Multi-stage transfer learning training method and device for rotary machine transfer diagnosis and medium

Abstract

The invention discloses a multi-stage transfer learning training method and device for rotary machine transfer diagnosis and a medium, and relates to the field of mechanical fault diagnosis, and the method comprises the steps: carrying out the data enhancement of sensing data of a rotary machine through employing an angular domain resampling data enhancement method, obtaining an order spectrum signal, and carrying out the data enhancement of the order spectrum signal; generating a source domain data set and a target domain data set; constructing an improved convolutional neural network; performing multi-stage training on the improved convolutional neural network based on the local data of the source domain data set and the global data of the source domain data set; adjusting parameters of the trained convolutional neural network by using fine tuning data of the target domain; and testing the parameter-adjusted convolutional neural network by adopting the test data of the target domain, and taking the parameter-adjusted convolutional neural network meeting the diagnosis accuracy requirement as a multi-stage parameter transfer learning network. The problems of insufficient training efficiency, insufficient data utilization efficiency and the like can be solved, and the performance of the transfer learning method based on parameter fine tuning is improved.

Description

Technical Field

[0001] The present application relates to the field of mechanical fault diagnosis, and particularly to a multi-stage transfer learning training method, device and medium for rotating machinery transfer diagnosis. Background Art

[0002] Fault diagnosis is considered one of the key technologies to ensure the safe and reliable operation of rotating machinery. With the advantage of adaptive feature extraction, deep learning-based fault diagnosis methods have been widely studied in the past decade. However, due to limitations such as the requirement that the training samples and test samples must follow the same distribution, deep learning-based fault diagnosis methods have not been effectively and widely applied in actual industrial applications.

[0003] The transfer learning method based on parameter fine-tuning can overcome the problem that the deep learning-based fault diagnosis method cannot perform accurate fault diagnosis when the distributions of the training samples and test samples are different. Although the transfer learning method based on parameter fine-tuning has achieved great success in transfer diagnosis, there are still certain limitations in the existing methods. First, there are limitations in the transfer training strategy of the existing transfer learning method based on parameter fine-tuning, and the learnable transfer features are not fully learned, which is particularly prominent when the number of available samples is small. On the other hand, the existing transfer learning method based on parameter fine-tuning has the problem of low utilization efficiency of vibration data.

[0004] Based on the above description, the existing transfer learning methods have the following disadvantages:

[0005] 1) There is a problem of insufficient training efficiency in the training strategy.

[0006] 2) The data is not fully utilized in the data preprocessing method.

[0007] 3) The current neural network used for transfer diagnosis has limitations in learning performance and generalization performance, resulting in difficulties in improving the transfer diagnosis accuracy.

[0008] 4) With the improvement of the requirements for transfer diagnosis accuracy, the training complexity and computational cost of the transfer diagnosis method are increasing significantly, seriously affecting the application prospect of transfer diagnosis. Summary of the Invention

[0009] Aiming at the problems of insufficient training efficiency and insufficient data utilization efficiency of the transfer diagnosis method based on parameter fine-tuning, the present application provides a multi-stage transfer learning training method, device and medium for rotating machinery transfer diagnosis, aiming to improve the performance of the transfer learning method based on parameter fine-tuning.

[0010] To achieve the above object, the present application provides the following solutions:

[0011] In a first aspect, the present application provides a multi-stage transfer learning training method for rotating machinery migration diagnosis, including:

[0012] Obtain the sensing data of the rotating machinery; the sensing data of the rotating machinery covers various health states and working conditions of the source domain and the target domain;

[0013] Adopt a data augmentation method of angular domain resampling to perform data augmentation on the sensing data of the rotating machinery to obtain an order spectrum signal, and generate a source domain dataset and a target domain dataset based on the order spectrum signal;

[0014] Construct an improved convolutional neural network;

[0015] Divide the source domain dataset into local data and global data, and divide the target domain dataset into fine-tuning data and test data;

[0016] Perform multi-stage training on the improved convolutional neural network based on the local data and the global data to obtain a trained convolutional neural network;

[0017] Adjust the parameters of the trained convolutional neural network using the fine-tuning data to obtain a convolutional neural network with adjusted parameters;

[0018] Test the convolutional neural network with adjusted parameters using the test data, and use the convolutional neural network with adjusted parameters that meets the diagnostic accuracy requirements as a multi-stage parameter transfer learning network; the multi-stage parameter transfer learning network is applied to the migration diagnosis of rotating machinery.

[0019] Optionally, adopting a data augmentation method of angular domain resampling to perform data augmentation on the sensing data of the rotating machinery to obtain an order spectrum signal includes:

[0020] Adopt an interpolation method to perform angular domain resampling on the original time-domain signal in the sensing data of the rotating machinery according to the sampling time points to obtain angular domain data;

[0021] Perform order spectrum analysis on the angular domain data to obtain the order spectrum signal.

[0022] Optionally, the determination process of the sampling time points is expressed as:

[0023] θ(t) = b1 + b2t + b3t 2 ;

[0024]

[0025] where θ(t) represents the angle corresponding to the sampling time point t of equal-angle sampling, b1, b2, and b3 all represent variables to be solved, θ(t1), θ(t2), and θ(t3) respectively represent the data points of adjacent time points t1, t2, and t3 of the original time-domain signal, and Δθ represents the angle interval of angular domain resampling, represents the time points corresponding to equiangular resampling, and m represents the m-th equiangular interval of the angular domain data.

[0026] Optionally, construct an improved convolutional neural network, including:

[0027] Construct a trainable parametric activation function and construct the trainable parametric activation function into a trainable parametric activation function layer;

[0028] Based on the trainable parametric activation function layer, construct an improved convolutional neural network, which is composed of four convolutional blocks, two fully connected layers, one trainable parametric activation function layer, and one Softmax layer.

[0029] Optionally, each convolutional block is composed of a one-dimensional convolutional layer, a batch normalization layer, a trainable parametric activation function layer, and a max pooling layer.

[0030] Optionally, the trainable parametric activation function is expressed as:

[0031]

[0032] In the formula, APMish(x i ) represents the trainable parametric activation function, a i and b i represent trainable parameters, x i represents the input variable of the i-th channel, t i represents the intermediate variable, c represents the total number of channels, and i represents the i-th channel.

[0033] Optionally, perform multi-stage training on the improved convolutional neural network based on local data and global data to obtain a trained convolutional neural network, including:

[0034] Perform local training stage, global training stage, and specific parameter training stage on the improved convolutional neural network based on local data and global data to obtain a trained convolutional neural network;

[0035] Among them, in the local training stage, use local data to perform local training on the parameters of the improved convolutional neural network except the trainable parametric activation function layer; in the global training stage, use global data to perform global training on the parameters of the improved convolutional neural network except the trainable parametric activation function layer; in the specific parameter training stage, use global data to train the parameters of the trainable parametric activation function layer.

[0036] Optionally, use fine-tuning data to adjust the parameters of the trained convolutional neural network to obtain a convolutional neural network with adjusted parameters, including:

[0037] Taking the fine-tuning data as input, fine-tuning the parameters of the trained convolutional neural network to obtain a convolutional neural network with adjusted parameters; where the fine-tuning process is expressed as:

[0038]

[0039] In the formula, θ e+f represents the parameters of the convolutional neural network after f fine-tuning cycles, η F represents the learning rate of the fast fine-tuning part, θ e represents the parameters of the trained convolutional neural network, represents the i F -th sample in the fine-tuning data, represents the i F -th sample in the fine-tuning data e,l-1 corresponding health category label, f represents the number of fine-tuning times, θ

[0040] In a second aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement the steps of the multi-stage transfer learning training method for rotating machinery transfer diagnosis provided above.

[0041] In a third aspect, the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the multi-stage transfer learning training method for rotating machinery transfer diagnosis provided above.

[0042] According to the specific embodiments provided by the present application, the present application has the following technical effects:

[0043] The present application provides a multi-stage transfer learning training method, device, and medium for rotating machinery transfer diagnosis, which can address problems such as low training efficiency, low data utilization efficiency, and high computational burden in the current transfer learning method based on parameter fine-tuning. Different types of data obtained by dividing the source domain dataset and the target domain dataset are used for separate training. Moreover, for these different types of divided data, a training strategy of multi-stage training is used to train the improved convolutional neural network, which can improve the performance of the transfer learning method based on parameter fine-tuning. Furthermore, when applied to rotating machinery transfer diagnosis, it can effectively improve the accuracy of fault diagnosis. Description of the Drawings

[0044] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0045] Figure 1 Flowchart of a multi-stage transfer learning training method for rotating machinery migration diagnosis provided by an embodiment of the present application;

[0046] Figure 2 Structural diagram of an improved convolutional neural network provided by an embodiment of the present application;

[0047] Figure 3 Flowchart of training an improved convolutional neural network provided by an embodiment of the present application;

[0048] Figure 4 Schematic diagram of the implementation process of a multi-stage transfer learning training method for rotating machinery migration diagnosis provided by an embodiment of the present application;

[0049] Figure 5 Schematic diagram of the diagnosis results of different transfer diagnosis methods in all transfer diagnosis scenarios provided by an embodiment of the present application;

[0050] Figure 6 Schematic diagram of the number of parameters and floating-point calculations of different transfer learning methods provided by an embodiment of the present application.

[0051] Figure 7 Structural diagram of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0053] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below in conjunction with the drawings and specific implementation manners.

[0054] In an exemplary embodiment, the present application provides a multi-stage transfer learning training method for rotating machinery migration diagnosis. This method is executed by a computer device, which can be specifically executed by a computer device such as a terminal or a server alone, or jointly executed by a terminal and a server. In the embodiments of the present application, this method is described by taking the application to a server as an example. As Figure 1 shown, the method includes:

[0055] Step 100: Obtain the sensing data of the rotating machinery. For example, use an acceleration sensor and a rotational speed sensor to collect the original time-domain signal of the rotating machinery to obtain the sensing data, covering various health states and working conditions of the source domain and the target domain.

[0056] Step 101: Adopt a data augmentation method of angular domain resampling to perform data augmentation on the sensing data of the rotating machinery to obtain an order spectrum signal, and generate a source domain dataset and a target domain dataset based on the order spectrum signal.

[0057] Step 102: Construct an improved convolutional neural network.

[0058] Step 103: Divide the source domain dataset into local data and global data, and divide the target domain dataset into fine-tuning data and test data.

[0059] Step 104: Perform multi-stage training on the improved convolutional neural network based on the local data and the global data to obtain a trained convolutional neural network.

[0060] Step 105: Adjust the parameters of the trained convolutional neural network by using the fine-tuning data to obtain a convolutional neural network with adjusted parameters.

[0061] Step 106: Test the convolutional neural network with adjusted parameters by using the test data, and use the convolutional neural network with adjusted parameters that meets the diagnostic accuracy requirements as a multi-stage parameter transfer learning network. The multi-stage parameter transfer learning network is applied to the migration diagnosis of rotating machinery.

[0062] By implementing the above steps 100 - step 106, the present application can effectively improve the fault diagnosis performance by using a new convolutional neural network trained with a multi-stage specific parameter training strategy for the problems of low training efficiency, low data utilization efficiency, and high computational burden in the current transfer learning method based on parameter fine-tuning.

[0063] In another exemplary embodiment of the present application, the present application may provide a data augmentation method based on angular domain resampling. This method resamples the original time-domain signal from time-domain data to angular domain data through angular domain resampling, and then obtains the order spectrum signal of the angular domain data through order spectrum analysis. Based on this, in the above step 101, the implementation process of using the data augmentation method based on angular domain resampling to perform data augmentation on the sensing data of the rotating machinery to obtain the order spectrum signal can be replaced by the following step 200-step 201.

[0064] Step 200: Use the interpolation method to perform angular domain resampling on the original time-domain signal in the sensing data of the rotating machinery according to the sampling time points to obtain angular domain data.

[0065] In the actual application process, the implementation process of this step may include:

[0066] Apply the angular domain resampling method to process the original time series information and calculate the time points of equiangular resampling. To calculate the sampling time points, the following calculation method is introduced in this embodiment:

[0067] θ(t) = b1 + b2t + b3t 2 .

[0068]

[0069] In the formula, θ(t) represents the angle corresponding to the time point t of equiangular sampling, b b1, b2 and b3 all represent variables to be solved, θ(t1), θ(t2) and θ(t3) respectively represent the data points of the adjacent time points t1, t2 and t3 of the original time-domain signal, Δθ represents the angular interval of angular domain resampling, represents the time point corresponding to equiangular resampling, and m represents the mth equiangular interval of the angular domain data.

[0070] Furthermore, use the interpolation method to resample the time-domain vibration data into an angular domain vibration signal to offset the influence brought by the rotational speed change.

[0071] Step 201: Perform order spectrum analysis on the angular domain data to obtain the order spectrum signal. In this step, order spectrum analysis is performed on the angular domain vibration signal to achieve data augmentation.

[0072] Furthermore, in the actual application process, in order to achieve data augmentation, the following process may be introduced in this embodiment:

[0073]

[0074] χ' = FFT(X').

[0075] In the formula, X represents the original time-domain vibration signal, The time points for resampling are denoted as, the cubic spline interpolation method is denoted as CS{·}, the fast Fourier transform is denoted as FFT(·), the signal X' represents the signal obtained by resampling the original time-domain vibration signal X according to the equal-angle resampling time points, and the order spectrum signal obtained after Fourier transform is denoted as χ'.

[0076] Based on the above process, the order spectrum signals of these different health states from the source domain and the target domain can be obtained similarly. On this basis, the source domain, the target domain training sets, and the test sets for multi-stage training are divided, and they are used as the data inputs of the multi-stage parameter transfer learning network proposed in this application. Among them, the multi-stage parameter transfer learning network can also be called the AR-MSPTL (Angular domain Resampling-assisted Multi-Stage Parameter Transfer Learning network) transfer learning model.

[0077] In another exemplary embodiment of this application, the implementation process of step 102 can be replaced by the following step 300 and step 301.

[0078] Step 300: Construct a trainable parameterized activation function (APMish), and construct the trainable parameterized activation function into a trainable parameterized activation function layer to achieve plug-and-play. Among them, the trainable parameterized activation function is expressed as:

[0079]

[0080] In the formula, APMish(x i ) represents the trainable parameterized activation function, a i and b i represent trainable parameters, x i represents the input variable of the i-th channel. t i represents an intermediate variable. The purpose of introducing t i is to make the above expression more concise. c represents the total number of channels, and i represents the i-th channel.

[0081] Step 301: Construct an improved convolutional neural network based on the trainable parameterized activation function layer. The improved convolutional neural network consists of four convolutional blocks (Conv Block), a fully connected

Full Connection, for example, corresponding to Figure 2 Dense(4096,64) in Figure 2

for example, corresponding to Figure 2Dense(64, Class)

[0082] Furthermore, as Figure 2 shown, each convolutional block is respectively composed of a one-dimensional convolutional (Conv1D) layer, a batch normalization (Batch Normalization, BN) layer, an APMish layer, and a max pooling layer. Among them, Figure 2 in Size represents the size of the max pooling layer, and Stride represents the moving step of the pooling window.

[0083] In another exemplary embodiment of the present application, a multi-stage training strategy as Figure 3 shown can be provided to train the improved convolutional neural network to achieve transfer diagnosis. Based on this, in this embodiment, the local data and global data divided in step 103 above can be expressed as:

[0084]

[0085] In the formula, represents the i L th sample of the local data S L , represents the i G th sample of the global data S G , and respectively represent the health category labels corresponding to the samples and , and H and C represent the length and total number of channels of the samples.

[0086] Furthermore, after dividing the source domain dataset, the improved convolutional neural network will be trained in different stages successively, including the local training stage, the global training stage, the specific parameter training stage, and the fast fine-tuning stage. Based on this, the implementation process of step 104 above can be replaced by: performing the local training stage, the global training stage, and the specific parameter training stage on the improved convolutional neural network based on the local data and the global data to obtain the trained convolutional neural network.

[0087] (1) In the local training stage, the parameters of the improved convolutional neural network except the training parametric activation function layer are locally trained using the local data. For example, taking each batch of local data as input, the improved convolutional neural network is iteratively trained multiple times so that the improved convolutional neural network can extract the deep features of the local data. Based on this, in this stage, the parameter optimization process of the improved convolutional neural network can be described as:

[0088]

[0089] Where, θ e,h represents the parameters of the improved convolutional neural network after local optimization, m L represents the quantity of local data, θ e represents the parameters of the improved convolutional neural network at the e-th iteration cycle, η L represents the learning rate in the local training phase, L L () represents the loss function in the local training phase, j represents the j-th training, and k represents the number of training times of the improved convolutional neural network for each batch of local data within each iteration cycle. represents the gradient of the parameter θ, θ e,j-1 represents the parameters of the improved convolutional neural network at the (j-1)-th training within the e-th iteration cycle.

[0090] The optimization objective in the local training phase can be expressed as:

[0091]

[0092] Where, L e,h represents the loss of the improved convolutional neural network for each batch of local data within each iteration cycle, and min represents taking the minimum value.

[0093] (2) In the global training phase, the parameters of the improved convolutional neural network except for the training parameterized activation function layer are globally trained using global data. For example, with the global data as the input, the loss is calculated once for the entire global data and the parameters of the improved convolutional neural network are updated, enabling the improved convolutional neural network after local training to extract the generalization features of the global data. The update process of the parameters of the improved convolutional neural network in the global training phase is as follows:

[0094]

[0095] Where, θ e+1,h represents the parameters of the improved convolutional neural network after global optimization, η G represents the learning rate in the global training phase, m G represents the quantity of global data, L G () represents the loss function in the global training phase.

[0096] Based on the above description, the optimization objective in the global training phase can be expressed as:

[0097]

[0098] Where, L e+1,hIndicates the loss between the predicted label and the true label of the neural network for the global data samples after the global training phase is completed.

[0099] (3) In the specific parameter training phase, the global data is used to train the parameters of the parameterized activation function layer. The specific parameter training phase is mainly to perform targeted updates on the parameters of the APMish layer. For example, the APMish layer is independently trained with the global data as the input, so that the parameters of the APMish layer can have better generalization performance. Based on this, the process of the specific parameter training phase is as follows:

[0100] The loss L between the predicted label and the actual label of the global data input by the parameters of the convolutional neural network improved through the local training phase and the global training phase AP () is expressed as:

[0101]

[0102] In the formula, Q represents the number of categories, y pred and y real represent the predicted label and the true label of this data, y iG represents the label of the iG-th data in the global data set, P(X iG ) represents the probability distribution of the model's prediction output for the data X iG .

[0103] Define the input of the r-th layer of APMish as z i,0 , and the output after passing through the APMish layer can be obtained from this:

[0104]

[0105] In the formula, (a i,0 , b i,0 ) represents the parameters of the r-th layer of APMish, and APMish() represents using the parameters (a i,0 , b i,0 ) of the r-th layer of APMish to calculate the output of the r-th layer of APMish for the input z i,0 with the APMish expression.

[0106] The parameters a i,0 and b i,0 The derivative of the loss L AP can be obtained through the following formula:

[0107]

[0108] Obtain the parameters a i,0 and b i,0 The loss L APAfter taking the derivative, the parameters a i,0 and b i,0 can be updated by the following formula:

[0109]

[0110] where a' i,0 and b' i,0 represent the updated parameters a i,0 and b i,0 , and η AP represents the learning rate of the parameters a i,0 and b i,0 .

[0111] Subsequently, repeating this process g times, a list containing g pairs can be obtained, denoted as:

[0112]

[0113] where p represents the repetition number, Ψ represents the list containing g pairs , represents the loss included in the first pair in the list, and represent the parameters (a , b i,0 , b i,0 ) of the first pair in the list, represents the loss included in the second pair and represent the parameters (a , b i,0 , b i,0 ) of the second pair in the list, and represent the parameters (a , b i,0 , b i,0 ) of the p-th pair

[0114] in the list. Finally, select the lowest loss as and use its corresponding parameters (a i,1 , b i,1 ) as the target parameters for updating in the specific parameter training stage. The update process is as follows:

[0115]

[0116] It should be noted that when updating the parameters at this stage, the parameters other than the APMish layer need to be frozen to ensure that only the parameters of the APMish layer are updated.

[0117] In another exemplary embodiment of the present application, after the foregoing local training stage, global training stage, and specific parameter training stage, the parameters of the trained convolutional neural network can be obtained. Subsequently, through a fast fine-tuning stage, using the fine-tuning data of the target domain as the input, the parameters of the trained convolutional neural network can be fine-tuned. Based on this, the implementation process of step 105 can be replaced by: using the fine-tuning data as the input, fine-tuning the parameters of the trained convolutional neural network to obtain the convolutional neural network with adjusted parameters. Among them, the fine-tuning process is expressed as:

[0118]

[0119] where θ e+f represents the parameters of the convolutional neural network after f cycles of fine-tuning, η F represents the learning rate of the fast fine-tuning part, θ e represents the parameters of the trained convolutional neural network, represents the i F -th sample in the fine-tuning data, represents the i F -th sample in the fine-tuning data e,l-1 corresponding to the healthy category label, f represents the number of fine-tuning times, and θ

[0120] In another exemplary embodiment of the present application, after adjusting the parameters of the trained convolutional neural network using the fine-tuning data, the test data for testing the model performance will be input into the convolutional neural network with adjusted parameters, and the final diagnostic classification result will be output. Among them, the test data obtained by dividing the target domain dataset in step 103 can be expressed as:

[0121]

[0122] where represents the m U -th sample in the test data T, and U represents the total number of samples in the test data.

[0123] The convolutional neural network with adjusted parameters will output a predicted label of the healthy category for each sample in the input test data. This process can be expressed as:

[0124]

[0125] where Softmax(·) represents the Softmax activation function, Denote the test data as T U and the m-th sample in it, with the predicted label Denote, taking the m-th sample U in the test data T as the input, the parameters of the fine-tuned convolutional neural network

[0126] Thus, prediction labels can be generated for all samples in the test data, and the diagnostic accuracy can be obtained by counting the predicted labels and the actual health category labels, thereby evaluating the performance of the convolutional neural network after parameter tuning. The calculation process is as follows:

[0127]

[0128] In the formula, A represents the average accuracy of the model in generating prediction labels for the test data and it is the m-th sample U in the test data T with the true label, and I(·) represents the indicator function

[0129] Based on the above description, the implementation process of this application can be described as follows:

[0130] First, use multiple sensors to collect the original time series signals of the vibration acceleration and rotational speed of the rotating machinery, covering various health states and working conditions in the source domain and the target domain

[0131] Second, perform order spectrum analysis assisted by angular domain resampling on the original time series signals: first set the equal angular intervals to be resampled, calculate the resampling time points through the principle of angular domain resampling, then calculate the resampled angular domain vibration data through the interpolation method, then perform order spectrum analysis on the resampled angular domain signals to obtain order signals, and divide the source domain dataset and the target domain dataset

[0132] Then, initialize the improved convolutional neural network. Design a trainable parameterized activation function, and construct parts such as the convolutional blocks and linear layers of the convolutional neural network

[0133] Again, divide the source domain dataset into local data and global data, and perform multi-stage training on the improved neural network: first use the local data for local training to enable the improved neural network to learn deep features, then use the global data for global training to enable the improved neural network to master generalization features, then perform specific parameter training with the global data as the input to improve the generalization performance and learning ability of the improved neural network, and then use the fine-tuning data to quickly fine-tune the trained neural network so that the trained neural network can learn the knowledge required for accurate transfer diagnosis

[0134] Finally, output the migration diagnosis results of the target domain. The test data on the target domain will be input into the tuned convolutional neural network, and the diagnosis classification results will be obtained after passing through multiple convolutional blocks and linear layers.

[0135] In summary, this application fully considers the impact of different working conditions in the industrial scenario on diagnosis, provides a new method for fault diagnosis, and provides important technical support for the safe and reliable operation of equipment. The specific implementation process can be referred to Figure 4 .

[0136] In another exemplary embodiment of this application, Figure 5 , where WDCNN represents a wide convolutional kernel convolutional neural network, DANN represents a domain adversarial neural network, TCNN represents a convolutional neural network with excellent migration diagnosis performance, and IS-DATN represents a domain adversarial migration network based on the InceptionV1 module. Based on this, the diagnosis accuracy of the AR-MSPTL migration learning model provided by this application in 8 migration diagnosis scenarios is quantified by comparing different migration diagnosis methods. As Figure 5 shown, the AR-MSPTL migration learning model provided by this application has achieved the best migration diagnosis accuracy in all migration diagnosis scenarios. It can be seen that compared with other migration learning methods, the method provided by this application has higher diagnosis accuracy.

[0137] As Figure 6 shown, by comparing the number of parameters and the number of floating-point calculations of WDCNN, DANN, TCNN, IS-DATN, and the AR-MSPTL migration learning model, it can be seen that the number of parameters of the AR-MSPTL migration learning model provided by this application is at a relatively low level, only 3.0*10 5 . It can be seen that the method provided by this application can achieve the best migration diagnosis accuracy with a relatively low number of parameters. The reason is that the data augmentation method assisted by angular domain resampling has a stronger enhancement effect on fault features than other data augmentations. At the same time, the designed APMish can also greatly improve the feature extraction ability and classification ability of the neural network with almost no increase in the number of parameters, so as to achieve better migration diagnosis with a relatively simple network. In terms of the number of floating-point calculations, which is a key indicator to measure the computational complexity of the network, the method provided by this application also has a relatively low number of floating-point calculations, only 2.6*10 5The number of floating-point operations is also benefited from the simplicity of the network structure. In addition, although APMish has a more complex expression than mainstream activation functions such as ReLU, the increase in the number of floating-point calculations is almost negligible compared to the change in the number of floating-point calculations caused by the change in the network structure. Based on the above content, it can be seen that compared with other transfer learning methods, the method provided in this application has a better network structure and can achieve better fault diagnosis performance.

[0138] Furthermore, in order to verify the diagnostic ability of the method provided in this application, an ablation experiment was also conducted to verify the effectiveness of the data enhancement method based on angular domain resampling, the trainable parameterized activation function, and the multi-stage parameter transfer training. In Table 1, AR-RMSPTL means replacing APMish with the ReLU activation function, AR-LRMSPTL means replacing APMish with the LeakyReLU activation function, AR-MMSPTL means replacing APMish with the Mish activation function, FFT-MSPTL means using only the data of the fast Fourier transform of the original data as input, and AR-BCNN means using the traditional transfer training method based on parameter fine-tuning for transfer training.

[0139] Table 1 Ablation experiment results of different methods in various migration tasks

[0140] Migration task AR-MSPTL AR-RMSPTL AR-LRMSPTL AR-MMSPTL FFT-MSPTL AR-BCNN <![CDATA[D1]]> 93.29±1.62％ 87.46±5.30％ 92.67±3.44％ 91.00±2.83％ 92.52±4.40％ 87.76±4.25％ <![CDATA[D2]]> 91.36±2.03％ 82.75±6.95％ 85.85±8.18％ 82.24±11.12％ 87.79±3.83％ 85.26±5.91％ <![CDATA[D3]]> 94.31±1.11％ 87.45±2.63％ 86.73±6.51％ 93.92±1.93％ 92.63±2.79％ 80.93±4.31％ <![CDATA[D4]]> 93.32±1.58％ 87.43±5.30％ 91.84±4.45％ 91.33±2.97％ 88.77±4.95％ 72.79±6.45％ <![CDATA[D5]]> 93.05±1.79％ 86.83±5.75％ 91.58±6.05％ 92.80±3.44％ 90.34±4.02％ 90.38±3.61％ <![CDATA[D6]]> 90.05±4.37％ 84.86±8.84％ 84.49±8.41％ 87.49±6.96％ 84.35±3.78％ 86.37±3.34％ <![CDATA[D7]]> 90.64±2.29％ 84.00±9.37％ 84.64±6.12％ 89.98±4.04％ 89.23±5.44％ 83.17±5.53％ <![CDATA[D8]]> 89.62±3.64％ 82.68±5.95％ 84.05±4.66％ 87.29±3.80％ 89.25±3.71％ 76.00±4.87％ Average precision 91.96％ 85.43％ 87.73％ 89.51％ 89.36％ 82.83％

[0141] In different migration tasks, as shown in Table 1, the method provided in this application (i.e., AR-MSPTL) has achieved the best migration diagnosis performance in all migration diagnosis tasks compared with AR-RMSPTL, AR-LRMSPTL, AR-MMSPTL, FFT-MSPTL, and AR-BCNN, and has significantly improved the diagnostic accuracy and diagnostic accuracy variance compared with other methods. This effect of the method provided in this application is due to the self-regularization and adaptive characteristics of APMish. The method provided in this application is 6.53%, 4.23% and 2.45% ahead of AR-RMSPTL, AR-LRMSPTL and AR-MMSPTL in average diagnostic accuracy, respectively, which proves the advancement of APMish provided in this application. Compared with FFT-MSPTL and AR-BCNN, the method provided in this application is 2.6% and 9.13% ahead in average diagnostic accuracy, respectively. Based on this, the effectiveness of the data enhancement method assisted by angular domain resampling and the multi-stage migration training method provided in this application can be demonstrated.

[0142] Furthermore, based on the above description, compared with the prior art, the method provided by the present application has the following advantages:

[0143] 1. The present application provides a data enhancement method for order spectrum analysis assisted by angular resampling. By angular resampling, the influence of frequency aliasing caused by variable rotational speed is offset. Order spectrum analysis can effectively enhance the domain-invariant features under different working conditions, and the effect of enhancing fault features is better, thus effectively enhancing the fault features.

[0144] 2. The present application provides a trainable parametric activation function (APMish) with characteristics such as self-regularization, adaptability, and curve continuity. By introducing trainable parameters, the activation function expression is adaptively adjusted with the network, bringing better feature extraction ability and classification ability to the network, while hardly increasing the parameters and computational complexity of the network.

[0145] 2. Compared with the existing neural network structures, the improved convolutional neural network architecture provided by the present application incorporates the proposed APMish, having a simpler structure, a smaller network scale, and lower computational complexity.

[0146] 3. Compared with the existing mainstream transfer learning methods, the multi-stage parameter training method for transfer diagnosis provided by the present application divides the source domain dataset and the target domain dataset, and conducts training in multiple stages such as local training, global training, specific parameter training, and fast fine-tuning, enabling the network to learn more deep features and generalization features. The feature extraction ability of the network is also improved through the method of specific parameter training. The fast fine-tuning stage can enable the network to quickly adapt to the feature distribution of the target domain, and at the same time enable specific key parameters to further improve the network performance, achieving better transfer learning.

[0147] In an exemplary embodiment, a computer device is provided. The computer device can be a server or a terminal, and its internal structure diagram can be as Figure 7 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store multi-stage transfer learning training data for the transfer diagnosis of rotating machinery. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it realizes the transfer diagnosis of rotating machinery.

[0148] Those skilled in the art can understand that Figure 7 The structure shown in Figure 7 is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0149] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0150] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0151] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0152] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tapes, floppy disks, flash memories, optical memories, high-density embedded non-volatile memories, resistive random access memories (ReRAMs), magnetoresistive random access memories (MRAMs), ferroelectric random access memories (FRAMs), phase change memories (PCMs), graphene memories, etc. Volatile memories can include random access memories (RAMs) or external cache memories, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0153] The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logics, data processing logics based on quantum computing, etc., without limitation.

[0154] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0155] Specific examples are used in this article to elaborate on the principles and implementation manners of the present application. The descriptions of the above embodiments are only used to help understand the methods and core ideas of the present application; at the same time, for those of ordinary skill in the art, according to the ideas of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A multi-stage transfer learning training method for rotating machinery migration diagnosis, characterized in that Including: Obtain the sensing data of the rotating machinery; The sensing data of the rotating machinery covers various health states and working conditions in the source domain and the target domain; Adopt a data augmentation method of angular domain resampling to perform data augmentation on the sensing data of the rotating machinery to obtain an order spectrum signal, and generate a source domain dataset and a target domain dataset based on the order spectrum signal; Construct an improved convolutional neural network; Divide the source domain dataset into local data and global data, and divide the target domain dataset into fine-tuning data and test data; Perform multi-stage training on the improved convolutional neural network based on the local data and the global data to obtain a trained convolutional neural network; Adjust the parameters of the trained convolutional neural network by using the fine-tuning data to obtain a convolutional neural network with adjusted parameters; Test the convolutional neural network with adjusted parameters by using the test data, and use the convolutional neural network with adjusted parameters that meets the diagnostic accuracy requirements as a multi-stage parameter transfer learning network; the multi-stage parameter transfer learning network is applied to the transfer diagnosis of the rotating machinery.

2. The multi-stage transfer learning training method for rotating machinery migration diagnosis according to claim 1, wherein Adopt a data augmentation method of angular domain resampling to perform data augmentation on the sensing data of the rotating machinery to obtain an order spectrum signal, including: Adopt an interpolation method to perform angular domain resampling on the original time-domain signal in the sensing data of the rotating machinery according to the sampling time points to obtain angular domain data; Perform order spectrum analysis on the angular domain data to obtain the order spectrum signal.

3. The multi-stage transfer learning training method for rotating machinery migration diagnosis according to claim 2, characterized in that, The determination process of the sampling time points is expressed as: θ(t) = b1 + b2t + b3t2; Where, θ(t) represents the angle corresponding to the time point t of equiangular sampling, b1, b2, and b3 all represent variables to be solved, θ(t1), θ(t2), and θ(t3) respectively represent the data points of adjacent time points t1, t2, and t3 of the original time-domain signal, and Δθ represents the angle interval of angular domain resampling. represents the time point corresponding to equiangular resampling, and m represents the m-th equiangular interval of the angular domain data.

4. The multi-stage transfer learning training method for rotating machinery migration diagnosis according to claim 1, wherein Construct an improved convolutional neural network, including: Construct a trainable parameterized activation function, and construct the trainable parameterized activation function into a training parameterized activation function layer; Based on the training parameterized activation function layer, construct an improved convolutional neural network. The improved convolutional neural network consists of four convolutional blocks, two fully connected layers, one training parameterized activation function layer, and one Softmax layer.

5. The multi-stage transfer learning training method for rotating machinery migration diagnosis according to claim 4, characterized in that Each convolutional block is composed of a one-dimensional convolutional layer, a batch normalization layer, a training parameterized activation function layer, and a max pooling layer.

6. The multi-stage transfer learning training method for rotating machinery migration diagnosis according to claim 4, characterized in that, The trainable parameterized activation function is expressed as: Wherein, APMish(x i ) represents a trained parametric activation function, a i and b i represent trainable parameters, x i represents the input variable of the i-th channel, t i represents an intermediate variable, c represents the total number of channels, and i represents the i-th channel.

7. The multi-stage transfer learning training method for rotating machinery migration diagnosis according to claim 4, characterized in that Perform multi-stage training on the improved convolutional neural network based on the local data and the global data to obtain a trained convolutional neural network, including: Perform local training stage, global training stage, and specific parameter training stage on the improved convolutional neural network based on the local data and the global data to obtain a trained convolutional neural network; Among them, in the local training stage, use the local data to perform local training on the parameters of the improved convolutional neural network except the training parameterized activation function layer; in the global training stage, use the global data to perform global training on the parameters of the improved convolutional neural network except the training parameterized activation function layer; in the specific parameter training stage, use the global data to train the parameters of the training parameterized activation function layer.

8. The multi-stage transfer learning training method for rotating machinery migration diagnosis according to claim 1, characterized in that Adjust the parameters of the trained convolutional neural network by using the fine-tuning data to obtain a convolutional neural network with adjusted parameters, including: Use the fine-tuning data as the input to fine-tune the parameters of the trained convolutional neural network to obtain a convolutional neural network with adjusted parameters; among them, the fine-tuning process is expressed as: Wherein, θ e+f represents the parameters of the convolutional neural network after fine-tuning for f cycles, η F represents the learning rate of the fast fine-tuning part, θ e represents the parameters of the trained convolutional neural network, represents the i-th F sample in the fine-tuning data, represents the i-th F sample in the fine-tuning data corresponding to the healthy class label, f represents the number of fine-tuning times, θ e,l-1 represents the parameters of the convolutional neural network after the (l - 1)-th fine-tuning cycle.

9. A computer device, comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program to implement the multi-stage transfer learning training method for rotating machinery migration diagnosis according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-stage transfer learning training method for rotating machinery migration diagnosis according to any one of claims 1-8.

Images