A planetary roller screw fault diagnosis model construction method based on federated learning and light-weight model

Abstract

The application discloses a kind of planetary roller screw fault diagnosis model construction methods based on federal learning and light weight model, it is related to the field of fault diagnosis.Vibration data of planetary roller screw normal state and fault state are collected, and data set is constructed;Data preprocessing;Lightweight model SResNet18 is built;Based on federal learning framework and light weight model, planetary roller screw joint fault diagnosis modeling is carried out;Finally, the size and complexity of the model, the accuracy of the model and the training time of the model under the federal learning framework are evaluated.The application effectively solves the problem that there is no planetary roller screw fault diagnosis model construction method at present;Under the premise of ensuring data privacy, the data of each client is fully utilized to jointly establish the planetary roller screw fault diagnosis model;The lightweight model SResNet18 proposed in the application reduces the training time of the model under the federal learning framework, and solves the problem of high transmission cost of federal learning.

Description

TECHNICAL FIELD

[0001] The application relates to the field of fault diagnosis, and in particular to a planetary roller screw fault diagnosis model construction method based on federated learning and a light model. BACKGROUND

[0002] The planetary roller screw is gradually applied in the fields of aviation, aerospace, navigation and some precise servo transmission occasions due to its high load capacity, high precision and high limit speed. However, the planetary roller screw usually exists in the form of mechanical single redundancy, and the reliability of the planetary roller screw determines the reliable work of the whole system, so it is urgent to study the planetary roller screw fault diagnosis model construction method.

[0003] Compared with the mature processing technology of gears and bearings, the processing technology of the planetary roller screw is complex, the fault implantation is difficult, the fault experiment cost is high, and the fault data collection period is long, so the problems of lack of planetary roller screw fault data and lack of planetary roller screw fault diagnosis model construction method are faced, and the data have privacy, so each institution will not choose to publicly release the obtained fault data, forming a data barrier, which further hinders the research on the planetary roller screw fault diagnosis model construction method.

[0004] Federated learning uses data distributed in each client to model, then uploads the model to the cloud, aggregates the models of each client in the cloud, finally obtains a global model, and then the parameter information of the global model is returned to each client. Federated learning can not only ensure the privacy of data, but also can fully utilize multi-party data to cooperatively train the model. However, when federated learning trains the model, the parameters need to be uploaded, aggregated and returned constantly, compared with local training, the time of federated learning to train the model is greatly increased, and it is difficult to be really applied to the actual industrial environment.

[0005] Regarding the planetary roller screw fault diagnosis model construction method, the article "Fault Diagnosis of Planetary Roller Screw Mechanism Based on Bird Swarm Algorithm and Support Vector Machine" published in the Journal of Physics: Conference Series in 2020 proposes a planetary roller screw fault diagnosis model construction method using bird swarm algorithm and support vector machine. The article only considers one kind of planetary roller screw fault state, is more suitable for single fault condition, is difficult to be applied to multi-class fault, and does not solve the problem of not sharing data of each client.

[0006] For the fault diagnosis model construction method using federated learning, the invention patent 202110644538.6 proposes a "photovoltaic power station joint fault diagnosis method based on asynchronous decentralized federated learning", which sends local model parameters to other participants to update the model, to a certain extent, improves the communication and training efficiency of the model, but this method requires constant exchange of models between participants, which requires high server performance of participants, does not reduce the local training time of the model, and does not solve the problem of high transmission cost between federated learning participants.

[0007] To solve the above problems, the present invention proposes a planetary roller screw fault diagnosis model construction method based on federated learning and lightweight model, which can train the planetary roller screw fault diagnosis model using the data of each client while ensuring data privacy, and can reduce the local training time of the model and the parameter transmission cost of the model. SUMMARY

[0008] In view of the above defects of the prior art, the technical problem to be solved by the present invention is the lack of a fault diagnosis model construction method for planetary roller screws at the present stage, and the problem of high transmission cost between federated learning participants.

[0009] To achieve the above purpose, the present invention proposes a planetary roller screw fault diagnosis model construction method based on federated learning and lightweight model, characterized by the following steps:

[0010] Step 1, data acquisition;

[0011] Further, the data collected in step 1 includes X, Y, Z three direction vibration data when the planetary roller screw is in normal, lubrication failure, and roller side tooth breaking three state operation, and the vibration sensor is installed on the nut of the planetary roller screw.

[0012] Step 2, data preprocessing;

[0013] Further, step 2 includes the following steps:

[0014] Step 2.1, divide the original data into training set and test set, and set different types of labels;

[0015] Step 2.2, randomly divide the training set of step 2.1 to form the training set of each client (Host and Guest) in federated learning, wherein the training sets of Host and Guest are non-overlapping and have different data amounts;

[0016] Step 2.3, normalize the training set and test set of Host and Guest clients of step 2.2 according to formula (1) to obtain x*;

[0017]

[0018] wherein x i is the current sample data value, min(x) is the minimum value of the current sample, max(x) is the maximum value of the current sample, M is the maximum value of the normalized data, and m is the minimum value of the normalized data, the data can be normalized to any range by using formula (1), the present application considers the directional characteristics of the vibration data when the planetary roller screw is running, and at the same time, in order to reduce the data span, the data is normalized to [-1, 1], that is, M is 1 and m is -1;

[0019] Step 2.4, in order to increase the number of training set samples, the normalized data in step 2.3 is subjected to data enhancement by window clipping, and the number p of samples obtained is shown in formula (2):

[0020]

[0021] wherein n is the number of data points, w is the length of the sampled signal, and s is the data sampling interval;

[0022] Step 2.5, the vibration signal of the planetary roller screw is a non-stationary signal, so the data in step 2.4 is subjected to wavelet packet transform by formula (3) to obtain the detail coefficient W i+1,2j of the high-frequency signal and the approximation coefficient W i+1,2j+1 of the low-frequency signal, and then the approximation coefficients and the detail coefficients of each frequency band in the last layer are arranged in rows to form a 64x64 coefficient matrix, and then the coefficient matrices of the X, Y and Z directions are stacked to obtain a 64x64x3 matrix as the input of the neural network;

[0023]

[0024] wherein h(·) is a high-pass orthogonal filter, g(·) is a low-pass orthogonal filter, W i,j (k), k = 1, 2, …, N / 2 i is the wavelet packet coefficient at the i-th layer and the j-th sub-node, and τ is the translation amount.

[0025] Step 3, because the vibration data in the X, Y and Z directions of the planetary roller screw in normal, lubrication failure, roller side tooth breaking and other states when running are not completely independent, there is a certain relationship, but as a time series signal, the signals at each time point are quite different, in view of the data characteristics, the traditional convolution layer is modified and is called a symmetric convolution layer, the kernels for multiplication and addition operation at each position in space are different, but a group of weight identical kernels are used in different channels, a symmetric convolution layer is used as the main network layer, the convolution layer in the traditional 18-layer residual neural network is replaced by a symmetric convolution layer, and a network structure as shown inFigure 3 The lightweight neural network model SResNet18 shown includes symmetric convolutional layers, Batch Normalization, ReLU activation function, max pooling layer, convolutional layer, symmetric convolutional residual layer 1, symmetric convolutional residual layer 2, global flat pooling layer, fully connected layer, and Softmax function. These will be described in detail below.

[0026] ① Symmetrical convolutional layers: such as Figure 4 As shown, the symmetric convolutional layer has multiple kernels with different weights in space, which perform multiplication and addition operations with the input at different locations in space. This allows it to adaptively extract more spatial information at different spatial locations. All channels share a set of kernels, which reduces the number of parameters and allows the symmetric convolutional layer to use larger kernels, thereby capturing long-range features. The calculation process on each channel is defined by equation (4):

[0027]

[0028] Where X is the input, I is the symmetric convolution kernel (all channels share the same set of symmetric convolution kernels with equal weights), u and v are the neighborhood offsets between the center positions of the symmetric convolution kernel and the input, K is the size of the symmetric convolution kernel, and C is the weight of the input kernel. i For the i-th channel;

[0029] There are multiple symmetric convolution kernels in space, and the weights of the symmetric convolution kernels at different positions are different. Therefore, it is necessary to dynamically generate different symmetric convolution kernels at different positions. This invention addresses this by... Figure 5 The bottleneck layer shown generates a symmetric convolutional kernel, where r is a parameter determining the size of the bottleneck layer. When the stride s of the symmetric convolutional kernel is greater than 1, the height H0 and width W0 of the input will change. The size of the symmetric convolutional kernel generated by only one bottleneck layer does not match the input, so a pooling layer is needed to reduce the height H0 and width W0 of the input. After a pooling layer, the height W1 and width W1 become H1 = H0 / s and W1 = W0 / s, respectively. When the stride s of the symmetric convolutional kernel is equal to 1, the size of the symmetric convolutional kernel generated by only one bottleneck layer matches the input, so no further pooling is needed. Figure 5 The structure within the dashed lines;

[0030] The symmetric convolutional layer performs multiplication and addition operations with the input C0 channels through a set of symmetric convolutional kernels to obtain an output with a single channel C0. The number of channels output by the symmetric convolutional layer is the same as the number of channels input. Therefore, a convolutional layer with a kernel size of 1×1 and a stride of 1 needs to be added after the symmetric convolutional layer to change the number of channels.

[0031] ②Batch Normalization: Batch Normalization can make the gradient descent algorithm more stable in the gradient descent process;

[0032] ③ReLU activation function: in order to improve the fitting ability of the neural network, the output needs to be de-linearized, and the ReLU activation function is selected, as shown in formula (5):

[0033]

[0034] Wherein, X is the input;

[0035] ④Max pooling layer: the max pooling layer traverses the entire input with a certain step, and outputs the maximum value of the data in the window;

[0036] ⑤Convolution layer: the convolution layer slides the convolution kernel on the input matrix, calculates the dot product of the convolution kernel and the input at each position, and the calculation process is defined by formula (6):

[0037]

[0038] Wherein, X is the input, F is the convolution kernel, u and v are the neighborhood offset of the center position of the convolution kernel and the input, K is the size of the convolution kernel, and C1 is the number of output channels;

[0039] ⑥Symmetric convolution residual layer: the symmetric convolution residual layer avoids gradient explosion and gradient disappearance by adding an identity mapping, when the step s of the first symmetric convolution layer is greater than 1, Figure 6 The output Y of the symmetric convolution residual layer 1 shown in the figure is not equal in height and width to the input X, so a convolution layer with a step s of 1 is needed to change the size of the input; when the step s of the first symmetric convolution layer is equal to 1, Figure 7 The output Y of the symmetric convolution residual layer 2 shown in the figure is equal in height and width to the input X, and can be directly added;

[0040] ⑦Global average pooling layer: the global average pooling layer takes the average of all inputs in the spatial direction, and the height and width of the output are both 1, using the global average pooling layer before the fully connected layer can greatly reduce the parameter quantity;

[0041] ⑧Fully connected layer and Softmax function: the fully connected layer and Softmax cooperate to output the fault diagnosis classification result;

[0042] Step 4, based on the federated learning framework and the lightweight model, the planetary roller screw joint fault diagnosis modeling is carried out. The purpose of the present application is to establish a planetary roller screw fault diagnosis model under the premise of ensuring data privacy, make full use of the planetary roller screw data of multiple clients, improve the performance of the model, and establish the lightweight model SResNet18 in step 3 for the problem of model parameter transmission cost under the federated learning framework. It is worth noting that in the method of the present application, each client does not need to provide local data, only needs to upload the trained model parameters, and obtains a model with high accuracy and fast training speed through the data motionless model motion method. In the process of jointly establishing a planetary roller screw fault diagnosis model by multiple clients, the following steps are mainly included:

[0043] Step 4.1, Host and Guest train local fault diagnosis models using local training sets;

[0044] Step 4.2, Host and Guest upload the parameters of the obtained local fault diagnosis models to the cloud;

[0045] Step 4.3, a global model is obtained by aggregating each model in the cloud, and the parameters of the global model are P C , and the aggregation process is defined by formula (7):

[0046]

[0047] Where P H and P G are the parameters of the Host and Guest models, respectively;

[0048] Step 4.4, the global model parameters are returned, and the parameters of the Host and Guest local fault diagnosis models are updated to P C ;

[0049] The above steps 4.1-4.4 are repeatedly performed until an excellent performance model is obtained.

[0050] Step 5, joint fault diagnosis modeling and lightweight model performance evaluation. In order to verify the effectiveness of federated learning for planetary roller screw fault diagnosis modeling and the effectiveness of lightweight model in improving the training speed under the federated learning framework, the present application carries out fault diagnosis experiments using the collected data of planetary roller screws in normal state, lubrication failure state and broken teeth on one side of the roller. In order to verify the effectiveness of the method of the present application, three aspects are evaluated, including:

[0051] ① Model size and complexity;

[0052] ② Model accuracy on test set;

[0053] ③ Training time of model under federated learning framework.

[0054] Compared with the prior art, the present application has the beneficial effects that the present application first proposes a processing method for planetary roller screw fault diagnosis data, first proposes a method of federated learning for planetary roller screw fault diagnosis, solves the problem of insufficient planetary roller screw fault diagnosis data held by a single institution under the premise of ensuring data privacy, improves the performance of the model, and proposes a lightweight model with less parameters according to the characteristics of the vibration signal. Under the federated learning framework, not only the local model training time is reduced, but also the model parameter upload, model aggregation and model parameter return time is reduced, which can shorten the time of planetary roller screw fault diagnosis modeling under the federated learning framework. In addition to planetary roller screws, the present application also has certain promoting effect on the development and application of intelligent diagnosis technology for other important mechanisms such as bearings and gearboxes. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 The structural diagram of the planetary roller screw fault diagnosis model construction method based on federated learning and lightweight model of the present application.

[0056] Figure 2 The generation schematic diagram of the wavelet packet transform coefficient matrix.

[0057] Figure 3 The structure diagram of SResNet18.

[0058] Figure 4 The forward propagation schematic diagram of the symmetric convolution layer.

[0059] Figure 5 The generation process schematic diagram of the symmetric convolution kernel.

[0060] Figure 6 The structure diagram of the symmetric convolution residual layer 1.

[0061] Figure 7 The structure diagram of the symmetric convolution residual layer 2.

[0062] Figure 8 The output result distribution diagram and the confusion matrix heat map of ResNet18 in the test set under the federated learning framework.

[0063] Figure 9 The output result distribution diagram and the confusion matrix heat map of SResNet18 in the test set under the federated learning framework. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be described clearly and completely below with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative effort belong to the scope of protection of the present application.

[0065] The present application aims at the problems of lack of planetary roller screw fault diagnosis model construction method, non-sharing of fault diagnosis data between clients, and high transmission cost under the federated learning framework, and proposes a planetary roller screw fault diagnosis model construction method based on federated learning and lightweight model as shown in the figure. Figure 1 The present application first introduces the construction method of the planetary roller screw fault diagnosis data set. Then, the processing method for the vibration data of the planetary roller screw is introduced. Next, the present application builds an SResNet18 model for training the fault diagnosis model, which has a small number of parameters and can reduce the transmission cost of parameters under the federated learning framework. Then, the present application carries out federated learning fault diagnosis modeling of the planetary roller screw. Finally, the present application carries out fault diagnosis experiment to evaluate the size and complexity of the fault diagnosis model, the accuracy of the model on the test set and the training time of the model under the federated learning framework, to verify the effectiveness of the method of the present application, which includes the following steps:

[0066] 1. Data acquisition: the present application proposes a planetary roller screw fault diagnosis model construction method for multiple client cooperative modeling. In which, the specific parameters of the planetary roller screw used by the present application are: screw diameter of 24mm, pitch of 2mm, number of heads of 5, number of rollers of 10, specific working conditions of screw speed of 104r / min, load of 9kN, and X, Y, Z three direction vibration data under normal state, lubrication failure and roller one side broken tooth two common fault states of planetary roller screw are collected, in which the vibration sensor is installed on the nut of the planetary roller screw.

[0067] 2. Data preprocessing: in the present application, each client performs the same data preprocessing on the data.

[0068] Firstly, the data is divided into training set and test set.

[0069] Then, considering that the data held by each client in the actual industrial scene is different and the data is not overlapped, the training set is randomly divided to obtain the training set of Host and Guest.

[0070] Next, in order to reduce the data span, the training and test sets of the Host and Guest clients are normalized. Since the vibration data has directionality, in order to preserve the directionality of the data and reduce the data span, the data is mapped to the range of [-1,1] according to Equation (8).

[0071]

[0072] Where, x i is the current sample data value, M is the maximum value of the normalized data, which is 1, and m is the minimum value of the normalized data, which is -1.

[0073] Subsequently, in order to improve model performance, the amount of data in the Host and Guest training sets was increased by using a window-cropping data augmentation method. The number of samples p after data augmentation is shown in formula (9):

[0074]

[0075] Where n is the number of data points, w is the length of the sampled signal, and s is the data sampling interval;

[0076] Finally, since the vibration signal of the planetary roller screw is a non-stationary signal, wavelet packet transformation is performed using formula (10). The selected wavelet function is "db1", and the wavelet packet decomposition level is 6. After wavelet packet transformation, the detail coefficients W of the high-frequency signal are obtained. i+1,2j The approximation coefficient W of the low-frequency signal i+1,2j+1 ,like Figure 2 As shown, the approximation coefficients of the low-frequency part and the detail coefficients of the high-frequency part are arranged in rows to form a 64×64 coefficient matrix. Then, the coefficient matrices in the X, Y, and Z directions are stacked to obtain a 64×64×3 matrix, which is used as the input of the neural network.

[0077]

[0078] Where h(·) is the high-pass orthogonal filter, g(·) is the low-pass orthogonal filter coefficient, and W i,j (k), k = 1, 2, ..., N / 2 i Let τ be the wavelet packet coefficient at the j-th child node of the i-th layer, and τ be the translation amount.

[0079] 3. Building a Lightweight Model. Because the vibration data in the X, Y, and Z directions of a planetary roller screw under various conditions such as normal operation, lubrication failure, and broken teeth on one side of the roller are not completely independent and are related, but as time-series signals, the signals at different time points differ significantly. To address these data characteristics, the traditional convolutional layer is modified into a symmetric convolutional layer. This allows different kernels to perform multiplication and addition operations at different spatial locations, but different channels use a set of kernels with identical weights. Using the symmetric convolutional layer as the main network layer, and replacing the convolutional layers in the traditional 18-layer residual neural network with symmetric convolutional layers, a lightweight model is built. Figure 3 The lightweight neural network model SResNet18 shown includes symmetric convolutional layers, BatchNormalization, ReLU activation function, max pooling layer, convolutional layer, symmetric convolutional residual layer 1, symmetric convolutional residual layer 2, global flat pooling layer, fully connected layer, and Softmax function. These will be described in detail below.

[0080] ① Symmetrical convolutional layers: such as Figure 4 As shown, the symmetric convolutional layer has multiple kernels with different weights in space, which perform multiplication and addition operations with the input at different locations in space. This allows it to adaptively extract more spatial information at different spatial locations. All channels share a set of kernels, which reduces the number of parameters and allows the symmetric convolutional layer to use larger kernels, thereby capturing long-range features. The calculation process on each channel is defined by equation (11):

[0081]

[0082] Where X is the input, I is the symmetric convolution kernel (all channels share the same set of symmetric convolution kernels with equal weights), u and v are the neighborhood offsets between the center positions of the symmetric convolution kernel and the input, K is the size of the symmetric convolution kernel, and C is the weight of the input kernel. i For the i-th channel;

[0083] There are multiple symmetric convolution kernels in space, and the weights of the symmetric convolution kernels at different positions are different. Therefore, it is necessary to dynamically generate different symmetric convolution kernels at different positions. This invention addresses this by... Figure 5 The bottleneck layer shown generates a symmetric convolutional kernel, where r is a parameter determining the size of the bottleneck layer. When the stride s of the symmetric convolutional kernel is greater than 1, the height H0 and width W0 of the input will change. The size of the symmetric convolutional kernel generated by only one bottleneck layer does not match the input, so a pooling layer is needed to reduce the height H0 and width W0 of the input. After one pooling layer, the height W1 and width W1 become H1 = H0 / s and W1 = W0 / s, respectively. When the stride s of the symmetric convolutional kernel is equal to 1, the size of the symmetric convolutional kernel generated by only one bottleneck layer matches the input, so no further pooling is needed. Figure 5 Structures within dashed lines;

[0084] The symmetric convolution layer respectively performs multiplication and addition operation with C0 input channels through a set of symmetric convolution kernels to obtain an output with one channel C0. The number of channels of the symmetric convolution layer output is the same as the number of input channels, so a convolution layer with a convolution kernel size of 1*1 and a step of 1 is needed after the symmetric convolution layer to change the number of channels.

[0085] ②Batch Normalization: Batch Normalization can make the gradient descent algorithm more stable during gradient descent;

[0086] ③ReLU activation function: In order to improve the fitting ability of the neural network, the output needs to be de-linearized, and the ReLU activation function is selected, as shown in formula (12):

[0087]

[0088] Wherein, X is the input;

[0089] ④Max pooling layer: The max pooling layer slides through the entire input with a certain step size and outputs the maximum value in the window;

[0090] ⑤Convolution layer: The convolution layer slides the convolution kernel on the input matrix, calculates the dot product of the convolution kernel and the input at each position, and the calculation process is defined by formula (13):

[0091]

[0092] Wherein, X is the input, F is the convolution kernel, u and v are the neighborhood offset of the center position of the convolution kernel and the input, K is the size of the convolution kernel, and C1 is the number of output channels;

[0093] ⑥Symmetric convolution residual layer: The symmetric convolution residual layer avoids gradient explosion and gradient disappearance by adding an identity mapping. When the step size s of the first symmetric convolution layer is greater than 1, Figure 6 The output Y of the symmetric convolution residual layer 1 shown in the figure is not equal in height and width to the input X, so a convolution layer with a step size s and a size of 1*1 is needed to change the size of the input; when the step size of the first symmetric convolution layer is equal to 1, Figure 7 The output Y of the symmetric convolution residual layer 2 shown in the figure is equal in height and width to the input X, and can be directly added;

[0094] ⑦Global average pooling layer: The global average pooling layer takes the average of all inputs in the spatial direction, and the height and width of the output are both 1. Using the global average pooling layer before the fully connected layer can greatly reduce the parameter quantity;

[0095] 8. Full connection layer and Softmax function: the full connection layer and Softmax cooperate to output the fault diagnosis classification result;

[0096] 4. Joint fault diagnosis modeling of planetary roller screw based on federated learning framework and lightweight model: Considering that the fault diagnosis data of a single mechanism held by the planetary roller screw is insufficient at the present stage, but the data is not shared among institutions, and the transmission cost is high under the federated learning framework, the present application proposes a planetary roller screw fault diagnosis model construction method based on federated learning and lightweight model to solve the above problems, which realizes the rapid joint fault diagnosis modeling of multiple clients by using the kubefate-docker-compose-v1.6.0 framework.

[0097] Firstly, the Host and the Guest train the local fault diagnosis model using the local training set;

[0098] Then, the Host and the Guest upload the parameters of the obtained local fault diagnosis model to the cloud;

[0099] Next, a global model is obtained by aggregating each model in the cloud, and the parameters of the global model are P C , and the aggregation process is defined by formula (14):

[0100]

[0101] Where P H and P G are the parameters of the Host and Guest models, respectively;

[0102] Then, the global model parameters are returned to update the parameters of the local fault diagnosis models of the Host and the Guest to P C .

[0103] Finally, the above steps are repeatedly performed until the global model converges to a higher accuracy, and the joint fault diagnosis modeling of multiple clients is completed.

[0104] 5. Joint fault diagnosis modeling and performance evaluation of lightweight model: In order to verify the effectiveness of federated learning for planetary roller screw fault diagnosis modeling and the effectiveness of lightweight model in improving the training speed under the federated learning framework, the present application uses the collected data of planetary roller screw in normal state, lubrication failure state and roller side tooth breakage for fault diagnosis experiment. In order to verify the effectiveness of the method of the present application, it is evaluated from three aspects, including:

[0105] ① The size and complexity of the model;

[0106] ② The accuracy of the model on the test set;

[0107] ③The training time of the model under the federated learning framework.

[0108] (1) Introduction of the data set

[0109] The specific description of the finally obtained data set is shown in Table 1.

[0110] Table 1: Data set

[0111]

[0112] (2) Experimental design

[0113] In order to verify the effectiveness of the method proposed in the application, comparative experiments are set: a. The traditional 18-layer residual neural network model ResNet18, b. The lightweight model SResNet18 proposed in the application. Compared with ResNet18, 4 layers of convolutional layers with a convolution kernel size of 1x1 and a step of 1 are added in SResNet18 to change the input channel. The specific experimental comparison includes:

[0114] ①In order to compare the size and complexity of the model, the parameter quantity and floating point operation number of ResNet18 and SResNet18 models are compared;

[0115] ②In order to verify the effectiveness of federated learning for modeling of planetary roller screw fault diagnosis, the accuracy of ResNet18 and SResNet18 local training models and federated learning models on the test set is compared;

[0116] ③In order to verify the effectiveness of the lightweight model for improving the training speed under the federated learning framework, the training time of ResNet18 and SResNet18 models under the federated learning framework is compared.

[0117] (3) Parameter setting

[0118] The specific network parameters of the experiments of the application are shown in Table 2.

[0119] Table 2: Parameter setting of each model

[0120] Model Number of network layers Learning rate Batch size Number of training rounds Optimizer ResNet18 18 0.01 64 100 Adam SResNet18 22 0.01 64 100 Adam

[0121] (4) Analysis of experimental results

[0122] The parameter quantity and floating point operation number of each model are shown in Table 3, wherein the model parameter quantity represents the number of parameters that need to be constantly updated by the model, which can measure the size of the model, and the floating point operation number represents the calculation amount of the model, which can measure the complexity of the model.

[0123] Table 3: Parameter quantity and floating point operation number of each model

[0124] Model Parameter quantity Floating point operation number ResNet18 11.2M 0.297G SResNet18 0.552M 0.0235G

[0125] In addition to the SResNet18 adding 4 layers of convolutional layers for changing the number of channels, the network depth of ResNet18 and SResNet18 is completely the same, as can be seen from Table 3, the parameter quantity and floating point operation number of SResNet18 are reduced by 95.07% and 92.09% respectively compared with ResNet18, so compared with the traditional ResNet18, the SResNet18 proposed in the application has great advantages in both the size of the model and the complexity of the model.

[0126] Because the training of the model is initiated by the Guest under the kubefate federated learning framework, the local training model refers to the model trained by the training set of the Guest, because the Host and the Guest both participate in the training under the federated learning framework, the accuracy of the training set is the accuracy of the Host and the Guest training set respectively, and the accuracy of each model on the test set is shown in Table 4.

[0127] Table 4: Accuracy of each model on the test set

[0128]

[0129] As can be seen from Table 4, during local training, although the training set accuracy of the traditional ResNet18 is 99.5%, the test set accuracy is only 67.5%, indicating that ResNet18 has serious overfitting during local training, but the test set accuracy and the training set accuracy of SResNet18 are less different, indicating that the lightweight model SResNet18 proposed in the application can reduce the overfitting phenomenon; during federated learning training of the model, the accuracy of ResNet18 on the Guest training set is only 37.1%, and the accuracy on the test set is only 36.3%, which is lower than the accuracy during local training, because ResNet18 has more parameters, and more rounds of training are needed to obtain a ResNet18 model with high accuracy, but setting a larger number of training rounds will increase the time cost and operation cost, causing unnecessary waste, on the contrary, the accuracy of SResNet18 on the Guest training set is 99.4%, and the accuracy on the test set is as high as 99.2%, both of which are higher than during local training, and there is almost no overfitting phenomenon. In summary, the performance of SResNet18 under the federated learning framework is the best, and the accuracy on the test set is the highest, and from the accuracy of SResNet18 on the test set during local training and federated learning framework, it can be seen that federated learning can reduce the overfitting phenomenon and improve the performance of the model.

[0130] The training time of each model under the federated learning framework is shown in Table 5.

[0131] Table 5 training time of each model under the federated learning framework

[0132] Model Model training time ResNet18 9039s SResNet18 3551s

[0133] As can be seen from Table 5, under the federated learning framework, the modeling time of SResNet18 is reduced by 60.71% compared with ResNet18, which shows that the lightweight SResNet18 model proposed in the application can significantly improve the training speed of the model and reduce the transmission cost under the federated learning framework.

Claims

1. A method for constructing a planetary roller screw fault diagnosis model based on federated learning and a lightweight model, characterized in that, Comprising the following steps: Step 1, data acquisition; The collected data include X, Y and Z direction vibration data when the planetary roller screw is in normal, lubrication failure and roller side broken tooth three states, and the vibration sensor is installed on the nut of the planetary roller screw; Step 2, data preprocessing; specifically including the following sub steps: Step 2.1, the original data is divided into training set and test set, and different types of labels are set; Step 2.2, the training set in step 2.1 is randomly divided to constitute the training set of each client Host and Guest in federated learning, wherein the training sets of Host and Guest are non overlapping and different in data amount; Step 2.3, the training set and test set of Host and Guest clients in step 2.2 are normalized according to formula (1) to obtain x*; wherein x d is the value of the dth data point of the current sample, min(x) is the minimum value of the current sample, max(x) is the maximum value of the current sample, M is the maximum value of the normalized data, and m is the minimum value of the normalized data. The data is normalized to an arbitrary range by using formula (1). Considering the directional characteristics of the vibration data when the planetary roller screw is running, the data is normalized to [-1, 1] in order to reduce the data span, that is, M is 1 and m is -1. Step 2.4, in order to increase the number of training set samples, the normalized data in step 2.3 is enhanced by window clipping to obtain the sample number p as shown in formula (2): Wherein, n is the number of data points, w is the sampling signal length, and Df is the data sampling interval; Step 2.

5. The vibration signal of the planetary roller screw is a non-stationary signal, so the data described in step 2.4 is subjected to wavelet packet transformation by formula (3) to obtain the transformed coefficient matrix WP. In the pth layer decomposition, the signal is divided into the detail coefficient WP of the high-frequency signal and the approximation coefficient WP of the low-frequency signal p+1,2q Step 2.

5. The vibration signal of the planetary roller screw is a non-stationary signal, so the data described in step 2.4 is subjected to wavelet packet transformation by formula (3) to obtain the transformed coefficient matrix WP. In the pth layer decomposition, the signal is divided into the detail coefficient WP of the high-frequency signal and the approximation coefficient WP of the low-frequency signal p+1,2q+1 Then the approximation coefficients and the detail coefficients of each frequency band in the last layer are arranged in rows to form a 64x64 coefficient matrix, and then the coefficient matrices in X, Y and Z directions are stacked to obtain a 64x64x3 tensor as the input of the neural network. where h(·) is a high-pass orthogonal filter, g(·) is a low-pass orthogonal filter, WP p,q (k), is a wavelet packet coefficient at the pth layer, the qth subnode, k is an index of the layer number, k = 1, 2, …, N / 2p, N is the length of the original signal, and τ is a translation amount. Step 3, the traditional convolution layer is modified, which is called symmetric convolution layer, so that the kernels in space are different, but a group of weights are completely the same, and the symmetric convolution layer is used as the main network layer, the convolution layer in the traditional 18 layer residual neural network is replaced by the symmetric convolution layer, and the lightweight neural network model SResNet18 is built, including symmetric convolution layer, Batch Normalization, ReLU activation function, maximum pooling layer, convolution layer, symmetric convolution residual layer 1, symmetric convolution residual layer 2, global average pooling layer, full connection layer and Softmax function: ①Symmetric convolution layer: the symmetric convolution layer has multiple kernels with different weights in space, which are multiplied and added with the input in different positions in space, and all channels share a group of kernels, and the calculation process in each channel is defined by formula (4): wherein, is the output of a symmetric convolution kernel, c is the channel index of the input and output, is the c-th channel of the input or the c-th channel of the output, X is the input, I is the symmetric convolution kernel, all channels share a set of weights of the symmetric convolution kernel, u s and v s is the neighborhood offset of the symmetric convolution kernel and the input, K s is the size of the symmetric convolution kernel; The symmetric convolution kernel has multiple kernels in space, and the weights of the symmetric convolution kernels in different positions are different, the symmetric convolution kernel is generated through a bottleneck layer, and the height H0 and width W0 of the input are reduced through a pooling layer, after the pooling layer, the height H1 and width W1 are changed to H1=H0 / s and W1=W0 / s respectively, when the step s of the symmetric convolution kernel moving is equal to 1, only the symmetric convolution kernel generated through a bottleneck layer has the same size as the input; The symmetric convolution layer is multiplied and added with C0 channels of input through a group of symmetric convolution kernels to obtain an output with C0 channels, the number of channels of the symmetric convolution layer output is the same as the number of channels of the input, so a convolution layer with kernel size of 1*1 and step of 1 is added after the symmetric convolution layer to change the number of channels; ②Batch Normalization: Batch Normalization makes the gradient descent algorithm more stable in the gradient descent process; ③ReLU activation function: as shown in formula (5): Wherein, X is the input; (4) Max-pooling layer: the max-pooling layer traverses the entire input with a certain step size, and outputs the maximum value of the data in the window; (5) Convolution layer: the convolution layer calculates the dot product of the convolution kernel and the input at each position by sliding the convolution kernel on the input matrix, and the calculation process is defined by formula (6): where F is the convolution kernel, C1 is the number of output channels; u and v are the neighborhood offsets of the convolution kernel and the input, and K is the size of the convolution kernel; (6) Symmetric convolution residual layer: the symmetric convolution residual layer avoids gradient explosion and gradient disappearance by adding an identity mapping. When the step size s of the first symmetric convolution layer is greater than 1, a convolution layer with a step size of s and a size of 1x1 is added to change the size of the input. When the step size s of the first symmetric convolution layer is equal to 1, the height and width of the output Y of the symmetric convolution residual layer are equal to those of the input X, and the addition operation is directly performed; (7) Global average pooling layer: the global average pooling layer takes the average of all inputs in the spatial direction, and the height and width of the output are both 1. Using the global average pooling layer before the fully connected layer greatly reduces the parameter quantity; (8) Fully connected layer and Softmax function: the fully connected layer and the Softmax function cooperate to output the fault diagnosis classification result; Step 4, based on the federated learning framework and the lightweight model, the planetary roller screw joint fault diagnosis modeling is carried out; under the premise of ensuring data privacy, the planetary roller screw fault diagnosis model is established, the data of the planetary roller screw of multiple clients is fully utilized, the performance of the model is improved, and the lightweight model SResNet18 described in step 3 is established to solve the problem of model parameter transmission cost under the federated learning framework; in the process of jointly establishing the planetary roller screw fault diagnosis model in multiple clients, the following steps are included: Step 4.1, the host and the guest train the local fault diagnosis model using the local training set; Step 4.2, the host and the guest upload the parameters of the obtained local fault diagnosis model to the cloud; Step 4.3, aggregate all the models in the cloud to get a global model, the parameters of the global model are P C The aggregation process is defined by equation (7): where P H and P G are the parameters of the Host and Guest models, respectively; Step 4.4, Global model parameters back propagation, update the parameters of Host and Guest local fault diagnosis models as P C ; The above steps 4.1-4.4 are repeatedly performed until a model with excellent performance is obtained; Step 5, joint fault diagnosis modeling and lightweight model performance evaluation.

2. The planetary roller screw fault diagnosis model construction method based on federated learning and lightweight model according to claim 1, characterized in that, In the step 5, the fault diagnosis modeling and the lightweight model performance evaluation are performed by using the collected data of the planetary roller screw in normal state, lubrication failure state and roller side tooth breakage for fault diagnosis experiment, which is evaluated from three aspects, including: (1) the size and complexity of the model; (2) the accuracy of the model on the test set; (3) the training time of the model under the federated learning framework.

Images