A method for predicting the life of turntable bearings based on neural network architecture search

Through the search method of neural network architecture based on Cell and wavelet packet decomposition technology, the problem of long search time and difficult to extract fault signals in turntable bearing life prediction is solved, and fast and accurate life prediction and model performance improvement is achieved.

CN116628482BActive Publication Date: 2024-11-05NANJING TECH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310561242.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-05-18
Publication Date
2024-11-05
Estimated Expiration
2043-05-18

AI Technical Summary

Technical Problem

The prior art has the problem of too long search time in the prediction of turntable bearing life, which makes the neural network architecture time-consuming and labor-intensive, and it is difficult for convolutional neural networks to effectively extract fault signals in low-speed environments.

Method used

Using the search method of neural network architecture based on Cell, the search space is turned into continuous space, and the gradient descent optimization method is used to find the optimal architecture, and the data set is constructed by decomposing the wavelet packet and extracting time domain features. The neural network architecture is automatically output to shorten it. Search time.

Benefits of technology

It realizes a fast and automated neural network architecture design, improves the accuracy and model performance of turntable bearing life prediction, reduces computing resource consumption, and improves the safety and economicality of engineering production.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116628482B_ABST
    Figure CN116628482B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of life prediction of turntable bearings, and in particular, to a life prediction method for turntable bearings based on neural network architecture search, comprising the following steps S1: performing wavelet packet decomposition and noise reduction on a turntable bearing vibration signal, and extracting time domain features to construct a turntable bearing data set; S2: constructing a search space based on Cell, wherein a Cell is a directed acyclic graph consisting of an ordered sequence of N nodes; S3: searching for an optimal Cell through the turntable bearing data set, and stacking the optimal Cell to form a final network architecture; S4: predicting the life of a turntable bearing using the network architecture obtained in S3; the present invention uses a neural network architecture search to automatically search for a neural network architecture that meets the requirements for life prediction of a turntable bearing, thereby freeing up manpower and improving the safety and economy in engineering production.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of turntable bearing life prediction, and in particular to a turntable bearing life prediction method based on neural network architecture search. Background Art

[0002] The slewing bearing is a large bearing that can bear axial load, radial load, overturning moment and other combined loads at the same time. It is mainly used in large equipment such as fan equipment, port machinery, engineering machinery, military machinery, etc. The working environment is relatively harsh and failures are also more frequent. Therefore, predicting its remaining service life can help predict failures in advance to reduce safety hazards.

[0003] Due to the diverse loads and low speed of turntable bearings, weak fault signals are easily masked by noise, and the research on the fault mechanism of turntable bearings is relatively scarce. The previous life prediction methods for ordinary bearings are not completely applicable. With the rapid development of deep learning technology, bearing life prediction technology based on deep learning has been widely used. For example, convolutional neural networks have a strong ability to discover implicit information in the original signal without knowing the damage mechanism of the component.

[0004] Convolutional neural networks perform well in predicting the life of turntable bearings, but the form and structure of neural networks vary according to different tasks and requirements. For example, classic convolutional neural network models such as AlexNet, VGGNet, and GoogLeNet require extensive hyperparameter tuning and architecture adjustments when facing different task requirements, otherwise the model performance will be greatly reduced. However, designing specific architectures for different tasks and requirements will take a lot of expert time, and the whole process is time-consuming and laborious.

[0005] Neural network architecture search aims to automatically generate network architecture using algorithms. Currently, there are mainly reinforcement learning-based methods and evolutionary algorithm-based methods for neural network architecture search. However, these two methods consume huge computing resources and require thousands of GPU days of search time, which are difficult to apply in practice. Therefore, it is necessary to solve the problem of long search time. The present invention proposes a differentiable neural network architecture search method, which converts the search space into a continuous space and the objective function into a differentiable function, so that a gradient-based optimization method can be used to find the optimal architecture, which greatly shortens the search time, and uses a Cell-based search space to stack Cells in multiple layers to increase the network depth and improve model performance. Summary of the invention

[0006] The purpose of the present invention is to provide a turntable bearing life prediction method based on neural network architecture search, automatically output the neural network architecture, realize the turntable bearing life prediction, liberate manpower, and improve the accuracy of prediction.

[0007] To achieve the above purpose, the technical solution provided by the present invention is as follows:

[0008] A method for predicting the life of a turntable bearing based on a neural network architecture search comprises the following steps:

[0009] S1: The vibration signal of the turntable bearing is decomposed and denoised by wavelet packet, and the time domain features are extracted to construct the turntable bearing dataset.

[0010] S2: Construct a Cell-based search space, where Cell is a directed acyclic graph consisting of an ordered sequence of N nodes.

[0011] S3: Search for the best Cell through the turntable bearing dataset and stack the best Cell to form the final network architecture.

[0012] S4: Use the network architecture obtained in S3 to predict the life of the turntable bearing.

[0013] Furthermore, step S1 includes the following steps:

[0014] S1.1: Decompose the turntable bearing vibration signal by wavelet packet to reduce noise. After the signal is decomposed by wavelet packet, the wavelet coefficient generated by the signal contains important information of the signal. After the signal is decomposed, the wavelet coefficient is larger and the wavelet coefficient of the noise is smaller. A threshold is selected, and the signal greater than the threshold is retained, and the signal less than the threshold is set to zero.

[0015] S1.2: Normalize the denoised signal to speed up the convergence of the neural network.

[0016] S1.3: Use the denoised signal to perform time domain feature extraction and extract 8 feature parameters: kurtosis, mean, variance, root mean square, waveform index, peak index, margin index and skewness index.

[0017] S1.4: Use the denoised signal and 8 time domain features to construct a dataset, and divide the dataset into training set, validation set, and test set.

[0018] Further, step S2 includes the following steps:

[0019] S2.1: Construct a Cell-based search space. A Cell is a directed acyclic graph consisting of an ordered sequence of N nodes. Each intermediate node is calculated based on all its previous nodes. A Cell is defined as having two input nodes and one output node. For a convolutional unit, the input node is defined as the unit output in the first two layers.

[0020] S2.2: After the definition of S2.1, the process of learning Cell can be simplified to the process of learning each edge operation. The operation on each edge of the directed acyclic graph is initially unknown. By placing mixed candidate operations on each edge to continuously relax the search space, the mixing probability and network weight are jointly optimized by solving the two-layer optimization problem, and finally the final architecture is determined by the mixing probability.

[0021] Furthermore, in step S2.1, the optional operations in the search space include max_pool_3×3, avg_pool_3×3, sep_conv_3×3, sep_conv_5×5, dil_conv_3×3, dil_conv_5×5 and skip_connect, which respectively represent maximum pooling 3×3, average pooling 3×3, separable convolution 3×3, separable convolution 5×5, dilated separable convolution 3×3, dilated separable convolution 5×5 and skip connection.

[0022] Further, step S3 includes the following steps:

[0023] S3.1: Mix the candidate operations in the search space with the softmax function, turn the search space into a continuous space, and turn the objective function into a differentiable function so that a gradient-based search strategy can be used to shorten the search time.

[0024] S3.2: Input the training set and validation set into the network for architecture search. During the architecture search process, the architecture α and the weight ω of the neural network need to be learned together. The training loss and validation loss α and ω are jointly determined, and the validation loss is optimized using gradient descent. The goal of architecture search is to find the minimum validation loss, which transforms the architecture search problem into a two-layer optimization problem, with α as the upper layer variable and ω as the lower layer variable. After the architecture search process is completed, only the most appropriate operations are retained between the two nodes, and the best Cell is finally output.

[0025] S3.3: Stack the best cells in a predefined way to form the final network architecture. There are two types of best cells output by the S3.2 architecture search, Normal Cell and Reduction Cell. The convolution kernel step size of Normal Cell is 1, which does not change the size of the feature map. The convolution kernel step size of Reduction Cell is 2, which reduces the length and width of the feature map to half of the original. Reduction Cell is located at 1 / 3 and 2 / 3 of the total depth of the network.

[0026] Furthermore, in step S3.2, the evaluation accuracy of the architecture will be given when the best Cell is output, but the evaluation accuracy of the architecture cannot represent the actual prediction accuracy of the model. In the process of architecture evaluation, in order to reduce the consumption of computing resources, the model will not be fully trained. Therefore, there is an error between the evaluation accuracy and the actual accuracy. In order to obtain the actual accuracy, the architecture needs to be used to retrain the model.

[0027] Furthermore, in step S4, the training set is input into the neural network for model retraining, and the prediction accuracy is tested using the test set. The best Epoch is determined according to the prediction accuracy under different Epochs, and the model trained with the best Epoch can be used to predict the life of the turntable bearing.

[0028] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0029] The present invention uses a neural network architecture search, selects a Cell-based search space, and searches through a gradient descent search strategy. It can automatically search for a neural network architecture that meets the requirements for turntable bearing life prediction, thereby freeing up manpower and improving the safety and economy in engineering production. Cells can be stacked to effectively enhance model performance, and it has the characteristics of low computing resource consumption and high prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0031] Figure 2 A schematic diagram of a turntable bearing test bench of the present invention;

[0032] Figure 3 This is a real picture of the turntable bearing test bench of the present invention;

[0033] Figure 4 This is a vibration acceleration signal diagram after noise reduction of the present invention;

[0034] Figure 5 It is a curve diagram of characteristic value of vibration signal of the present invention;

[0035] Figure 6 It is an overview diagram of the search space of the present invention;

[0036] Figure 7 The best Normal Cell of the present invention;

[0037] Figure 8 The best reduction cell of the present invention;

[0038] Fig. 9 A schematic diagram of the network architecture of the present invention;

[0039] Fig.10 is the confusion matrix of the recognition result of the present invention;

[0040] Fig.11 This is a comparison chart of the recognition accuracy of the present invention. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0042] Example 1

[0043] This embodiment discloses a turntable bearing life prediction method based on neural network architecture search, such as Figure 1 , including the following steps:

[0044] S1: The vibration signal of the turntable bearing is decomposed and denoised by wavelet packet, and the time domain features are extracted to construct the turntable bearing dataset.

[0045] S2: Construct a Cell-based search space, where Cell is a directed acyclic graph consisting of an ordered sequence of N nodes.

[0046] S3: Search for the best Cell through the turntable bearing dataset and stack the best Cell to form the final network architecture.

[0047] S4: Use the network architecture obtained in S3 to predict the life of the turntable bearing.

[0048] In step S1, in order to obtain the vibration signal of the turntable bearing, an accelerated life test of the turntable bearing is carried out, which mainly includes mechanical, hydraulic and measurement and control parts. The test bench is driven by a hydraulic motor, and the turntable bearing meshing with it is driven to rotate through a small gear. The schematic diagram of the test bench is shown in Figure 2 As shown in the actual picture Figure 3 shown.

[0049] In step S1, an axial force of 96 kN and an overturning force of 240 kNm are selected as the limit loads according to factors such as the specifications and load-bearing capacity of the turntable bearing. The experiment lasted for 11 days. All components of the turntable bearing showed performance degradation until they got stuck. At this time, the turntable bearing was in a state of wear and failure. The vibration signal of the turntable bearing is decomposed and denoised by wavelet packets. After the signal is decomposed by wavelet packets, the wavelet coefficients generated by the signal contain important information about the signal. After the signal is decomposed, the wavelet coefficients are larger and the wavelet coefficients of the noise are smaller. A threshold is selected, and signals greater than the threshold are retained, and signals less than the threshold are set to zero. The sym wavelet function is selected to perform a 5-layer wavelet packet decomposition on the signal, and the soft threshold function is selected to quantize the coefficients of each layer. Finally, the signal is reconstructed using the processed coefficients. The denoised vibration signal is as follows: Figure 4 shown.

[0050] The signal after noise reduction has several obvious amplitude mutation points, and they are all located at the dividing point of two days. This is caused by the sampling method of this experiment. In this experiment, the acceleration sensor samples once a day, the sampling frequency is 2048Hz, and each sampling lasts for 10 minutes, that is, the interval between each sampling is 23 hours and 50 minutes. The turntable bearing wears during the sampling interval, and the change in vibration amplitude will only be reflected in the next sampling. Due to the low speed of the turntable bearing, the bearing vibration signal is easily masked by noise, so it is difficult to see these amplitude mutation points in the vibration signal before noise reduction.

[0051] The denoised signal is used for time domain feature extraction, and 8 feature parameters are extracted: kurtosis, mean, variance, root mean square, waveform index, peak index, margin index and skewness index. Some of the characteristic value curves are shown in Figure 2. Figure 5 shown.

[0052] The data set is constructed using the denoised signal and 8 time domain features. 200 samples of vibration signals are selected every day, and a total of 2200 samples are selected for 11 days. The data set is divided into a training set, a validation set, and a test set in a ratio of 6:2:2 using a stratified sampling method. According to the sampling method of this experiment, a classification model is selected, and the labels of all samples on the first day are set to 10, indicating that their remaining service life is 10 days. Similarly, the label on the 11th day is 0, so the model in this embodiment is an 11-classification model.

[0053] In step S2, the optional operations in the search space include max_pool_3×3, avg_pool_3×3, sep_conv_3×3, sep_conv_5×5, dil_conv_3×3, dil_conv_5×5 and skip_connect, which represent maximum pooling 3×3, average pooling 3×3, separable convolution 3×3, separable convolution 5×5, dilated separable convolution 3×3, dilated separable convolution 5×5 and skip connection, respectively. Each Cell consists of 7 nodes, where the input node is the output of the previous two Cells, and the output node is the deep concatenation of all intermediate nodes. The process of learning Cell can be simplified to the process of learning each edge operation, such as Figure 6 As shown, the operations on each edge of the directed acyclic graph are initially unknown (see Figure 6 (a)), by placing a mix of candidate operations on each edge and continuously relaxing the search space (see Figure 6 (b)), by solving a two-level optimization problem, the mixing probability and network weights are jointly optimized (see Figure 6 (c)), and finally the final architecture is determined by mixing the probability (see Figure 6 (d)).

[0054] In the specific implementation process, the RTX3090 GPU, Python3.8, deep learning framework PyTorch, and server version 1.8.0 were used for experiments. The experiment is mainly divided into two parts: neural network architecture search and model retraining. The neural network architecture search is used to automatically obtain the best Cell, and the Cell is stacked to form the final network architecture. The obtained architecture is then used to retrain the model to verify the performance of the architecture. The architecture search process takes 2 hours and 40 minutes, and the model retraining takes 3 hours and 30 minutes.

[0055] In step S3, the candidate operations in the search space are mixed with the softmax function, the search space is turned into a continuous space, and the objective function is turned into a differentiable function, so that a gradient-based search strategy can be used to shorten the search time. The training set and the validation set are input into the network for architecture search. During the architecture search process, the architecture α and the weight ω of the neural network need to be learned together. The training loss and validation loss α and ω are jointly determined, and gradient descent is used to optimize the validation loss. The goal of the architecture search is to find the minimum validation loss, which transforms the architecture search problem into a two-layer optimization problem, where α is the upper-layer variable and ω is the lower-layer variable. After the architecture search process is completed, only the most suitable operations are retained between the two nodes, and the best Cell is output. There are two types of best Cells, Normal Cell and Reduction Cell. The convolution kernel step size of Normal Cell is 1, which does not change the size of the feature map. The convolution kernel step size of Reduction Cell is 2, which reduces the length and width of the feature map to half of the original. The best Normal Cell and Reduction Cell searched in the experiment are as follows Figure 7 , 8 shown.

[0056] The best cells are stacked in a predefined manner to form the final network architecture, where the Reduction Cells are located at 1 / 3 and 2 / 3 of the total depth of the network, and the remaining layers are Normal Cells. For convolutional neural networks, within a certain threshold range, as the number of network layers increases, the network performance will also increase, but the network training time will also increase. This embodiment uses 20 cells to stack to form the final network architecture, where the Reduction Cells are located at the 7th and 14th layers of the network, and the remaining 18 layers are Normal Cells. The network architecture diagram of this embodiment is shown in the figure. Fig. 9 shown.

[0057] To verify the performance of the model, the searched neural network architecture is used to retrain the model, the training set is used to train the model, and its accuracy is verified on the test set. The model recognition accuracy will change with the increase of Epoch. In this embodiment, the accuracy is basically stable after Epoch exceeds 80. At this time, increasing Epoch will cause overfitting and reduce the generalization performance of the model. Therefore, Epoch is selected as 80.

[0058] The Epoch is set to 80 for model training. The confusion matrix of the recognition results on the test set is as follows Fig.10 As shown. The horizontal axis represents the expected output of the sample, and the vertical axis represents the actual output. It can be seen from the confusion matrix that the recognition accuracy is higher in the middle and late stages of the life of the turntable bearing, that is, the period of intensified faults and severe damage, while the recognition accuracy is slightly lower during the normal use period. The reason is that the amplitude fluctuation of the vibration signal during the normal use period is not obvious, which increases the recognition difficulty of the neural network. The overall recognition accuracy of the model in this embodiment has reached 95.9%, which can meet industrial needs.

[0059] The neural network architecture search model is compared with the manually designed classic neural network models AlexNet, VGGNet and GoogLeNet. The comparison results are as follows Fig.11 As shown, it can be seen that the recognition accuracy of the neural network architecture search model is significantly higher than that of the manually designed classic neural network model. If these classic neural network models undergo extensive hyperparameter tuning and architecture adjustment, their recognition accuracy will increase, but this will take a lot of expert time. In contrast, the neural network architecture search can automatically search for the architecture, which greatly liberates manpower and also achieves a higher recognition accuracy, making it more competitive in all aspects.

[0060] The above are only preferred embodiments of the present invention. Those skilled in the art will appreciate that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the protection scope of the present invention.

Claims

1. A method for predicting the life of a turntable bearing based on a neural network architecture search, characterized in that: The steps include: S1: Decompose the turntable bearing vibration signal by wavelet packet to reduce noise, and extract the time domain features to construct the turntable bearing data set; S2: Construct a Cell-based search space, where Cell is a directed acyclic graph consisting of an ordered sequence of N nodes; S3: Search for the best cell through the turntable bearing dataset, and stack the best cells to form the final network architecture; S4: Use the network architecture obtained in S3 to predict the life of the turntable bearing; Wherein, the step S1 specifically includes: S1.1: Decompose the turntable bearing vibration signal by wavelet packet to reduce noise. After the signal is decomposed by wavelet packet, the wavelet coefficient of the signal is larger and the wavelet coefficient of the noise is smaller. A threshold is selected, and the signal greater than the threshold is retained, and the signal less than the threshold is set to zero; S1.2: normalize the denoised signal to speed up the convergence of the neural network; S1.3: Use the denoised signal to perform time domain feature extraction and extract 8 feature parameters: kurtosis, mean, variance, root mean square, waveform index, peak index, margin index and skewness index; S1.4: Use the denoised signal and 8 time domain features to construct a data set, and divide the data set into a training set, a validation set, and a test set; The step S2 specifically includes: S2.1: Construct a cell-based search space. A cell is a directed acyclic graph consisting of an ordered sequence of N nodes. Each intermediate node is calculated based on all its previous nodes. S2.2: The process of learning Cell is simplified to the process of learning each edge operation. The operations on each edge of the directed acyclic graph are initially unknown. The search space is continuously relaxed by placing mixed candidate operations on each edge. By solving the two-layer optimization problem, the mixing probability and network weights are jointly optimized. Finally, the final architecture is determined by the mixing probability.

2. The method for predicting the life of a turntable bearing based on a neural network architecture search according to claim 1, characterized in that: In step S1.4, a stratified sampling method is used to divide the data set into a training set, a validation set, and a test set in a ratio of 6:2:

2.

3. The method for predicting the life of a turntable bearing based on a neural network architecture search according to claim 1, characterized in that: In step S2.1, a Cell is defined as having two input nodes and one output node. For a convolutional cell, the input node is defined as the cell output in the first two layers.

4. The method for predicting the life of a turntable bearing based on a neural network architecture search according to claim 1, characterized in that: In the step S2.1, the optional operations in the search space include max_pool_3×3, avg_pool_3×3, sep_conv_3×3, sep_conv_5×5, dil_conv_3×3, dil_conv_5×5 and skip_connect, which respectively represent maximum pooling 3×3, average pooling 3×3, separable convolution 3×3, separable convolution 5×5, expanded separable convolution 3×3, expanded separable convolution 5×5 and skip connection.

5. The method for predicting the life of a turntable bearing based on a neural network architecture search according to claim 1, characterized in that: The step S3 specifically includes the following steps: S3.1: Mix the candidate operations in the search space with the softmax function to make the search space a continuous space and the objective function a differentiable function; S3.2: Input the training set and validation set into the network for architecture search. After the architecture search process is completed, only the most suitable operation is retained between the two nodes, and the best cell is output; S3.3: Stack the best cells in a predefined manner to form the final network architecture.

6. The method for predicting the life of a turntable bearing based on a neural network architecture search according to claim 5, characterized in that: In the step S3.2, there are two kinds of optimal cells output, Normal Cell and Reduction Cell. The convolution kernel step size of Normal Cell is 1, which does not change the size of the feature map. The convolution kernel step size of Reduction Cell is 2, which reduces the length and width of the feature map to half of the original. Reduction Cell is located at 1 / 3 and 2 / 3 of the total depth of the network.

7. The method for predicting the life of a turntable bearing based on a neural network architecture search according to claim 1, characterized in that: In step S4, the training set is input into the neural network for model retraining, and the prediction accuracy thereof is tested using the test set, and the best Epoch is determined according to the prediction accuracy under different Epochs.