CN116561684A - Planetary roller screw fault diagnosis model construction method based on federal learning and lightweight model - Google Patents
Planetary roller screw fault diagnosis model construction method based on federal learning and lightweight model Download PDFInfo
- Publication number
- CN116561684A CN116561684A CN202310481747.2A CN202310481747A CN116561684A CN 116561684 A CN116561684 A CN 116561684A CN 202310481747 A CN202310481747 A CN 202310481747A CN 116561684 A CN116561684 A CN 116561684A
- Authority
- CN
- China
- Prior art keywords
- model
- layer
- data
- convolution
- fault diagnosis
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000003745 diagnosis Methods 0.000 title claims abstract description 71
- 238000010276 construction Methods 0.000 title claims abstract description 15
- 238000012549 training Methods 0.000 claims abstract description 60
- 238000000034 method Methods 0.000 claims abstract description 33
- 230000005540 biological transmission Effects 0.000 claims abstract description 11
- 238000007781 pre-processing Methods 0.000 claims abstract description 5
- 238000011176 pooling Methods 0.000 claims description 27
- 238000012360 testing method Methods 0.000 claims description 22
- 230000006870 function Effects 0.000 claims description 15
- 239000011159 matrix material Substances 0.000 claims description 13
- 238000010606 normalization Methods 0.000 claims description 10
- 230000008859 change Effects 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 8
- 238000013528 artificial neural network Methods 0.000 claims description 8
- 238000005461 lubrication Methods 0.000 claims description 8
- 238000002474 experimental method Methods 0.000 claims description 7
- 238000003062 neural network model Methods 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000011156 evaluation Methods 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 4
- 230000009466 transformation Effects 0.000 claims description 4
- 230000008034 disappearance Effects 0.000 claims description 3
- 238000004880 explosion Methods 0.000 claims description 3
- 238000006116 polymerization reaction Methods 0.000 claims description 3
- 230000004931 aggregating effect Effects 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 9
- 238000007667 floating Methods 0.000 description 6
- 230000007246 mechanism Effects 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000003672 processing method Methods 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 102100029469 WD repeat and HMG-box DNA-binding protein 1 Human genes 0.000 description 1
- 101710097421 WD repeat and HMG-box DNA-binding protein 1 Proteins 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013400 design of experiment Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002513 implantation Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2433—Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a planetary roller screw fault diagnosis model construction method based on federal learning and a lightweight model, and relates to the field of fault diagnosis. Collecting vibration data of a normal state and a fault state of the planetary roller screw, and constructing a data set; preprocessing data; constructing a lightweight model SResNet18; developing planetary roller screw joint fault diagnosis modeling based on a federal learning frame and a lightweight model; and finally, evaluating the size and complexity of the model, the accuracy of the model and the training time of the model under the federal learning framework. The method effectively solves the problem that a method for constructing a fault diagnosis model of the planetary roller screw is lacked in the current stage; on the premise of ensuring data privacy, fully utilizing the data of each client to establish a planetary roller screw fault diagnosis model; the lightweight model SReNet 18 provided by the invention reduces the training time of the model under the federal learning framework and solves the problem of high federal learning transmission cost.
Description
Technical Field
The invention relates to the field of fault diagnosis, in particular to a planetary roller screw fault diagnosis model construction method based on federal learning and a lightweight model.
Background
The planetary roller screw has the advantages of strong bearing capacity, high precision, high limit rotation speed and the like, and is increasingly applied to the fields of aviation, aerospace, navigation and the like and some occasions requiring precise servo transmission, but the planetary roller screw usually exists in a mechanical single redundancy mode, and the reliability of the planetary roller screw determines the reliable operation of the whole system, so that the research on a fault diagnosis model construction method of the planetary roller screw is urgently needed.
Compared with the mature processing technology of the gears and the bearings, the processing technology of the planetary roller screw is complex, the fault implantation difficulty is high, the fault experiment cost is high, and the fault data acquisition period is long, so that the problems of lack of fault data of the planetary roller screw and lack of a construction method for a fault diagnosis model of the planetary roller screw are faced, and the data has privacy, so that each mechanism does not select to disclose the obtained fault data, a data barrier is formed, and the research of the construction method of the fault diagnosis model of the planetary roller screw is further hindered.
The federal learning is modeled by data scattered at all clients, the models are uploaded to the cloud, all client models are aggregated at the cloud to finally obtain a global model, and parameter information of the global model is returned to all clients. The federal learning ensures the privacy of data and can fully utilize multiparty data to cooperatively train the model, but parameters need to be continuously uploaded, aggregated and returned when the federal learning trains the model, so that compared with local training, the time of the federal learning trains the model is greatly increased, and the federal learning training model is difficult to be truly applied to actual industrial environments.
Regarding a method for constructing a fault diagnosis model of a planetary roller screw, in article Fault Diagnosis of Planetary Roller Screw Mechanism Based on Bird Swarm Algorithm and Support Vector Machine published in Journal of Physics: conference Series in 2020, a method for constructing a fault diagnosis model of a planetary roller screw by adopting a shoal algorithm and a support vector machine is proposed, and the article only considers a fault state of the planetary roller screw, is more suitable for a single fault condition, is difficult to apply to multiple types of faults, and does not solve the problem that data of each client is not shared.
Regarding a fault diagnosis model construction method adopting federal learning, the invention patent 202110644538.6 provides a photovoltaic power station combined fault diagnosis method based on asynchronous decentralization federal learning, and the method updates the model by sending local model parameters to other participants, so that the communication and training efficiency of the model is improved to a certain extent, but the method requires continuous exchange of the model among the participants, has higher requirements on the server performance of the participants, does not reduce the time of local training of the model, and does not solve the problem of high transmission cost among the participants of federal learning.
Aiming at the problems, the patent provides a construction method of a planetary roller screw fault diagnosis model based on federal learning and a lightweight model, which can train the planetary roller screw fault diagnosis model by utilizing the data of each client while ensuring the data privacy, and can reduce the local training time of the model and the parameter transmission cost of the model.
Disclosure of Invention
In view of the above-mentioned drawbacks of the prior art, the present invention aims to solve the technical problems that a method for constructing a fault diagnosis model for a planetary roller screw is lacking in the present stage, data between clients is not shared, and federal learning transmission cost is high.
In order to achieve the above purpose, the invention provides a planetary roller screw fault diagnosis model construction method based on federal learning and lightweight model, which is characterized by comprising the following steps:
step 1, data acquisition;
further, the data collected in the step 1 comprise vibration data in three directions of X, Y, Z when the planetary roller screw is in normal state, lubrication fails and one side of the roller is in gear breakage operation, and the vibration sensor is arranged on a nut of the planetary roller screw.
Step 2, data preprocessing;
further, step 2 includes the steps of:
step 2.1, dividing original data into a training set and a testing set, and setting different types of labels;
step 2.2, randomly dividing the training set in the step 2.1 to form training sets of all clients (Host and Guest) in federal learning, wherein the training sets of the Host and Guest are not overlapped and the data quantity is different;
step 2.3, normalizing the training set and the testing set of the Host and the Guest clients described in step 2.2 according to formula (1) to obtain x;
wherein x is i For 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 normalized data, M is the minimum value of normalized data, and the data can be normalized to any range by using the formula (1), the invention considers the characteristic that vibration data has directivity when the planetary roller screw operates, and simultaneously normalizes the data to [ -1,1 in order to reduce the data span]I.e., M is 1, M is-1;
step 2.4, in order to increase the number of samples of the training set, data enhancement is performed on the normalized data in step 2.3 in a window clipping manner, and the number p of the obtained samples is shown in formula (2):
wherein n is the number of data points, w is the sampling signal length, and s is the sampling interval of data;
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 through formula (3) to obtain a detail coefficient W of the high-frequency signal i+1,2j And the approximation coefficient W of the low frequency signal i+1,2j+1 Then the approximation coefficients and detail coefficients of the last layer of each frequency band are arranged in rows to form a 64×64 coefficient matrix, and thenX, Y, Z the coefficient matrixes in three directions are stacked to obtain a matrix of 64 multiplied by 3, and the matrix is used as the input of the neural network;
wherein h (·) is a high-pass quadrature filter, g (·) is a low-pass quadrature filter, W i,j (k),k＝1,2,…,N/2 i For the ith layer, the wavelet packet coefficients at the jth child node, τ is the shift amount.
Step 3, because vibration data in X, Y, Z directions are not completely independent in the running process of the planetary roller screw in various states such as normal state, lubrication failure state, broken teeth at one side of the roller and the like, a certain relationship exists, but as a time sequence signal, signals at various time points have large differences, a traditional convolution layer is modified according to the data characteristics, namely a symmetrical convolution layer, so that kernels for multiplying and adding operation at various positions in space are different, a group of kernels with identical weights are used for different channels, the symmetrical convolution layer is taken as a main network layer, a lightweight neural network model SResNet18 shown in fig. 3 is built by replacing the convolution layer in the traditional 18-layer residual neural network with the symmetrical convolution layer, and the lightweight neural network model SResNet18 comprises the symmetrical convolution layer, batch Normalization, a ReLU activation function, a maximum pooling layer, the convolution layer, a symmetrical convolution residual layer 1, a symmetrical convolution residual layer 2, a global tie pooling layer, a full connection layer and a Softmax function;
(1) symmetrical convolution layer: as shown in fig. 4, the symmetric convolution layer has multiple kernels with different weights in space, performs multiply-add operation with inputs at different positions in space, can adaptively extract more spatial information in different spatial positions, and all channels share a set of kernels, reducing the number of parameters, so that the symmetric convolution layer can use kernels with larger size, thereby capturing long-distance features, and the calculation process on each channel is defined by equation (4):
wherein X is input, I is symmetric convolution kernel, all channels share a group of symmetric convolution kernels with the same weight, u and v are the symmetric convolution kernels and neighborhood offset of the central position of the input, K is the size of the symmetric convolution kernels, and C i Is the ith channel;
the symmetrical convolution kernels are multiple in space, and weights of the symmetrical convolution kernels at different positions are different, so that different symmetrical convolution kernels are required to be dynamically generated at different positions, the symmetrical convolution kernels are generated through a bottleneck layer shown in fig. 5, r is a parameter for determining the size of the bottleneck layer, and when the step size s of the movement of the symmetrical convolution kernels is greater than 1, the input height H is equal to the input height H 0 And width W 0 Will change and the size of the symmetric convolution kernel generated by only one bottleneck layer will not match the input, so it is necessary to reduce the input height H by one pooling layer 0 And width W 0 After passing through a pooling layer, height W 1 And width W 1 Respectively become H 1 ＝H 0 /s、W 1 ＝W 0 S, when the step size s of the symmetrical convolution kernel movement is equal to 1, the size of the symmetrical convolution kernel generated by only one bottleneck layer is matched with the input, so that the structure in the broken line of fig. 5 is not needed;
the symmetrical convolution layer is connected with the input C through a group of symmetrical convolution kernels 0 The multiplication and addition operations are respectively carried out on the channels to obtain a channel C 0 The number of channels outputted by the symmetrical convolution layer is the same as the number of channels inputted, so that a convolution layer with a convolution kernel size of 1 multiplied by 1 and a step length of 1 is added after the symmetrical convolution layer to change the number of channels;
(2) batch Normalization: batch Normalization can make the gradient descent algorithm more stable during gradient descent;
(3) ReLU activation function: in order to improve the fitting capability of the neural network, the output needs to be subjected to a de-linearization operation, and the ReLU activation function is selected as shown in a formula (5):
wherein X is an input;
(4) maximum pooling layer: the maximum pooling layer slides and traverses the whole input by a certain step length 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 over the input matrix, the calculation process being defined by equation (6):
wherein X is input, F is convolution kernel, u and v are convolution kernel and neighborhood offset of central position of input, K is convolution kernel size, C 1 The number of channels for output;
(6) symmetrical convolution residual layer: the symmetrical convolution residual layers avoid gradient explosion and gradient disappearance by adding identity mapping, when the step length s of the first symmetrical convolution layer is larger than 1, the height and width of the output Y and the input X of the symmetrical convolution residual layer 1 shown in fig. 6 are unequal, and addition operation cannot be performed, so that a convolution layer with the step length s and the size of 1 multiplied by 1 needs to be added to change the input size; when the step s of the first symmetrical convolution layer is equal to 1, the height and width of the output Y and the input X of the symmetrical convolution residual layer 2 shown in fig. 7 are equal, and the addition operation can be directly performed;
(7) global tie pooling layer: the global average pooling layer averages all inputs in the space direction, the height and the width of the output are 1, and the parameter quantity can be reduced to a great extent by using the global average pooling layer before the full connection layer;
(8) full connectivity layer vs Softmax function: the full connection layer and the Softmax cooperate to output a fault diagnosis classification result;
and 4, developing planetary roller screw joint fault diagnosis modeling based on the federal learning frame and the lightweight model. The invention aims to establish a planetary roller screw fault diagnosis model on the premise of ensuring data privacy, fully utilize planetary roller screw data of a plurality of clients, improve the performance of the model, and establish a lightweight model SResNet18 in step 3 aiming at the problem of model parameter transmission cost under a federal learning framework. It is worth noting that in the method of the invention, each client does not need to provide local data, only needs to upload the trained model parameters, and obtains the model with high accuracy and training speed by a method of data motionless model movement. In the process of jointly establishing a planetary roller screw fault diagnosis model by a plurality of clients, the method mainly comprises the following steps of:
step 4.1, the Host and the Guest train a local fault diagnosis model by utilizing a local training set;
step 4.2, uploading the obtained parameters of the local fault diagnosis model to the cloud by the Host and the Guest;
step 4.3, aggregating all models in the cloud to obtain a global model, wherein the parameters of the global model are P C The polymerization process is defined by formula (7):
wherein P is H And P G Parameters of the Host and Guest models respectively;
step 4.4, returning the parameters of the global model, and updating the parameters of the Host and Guest local fault diagnosis models to be P C ；
And continuously repeating the steps 4.1-4.4 until a model with excellent performance is obtained.
And 5, performing joint fault diagnosis modeling and light-weight model performance evaluation. In order to verify the effectiveness of federal learning to model the fault diagnosis of a planetary roller screw and the effectiveness of a lightweight model to improve the training speed under a federal learning framework, the invention utilizes collected data of the normal state, the lubrication failure state and the broken teeth at one side of the roller of the planetary roller screw to carry out fault diagnosis experiments, and in order to verify the effectiveness of the method, the invention respectively evaluates from three layers, and specifically comprises the following steps:
(1) the size and complexity of the model;
(2) the accuracy of the model in the test set;
(3) training time of model under federal learning framework.
Compared with the prior art, the invention has the beneficial effects that: the invention provides a processing method of the planetary roller screw fault diagnosis data for the first time, and provides a processing method of the planetary roller screw fault diagnosis data for the first time, wherein the planetary roller screw fault diagnosis data is diagnosed by adopting a federal learning method, the problem of insufficient planetary roller screw fault diagnosis data held by a single mechanism is solved on the premise of guaranteeing data privacy, the performance of a model is improved, and a lightweight model with small parameter quantity is provided according to the characteristics of vibration signals. Besides the planetary roller screw, the invention has certain promotion effect on the development and application of intelligent diagnosis technologies of other important mechanisms such as bearings, gear boxes and the like.
Drawings
Fig. 1 is a block diagram of a method for constructing a planetary roller screw fault diagnosis model based on federal learning and lightweight models.
Fig. 2 is a schematic diagram of wavelet packet transform coefficient matrix generation.
Fig. 3 is a structural diagram of SResNet18.
Fig. 4 is a schematic diagram of symmetric convolutional layer forward propagation.
Fig. 5 is a schematic diagram of a symmetric convolution kernel generation process.
Fig. 6 is a diagram of a symmetrical convolution residual layer 1 structure.
Fig. 7 is a diagram of a symmetrical convolution residual layer 2 structure.
Fig. 8 is a graph of the output results of res net18 in the test set and a confusion matrix thermodynamic diagram under the federal learning framework.
Fig. 9 is a graph of the output results of SResNet18 in the test set and confusion matrix thermodynamic diagram under the federal learning framework.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without any inventive effort, are intended to be within the scope of the invention.
Aiming at the problem that the failure diagnosis data among all clients are not shared due to lack of the construction method of the failure diagnosis model of the planetary roller screw, and the transmission cost is high under the federal learning framework, the invention provides the construction method of the failure diagnosis model of the planetary roller screw based on the federal learning and lightweight model shown in the figure 1. The invention firstly introduces a construction method of a planetary roller screw fault diagnosis data set. Next, a processing method for the planetary roller screw vibration data is described. Then, the invention builds the SResNet18 model for training the fault diagnosis model, which has less parameters and can reduce the transmission cost of the parameters under the federal learning framework. Then, federal learning fault diagnosis modeling of the planetary roller screw is carried out based on the invention. Finally, the invention carries out a fault diagnosis experiment to evaluate the size and complexity of a fault diagnosis model, the accuracy of the model in a test set and the training time of the model under a federal learning framework, and the method is used for verifying the effectiveness of the method and specifically comprises the following steps:
1. and (3) data acquisition: the invention provides a method for constructing a planetary roller screw fault diagnosis model through collaborative modeling of a plurality of clients. The specific parameters of the planetary roller screw used in the invention are as follows: the diameter of the screw rod is 24mm, the screw pitch is 2mm, the number of heads is 5, the number of rollers is 10, the specific working condition is that the screw rod rotating speed is 104r/min, the load is 9kN, and the vibration data of X, Y, Z directions under the normal state of the planetary roller screw rod, lubrication failure and broken teeth at one side of the rollers are collected, wherein the vibration sensor is arranged on a nut of the planetary roller screw rod.
2. Data preprocessing: in the invention, each client performs the same data preprocessing on the data.
First, the data is divided into training and testing sets.
And then, considering that the data of different data amounts held by all clients in an actual industrial scene are not overlapped, randomly dividing the training set to obtain the training set of Host and Guest.
Then, in order to reduce the data span, carrying out normalization processing on training sets and test sets of the Host and Guest clients, and mapping the data into the range of [ -1,1] according to a formula (8) in order to preserve the directionality of the data due to the directionality of the vibration data and reduce the span of the data;
wherein x is i For the current sample data value, M is the maximum value of the normalized data, 1, M is the minimum value of the normalized data, and-1.
Then, in order to improve the performance of the model, the data volume of the Host and Guest training sets is increased by a data enhancement method of window clipping, and the number p of samples after data enhancement is shown in a formula (9):
wherein n is the number of data points, w is the sampling signal length, and s is the sampling interval of data;
finally, since the vibration signal of the planetary roller screw is a non-stationary signal, wavelet packet transformation is performed by the formula (10), the selected wavelet function is "db1", the number of wavelet packet decomposition layers is 6, and the detail coefficient W of the high-frequency signal is obtained after the wavelet packet transformation i+1,2j And the approximation coefficient W of the low frequency signal i+1,2j+1 As shown in fig. 2, 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, and then the coefficient matrices in the three directions of X, Y, Z are stacked to obtain a 64×64×3 matrix, which is used as the input of the neural network;
wherein h (·) is a high-pass orthogonal filter, g (·) is a low-pass orthogonal filter coefficient, W i,j (k),k＝1,2,…,N/2 i For the ith layer, the wavelet packet coefficients at the jth child node, τ is the shift amount.
3. And (5) constructing a lightweight model. Because vibration data in X, Y, Z directions of planetary roller screws in various states such as normal states, lubrication failure states, broken teeth at one side of rollers and the like are not completely independent, a certain relationship exists, signals at each time point have large differences as time sequence signals, a traditional convolution layer is modified according to the data characteristics, the traditional convolution layer is called a symmetrical convolution layer, the kernels for multiplying and adding operation at each position in space are different, a group of kernels with identical weights are used for different channels, the symmetrical convolution layer is taken as a main network layer, a lightweight neural network model SResNet18 shown in fig. 3 is built by replacing the convolution layer in the traditional 18-layer residual neural network with the symmetrical convolution layer, and the lightweight neural network model SResNet18 comprises the symmetrical convolution layer, batch Normalization, a ReLU activation function, a maximum pooling layer, the convolution layer, a symmetrical convolution residual layer 1, a symmetrical convolution residual layer 2, a global tie pooling layer, a full connection layer and a Softmax function, and is specifically described below;
(1) symmetrical convolution layer: as shown in fig. 4, the symmetric convolution layer has multiple kernels with different weights in space, performs multiply-add operation with inputs at different positions in space, can adaptively extract more spatial information in different spatial positions, and all channels share a set of kernels, reducing the number of parameters, so that the symmetric convolution layer can use kernels with larger size, thereby capturing long-distance features, and the calculation process on each channel is defined by equation (11):
wherein X is input, I is symmetric convolution kernel, and all channels share a group of symmetric with the same weightThe convolution kernel, u and v are the symmetrical convolution kernel and the neighborhood offset of the input center position, K is the symmetrical convolution kernel size, C i Is the ith channel;
the symmetrical convolution kernels are multiple in space, and the weights of the symmetrical convolution kernels at different positions are different, so that different symmetrical convolution kernels are required to be dynamically generated at different positions, the symmetrical convolution kernels are generated through a bottleneck layer shown in fig. 5, r is a parameter for determining the size of the bottleneck layer, and when the step size s of the movement of the symmetrical convolution kernels is greater than 1, the input height H is equal to the input height H 0 And width W 0 Will change and the size of the symmetric convolution kernel generated by only one bottleneck layer will not match the input, so it is necessary to reduce the input height H by one pooling layer 0 And width W 0 Height W after passing through a pooling layer 1 And width W 1 Respectively become H 1 ＝H 0 /s、W 1 ＝W 0 S, when the step size s of the symmetrical convolution kernel movement is equal to 1, the size of the symmetrical convolution kernel generated by only one bottleneck layer is matched with the input, so that the structure in the broken line of fig. 5 is not needed;
the symmetrical convolution layer is connected with the input C through a group of symmetrical convolution kernels 0 The multiplication and addition operations are respectively carried out on the channels to obtain a channel C 0 The number of channels outputted by the symmetrical convolution layer is the same as the number of channels inputted, so that a convolution layer with a convolution kernel size of 1 multiplied by 1 and a step length of 1 is added after the symmetrical convolution layer to change the number of channels;
(2) batch Normalization: batch Normalization can make the gradient descent algorithm more stable during gradient descent;
(3) ReLU activation function: to improve the fitting ability of the neural network, the output needs to be subjected to a de-linearization operation, and the ReLU activation function is selected as shown in a formula (12):
wherein X is an input;
(4) maximum pooling layer: the maximum pooling layer slides and traverses the whole input by a certain step length and outputs the maximum value of the data in the window;
(5) convolution layer: the convolution layer computes the dot product of the convolution kernel and the input at each location by sliding the convolution kernel over the input matrix, the computation being defined by equation (13):
wherein X is input, F is convolution kernel, u and v are convolution kernel and neighborhood offset of central position of input, K is convolution kernel size, C 1 The number of channels for output;
(6) symmetrical convolution residual layer: the symmetrical convolution residual layers avoid gradient explosion and gradient disappearance by adding identity mapping, when the step length s of the first symmetrical convolution layer is larger than 1, the height and width of the output Y and the input X of the symmetrical convolution residual layer 1 shown in fig. 6 are unequal, and addition operation cannot be performed, so that a convolution layer with the step length s and the size of 1 multiplied by 1 needs to be added to change the input size; when the step length of the first symmetrical convolution layer is equal to 1, the height and width of the output Y and the input X of the symmetrical convolution residual layer 2 shown in fig. 7 are equal, and the addition operation can be directly performed;
(7) global tie pooling layer: the global average pooling layer averages all inputs in the space direction, the height and the width of the output are 1, and the parameter quantity can be reduced to a great extent by using the global average pooling layer before the full connection layer;
(8) full connectivity layer vs Softmax function: the full connection layer and the Softmax cooperate to output a fault diagnosis classification result;
4. planetary roller screw joint fault diagnosis modeling is carried out based on a federal learning framework and a lightweight model: in consideration of the defect of insufficient fault diagnosis data of a planetary roller screw held by a single mechanism at the present stage, but the data of all mechanisms are not shared, and the transmission cost is high under a federal learning framework, the invention provides a construction method of a planetary roller screw fault diagnosis model based on federal learning and a lightweight model to solve the problems, and a kubrerate-docker-compound-v 1.6.0 framework is utilized to realize rapid joint fault diagnosis modeling of a plurality of clients.
Firstly, host and Guest train a local fault diagnosis model by using a local training set;
then, the Host and the Guest upload the obtained parameters of the local fault diagnosis model to the cloud;
then, all the models are aggregated in the cloud to obtain a global model, and the parameters of the global model are P C The polymerization process is defined by formula (14):
wherein P is H And P G Parameters of the Host and Guest models respectively;
then, the global model parameters are returned, and the parameters of the Host and Guest local fault diagnosis models are updated to be P C ；
And finally, continuously repeating the steps until the global model converges to a higher accuracy rate, and completing the joint fault diagnosis modeling of a plurality of clients.
5. And (3) joint fault diagnosis modeling and light-weight model performance evaluation: in order to verify the effectiveness of federal learning to model the fault diagnosis of a planetary roller screw and the effectiveness of a lightweight model to improve the training speed under a federal learning framework, the invention utilizes collected data of the normal state, the lubrication failure state and the broken teeth at one side of the roller of the planetary roller screw to carry out fault diagnosis experiments, and in order to verify the effectiveness of the method, the invention respectively evaluates from three layers, and specifically comprises the following steps:
(1) the size and complexity of the model;
(2) the accuracy of the model in the test set;
(3) training time of model under federal learning framework.
(1) Introduction to data set
A detailed description of the resulting dataset is shown in table 1.
Table 1 dataset
(2) Design of experiment
In order to verify the effectiveness of the method proposed by the invention, a comparative experiment was set up: a. a traditional 18-layer residual neural network model ResNet18, and b. The lightweight model SResNet18 provided by the invention. Compared with ResNet18, the 4-layer convolution kernel size of 1×1 and step size of 1 is added to SResNet18 to change the input channel. Specific experimental comparisons include:
(1) to compare the size and complexity of the models, the parameter and floating point operands of the ResNet18 and SResNet18 models are compared;
(2) in order to verify the effectiveness of federal learning in modeling the fault diagnosis of the planetary roller screw, the accuracy of the ResNet18 and SResNet18 local training models and the federal learning models in a test set are compared;
(3) to verify the effectiveness of the lightweight model for the increase in training speed under the federal learning framework, the training time of the ResNet18 and SResNet18 models under the federal learning framework was compared.
(3) Parameter setting
The specific network parameters of the experiments of the present invention are shown in table 2.
Table 2 model parameter settings
| Model | Network layer number | Learning rate | Batch size | Number of training wheels | Optimizer |
| ResNet18 | 18 | 0.01 | 64 | 100 | Adam |
| SResNet18 | 22 | 0.01 | 64 | 100 | Adam |
(4) Analysis of experimental results
The parameter number and floating point operand of each model are shown in table 3, wherein the parameter number of the model represents the number of parameters which need to be updated continuously, the size of the model can be measured, the calculation amount of the model is represented by the floating point operand, and the complexity of the model can be measured.
TABLE 3 parameter and floating point operands for each model
| Model | Quantity of parameters | Floating point operand |
| ResNet18 | 11.2M | 0.297G |
| SResNet18 | 0.552M | 0.0235G |
Except that SResNet18 is added with 4 layers of convolution layers for changing the channel number, the network depth of the ResNet18 and the network depth of the SResNet18 are identical, and as can be seen from table 3, the reference number and floating point operand of the SResNet18 are reduced by 95.07% and 92.09% respectively compared with the ResNet18, so that the SResNet18 provided by the invention has great advantages compared with the traditional ResNet18 regardless of the size of the model or the complexity of the model.
Because the training of the model is initiated by the Guest under the kubuefate federate learning framework, the local training model refers to a model trained by a Guest training set, and because the Host and the Guest participate in the training under the federate learning framework, the accuracy of the training set is respectively the accuracy of the Host and the Guest training sets, and the accuracy of each model in the test set is shown in table 4.
Table 4 accuracy of each model in test set
As can be seen from Table 4, during local training, although the accuracy of the conventional ResNet18 training set is 99.5%, the accuracy of the test set is only 67.5%, which indicates that the ResNet18 is subjected to severe overfitting during the local training, but the difference between the accuracy of the SResNet18 test set and the accuracy of the training set is smaller, which indicates that the lightweight model SResNet18 provided by the invention can reduce the overfitting phenomenon; when the model is trained in federal learning, the accuracy of ResNet18 in a Guest training set is only 37.1%, the accuracy of the test set is only 36.3%, and the accuracy is lower than that in local training, because the ResNet18 parameters are more, more training rounds are needed to obtain a ResNet18 model with high accuracy, but setting a larger training round number can increase time cost and operation cost, so that unnecessary waste is caused, conversely, the accuracy of SResNet18 in the Guest training set is 99.4%, the accuracy of the test set is up to 99.2%, and the accuracy of the test set is higher than that in local training, and almost no overfitting phenomenon exists. By combining the above, the SResNet18 has the best performance under the federal learning framework, the highest accuracy of the test set, and the local training of the SResNet18 and the accuracy of the test set under the federal learning framework can be seen, the federal learning can alleviate the over-fitting phenomenon, and the performance of the model is improved.
The training time of each model under the federal learning framework is shown in table 5.
Table 5 training time of each model under federal learning framework
| Model | Model training time |
| ResNet18 | 9039s |
| SResNet18 | 3551s |
From table 5, it can be seen that under the federal learning framework, the modeling time of SResNet18 is reduced by 60.71% compared with that of ResNet18, which indicates that the lightweight SResNet18 model provided by the invention can significantly improve the training speed of the model and reduce the transmission cost under the federal learning framework.
Claims (2)
1. A planetary roller screw fault diagnosis model construction method based on federal learning and lightweight models is characterized by comprising the following steps:
step 1, data acquisition;
the collected data comprise vibration data in three directions of normal, lubrication failure and X, Y, Z when the planetary roller screw runs in a gear-breaking state at one side of the roller, and the vibration sensor is arranged on a nut of the planetary roller screw;
step 2, data preprocessing; the method specifically comprises the following substeps:
step 2.1, dividing original data into a training set and a testing set, and setting different types of labels;
step 2.2, randomly dividing the training set in the step 2.1 to form training sets of all clients (Host and Guest) in federal learning, wherein the training sets of the Host and Guest are not overlapped and the data quantity is different;
step 2.3, normalizing the training set and the testing set of the Host and the Guest clients described in step 2.2 according to formula (1) to obtain x;
wherein x is i For 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 normalized data, M is the minimum value of normalized data, and the data can be normalized to any range by using the formula (1), the invention considers the characteristic that vibration data has directivity when the planetary roller screw operates, and simultaneously normalizes the data to [ -1,1 in order to reduce the data span]I.e., M is 1, M is-1;
step 2.4, in order to increase the number of samples of the training set, data enhancement is performed on the normalized data in step 2.3 in a window clipping manner, and the number p of the obtained samples is shown in formula (2):
wherein n is the number of data points, w is the sampling signal length, and s is the sampling interval of data;
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 through formula (3) to obtain a detail coefficient W of the high-frequency signal i+1,2j And the approximation coefficient W of the low frequency signal i+1,2j+1 Then the approximate coefficient and the detail coefficient of each frequency band of the last layer are arranged in rows to form a 64×64 coefficient matrix, and then the coefficient matrices in the X, Y, Z directions are stacked to obtain a 64×64×3 matrix which is used as the input of the neural network;
wherein h (·) is a high-pass quadrature filter, g (·) is a low-pass quadrature filter, W i,j (k),k＝1,2,…,N/2 i For the ith layer, the wavelet packet coefficients at the jth child node, τ is the shift amount.
Step 3, modifying a traditional convolution layer, namely a symmetrical convolution layer, so that kernels for performing multiply-add operation at each position in space are different, but different channels use a group of kernels with identical weights, the symmetrical convolution layer is used as a main network layer, and a lightweight neural network model SReNet 18 is built by replacing the convolution layer in the traditional 18-layer residual neural network with the symmetrical convolution layer, wherein the lightweight neural network model SReNet comprises the symmetrical convolution layer, batch Normalization, a ReLU activation function, a maximum pooling layer, a convolution layer, a symmetrical convolution residual layer 1, a symmetrical convolution residual layer 2, a global tie pooling layer, a full connection layer and a Softmax function:
(1) symmetrical convolution layer: the symmetrical convolution layer has a plurality of kernels with different weights in space, performs multiply-add operation with inputs at different positions in space, all channels share a set of kernels, and the computation process on each channel is defined by equation (4):
wherein,,x is input, I is a symmetrical convolution kernel, all channels share a group of symmetrical convolution kernels with the same weight, u and v are the symmetrical convolution kernels and neighborhood offset of the central position of the input, K is the size of the symmetrical convolution kernels, and C i Is the ith channel;
the symmetrical convolution kernels are multiple in space, the weights of the symmetrical convolution kernels at different positions are different, the symmetrical convolution kernels are generated through a bottleneck layer, and the height H of the input is reduced through a pooling layer 0 And width W 0 After passing through a pooling layer, height W 1 And width W 1 Respectively become H 1 ＝H 0 /s、W 1 ＝W 0 S, when the step size s of the symmetrical convolution kernel movement is equal to 1, the size of the symmetrical convolution kernel generated by only one bottleneck layer is matched with the input;
the symmetrical convolution layer is connected with the input C through a group of symmetrical convolution kernels 0 The multiplication and addition operations are respectively carried out on the channels to obtain a channel C 0 The number of channels outputted by the symmetrical convolution layer is the same as the number of channels inputted, so that a convolution layer with a convolution kernel size of 1 multiplied by 1 and a step length of 1 is added after the symmetrical convolution layer to change the number of channels;
(2) batch Normalization: batch Normalization can make the gradient descent algorithm more stable during gradient descent;
(3) ReLU activation function: as shown in formula (5):
wherein X is an input;
(4) maximum pooling layer: the maximum pooling layer slides and traverses the whole input by a certain step length 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 over the input matrix, the calculation process being defined by equation (6):
wherein X is input, F is convolution kernel, u and v are convolution kernel and neighborhood offset of central position of input, K is convolution kernel size, C 1 The number of channels for output;
(6) symmetrical convolution residual layer: the symmetrical convolution residual layers avoid gradient explosion and gradient disappearance by adding identity mapping, and when the step length s of the first symmetrical convolution layer is larger than 1, one convolution layer with the step length s and the size of 1 multiplied by 1 is added to change the input size; when the step length s of the first symmetrical convolution layer is equal to 1, the height and width of the output Y and the input X of the symmetrical convolution residual layer 2 are equal, and the addition operation can be directly performed;
(7) global tie pooling layer: the global average pooling layer averages all inputs in the space direction, the height and the width of the output are 1, and the parameter quantity can be reduced to a great extent by using the global average pooling layer before the full connection layer;
(8) full connectivity layer vs Softmax function: the full connection layer and the Softmax cooperate to output a fault diagnosis classification result;
and 4, developing planetary roller screw joint fault diagnosis modeling based on the federal learning frame and the lightweight model. The invention aims to establish a planetary roller screw fault diagnosis model on the premise of ensuring data privacy, fully utilize planetary roller screw data of a plurality of clients, improve the performance of the model, and establish a lightweight model SResNet18 in step 3 aiming at the problem of model parameter transmission cost under a federal learning frame; in the process of jointly establishing a planetary roller screw fault diagnosis model by a plurality of clients, the method mainly comprises the following steps of:
step 4.1, the Host and the Guest train a local fault diagnosis model by utilizing a local training set;
step 4.2, uploading the obtained parameters of the local fault diagnosis model to the cloud by the Host and the Guest;
step 4.3, aggregating all models in the cloud to obtain a global model, wherein the parameters of the global model are P C The polymerization process is defined by formula (7):
wherein P is H And P G Parameters of the Host and Guest models respectively;
step 4.4, returning the parameters of the global model, and updating the parameters of the Host and Guest local fault diagnosis models to be P C ；
And continuously repeating the steps 4.1-4.4 until a model with excellent performance is obtained.
And 5, performing joint fault diagnosis modeling and light-weight model performance evaluation.
2. The method for constructing a planetary roller screw fault diagnosis model based on federal learning and lightweight model according to claim 1, wherein in the step 5, when performing fault diagnosis modeling and lightweight model performance evaluation, fault diagnosis experiments are performed by using collected data of a normal state, a lubrication failure state and a broken tooth at one side of a roller of the planetary roller screw, and the method comprises the following steps:
(1) the size and complexity of the model;
(2) the accuracy of the model in the test set;
(3) training time of model under federal learning framework.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310481747.2A CN116561684A (en) | 2023-04-28 | 2023-04-28 | Planetary roller screw fault diagnosis model construction method based on federal learning and lightweight model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310481747.2A CN116561684A (en) | 2023-04-28 | 2023-04-28 | Planetary roller screw fault diagnosis model construction method based on federal learning and lightweight model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116561684A true CN116561684A (en) | 2023-08-08 |
Family
ID=87497554
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310481747.2A Pending CN116561684A (en) | 2023-04-28 | 2023-04-28 | Planetary roller screw fault diagnosis model construction method based on federal learning and lightweight model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116561684A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117589444A (en) * | 2024-01-18 | 2024-02-23 | 湖南科技大学 | Wind driven generator gear box fault diagnosis method based on federal learning |
-
2023
- 2023-04-28 CN CN202310481747.2A patent/CN116561684A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117589444A (en) * | 2024-01-18 | 2024-02-23 | 湖南科技大学 | Wind driven generator gear box fault diagnosis method based on federal learning |
| CN117589444B (en) * | 2024-01-18 | 2024-04-02 | 湖南科技大学 | Wind driven generator gear box fault diagnosis method based on federal learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108197648B (en) | Hydroelectric generating set fault diagnosis method and system based on LSTM deep learning model | |
| CN111369563B (en) | Semantic segmentation method based on pyramid void convolutional network | |
| CN110263925B (en) | Hardware acceleration implementation device for convolutional neural network forward prediction based on FPGA | |
| CN108334948B (en) | Mechanical bearing fault diagnosis technology based on wide residual error network learning model | |
| CN109919108A (en) | Remote sensing images fast target detection method based on depth Hash auxiliary network | |
| CN104714925B (en) | A kind of gear transmission noises analysis method based on Fourier Transform of Fractional Order and SVMs | |
| CN113723491A (en) | Rolling bearing with semi-supervised model contrast migration and gear fault diagnosis method | |
| CN111458148A (en) | CBAM-based convolutional neural network rolling bearing fault diagnosis method | |
| CN116561684A (en) | Planetary roller screw fault diagnosis model construction method based on federal learning and lightweight model | |
| CN114912705A (en) | Optimization method for heterogeneous model fusion in federated learning | |
| Miao et al. | Deep sparse representation network for feature learning of vibration signals and its application in gearbox fault diagnosis | |
| CN114091667A (en) | Federal mutual learning model training method oriented to non-independent same distribution data | |
| CN112651426A (en) | Fault diagnosis method for rolling bearing of wind turbine generator | |
| CN110879927A (en) | Sea clutter amplitude statistical distribution field modeling method for sea target detection | |
| CN114491828A (en) | Compound fault diagnosis method based on multi-gating hybrid expert multi-task learning | |
| CN103698809A (en) | Kirchhoff pre-stack time migration parallel method with no speed-up ratio bottleneck | |
| CN114972232A (en) | No-reference image quality evaluation method based on incremental meta-learning | |
| CN117633448A (en) | Cloud-edge collaboration-based equipment cluster health monitoring method | |
| CN112270259B (en) | SAR image ship target rapid detection method based on lightweight convolutional neural network | |
| CN117171544B (en) | Motor vibration fault diagnosis method based on multichannel fusion convolutional neural network | |
| CN114186602A (en) | Sparse filtering domain-based mechanical fault diagnosis method for anti-neural network | |
| CN117312948A (en) | Rotary machine fault diagnosis method based on multi-sensor fusion and ECA-CNN | |
| CN116754231A (en) | Method for rapidly diagnosing faults of rolling bearing based on RegNet-SES | |
| CN111461184A (en) | XGB multi-dimensional operation and maintenance data anomaly detection method based on multivariate feature matrix | |
| CN114781427A (en) | Wind generating set rolling bearing fault diagnosis system based on information fusion technology |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |