[0045] S22. To accelerate the training process and improve the stability of the model, batch normalization layers (i.e., the first BN layer and the second BN layer) are used after convolutional layer C1 and convolutional layer C3 to standardize the first feature and the second feature, respectively, to obtain the first standard feature and the second standard feature. The formulas are as follows:

[0046]

[0047] Among them, u represents the batch mean, σ 2 represents the batch variance, x i represents the i-th sample in the batch, γ represents the scaling parameter, β represents the translation parameter, ε represents a small positive number, and y i Represents the batch normalization result.

[0048] S23. Use pooling layer S1 and pooling layer S3 to perform 2×2 pooling operations on the first standard feature and the second standard feature respectively, retain the maximum value in each 2×2 area, and reduce the computational complexity to obtain the first pooled feature and the second pooled feature.

[0049] S24. Use convolution layer C2 and convolution layer C4 to obtain the third feature and the fourth feature respectively from the first pooled feature and the second pooled feature.

[0050] S25. Use pooling layer S2 and pooling layer S4 to perform 2×2 pooling operations on the third feature and the fourth feature respectively to obtain a third pooled feature and a fourth pooled feature.

[0051] S26. Use the first global average pooling layer and the second global average pooling layer to perform average pooling on the third pooled features and the fourth pooled features respectively, converting each feature map into a single value and retaining important global information.

[0052] S27. Concatenate the outputs of the first and second global average pooling layers into a long vector, and use a fully connected layer with 512 neurons to perform feature fusion to generate the final output.

[0053] Specifically, the PReLU function is used as the activation function in the dual-branch convolutional neural network, and the formula is:

[0054]

[0055] Among them, a i represents a learnable parameter, z i Represents the input of the nonlinear activation function f on the i-th channel.

[0056] Specifically, in order to reduce the overfitting tendency of the neural network, a Dropout layer is used to randomly drop 30% of the neurons during training.

[0057] Specifically, the Adam algorithm is used to train and optimize the two-branch convolutional neural network.

[0058] S3. Build an improved cascade forest and use feature data for training. The trained improved cascade forest is combined with a two-branch convolutional neural network to obtain a fault diagnosis model.

[0059] Specifically, based on the existing cascade forest structure, the improved cascade forest includes multiple cascade layers, each layer includes a random forest (Random Forest), two extreme forests (Extra-Trees) and an extreme gradient boosting tree (XGBoost).

[0060] Specifically, the performance of the entire improved cascade forest will adopt the k-fold cross-validation method. A performance test will be done after each layer, and then the next layer will be generated. After expanding a new level, the performance of the entire cascade will be estimated on the validation set. If there is no significant performance gain, the training process will be stopped.

[0061] S4. Input the data to be diagnosed into the fault diagnosis model to obtain the diagnosis result.

[0062] In one embodiment, different data sets are used to implement the content of this method to verify the effectiveness of this method. As shown in the following table:

[0063] Table 1 Test accuracy of Jiangnan University bearing dataset

[0064]

[0065] Table 2 Test accuracy of the University of Ottawa bearing dataset

[0066]

[0067] Table 3 Test accuracy of the bearing dataset of the Association for Mechanical Failure Prevention Technology

[0068]

[0069] The above test results were obtained using the steps of this invention. The first row in the table above indicates the ratio of the number of training samples to the total number of samples; for example, 10% indicates that the number of training samples is 10% of the total number of samples. The data items in the table represent the accuracy rate, such as 0.8476, which indicates that the accuracy rate of that item is 84.76%. For the various convolutional neural networks compared, the parameter settings were set to the original configurations in the original papers. MA1DCNN is the network proposed in the paper "Understanding and Learning Discriminant Features based on Multi-Attention1DCNN for Wheelset Bearing Fault Diagnosis," CNN-gcForest refers to the network proposed in "A hybrid deep-learning model for fault diagnosis of rolling bearings," DRSN is provided by "Deep Residual Shrinkage Networks for Fault Diagnosis," DFCNN is provided by "A new bearing fault diagnosis method based on modified convolutional neural networks," ADCNN is the network architecture proposed in "Automated bearing fault diagnosis scheme using 2D representation of wavelet packet transform and deep convolutional neural network," and QCNN is the network proposed in "Attention-embedded Quadratic Network (Qttention) for Effective and Interpretable Bearing Fault Diagnosis." The optimal batch size and epochs obtained and set through grid search were uniformly set to 128 and 50, respectively. The table above shows that: 1. As an improvement over Deep Forest, our method achieves excellent results on all datasets, improving accuracy by over 10% compared to Deep Forest, particularly when the training sample ratio is low. 2. Our method achieves high accuracy across all datasets, far exceeding other convolutional neural networks, demonstrating its significant advantages on small datasets.

[0070] In the present invention, unless otherwise clearly stipulated and limited, the terms "installation", "setting", "connection", "fixation", "rotation" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0071] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A fault diagnosis method based on convolutional neural network and improved cascade forest, characterized in that: The following steps are involved: S1. Acquire one-dimensional vibration data and preprocess it to obtain a preprocessed dataset, and divide the preprocessed dataset into a training set, a test set, and a validation set in proportion; S2. Construct a two-branch convolutional neural network, train the two-branch convolutional neural network using the training set and obtain feature data; S3. Construct an improved cascade forest and train it using feature data. The trained improved cascade forest is combined with a two-branch convolutional neural network to obtain a fault diagnosis model. S4. Input the data to be diagnosed into the fault diagnosis model to obtain the diagnosis result.

2. A fault diagnosis method based on convolutional neural network and improved cascade forest according to claim 1, characterized in that: The preprocessing of step S1 includes: S11. Cut the one-dimensional vibration data every 1024 data points to obtain multiple cutting data; S12. Normalizing each cutting data to obtain normalized data; S13. Reshape each normalized data into an integer grayscale value in the range [0, 255]; S14. Normalize each integer grayscale image value and reshape it into a sample that meets the input of the convolutional neural network; S15. All samples form a preprocessed dataset, which is divided into a training set, a test set, and a validation set in proportion.

3. A fault diagnosis method based on convolutional neural network and improved cascade forest according to claim 1, characterized in that: The dual-branch convolutional neural network includes a first branch, a second branch and a fully connected layer; The first branch includes convolutional layer C1, the first BN layer, pooling layer S1, convolutional layer C2, pooling layer S2, and the first global average pooling layer; The second branch includes convolution layer C3, the second BN layer, pooling layer S3, convolution layer C4, pooling layer S4, and the second global average pooling layer.

4. A fault diagnosis method based on convolutional neural network and improved cascade forest according to claim 3, characterized in that: Convolutional layer C1 includes 64 5×5 convolution kernels, convolutional layer C2 includes 128 3×3 convolution kernels, convolutional layer C3 includes 64 7×7 convolution kernels, and convolutional layer C4 includes 128 4×4 convolution kernels. Pooling layer S1, pooling layer S2, pooling layer S3 and pooling layer S4 all use a 2×2 kernel size.

5. A fault diagnosis method based on convolutional neural network and improved cascade forest according to claim 3, characterized in that: The PReLU function is used as the activation function in the two-branch convolutional neural network.

6. A fault diagnosis method based on convolutional neural network and improved cascade forest according to claim 1, characterized in that: The Adam algorithm is used to train and optimize the two-branch convolutional neural network.

7. A fault diagnosis method based on convolutional neural network and improved cascade forest according to claim 1, characterized in that: Each layer in the improved cascade forest consists of a random forest, two extreme forests, and an extreme gradient boosted tree.

Images