CN110657984B - Planetary gearbox fault diagnosis method based on reinforced capsule network - Google Patents

Planetary gearbox fault diagnosis method based on reinforced capsule network Download PDF

Info

Publication number
CN110657984B
CN110657984B CN201910925753.6A CN201910925753A CN110657984B CN 110657984 B CN110657984 B CN 110657984B CN 201910925753 A CN201910925753 A CN 201910925753A CN 110657984 B CN110657984 B CN 110657984B
Authority
CN
China
Prior art keywords
layer
capsule
convolution
convolution layer
network
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.)
Active
Application number
CN201910925753.6A
Other languages
Chinese (zh)
Other versions
CN110657984A (en
Inventor
李琪康
汤宝平
余晓霞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing University
Original Assignee
Chongqing University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Chongqing University filed Critical Chongqing University
Priority to CN201910925753.6A priority Critical patent/CN110657984B/en
Publication of CN110657984A publication Critical patent/CN110657984A/en
Application granted granted Critical
Publication of CN110657984B publication Critical patent/CN110657984B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/02—Gearings; Transmission mechanisms
    • G01M13/021—Gearings
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/02—Gearings; Transmission mechanisms
    • G01M13/025—Test-benches with rotational drive means and loading means; Load or drive simulation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention discloses a fault diagnosis method for a planetary gearbox based on an enhanced capsule network, which comprises the following steps of: s1, acquiring data to be diagnosed; s2, inputting data to be diagnosed into a reinforced capsule network, wherein the reinforced capsule network comprises a convolution layer, a Primary capsule layer, a Digit capsule layer and a decoding layer, the convolution layer comprises a first convolution layer and a second convolution layer, the void factor of the first convolution layer is larger than 1, and the void factor of the second convolution layer is equal to 1; and S3, outputting a fault diagnosis result by the reinforced capsule network. Compared with the prior art, the reinforced capsule network firstly adopts the cavity convolution (the first convolution layer) to extract the characteristics, and increases the sensing visual field of the convolution kernel on the premise of ensuring that the parameter quantity is not changed so as to enhance the nonlinear capability of the reinforced capsule network, thereby improving the information extraction and characteristic learning capability in the fault diagnosis of the planetary gearbox.

Description

Planetary gearbox fault diagnosis method based on reinforced capsule network
Technical Field
The invention relates to the technical field of simulation analysis, in particular to a fault diagnosis method for a planetary gearbox based on an enhanced capsule network.
Background
The planetary gear box has the advantages of large transmission ratio, high transmission efficiency and the like, and is widely applied to the heavy industries of aerospace, automobiles and the like. However, the planetary gear box has a complex structure, is easy to break down when working under the severe working conditions of variable rotating speed and variable load for a long time, and the health state of the planetary gear box directly influences the personal safety and the economic benefit of enterprises. The development of fault diagnosis research of the planetary gearbox is of great significance.
Deep learning is a research hotspot in the field of artificial intelligence at present, and a deep network structure of the deep learning is highly nonlinear and has high adaptive feature learning capability, so that the uncertainty of manually extracted features is avoided, and the deep learning is applied to fault diagnosis of the planetary gear box. Researchers carry out collective empirical mode decomposition on the vibration data, and then a Deep Belief Network (DBN) is combined to be applied to fault diagnosis of the planetary gearbox; researchers also obtain frequency domain change characterization information by extracting the envelope of the low-level frequency domain signal, and the frequency domain change characterization information is combined with a Deep Auto-Encoder (DAE) to carry out fault diagnosis on the gearbox; and researchers acquire time-frequency characteristics of the vibration signals through wavelet packet transformation, and then perform fault diagnosis on the planetary gearbox by taking the time-frequency characteristics as input of CNN (volumetric Neural network). However, the DBN and DAE are usually fully connected, the training parameters are more, and the model is not easy to train; the CNN network has a weight sharing characteristic by adopting a local connection mode, network parameters can be greatly reduced through a pooling layer structure, the network training efficiency is effectively improved, and the defects of DBN and DAE are overcome to a certain extent.
A Capsule Network (CapsNet) is a new type of supervised learning Network. Different from the traditional CNN, the Capsule network adopts a vector to replace a scalar to represent the characteristics, updates the capsule layer parameters through a dynamic routing mechanism, further increases the coupling coefficient of the child node and the father node, and fully utilizes local position information to enrich the characteristic representation capability and the information content; and the CapsNet structure has translation invariance, can extract the relative position relation of the input features of the planetary gear box, and is favorable for improving the identification accuracy. Based on the advantages, researchers can obtain time-frequency characteristics as input of the CapsNet network to diagnose bearing faults by performing short-time Fourier transform processing on the original vibration signals. However, in the network mapping process, the CapsNet adopts a method of dividing scalar feature into vector features discretely, so that the correlation information between adjacent feature layers cannot be stored, and meanwhile, the view range of the convolutional layer of the CapsNet is limited, so that the fault feature extraction capability of the CapsNet is further inhibited.
Therefore, how to improve the information extraction and feature learning capability in the fault diagnosis of the planetary gearbox becomes a problem which needs to be solved urgently by the technical personnel in the field.
Disclosure of Invention
In summary, the practical problems to be solved by the present invention are: how to improve information extraction and characteristic learning ability in the fault diagnosis of the planetary gearbox.
In order to solve the problems, the invention adopts the following technical scheme:
a fault diagnosis method for a planetary gearbox based on an enhanced capsule network comprises the following steps:
s1, acquiring data to be diagnosed;
s2, inputting data to be diagnosed into a reinforced capsule network, wherein the reinforced capsule network comprises a convolution layer, a Primary capsule layer, a Digit capsule layer and a decoding layer, the convolution layer comprises a first convolution layer and a second convolution layer, the void factor of the first convolution layer is larger than 1, and the void factor of the second convolution layer is equal to 1;
and S3, outputting a fault diagnosis result by the reinforced capsule network.
Preferably, step S1 includes:
s101, acquiring original data;
s102, decomposing original data by adopting a wavelet packet and constructing a time-frequency coefficient matrix;
and S103, taking the time-frequency coefficient matrix as data to be diagnosed.
Preferably, the convolution kernel size calculation expression of the first hole convolution layer is:
N＝k+(k-1)×(d-1)
where N denotes the convolution kernel size of the first convolution layer, k denotes the convolution kernel size of the second convolution layer, and d denotes the void factor.
Preferably, the method further comprises the following steps: introducing an overlap coefficient theta into a Primary capsule layer, and reusing a part of scalar characteristic layer in the process of carrying out discrete segmentation on the original scalar characteristic;
the calculation expression of the Primary capsule layer after the overlap coefficient theta is introduced is as follows:
Figure BDA0002218903870000021
Figure BDA0002218903870000022
in the formula (I), the compound is shown in the specification,
Figure BDA0002218903870000023
i capsule as the l capsule layer, MkIs the layerK-th mapping region of (1), uijA vector matrix, w, for mapping the jth feature in the upper convolutional feature layer to the ith capsule regionijConvolution kernel matrix of corresponding eigenvector for the mapping process, bijFor the corresponding added bias matrix, N represents the number of convolution feature layer layers, Nc represents the capsule length, θ represents the overlap factor, and Z represents the total number of capsule layers.
Preferably, the value range of theta is a discrete value sequence,
Figure BDA0002218903870000031
Figure BDA0002218903870000032
wherein L is an integer value between 0 and L, L representing the capsule length.
In summary, the invention discloses a fault diagnosis method for a planetary gearbox based on an enhanced capsule network, which comprises the following steps: s1, acquiring data to be diagnosed; s2, inputting data to be diagnosed into a reinforced capsule network, wherein the reinforced capsule network comprises a convolution layer, a Primary capsule layer, a Digit capsule layer and a decoding layer, the convolution layer comprises a first convolution layer and a second convolution layer, the void factor of the first convolution layer is larger than 1, and the void factor of the second convolution layer is equal to 1; and S3, outputting a fault diagnosis result by the reinforced capsule network. Compared with the prior art, the reinforced capsule network firstly adopts the cavity convolution (the first convolution layer) to extract the characteristics, and increases the sensing visual field of the convolution kernel on the premise of ensuring that the parameter quantity is not changed so as to enhance the nonlinear capability of the reinforced capsule network, thereby improving the information extraction and characteristic learning capability in the fault diagnosis of the planetary gearbox.
Drawings
For purposes of promoting a better understanding of the objects, aspects and advantages of the invention, reference will now be made in detail to the present invention as illustrated in the accompanying drawings, in which:
FIG. 1 is a flow chart of an embodiment of a planetary gearbox fault diagnosis method based on an enhanced capsule network disclosed by the invention;
FIG. 2 is a schematic diagram of a conventional capsule network;
FIG. 3 is a schematic view of a second convolutional layer;
FIG. 4 is a schematic view of a first winding layer;
FIG. 5 is a diagram of an information enhancement structure incorporating an overlap factor θ;
FIG. 6 is a schematic illustration of a noisy vibration signal;
FIG. 7 is a graph showing a comparison of test accuracies for different overlay coefficients;
FIG. 8 is a graphical illustration of a comparison of test accuracy for different methods;
FIG. 9 is a schematic diagram of a sample test result confusion matrix;
FIG. 10 is a schematic view of capsule layer input visualization;
FIG. 11 is a schematic view of capsule layer output visualization.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings.
As shown in FIG. 1, the invention discloses a fault diagnosis method for a planetary gearbox based on an enhanced capsule network, which comprises the following steps:
s1, acquiring data to be diagnosed;
s2, inputting data to be diagnosed into a reinforced capsule network, wherein the reinforced capsule network comprises a convolution layer, a Primary capsule layer, a Digit capsule layer and a decoding layer, the convolution layer comprises a first convolution layer and a second convolution layer, the void factor of the first convolution layer is larger than 1, and the void factor of the second convolution layer is equal to 1;
and S3, outputting a fault diagnosis result by the reinforced capsule network.
The capsNet is a novel neural network model, and the main structure of the capsNet comprises four parts: a convolution layer, a Primary capsule layer, a Digit capsule layer and a decoding layer. Wherein the convolutional layer adaptively extracts high-dimensional signal features layer by layer through convolutional kernels with different set sizes; the Primary capsule layer maps the high-dimensional abstract features extracted by the convolutional layer into capsules; the Digit capsule layer calculates the upper layer capsule through dynamic routing to obtain a final output capsule, and the output is respectively used as a classification part and a decoding part; the decoding layer reconstructs the output capsule into an input signal through the multiple fully connected layers. The structure of the CapsNet is shown in FIG. 2.
In the CapsNet structure, scalar neuron features extracted from the convolutional layer are mapped to vector neuron (capsule) features, and final output is obtained through capsule level calculation. Mapping capsule UiFirst pass through a transformation matrix WijCalculating to obtain the intermediate capsule uj|iAnd then by the coupling coefficient cijWeighted computing output capsule SjThe specific calculation expression is as follows:
uj|i＝Wij·Ui
Figure BDA0002218903870000041
because input and output of the capsNet capsule layer are vectors, a squaring nonlinear function is adopted as an activation function in the capsule calculation process, so that the length of an output vector is compressed between 0 and 1, and the function expression is as follows:
Figure BDA0002218903870000042
in the formula SjRepresenting the weighted sum of the outputs from the upper capsule, VjDenotes the V thjOutput vector of layer capsule.
In CapsNet, the network parameters of the capsule layer are calculated by a Dynamic routing (Dynamic routing) mechanism. The dynamic routing mechanism is through the intermediate capsule uj|iAnd the output vector VjThe parameters are updated and only three iterations are needed to obtain the better network parameters. Setting an intermediate parameter b during the first iteration calculation ij0, calculating the initial weight c by a softmax functionijThen according to the intermediate variable uj|iAnd VjTo update the intermediate parameter bij. The calculation expression is as follows:
Figure BDA0002218903870000051
bij＝bij+uj|i·Vj
the loss calculation of the CapsNet is divided into two parts of classification error and reconstruction error. Since the capsule layer data in the network structure is calculated in the form of vectors, the edge loss (margin loss) is adopted as a loss function in the process of back propagation of classification errors, and the mathematical expression is as follows:
Lk＝Tk·max(0,m+-||vk||)2+λ(1-Tk)max(0,||vk||-m-)2
wherein k is a classification category; t iskA label for classification (presence is 1, absence is 0); m is+An upper bound for penalizing false positives; m is-A lower boundary for penalizing false negatives; λ is the penalty proportionality coefficient for two errors, LkRepresenting the edge loss function, vkRepresenting the predicted output value of the capsule network.
The method is characterized in that a large convolution kernel is used for feature extraction, wide feature information is obtained in a mode of large perception visual field, feature size is reduced, follow-up calculation consumption is reduced, a hole convolution (a first convolution layer) is used for increasing the perception visual field on the premise of keeping the size of an original convolution kernel (a second convolution layer), and meanwhile, a double convolution feature extraction layer combining hole convolution and common convolution is used for enhancing the nonlinear capacity of a model. As shown in fig. 3 and 4, the circle is the convolution kernel and the shaded box is the convolution kernel perception field of view.
In specific implementation, step S1 includes:
s101, acquiring original data;
s102, decomposing original data by adopting a wavelet packet and constructing a time-frequency coefficient matrix;
and S103, taking the time-frequency coefficient matrix as data to be diagnosed.
In order to avoid the defect of single time domain or frequency domain signal fault characterization capability, the invention adopts a sparse self-coding algorithm based on a spectrum envelope curve and an application [ J ] vibration and impact in gearbox fault diagnosis, 2018,37(04):249-256 preprocessing method to perform wavelet packet transformation on original vibration data and extract the time-frequency characteristics of signals. And performing 6-layer decomposition on the original data by adopting a DB1 wavelet packet, and constructing a time-frequency coefficient matrix with the size of 64 multiplied by 64 to be used as the input of the reinforced capsule network.
In specific implementation, the convolution kernel size calculation expression of the first void convolution layer is as follows:
N＝k+(k-1)×(d-1)
where N denotes the convolution kernel size of the first convolution layer, k denotes the convolution kernel size of the second convolution layer, and d denotes the void factor.
When the concrete implementation, still include: introducing an overlap coefficient theta into a Primary capsule layer, and reusing a part of scalar characteristic layer in the process of carrying out discrete segmentation on the original scalar characteristic;
the calculation expression of the Primary capsule layer after the overlap coefficient theta is introduced is as follows:
Figure BDA0002218903870000061
Figure BDA0002218903870000062
in the formula (I), the compound is shown in the specification,
Figure BDA0002218903870000063
i capsule as the l capsule layer, MkFor the k-th mapped region of the layer, uijA vector matrix, w, for mapping the jth feature in the upper convolutional feature layer to the ith capsule regionijConvolution kernel matrix of corresponding eigenvector for the mapping process, bijFor the corresponding added bias matrix, N represents the number of convolution feature layer layers, Nc represents the capsule length, θ represents the overlap factor, and Z represents the total number of capsule layers.
After the features are extracted by the convolutional layers, scalar features are mapped into vector features, so that the extracted features have richer information, but the discrete segmentation mode adopted by the CapsNet in the capsule mapping stage, namely all convolutional feature layers are averagely divided into a plurality of capsule layers, so that the original relevance between the edge feature layers of each capsule is lost. Therefore, in order to fully utilize the correlation information between scalar feature layers, in the enhanced capsule network of the present invention, an overlap coefficient θ can be introduced into the Primary capsule layer, and in the process of performing discrete segmentation on the original scalar features, part of the scalar feature layer is reused, so as to enhance the information amount of the extracted capsule features while retaining the scalar feature layer correlation information, and the information enhancement structure is shown in fig. 5. After the overlapping coefficient is introduced, the number of layers of the Primary capsules is increased compared with the original number.
In practice, the value range of theta is a discrete value sequence, and the correlation system is determined according to the capsule length L
Figure BDA0002218903870000064
Figure BDA0002218903870000065
In the formula, L takes an integer value between 0 and L.
The size of the overlap factor has a significant impact on the network structure. If the overlapping coefficient is too large, the amount of information extracted by the network is larger, but the redundancy of the information caused by too much overlapping of the feature layers is too high, and the generalization capability of the model can be reduced. On the contrary, if the overlapping coefficient is too small, the information utilization rate is possibly insufficient, and the identification precision is not greatly improved, so that the overlapping coefficient is determined by adopting the formula, and the identification precision is ensured.
In the present invention, the reinforced capsule network needs to be trained first, and the specific way may be:
1) and selecting acquisition parameters, and acquiring vibration signals of different health states of the planetary gearbox of the test bed by using the acceleration sensor.
2) And selecting a wavelet packet base to carry out wavelet packet decomposition on the original vibration signal, obtaining a coefficient matrix representing time-frequency characteristics after decomposition, and building a training data set and a test data set required by the model according to a proportion.
3) Initializing parameters of E-Capsule network, setting superparameters such as training times, learning rate and batch training size, taking a preprocessed training data set as input of the E-Capsule network in a batch training mode, obtaining output through network calculation, and calculating error loss to optimize and adjust network parameters to obtain a better structural model.
4) And taking the test data set as the input of the trained model, comparing and analyzing the obtained output result with the test label to verify the effectiveness of the fault diagnosis model of the planetary gearbox based on the reinforced capsule network, and finishing the training when the effectiveness meets the requirement.
In order to verify the effectiveness of the method disclosed by the invention, a power transmission comprehensive experiment table is used for acquiring vibration signals of the planetary gearbox under different faults. The experiment table mainly comprises a driving motor, a two-stage planetary gear box, a two-stage parallel gear box and a magnetic powder brake. An NI9234 signal acquisition card and a vibration acceleration sensor (the model is PCB352C03) are adopted to acquire vibration signals of the planetary gearbox in the horizontal and vertical directions at the sampling frequency of 25600 Hz. The invention sets three working conditions aiming at the running state of the planetary gearbox, and the specific working conditions are set as shown in table 1. Considering the case of the planetary gearbox in practical engineering applications, the motor rotation frequency is set to accelerate from 20Hz to 36Hz for 16s, wherein the load conditions are 0, 4hp and 8hp respectively.
TABLE 1 Condition set description
Figure BDA0002218903870000071
The present invention contemplates 9 planetary gearbox failure categories as shown in table 2. In order to obtain a sufficient number of samples, 4 sets of vibration data are respectively measured under each working condition, the sampling time of each set of vibration data is 16 seconds, the sampling length is 400K (1K is 1024), the sampling length is divided into 100 samples, and each sample is 4096 points. The sample size for each category is 1200, 1000 samples are randomly drawn as a training set, and 200 samples are taken as a test set. A total of 9000 samples were collected for training and a total of 1800 samples were collected for testing.
TABLE 2 planetary gearbox data set description
Figure BDA0002218903870000072
Figure BDA0002218903870000081
Meanwhile, in order to simulate the vibration state of the planetary gearbox under the real condition, Gaussian noise with the signal-to-noise ratio of 4dB is added into each original vibration data. The time domain waveform of the vibration after each category is added with noise is shown in FIG. 6, and corresponds to the state category of the planetary gearbox described in Table 2.
The E-CapsNet network structure provided by the invention can specifically have 2 convolution layers in total, the first cavity convolution layer has 16 convolution kernels, the cavity rate is 2, and the convolution kernel sliding window is 1; 32 convolution kernels in the second common convolution layer, wherein the void rate is 1, and the convolution kernel sliding window is 2; the Primary capsule layer maps shallow features into capsules by adopting 7 three-dimensional convolution kernels with the thickness of 8, and the size of the Primary capsules is 8 multiplied by 1, namely each capsule comprises 8 neurons; the size of the digit capsule is 16 multiplied by 1; the convolution kernel sizes in the network are all 9 x 9. The reconstruction layer comprises three fully connected layers with sizes of 1024, 2048 and 4096. The data were trained 80 times in total, with a learning rate set to 0.001 and a batch size minipatch set to 50. Specific network structure parameters and calculation process characteristic parameters are shown in table 3.
TABLE 3 Capsule network architecture parameters
Figure BDA0002218903870000082
In the E-CapsNet proposed by the invention, the cited overlapping coefficient size has a great influence on the network structure. If the overlapping coefficient is too large, the amount of information extracted by the network is larger, but the redundancy of the information caused by too much overlapping of the feature layers is too high, and the generalization capability of the model can be reduced. On the contrary, if the overlap factor is too small, the information utilization rate may be insufficient, and the recognition accuracy of the model may not be improved much. Therefore, to provide more comprehensive results, the selection of the overlap factor needs to be experimentally verified. Assuming that the Primary capsule length is 8, the overlap factor θ is chosen as follows:
Figure BDA0002218903870000091
ten random experiments are respectively carried out on the 7 overlapping coefficients, the average value of the test accuracy is taken as the final result, and the test precision is shown in fig. 7. As can be seen from the figure, when the overlap factor is 0.5, the model test accuracy is the highest, and the average precision is 96.2%, so the overlap factor θ is selected to be 0.5.
To verify the advantages of the proposed method, the CNN, and the original CapsNet method were analyzed in comparison. The experimental result is shown in table 4, on the test set, the E-CapsNet method has higher test accuracy, which shows that the method has stronger feature learning ability, and the structural model can further improve the fault identification precision of the planetary gearbox.
Table 4 experimental results of different algorithms
Figure BDA0002218903870000092
In order to verify the robustness of the reinforced capsule network of the present invention, the original data set was randomly shuffled to reconstruct the training set and the test set using the methods of documents Han Y, Tang B, ding l.enhanced contained neural network with enhanced reconstructed fields for fault diagnosis of display objects [ J ]. Computers in Industry,2019,107:50-58, and 10 times of training and testing were performed, with the test accuracy of ten times of verification by each method being shown in fig. 8. Experimental results show that the E-CapsNet method provided by the invention is superior to CNN and CapsNet in ten verification, and has stable fault identification precision and good robustness.
The confusion matrix of the results of the last random test of the E-CapsNet on the test set is shown in FIG. 9, wherein the horizontal axis represents the true label of the test set, the vertical axis represents the prediction label of the model on the test set, and the main diagonal is the number of samples for which the model predicts the correct. As can be seen from the graph, the E-CapsNet model has high accuracy in identifying various health states of the planetary gearbox, and only few samples are misjudged.
In order to show the learning capability of the gelatin capsule layer to different types of features, the invention respectively reduces the output feature dimensions before and after the capsule layer to two-dimensional characterization by using t-SNE (t-distributed stored geometric neighbor embedding) in manifold learning, and performs feature visualization on the obtained characterization data, wherein the visualization result is shown in FIG. 10 and FIG. 11. FIG. 10 shows the input feature distribution result of primary capsule layer, the original input signal is learned by two convolutional layers, and the samples in each category (between classes) are still mixed into a whole, but the distance between the samples in each category (within classes) has a significant "gathering" trend, which indicates that the convolutional layer has extracted the features of the original signal; fig. 11 shows the output characteristic distribution result of the digit capsule layer, and through the learning of the two capsule layers, the inter-class distance is significantly increased and is easy to distinguish, the intra-class distance is contracted, and the other samples are distributed in the corresponding ranges except for a very small number of heterogeneous samples.
The invention provides a fault diagnosis method based on E-CapsNet, aiming at the problems of nonlinearity and non-stationarity of a vibration signal of a planetary gear box. The nonlinearity and the feature extraction capability of the model are enhanced by introducing the cavity convolution and constructing the reinforced capsule layer, and the experimental result shows that:
(1) by combining the advantages of nonlinear and non-stationary signal processing by wavelet packet transformation and the strong characteristic learning capability of the E-CapsNet, the fault diagnosis of the planetary gearbox can be realized with high precision, and meanwhile, the robustness is good.
(2) Compared with the traditional CNN and original CapsNet method, the E-CapsNet has higher diagnosis accuracy, and the introduction of the cavity convolution and the overlapping coefficient enhances the adaptive learning capability of the model.
Finally, it is noted that the above-mentioned embodiments illustrate rather than limit the invention, and that, while the invention has been described with reference to preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (4)

1. A fault diagnosis method for a planetary gearbox based on an enhanced capsule network is characterized by comprising the following steps:
s1, acquiring data to be diagnosed;
s2, inputting data to be diagnosed into a reinforced capsule network, wherein the reinforced capsule network comprises a convolution layer, a Primary capsule layer, a Digit capsule layer and a decoding layer, the convolution layer comprises a first convolution layer and a second convolution layer, the void factor of the first convolution layer is larger than 1, and the void factor of the second convolution layer is equal to 1; the convolution kernel size calculation expression of the first void convolution layer is N ═ k + (k-1) x (d-1), wherein N represents the convolution kernel size of the first convolution layer, k represents the convolution kernel size of the second convolution layer, and d represents the void factor of the first convolution layer;
and S3, outputting a fault diagnosis result by the reinforced capsule network.
2. The method for diagnosing faults of an epicyclic gearbox based on an intensified capsule network according to claim 1, wherein the step S1 comprises:
s101, acquiring original data;
s102, decomposing original data by adopting a wavelet packet and constructing a time-frequency coefficient matrix;
and S103, taking the time-frequency coefficient matrix as data to be diagnosed.
3. The method for diagnosing faults of an epicyclic gearbox based on an intensified capsule network according to claim 1, further comprising: introducing an overlap coefficient theta into a Primary capsule layer, and reusing a part of scalar characteristic layer in the process of carrying out discrete segmentation on the original scalar characteristic;
the calculation expression of the Primary capsule layer after the overlap coefficient theta is introduced is as follows:
Figure FDA0002460107650000011
Figure FDA0002460107650000012
in the formula (I), the compound is shown in the specification,
Figure FDA0002460107650000013
i capsule as the l capsule layer, MkFor the k-th mapped region of the layer, uijA vector matrix, w, for mapping the jth feature in the upper convolutional feature layer to the ith capsule regionijConvolution kernel matrix of corresponding eigenvector for the mapping process, bijFor the corresponding added bias matrix, N represents the number of convolution feature layer layers, Nc represents the capsule length, θ represents the overlap factor, and Z represents the total number of capsule layers.
4. The planetary gearbox fault diagnosis method based on the reinforced capsule network as claimed in claim 3, wherein the value range of θ is a discrete value sequence,
Figure FDA0002460107650000014
Figure FDA0002460107650000015
wherein L is an integer value between 0 and L, L representing the capsule length.
CN201910925753.6A 2019-09-27 2019-09-27 Planetary gearbox fault diagnosis method based on reinforced capsule network Active CN110657984B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910925753.6A CN110657984B (en) 2019-09-27 2019-09-27 Planetary gearbox fault diagnosis method based on reinforced capsule network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910925753.6A CN110657984B (en) 2019-09-27 2019-09-27 Planetary gearbox fault diagnosis method based on reinforced capsule network

Publications (2)

Publication Number Publication Date
CN110657984A CN110657984A (en) 2020-01-07
CN110657984B true CN110657984B (en) 2020-06-02

Family

ID=69039641

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910925753.6A Active CN110657984B (en) 2019-09-27 2019-09-27 Planetary gearbox fault diagnosis method based on reinforced capsule network

Country Status (1)

Country Link
CN (1) CN110657984B (en)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109085436A (en) * 2018-08-17 2018-12-25 南京航空航天大学 Non-intrusion type three-phase full-bridge inverter combined failure diagnostic method
CN111170108B (en) * 2020-02-13 2021-11-23 浙江新再灵科技股份有限公司 Method for detecting out-of-door elevator stopping based on multi-source sensor fault label sample data gate
CN111562105B (en) * 2020-03-25 2021-11-23 浙江工业大学 Wind turbine generator gearbox fault diagnosis method based on wavelet packet decomposition and convolutional neural network
CN111460818B (en) * 2020-03-31 2023-06-30 中国测绘科学研究院 Webpage text classification method based on enhanced capsule network and storage medium
CN111626361B (en) * 2020-05-28 2023-08-11 辽宁大学 Bearing sub-health identification method for improving capsule network optimization hierarchical convolution
CN111766513B (en) * 2020-06-18 2021-08-20 电子科技大学 Capsule network-based variable-working-condition multi-fault diagnosis method for three-phase induction motor
CN112364920B (en) * 2020-11-12 2023-05-23 西安电子科技大学 Thyroid cancer pathological image classification method based on deep learning
CN113255882B (en) * 2021-04-30 2022-11-04 南通大学 Bearing fault diagnosis method based on improved convolution capsule network

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9863917B2 (en) * 2006-03-20 2018-01-09 Clarkson University Method and system for real-time vibroacoustic condition monitoring and fault diagnostics in solid dosage compaction presses
CN109029974A (en) * 2018-06-22 2018-12-18 上海电力学院 A kind of epicyclic gearbox fault detection method based on one-dimensional convolutional neural networks

Also Published As

Publication number Publication date
CN110657984A (en) 2020-01-07

Similar Documents

Publication Publication Date Title
CN110657984B (en) Planetary gearbox fault diagnosis method based on reinforced capsule network
Ding et al. A novel time–frequency Transformer based on self–attention mechanism and its application in fault diagnosis of rolling bearings
Luo et al. A case study of conditional deep convolutional generative adversarial networks in machine fault diagnosis
Li et al. Intelligent fault diagnosis of rolling bearings under imbalanced data conditions using attention-based deep learning method
Wang et al. A novel method for intelligent fault diagnosis of bearing based on capsule neural network
Tao et al. Bearing fault diagnosis method based on stacked autoencoder and softmax regression
CN112926641B (en) Three-stage feature fusion rotating machine fault diagnosis method based on multi-mode data
CN112964469B (en) Online fault diagnosis method for rolling bearing under variable load of transfer learning
CN112906644A (en) Mechanical fault intelligent diagnosis method based on deep migration learning
CN110705525A (en) Method and device for diagnosing rolling bearing fault
CN113203566B (en) Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN
Li et al. Multiscale dynamic fusion prototypical cluster network for fault diagnosis of planetary gearbox under few labeled samples
Wang et al. Intelligent fault diagnosis for planetary gearbox using transferable deep q network under variable conditions with small training data
CN114048769A (en) Multi-source multi-domain information entropy fusion and model self-optimization method for bearing fault diagnosis
Liu et al. Intelligent fault diagnosis under small sample size conditions via Bidirectional InfoMax GAN with unsupervised representation learning
CN112132102B (en) Intelligent fault diagnosis method combining deep neural network with artificial bee colony optimization
Ma et al. An unsupervised domain adaptation approach with enhanced transferability and discriminability for bearing fault diagnosis under few-shot samples
EP4273754A1 (en) Neural network training method and related device
CN116502175A (en) Method, device and storage medium for diagnosing fault of graph neural network
CN110930378A (en) Emphysema image processing method and system based on low data demand
CN115290326A (en) Rolling bearing fault intelligent diagnosis method
CN113705396A (en) Motor fault diagnosis method, system and equipment
CN114565594A (en) Image anomaly detection method based on soft mask contrast loss
CN114970326A (en) Oil well fault diagnosis method based on improved capsule neural network
CN116702076A (en) Small sample migration learning fault diagnosis method, system, computer and storage medium based on CNN feature fusion

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
GR01 Patent grant
GR01 Patent grant