A method for constructing a life prediction model for rotating parts and its application

Abstract

The present invention belongs to the field of life prediction of rotating machinery components, and specifically relates to a method for constructing a rotating component life prediction model and its application, including: constructing a labeled sample set for working condition one and an unlabeled sample set for working condition two; constructing time-frequency domain features for all samples in the two sample sets; using the constructed time-frequency domain features to perform transfer adversarial training on a neural network model to obtain a rotating component life prediction model for working condition two, wherein, when the parameters of a data source discriminator are fixed during the transfer adversarial training, the maximum mean difference of the output of the neural network for the two working condition data is introduced into the loss function, and the parameters of the life prediction model can be updated based on the data under the two working conditions at the same time during the training process. Reducing the maximum mean difference of the output can reduce the difference between the features extracted in the fully connected layer, thereby improving the prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of life prediction of rotating machinery components, and more specifically, relates to a method for constructing a life prediction model for rotating components and its application. Background Art

[0002] With the continuous development of industry, mechanical equipment is widely used in all aspects of industrial production, and is becoming increasingly complex and intelligent, greatly improving social productivity. At the same time, the safety of mechanical equipment has become particularly important. Bearings are key components of rotating machinery, and their health is directly related to the failure of the entire machinery. If bearing failures can be detected early, major mechanical failures can be effectively avoided and losses can be minimized.

[0003] With the continuous development of information technology and artificial intelligence, intelligent algorithms and technologies are increasingly being integrated with traditional industrial applications, creating new avenues for resolving these challenges. Applying advanced AI algorithms to the industrial sector is crucial for resolving challenges in predicting the remaining life of mechanical equipment, which are difficult to address using traditional signal processing techniques. This, in turn, improves the intelligence of mechanical equipment and its operational reliability.

[0004] Predicting the remaining life of bearings is essentially a regression prediction problem, and the use of neural networks for regression prediction has been widely applied in various fields. However, in practical applications, bearings are often used under different operating conditions, and the characteristics of bearing degradation under different operating conditions often vary. A neural network model trained using bearing data from one operating condition is unlikely to provide a good representation of bearing life under other conditions. Furthermore, in practice, it is difficult to obtain a large amount of labeled data under all operating conditions to optimize the life prediction model. Therefore, improving the efficiency and accuracy of bearing life prediction for various operating conditions remains an unresolved issue. Summary of the Invention

[0005] In response to the defects of the existing technology and the need for improvement, the present invention provides a method for constructing a rotating component life prediction model and its application, the purpose of which is to improve the efficiency and accuracy of rotating component life prediction under target working conditions in the context of a lack of sample label data.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for constructing a life prediction model for a rotating component is provided, comprising:

[0007] Construct a labeled sample set for operating condition 1 and an unlabeled sample set for operating condition 2. Construct time-frequency domain features for all samples in the two sample sets. Use the constructed time-frequency domain features to perform transfer adversarial training on the neural network model to obtain a rotating component life prediction model for operating condition 2.

[0008] Among them, when the parameters of the data source discriminator are fixed during the transfer adversarial training process, the following loss function loss3 is used to adjust the parameters of the lifespan prediction model:

[0009]

[0010] Where loss1 is the life prediction loss function of the life prediction model for working condition 1, loss2 is the data source identification loss function of the data source discriminator, The predicted output corresponding to the input data of working condition 1 and working condition 2 respectively and The maximum mean difference between them, λ1, λ2 and λ3 are the weights of each loss value, and the values ​​are positive. In the process of transfer adversarial training, the value of λ2 is adjusted after each iteration, and the value of λ2 decreases exponentially with the current number of iterations of adversarial training.

[0011] Furthermore, wavelet packet decomposition is used to construct time-frequency domain features.

[0012] Furthermore, the specific construction method of the time-frequency domain features of each sample is as follows:

[0013] Randomly select m signal data segments that can overlap from each sample signal of the rotating component For each signal data segment Perform wavelet packet decomposition:

[0014] The kth signal data segment for According to the formula Get the coefficient of the tth node in the nth layer Where t=1,2,...,2 n-1 , n is the number of wavelet packet decomposition layers, l is the length of each signal data segment; G n is the low-pass filter coefficient matrix corresponding to the n-th layer wavelet packet decomposition; H n is the high-pass filter coefficient matrix corresponding to the n-th layer wavelet packet decomposition;

[0015] Perform threshold filtering on each element in each node coefficient of the nth layer, expressed as Where z is the set threshold, is the node coefficient after filtering, for The u-th element in ;

[0016] Sum the elements of each node coefficient in the nth layer after threshold filtering, and perform the calculation on the nth layer 2 n-1 The sum values ​​corresponding to the node coefficients are arranged to form the eigenvector f k , for the eigenvector f k Perform L2 norm normalization to obtain the normalized feature vector

[0017] The normalized feature vectors corresponding to each signal data segment are arranged to form a two-dimensional feature matrix as the time-frequency domain features of the current sample signal.

[0018] Furthermore, the value of λ2 is: Where I is the total number of iterations of adversarial training, iter is the number of current iterations of adversarial training, α and β are coefficients, and α < 0.

[0019] Furthermore, the loss function of the data source discriminator in the adversarial network uses Wasserstein distance, and the discriminator loss function is in, are the outputs of the data source discriminator for the rth group of time-frequency domain features of working condition 1 and the sth group of time-frequency domain features of working condition 2; g(·) is the sigmoid function.

[0020] The present invention also provides a method for predicting the life of a rotating component, which uses a rotating component life prediction model constructed by the method for constructing a rotating component life prediction model as described above to predict the life of the rotating component.

[0021] The present invention also provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the method for constructing a rotating component life prediction model and / or the method for predicting the life of a rotating component as described above.

[0022] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0023] (1) Due to the lack of data labels, the unlabeled data under working condition 2 cannot be used to train the fully connected layer (i.e., the prediction unit in the life prediction model). Therefore, the parameters of the fully connected layer cannot be updated to reduce the difference between the high-level features under the two working conditions extracted in the fully connected layer, resulting in a decrease in the accuracy of life prediction for working condition 2. When the parameters of the data source discriminator are fixed during the transfer adversarial training process, the maximum mean difference of the output of the neural network for the two working condition data is introduced into the loss function. The parameters of the life prediction model can be updated based on the data under the two working conditions at the same time during the training process. Reducing the maximum mean difference of the output can reduce the difference between the features extracted in the fully connected layer, thereby improving the prediction accuracy.

[0024] (2) Wavelet packet decomposition can decompose the signal into different frequency bands. Compared with directly inputting the time domain signal into the neural network for life prediction, the samples constructed by wavelet packet decomposition can simultaneously reflect the time domain and frequency domain characteristics of the signal. Using wavelet packet decomposition to preprocess the signal before using the neural network for feature extraction is conducive to better feature extraction by the neural network.

[0025] (3) Since the noise in the mechanical component signal can be considered as Gaussian white noise, its corresponding wavelet packet node coefficient is relatively small. Therefore, in the wavelet packet decomposition, it can be filtered out by threshold filtering, thereby enhancing the anti-noise interference ability of the remaining life prediction method.

[0026] (4) The requirements for the data source identifier and the feature extractor in the prediction model are different during the adversarial training process. In the early stage of adversarial training, the data source identifier needs to be able to quickly obtain parameters that can be effectively identified, that is, the update step size is large, and the weight λ2 in the loss function loss3 needs to be large. As the adversarial training progresses, the network convergence speed slows down, and the update step size of the data source identifier needs to be appropriately shortened to prevent overtraining, that is, the weight λ2 in the loss function loss3 is reduced. Therefore α<0. Updating the weight λ2 according to the number of iterations can keep the life prediction neural network stable during adversarial training and prevent overtraining.

[0027] (5) When the data distributions of the two working conditions do not overlap, the adversarial network using JS divergence as the discriminator loss function cannot measure the difference between the two distributions, resulting in problems such as gradient vanishing and unstable training. However, the Wasserstein distance can reflect the difference between the two working conditions even when there is no overlap between the data distributions, which is helpful in solving the instability problem in adversarial training. In addition, since the value of the discriminator loss function may be very large during training, the proportion of loss1 and MMD is very small, making it impossible to reflect the loss of these two parts, affecting the training effect. Therefore, the sigmoid function is introduced to limit the value of the discriminator loss. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A flow chart of a method for constructing a life prediction model for rotating components provided by an embodiment of the present invention;

[0029] Figure 2 A diagram showing the neural network adversarial training structure provided by an embodiment of the present invention;

[0030] Figure 3 A flowchart of neural network adversarial training provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0032] Example 1

[0033] A method for constructing a life prediction model for rotating parts, such as Figure 1 Shown, including:

[0034] Construct a labeled sample set for operating condition 1 and an unlabeled sample set for operating condition 2. Construct time-frequency domain features for all samples in the two sample sets. Use the constructed time-frequency domain features to perform transfer adversarial training on the neural network model to obtain a rotating component life prediction model for operating condition 2.

[0035] Among them, when the parameters of the data source discriminator are fixed during the transfer adversarial training process, the following loss function loss3 is used to adjust the parameters of the lifespan prediction model:

[0036]

[0037] Where loss1 is the life prediction loss function of the life prediction model for working condition 1, loss2 is the data source identification loss function of the data source discriminator, The predicted output corresponding to the input data of working condition 1 and working condition 2 respectively and The maximum mean difference between them, λ1, λ2 and λ3 are the weights of each loss value, and the values ​​are positive. In the process of transfer adversarial training, the value of λ2 is adjusted after each iteration, and the value of λ2 decreases exponentially with the current number of iterations of adversarial training.

[0038] This method combines the idea of ​​transfer learning, uses adversarial network technology, and adds maximum mean difference to train the neural network. When conducting adversarial training, the sample signal can be a vibration signal, and the following method can be used:

[0039] A two-dimensional convolutional neural network is used as the feature extraction unit for the rotating component sample signal; a fully connected neural network is used as the life prediction unit; and another fully connected neural network is used as the data source identifier.

[0040] Using multiple sets of time-frequency domain features and their corresponding life labels from working condition 1, the parameters of the feature extraction unit and the life prediction unit are updated and trained to obtain a rotating component life prediction model for working condition 1.

[0041] The parameters of the feature extraction unit and the life prediction unit are fixed, and multiple sets of time-frequency domain features under working conditions 1 and 2 are input into the feature extraction unit respectively. The outputs of the feature extraction unit and the life prediction unit are used as inputs of the data source discriminator, and the parameters of the data source discriminator are updated and trained;

[0042] The parameters of the data source discriminator are fixed, and multiple sets of time-frequency domain features under working conditions one and two are input into the feature extraction unit. The parameters of the feature extraction unit and the life prediction unit are updated and trained using the label data of working condition one and the results corresponding to the multiple sets of time-frequency domain features under working conditions one and two output by the data source discriminator.

[0043] When implementing it specifically, Figure 2 As shown in the figure, two weight-sharing two-dimensional convolutional neural networks Conv1 and Conv2 can be used as feature extraction units for working conditions 1 and 2 respectively; two weight-sharing fully connected neural networks FC1 and FC2 can be used as life prediction units for working conditions 1 and 2 respectively, and the two-dimensional convolutional neural network of each working condition is connected to the corresponding fully connected neural network; another fully connected neural network FC3 is used as a data discriminator. The outputs of Conv1 and Conv2 are arranged in one dimension and then form a fully connected relationship with FC1, FC2 and FC3 respectively. Figure 3As shown, the characteristic matrix of the working condition 1 is input into Conv1, and the parameters of Conv1 and FC1 are updated and trained to obtain the rotating component life prediction model for working condition 1; the parameters of Conv1 and Conv2 as well as FC1 and FC2 are fixed, and the characteristic matrix of the working condition 1 and the characteristic matrix of the working condition 2 are input into Conv1 and Conv2 respectively, and the outputs of Conv1 and Conv2 are used as the input of FC3, and the parameters of FC3 are updated and trained; the parameters of FC3 are fixed, and the characteristic matrix of the working condition 1 and the characteristic matrix of the working condition 2 are input into Conv1 and Conv2 respectively, and the parameters of Conv1 and FC1 as well as the parameters of Conv2 and FC2 are updated and trained.

[0044] Additionally, the maximum mean difference is calculated as:

[0045]

[0046] Where k(·,·) is the kernel function, which is taken as Gaussian kernel function

[0047] Preferably, wavelet packet decomposition is used to construct time-frequency domain features.

[0048] Preferably, the specific construction method of the time-frequency domain features of each sample is:

[0049] Randomly select m signal data segments that can overlap from each sample signal of the rotating component For each signal data segment Perform wavelet packet decomposition:

[0050] The kth signal data segment for According to the formula Get the coefficient of the tth node in the nth layer Where t=1,2,...,2 n-1 , n is the number of wavelet packet decomposition layers, l is the length of each signal data segment; G n is the low-pass filter coefficient matrix corresponding to the n-th layer wavelet packet decomposition; H n is the high-pass filter coefficient matrix corresponding to the n-th layer wavelet packet decomposition;

[0051] Perform threshold filtering on each element in each node coefficient of the nth layer, expressed as Where z is the set threshold, is the node coefficient after filtering, for The u-th element in ;

[0052] Sum the elements of each node coefficient in the nth layer after threshold filtering, and perform the calculation on the nth layer 2n-1 The sum values ​​corresponding to the node coefficients are arranged to form the eigenvector f k , for the eigenvector f k Perform L2 norm normalization to obtain the normalized feature vector

[0053] The normalized feature vectors corresponding to each signal data segment are arranged to form a two-dimensional feature matrix as the time-frequency domain features of the current sample signal.

[0054] Preferably, the value of λ2 is: Where I is the total number of iterations of adversarial training, iter is the number of current iterations of adversarial training, α and β are coefficients, and α<0.

[0055] Preferably, in, are the outputs of the data source discriminator for the rth group of time-frequency domain features of working condition 1 and the sth group of time-frequency domain features of working condition 2; g(·) is the sigmoid function.

[0056] In order to better illustrate the present invention, the following examples are given:

[0057] This example uses the rolling bearing remaining life prediction vibration dataset provided by the IEEE Reliability Society and the FEMTO-ST Institute as an example to introduce the specific implementation of the proposed method.

[0058] The vibration signals were collected under three operating conditions: 4000N load and 1800rpm speed, 4200N load and 1650rpm speed, and 4200N load and 1650rpm speed. The sampling frequency was 25.6kHz. The data was collected every 10 seconds, 0.1s per time, that is, 2560 data points were collected at a time. The data every 0.1s was used as the vibration data x i , and convert its corresponding remaining life into a percentage. Bearing 2_1 is used as the data for bearing X1 in condition 1, where the remaining life is known, and bearing 3_1 is used as the data for condition 2, where the remaining life is unknown. A neural network is trained and the trained network is used to predict the remaining life of bearing 3_3 in condition 3. Assume α = -1 and β = 0.

[0059] The specific implementation is as follows:

[0060] S1. Randomly select 32 segments of 1280 data points from the vibration data of bearing2_1 and bearing3_1, and perform a 5-layer wavelet packet transform to obtain the corresponding wavelet packet coefficients;

[0061] S2. Sum the wavelet packet coefficients and perform L2 normalization to obtain eigenvectors. Arrange the eigenvectors corresponding to the same vibration data to construct a 32×32 eigenmatrix.

[0062] S3. Construct neural networks Network 1, Network 2, and Network 3. Conv1 and Conv2 are both two-layer convolutional networks. The convolution kernel size of the first convolutional network is 5×5 with a stride of 3, and the convolution kernel size of the second convolutional network is 4×4 with a stride of 3. The activation function of both layers is the ReLU function. FC1, FC2, and FC3 are all two-layer fully connected networks. The number of neurons in the first and second layers are 10 and 1 respectively. The activation function of the first layer is ReLU, and the activation function of the second layer is sigmoid.

[0063] S4. Fix the parameters of network 3, input the feature matrix set and life corresponding to bearing2_1 into the neural network, use the mean square error function between the predicted remaining life value and the actual remaining life value as the loss function λ1·loss1, and use the Adam algorithm to optimize network 1 and network 2;

[0064] S5. Fix the parameters of networks 1 and 2, input the feature matrix set and lifespan corresponding to bearing2_1 and bearing3_1 into the neural network, and train network 3 to distinguish the data of bearing2_1 and bearing3_1 by minimizing λ2·loss2;

[0065] S6. Fix the parameters of network 3, input the feature matrix set corresponding to bearing2_1 and bearing3_1 and the life span into the neural network to minimize loss3 and train networks 1 and 2;

[0066] S7. Repeat S5 and S6, and update λ2 at each repetition until the number of training iterations is reached. Then stop training and obtain a trained neural network.

[0067] S8. Process the vibration data in bearing3_3 according to S1 and S2 and then input it into the trained neural network to predict the remaining life of bearing3_3.

[0068] Example 2

[0069] A method for predicting the life of a rotating component uses a rotating component life prediction model constructed by the method for constructing a rotating component life prediction model as described in Example 1 to predict the life of the rotating component.

[0070] The relevant technical solutions are the same as those in Example 1 and will not be described again here.

[0071] Example 3

[0072] A computer-readable storage medium includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute a method for constructing a rotating component life prediction model as described in Example 1 and / or a rotating component life prediction method as described in Example 2.

[0073] The relevant technical solutions are the same as those in Example 1 and will not be described again here.

[0074] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing a life prediction model for rotating parts, characterized in that: include: Construct a labeled sample set for operating condition 1 and an unlabeled sample set for operating condition 2. Construct time-frequency domain features for all samples in the two sample sets. Use the constructed time-frequency domain features to perform transfer adversarial training on the neural network model to obtain a rotating component life prediction model for operating condition 2. Among them, when the parameters of the data source discriminator are fixed during the transfer adversarial training process, the following loss function loss3 is used to adjust the parameters of the lifespan prediction model: Where loss1 is the life prediction loss function of the life prediction model for working condition 1, loss2 is the data source identification loss function of the data source discriminator, The predicted output corresponding to the input data of working condition 1 and working condition 2 respectively and The maximum mean difference between them, λ1, λ2 and λ3 are the weights of each loss value, and the values ​​are positive. In the process of transfer adversarial training, the value of λ2 is adjusted after each iteration. The value of λ2 decreases exponentially with the number of current iterations of adversarial training; Among them, the value of λ2 is: Where I is the total number of iterations of adversarial training, iter is the number of current iterations of adversarial training, α and β are coefficients, and α < 0.

2. The construction method according to claim 1, characterized in that Wavelet packet decomposition is used to construct time-frequency domain features.

3. The construction method according to claim 2, characterized in that The specific construction method of the time-frequency domain features of each sample is: Randomly select m signal data segments that can overlap from each sample signal of the rotating component For each signal data segment Perform wavelet packet decomposition: The kth signal data segment for According to the formula Get the coefficient of the tth node in the nth layer Where t=1,2,...,2 n-1 , n is the number of wavelet packet decomposition layers, l is the length of each signal data segment; G n is the low-pass filter coefficient matrix corresponding to the n-th layer wavelet packet decomposition; H n is the high-pass filter coefficient matrix corresponding to the n-th layer wavelet packet decomposition; Perform threshold filtering on each element in each node coefficient of the nth layer, expressed as Where z is the set threshold, is the node coefficient after filtering, for The u-th element in ; Sum the elements of each node coefficient in the nth layer after threshold filtering, and perform the calculation on the nth layer 2 n-1 The sum values ​​corresponding to the node coefficients are arranged to form the eigenvector f k , for the eigenvector f k Perform L2 norm normalization to obtain the normalized feature vector The normalized feature vectors corresponding to each signal data segment are arranged to form a two-dimensional feature matrix as the time-frequency domain features of the current sample signal.

4. The construction method according to claim 1, characterized in that The loss function of the data source discriminator in the adversarial network uses Wasserstein distance, and the discriminator loss function is in, are the outputs of the data source discriminator for the rth group of time-frequency domain features of working condition 1 and the sth group of time-frequency domain features of working condition 2; g(·) is the sigmoid function.

5. A method for predicting the life of a rotating component, characterized in that: The life prediction of a rotating component is performed using a rotating component life prediction model constructed by the method for constructing a rotating component life prediction model according to any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute a method for constructing a rotating component life prediction model as described in any one of claims 1 to 4 and / or a rotating component life prediction method as described in claim 5.

Images