CN114386453A - Fault diagnosis method for three-phase asynchronous motor and computer readable medium - Google Patents

Fault diagnosis method for three-phase asynchronous motor and computer readable medium Download PDF

Info

Publication number
CN114386453A
CN114386453A CN202111487994.0A CN202111487994A CN114386453A CN 114386453 A CN114386453 A CN 114386453A CN 202111487994 A CN202111487994 A CN 202111487994A CN 114386453 A CN114386453 A CN 114386453A
Authority
CN
China
Prior art keywords
layer
fault
capsule
phase asynchronous
asynchronous motor
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
Application number
CN202111487994.0A
Other languages
Chinese (zh)
Inventor
王超群
胡江
李彬彬
焦斌
程鹏宇
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.)
Shanghai Dianji University
Original Assignee
Shanghai Dianji 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 Shanghai Dianji University filed Critical Shanghai Dianji University
Priority to CN202111487994.0A priority Critical patent/CN114386453A/en
Publication of CN114386453A publication Critical patent/CN114386453A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G06F18/24137—Distances to cluster centroïds
    • G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction

Landscapes

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

Abstract

The invention relates to a fault diagnosis method for a three-phase asynchronous motor and a computer readable medium, wherein the fault diagnosis method for the motor comprises the following steps: acquiring motor vibration signal data, and performing data enhancement processing on the data; building a capsule network model, wherein the model comprises a one-dimensional convolution layer, a convolution layer with a multi-layer structure, a residual error layer, a Concat layer and a digital capsule layer which are sequentially connected; training the built capsule network model; and performing fault diagnosis on the acquired motor vibration signal by using the trained capsule network model, and outputting a fault diagnosis result. Compared with the prior art, the method has the advantages of good robustness, strong generalization capability and the like.

Description

Fault diagnosis method for three-phase asynchronous motor and computer readable medium
Technical Field
The invention relates to the technical field of motor fault diagnosis, in particular to a three-phase asynchronous motor fault diagnosis method based on an improved capsule network and a computer readable medium.
Background
The conventional motor fault diagnosis method generally adopts a signal processing method. Most motor fault diagnosis uses signals such as vibration, current, infrared and sound. Due to the influence of a large amount of redundant information, noise and other factors on the collected original signals, fault characteristics are hidden deeply in the original signals and are not easy to find out. Therefore, a signal processing method is needed to manually extract the fault features. Common signal processing methods include fourier transform, wavelet transform, hilbert transform, empirical mode decomposition, and the like.
With the vigorous development of the computer field, researchers propose another new intelligent algorithm. More and more scholars apply the fault diagnosis method to the field of fault diagnosis and achieve certain effect. Among them, fuzzy theory, BP neural network, support vector machine, random forest and other methods are typical.
Conventional methods require extensive signal processing and mathematical knowledge and often only consider a small fraction of the fault types, which fail when the number of faults increases. Some new algorithms based on the artificial intelligence technology classify the extracted fault features by analyzing and preprocessing fault data and then by a deep learning method. The method does not need to know factors such as fault characteristic frequency and the like deeply influenced by the environment, can eliminate the interference caused by manually extracting the characteristics, completes the whole process from extracting the fault characteristics to fault classification, and automatically realizes the end-to-end intelligent fault diagnosis and classification. However, these methods still have some problems. Most of the existing methods only use the original state of a motor for experiments due to the limitation of experimental conditions and the like, and do not consider the problems possibly encountered in the actual industrial environment. In an actual industrial environment, the motor may be in a strong noise environment, such as a vibration signal and the like may be submerged by noise; in addition, when the motor is in different load states, the rotating speed of the motor can change, so that the vibration signal and the current signal are influenced. It is far from sufficient to consider the failure state of a motor under the original conditions alone.
Disclosure of Invention
The present invention is directed to overcome the above-mentioned drawbacks of the prior art, and provides a method for diagnosing a fault of a three-phase asynchronous motor with good robustness and high generalization capability, and a computer readable medium.
The purpose of the invention can be realized by the following technical scheme:
a fault diagnosis method for a three-phase asynchronous motor comprises the following steps:
step 1: acquiring motor vibration signal data, and performing data enhancement processing on the data;
step 2: building a capsule network model, wherein the model comprises a one-dimensional convolution layer, a convolution layer with a multi-layer structure, a residual error layer, a Concat layer and a digital capsule layer which are sequentially connected;
and step 3: training the capsule network model built in the step 2;
and 4, step 4: and 3, performing fault diagnosis on the motor vibration signal obtained in the step 1 by using the capsule network model trained in the step 3, and outputting a fault diagnosis result.
Preferably, the one-dimensional convolutional layer in the capsule network model is specifically:
Figure BDA0003398106670000021
wherein the content of the first and second substances,
Figure BDA0003398106670000022
is the output of the jth neuron in the current layer, i.e. layer l;
Figure BDA0003398106670000023
is the output of the ith neuron in the previous layer, i.e. layer l-1;
Figure BDA0003398106670000024
all input features;
Figure BDA0003398106670000025
convolution kernels from the ith neuron in layer l-1 to the jth neuron in layer l;
Figure BDA0003398106670000026
bias for the jth neuron at layer l; f is the activation function.
More preferably, the output end of the one-dimensional convolution layer is provided with a one-dimensional pooling layer, specifically:
Figure BDA0003398106670000027
wherein the content of the first and second substances,
Figure BDA0003398106670000028
is the output of the ith neuron in the current layer, i.e. the layer l; maxporoling () is a downsampling function; sscaleIs a pooling scale; sstrideIs the pooling step size.
Preferably, the convolution with the multilayer structure performs network hop Connection by using a Skip Connection method.
Preferably, the residual error layer specifically includes:
F(x)＝x+f(x)
wherein x is the input of the residual layer; (x) is the output after lossy compression; f (x) is the input of the subsequent network.
Preferably, the digital capsule layer is specifically:
the inputs to the digital capsule layer are:
Figure BDA0003398106670000031
Figure BDA0003398106670000032
wherein u isiIs the output of the previous capsule layer; wijIs a weight matrix;
Figure BDA0003398106670000033
is a prediction vector; c. CijIs the coupling coefficient; sjIs an intermediate vector;
sjobtaining an output vector through a squaring activation function of the digital capsule layer, specifically:
Figure BDA0003398106670000037
more preferably, the digital capsule layer searches for the optimal coupling coefficient c through a dynamic routing algorithmij(ii) a The coupling coefficient is formed by a prediction vector
Figure BDA0003398106670000034
And the output vector vjUpdating, namely:
Figure BDA0003398106670000035
Figure BDA0003398106670000036
more preferably, the loss function of the digital capsule layer is:
Lc＝Tcmax(0,m+-||vc||)2+λ(1-Tc)max(0,||vc||-m-)2
wherein, TcFor a true tag, T when the input sample class coincides with ccTaking 1, otherwise, taking 0; | v | (V)c| | is the modular length of the vector, i.e., the probability of the fault type; m is+And m-Respectively, an upper boundary and a lower boundary, i.e. when | | | vc||＞m+Or vc||＜m-The loss function is 0; and lambda is a ratio parameter and is used for adjusting the ratio of the two terms.
Preferably, said step 3 optimizes the total loss using an Adam optimizer.
A computer readable medium, wherein the computer readable medium stores therein the fault diagnosis method for the three-phase asynchronous motor.
Compared with the prior art, the invention has the following beneficial effects:
good robustness and strong generalization ability: the three-phase asynchronous motor fault diagnosis method adopts the one-dimensional convolutional neural network with a multilayer structure to quickly extract the characteristics, and then utilizes the residual error network to reduce the overfitting risk. Finally, the capsule network is used for completing fault classification, the average accuracy of the method can still reach 91.9% in a strong noise environment (-10dB), the average accuracy of the method can reach 90.82% in a variable load environment, and compared with the prior art, the network model fully proves the robustness and the generalization capability of the method.
Drawings
FIG. 1 is a schematic flow chart of a fault diagnosis method for a three-phase asynchronous motor according to the invention;
FIG. 2 is a schematic diagram of the improved capsule network model of the present invention;
FIG. 3 is a graphical illustration of accuracy and loss on a validation set in an embodiment of the present invention;
wherein FIG. 3(a) is the accuracy on the validation set and FIG. 3(b) is the loss on the validation set;
FIG. 4 is a diagram of an exemplary confusion matrix;
FIG. 5 is a diagram illustrating a noise signal according to an embodiment of the present invention;
wherein, fig. 5(a) is the original signal, fig. 5(b) is the gaussian white noise signal, fig. 5(c) is the signal after noise addition;
FIG. 6 is a schematic diagram of the accuracy of model identification in a strong noise environment according to an embodiment of the present invention;
where fig. 6(a) is the accuracy at no load, fig. 6(b) is the accuracy at 25% load, fig. 6(c) is the accuracy at 50% load, and fig. 6(d) is the accuracy at 75% load.
FIG. 7 is a signal diagram of different loads according to an embodiment of the present invention;
where fig. 7(a) is the signal when the training set is empty, fig. 7(b) is the signal when the training set is 25% loaded, fig. 7(c) is the signal when the training set is 50% loaded, and fig. 7(d) is the signal when the training set is 75% loaded;
FIG. 8 is a schematic diagram of the accuracy of model identification in a variable load environment according to an embodiment of the present invention;
fig. 8(a) shows the accuracy when the training set is empty, fig. 8(b) shows the accuracy when the training set is 25% loaded, fig. 8(c) shows the accuracy when the training set is 50% loaded, and fig. 8(d) shows the accuracy when the training set is 75% loaded.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, not all, embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, shall fall within the scope of protection of the present invention.
The embodiment provides a fault diagnosis method for a three-phase asynchronous motor based on an improved capsule network, aiming at the limitations of a signal processing method and the problems of the motor under the strong noise and variable load environment. The method uses a one-dimensional convolutional neural network as a model input, and adopts a multi-scale strategy on the basis of the one-dimensional convolutional neural network, wherein the strategy can improve the feature expression capability of the Convolutional Neural Network (CNN) from multiple layers. The model is further processed by a residual layer to prevent gradient explosion. And finally, further extracting detail features through a capsule network, and finishing fault classification. Practical experiments and comparative experiments show that the method has better robustness and generalization capability.
A three-phase asynchronous motor fault diagnosis method based on an improved capsule network is provided, the flow of which is shown in figure 1, and the method comprises the following steps:
step 1: acquiring motor vibration signal data, and performing data enhancement processing on the data;
step 2: building a capsule network model, wherein the model comprises a one-dimensional convolution layer, a convolution layer with a multi-layer structure, a residual error layer, a Concat layer and a digital capsule layer which are sequentially connected;
and step 3: training the capsule network model built in the step 2;
and 4, step 4: and 3, performing fault diagnosis on the motor vibration signal obtained in the step 1 by using the capsule network model trained in the step 3, and outputting a fault diagnosis result.
The structure of the capsule network model in this embodiment is shown in fig. 2, and the following layers are described respectively:
(1) the first layer of the model is the convolutional layer:
since the vibration signal of the motor is time-series, a one-dimensional convolutional neural network is used as a model input. Compared with a two-dimensional convolutional neural network, a one-dimensional convolutional neural network changes only the dimension of the feature map. The one-dimensional convolutional neural network mainly comprises a one-dimensional convolutional layer, a pooling layer, a full-link layer and the like. One-dimensional convolutional layers: the purpose of convolutional layer is to enhance the characteristics of the original input signal and eliminate part of the noise in the original input, which is also one of the reasons why the original signal is directly input into the network without any characteristic extraction. Its formula is shown as follows:
Figure BDA0003398106670000051
wherein the content of the first and second substances,
Figure BDA0003398106670000052
is the output of the jth neuron in the current layer, i.e. layer l;
Figure BDA0003398106670000053
is the output of the ith neuron in the previous layer, i.e. layer l-1;
Figure BDA0003398106670000054
all input features;
Figure BDA0003398106670000055
convolution kernels from the ith neuron in layer l-1 to the jth neuron in layer l;
Figure BDA0003398106670000056
bias for the jth neuron at layer l; f is the activation function.
A one-dimensional pooling layer: the purpose of the pooling layer is to reduce the size of the feature matrix without changing the depth of the feature matrix, thereby effectively reducing the number of parameters in the entire network. Its formula is shown as follows:
Figure BDA0003398106670000057
wherein the content of the first and second substances,
Figure BDA0003398106670000058
is the output of the ith neuron in the current layer, i.e. the layer l; maxporoling () is a downsampling function; sscaleIs a pooling scale; sstrideIs the pooling step size.
To obtain more useful information, the first convolutional layer uses a wider convolutional kernel of size 127, with 32 convolutional kernels and a step size of 8. The larger receptive field can improve the anti-noise capability of the model to a certain extent. In addition, ReLU is used as the activation function in the convolutional layer. And then adding a batch normalization layer in the network, wherein the purpose of adding the batch normalization layer is to prevent the disappearance or explosion of the gradient and accelerate the training speed. And then adding a pooling layer in the network, so as to reduce the parameter quantity in the whole network and prevent overfitting.
(2) The second layer of the model is a multilayer structure:
multilayer structure: smaller convolution kernels contain more detailed features and corresponding location information, but their receptive field is small, i.e., the convolution kernels contain less information. Conversely, a larger convolution kernel has a larger field of view and can extract more information, but relatively speaking, it will have fewer detail features than a smaller convolution kernel. Therefore, the convolution kernels with various sizes can be fused to better express the characteristic information contained in the signals. In this embodiment, network hopping is performed by using a Skip Connection method, that is, a multi-layer feature is fused first, and then the next operation is performed, where the next operation is a Concat operation.
Three convolution kernels with the sizes of 3, 5 and 7 are respectively used in the layer, the number of the convolution kernels is 32, and the step length is 2. Fusing convolution kernels of different sizes may better express the characteristic information contained in the signal than convolution kernels of a single size. The subsequent operation is the same as the previous layer. ReLU is also used as the activation function in convolutional layers. And then adding a batch normalization layer and a pooling layer in the network.
(3) The third layer of the model is the residual layer
In general, as the degree of the network increases, the performance of the network becomes better. However, the deeper the network structure is, the better the network structure is, when the network reaches a certain depth, the gradient explosion and gradient disappearance may occur when the network layer number is increased. In order to solve the problems that the learning rate becomes low and the accuracy cannot be effectively improved due to the deepening of the network depth, the embodiment introduces a residual error layer, specifically:
F(x)＝x+f(x)
wherein x is the input of the residual layer; (x) is the output after lossy compression; f (x) is the input of the subsequent network. In this way the network can learn more signal characteristics.
In this embodiment, the residual layer uses a residual block with a convolution kernel size of 3 and a convolution kernel number of 32, so as to avoid the problem of accuracy degradation while deepening the network.
(4) The fourth layer of the model is the Concat layer:
the method has the function of converting the scalar features extracted from the layers into vector features and then splicing the vector features. Vectors are formed by merging the channels of the feature layers into one capsule unit. In this layer, there are a total of 1536(16 × 32 × 3) primary capsules, since the output of the previous layer residual layer is 16 × 32, and there are a total of three such feature numbers. To facilitate subsequent processing, it is reshaped into a 192 × 8 vector.
(5) The fifth layer of the model is a digital capsule layer:
the function of the digital capsule layer is equal to that of the fully-connected layer in the convolutional neural network, and the digital capsule layer is used for distinguishing various fault types. Since the number of fault types to be identified is 8, the number of capsules on the layer is 8, the vector dimension is set to be 8, and the modular length of the vector is the probability of a certain fault type.
The inputs to the digital capsule layer are:
Figure BDA0003398106670000071
Figure BDA0003398106670000072
wherein u isiIs the output of the previous capsule layer; wijIs a weight matrix;
Figure BDA0003398106670000073
is a prediction vector; c. CijIs the coupling coefficient; sjIs an intermediate vector;
sjobtaining an output vector through a squaring activation function of the digital capsule layer, specifically:
Figure BDA0003398106670000074
cijdetermining the degree of closeness of the current capsule layer and the capsules of the previous layer, which is formed by bijNormalized by softmax.
The digital capsule layer searches the optimal coupling coefficient c through a dynamic routing algorithmij(ii) a The coupling coefficient is formed by a prediction vector
Figure BDA0003398106670000075
And the output vector vjUpdating, namely:
Figure BDA0003398106670000076
Figure BDA0003398106670000077
the loss function for the digital capsule layer is:
Lc＝Tcmax(0,m+-||vc||)2+λ(1-Tc)max(0,||vc||-m-)2
wherein, TcFor a true tag, T when the input sample class coincides with ccTaking 1, otherwise, taking 0; | v | (V)c| | is the modular length of the vector, i.e., the probability of the fault type; m is+And m-Respectively, an upper boundary and a lower boundary, i.e. when | | | vc||＞m+Or vc||＜m-The loss function is 0, m in this example+And m-Respectively taking 0.9 and 0.1; λ is a scaling parameter for adjusting the ratio of the two terms, and λ is 0.5 in this embodiment.
The parameter setting method of the capsule network model comprises the following steps:
this example is done in the Google deep learning framework TensorFlow. The batch size was set to 32 and the number of training sessions for all samples was 10. The number of iterations of the dynamic routing algorithm in the capsule layer is 2. In addition, the Adam optimizer is used for optimizing the total loss, the learning rate is set to be 0.001, the dynamic attenuation mode is adopted, and the attenuation rate is 10-8. To verify the validity of the model, the following four networks were used as comparative experiments.
(1) No residual block is used. This comparison model does not use a residual block in the network, and the remaining parameters are the same as the proposed network.
(2) A multi-level structure is not used. The comparison model does not use a multi-level structure in the network, only uses a convolution kernel with the size of 5 in the second convolution layer, and the rest parameters are the same as the proposed network.
(3) No residual block and no multi-level structure are used. The structure of the first layer of the one-dimensional convolutional neural network of the comparison model is the same as that of the proposed network, a multilayer structure is not used in the second layer, a convolutional layer with a convolutional kernel size of 5 is selected, and a residual error network is not used in the third layer, and then the convolutional layer passes through a capsule layer. The remaining parameters are the same as for the proposed network.
(4) A one-dimensional convolutional neural network. The structure of the first layer of the one-dimensional convolutional neural network of the comparison model is the same as that of the proposed network, a multilayer structure is not used in the second layer, and a convolutional layer with a convolutional kernel size of 5 is selected and then passes through three full-connection layers with the number of nodes of 512, 256 and 8 respectively. The remaining parameters are the same as for the proposed network. The loss function uses a cross-entropy loss function.
The present embodiment also relates to a computer-readable medium having stored therein any one of the motor failure diagnosis methods described above.
To verify the validity of the model, a validation was performed using fault data collected by a laboratory motor data collection platform. The platform consists of a 3kW asynchronous motor, a DC generator, a motor data acquisition box, a protection circuit and a plurality of sensors. And selecting a vibration signal of the axial direction of the driving end of the motor as experimental data. Different loads are realized by adding corrugated resistors with different resistance values at two ends of the direct current generator, and the four states of no load, 25% load, 50% load and 75% load are realized respectively. The experiment studied 5 models, respectively, Normal State (Normal), rotor Bar out (RBB), stator turn-to-turn short (divided into short 1 turn (SC1T) and short 3 turn (SC3T), eccentricity (AE) and bearing failure (bearing failure divided into outer race wear (ROO), inner race wear (ROI), cage breakage RCB.) together obtained 8 state signals for the electric machine The sliding step size and the number of samples must satisfy the inequality: l + (N-1) x s is less than or equal to T. In the experiment, the length of each sample was set to 1024. The total collected data is 100000. The sliding step is set to 99 and brought into equation 9. Finally, the number of collected samples can be calculated to be 1000(1024+ (1000-1) × 99 ≦ 99925 ≦ 100000). Because there are 8 states, the total number of samples obtained is 8000, namely the number of samples in normal state and 7 fault states is 1000 respectively, and the corresponding label is 0-7. 8000 data were recorded as 7: 2: the scale of 1 is divided into training set, validation set and test set, i.e. containing 5600 training samples, 1600 validation samples and 800 test samples. The final settings are shown in table 1.
Analysis of results in the initial state: the experiment was performed using four states of no load, 25% load, 50% load, and 75% load, respectively. The results are the average of 5 experiments. The accuracy and loss values for the validation set are shown in fig. 3. For the test set samples, the confusion matrix shown in FIG. 4 and the fault diagnosis reports shown in Table 2 were used to evaluate the performance of the model. As can be seen in fig. 3, after approximately 7-8 iterations, the accuracy on the validation set is as high as 100%. After 10 iterations the loss of the model is only around 0.002 and the time required to diagnose one sample is only 18 ms. This shows that the convergence rate of the model is fast and the convergence process is stable. The accuracy of the prediction of each fault can be seen on the confusion matrix and fault diagnosis report. As can be seen from fig. 4, in the original state, the model fault proposed by the present embodiment is all diagnosed correctly under all faults. From table 2, the precision, recall, and F1 Score were 1.0 for each fault. This also fully accounts for the effectiveness of the model. In contrast to the others: the results are shown in Table 3. On one hand, the models do not need data processing, are directly input into a network, and can have high identification accuracy. This fully accounts for the superiority of deep learning. On the other hand, the network model proposed in this embodiment has no advantage due to the superiority of the capsule network. But 0.33% higher than the one-dimensional convolution network model.
TABLE 1 sample data set
Label (R) Location of failure Training set Verification set Test set
0 Is normal 700 200 100
1 Rotor broken bar 700 200 100
2 Short circuit 1 turn between turns 700 200 100
3 3 turns of turn-to-turn short circuit 700 200 100
4 Eccentric fault 700 200 100
5 Wear of bearing inner race 700 200 100
6 Wear of bearing outer race 700 200 100
7 Bearing cage fracture 700 200 100
TABLE 2 Fault diagnosis report
Figure BDA0003398106670000091
Figure BDA0003398106670000101
TABLE 3 accuracy in raw State
Figure BDA0003398106670000102
Analysis of results in the noise state: gaussian white noise is selected as noise interference, original signals are used in a training set for better verifying the anti-noise capability of a model, and Gaussian white noise with different signal-to-noise ratios (SNR) is added in a verification set and a test set. Taking the vibration time domain signal of the rotor broken bar under 50% load as an example, as shown in fig. 5, when the signal-to-noise ratio is-10 dB, the noise-added signal under noise pollution is greatly changed compared with the original signal, and the difficulty of extracting the fault feature from the noise-added signal is large. Therefore, it is important for the model to maintain a high accuracy even in a noisy environment. Experiments with signal-to-noise ratios of-10 dB to 4dB are respectively carried out under different load conditions, and each result is an average value of 5 experiments. The results of the experiment are shown in FIG. 6. As can be seen from the figure, on the one hand, the accuracy of each model is increased along with the reduction of the noise signal, and when the load is 75% and the signal-to-noise ratio is 4dB, the accuracy of the model proposed in the present embodiment reaches 100%; on the other hand, whether empty or loaded at 25%, 50%, 75%, with the addition of the multi-level structure and the residual block, the present embodiment suggests that the accuracy of the proposed network is the highest in the compared network, which is 0.325%, 0.454%, 0.338%, 5.816% higher than the four comparative models respectively at empty, 0.417%, 0.296%, 0.54%, 6.581% higher than the four comparative models respectively at 25%, 0.444%, 0.519%, 0.596%, 7.0% higher than the four comparative models respectively at 50%, 0.196%, 0.312%, 0.354%, 6.283% higher than the four comparative models respectively at 75%. The model provided by the embodiment is the most higher than convolution, which not only shows that the vector in the capsule network can extract more detailed features than a scalar in the convolution network, but also shows that the model can still directly use the original signal input under a strong noise environment, thereby realizing end-to-end motor fault diagnosis.
And (3) analyzing results in a variable load state: the vibration time domain signals of the bearing retainer fracture under different load conditions are taken as an example. As shown in fig. 7, the signal waveforms of the same fault under different load conditions are also very different, so that the model cannot distinguish the extracted features, thereby affecting the accuracy of identification. Therefore, fault diagnosis in a variable load environment is also important. It was found experimentally that under varying load conditions, the larger convolution kernel was used less often than the smaller convolution kernel, so the convolution kernel size in the first layer of convolutional layers was changed to 21 in varying load conditions, with the remaining parameters remaining unchanged. The no-load, 25%, 50% and 75% load are respectively used as training set, and the rest are used as verification set and test set. There are 12 cases in total, which are 0-25, 0-50, 0-75, 25-0, 25-50, 25-75, 50-0, 50-25, 50-75, 75-0, 75-25, 75-50, respectively, and the experimental results are shown in FIG. 8. On the one hand, in most cases, as the load difference decreases, the accuracy also increases, and the difference between the vibration signals with load and without load is the greatest. On the other hand, except for the case from 25% load to no load, the accuracy of the network proposed in this example is highest in all the comparative networks under other conditions, 0.395%, 0.663%, 0.915%, 2.301% higher than those of the four comparative experiments under no load to other load conditions, and 0.495%, 0.125%, 0.298%, 7.49% higher than those of the four comparative experiments under 25% load to other load conditions. The test sample is 1.14%, 1.197%, 1.391% and 3.331% higher than the four groups of comparative experiments under the condition of 50% load to other loads, and is 0.76%, 0.292%, 0.667% and 5.185% higher than the four groups of comparative experiments under the condition of 75% load to other loads. Comparing the model of the embodiment with the comparative model shows that the model has better cross-load diagnosis capability and can adapt to complex and variable industrial environment.
While the invention has been described with reference to specific embodiments, the invention is not limited thereto, and various equivalent modifications and substitutions can be easily made by those skilled in the art within the technical scope of the invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (10)

1. A fault diagnosis method for a three-phase asynchronous motor is characterized by comprising the following steps:
step 1: acquiring motor vibration signal data, and performing data enhancement processing on the data;
step 2: building a capsule network model, wherein the model comprises a one-dimensional convolution layer, a convolution layer with a multi-layer structure, a residual error layer, a Concat layer and a digital capsule layer which are sequentially connected;
and step 3: training the capsule network model built in the step 2;
and 4, step 4: and 3, performing fault diagnosis on the motor vibration signal obtained in the step 1 by using the capsule network model trained in the step 3, and outputting a fault diagnosis result.
2. The method for diagnosing the fault of the three-phase asynchronous motor according to claim 1, wherein the one-dimensional convolution layer in the capsule network model is specifically:
Figure FDA0003398106660000011
wherein the content of the first and second substances,
Figure FDA0003398106660000012
is the output of the jth neuron in the current layer, i.e. layer l;
Figure FDA0003398106660000013
is the output of the ith neuron in the previous layer, i.e. layer l-1;
Figure FDA0003398106660000014
all input features;
Figure FDA0003398106660000015
convolution kernels from the ith neuron in layer l-1 to the jth neuron in layer l;
Figure FDA0003398106660000016
bias for the jth neuron at layer l; f is the activation function.
3. The method for diagnosing the fault of the three-phase asynchronous motor according to claim 2, wherein a one-dimensional pooling layer is arranged at the output end of the one-dimensional convolution layer, and specifically comprises the following steps:
Figure FDA0003398106660000017
wherein the content of the first and second substances,
Figure FDA0003398106660000018
is the output of the ith neuron in the current layer, i.e. the layer l; maxporoling () is a downsampling function; sscaleIs a pooling scale; sstrideIs the pooling step size.
4. The method for diagnosing the fault of the three-phase asynchronous motor according to claim 1, wherein the convolution with the multilayer structure carries out network jump Connection by using a Skip Connection method.
5. The method for diagnosing the fault of the three-phase asynchronous motor according to claim 1, wherein the residual error layer is specifically as follows:
F(x)＝x+f(x)
wherein x is the input of the residual layer; (x) is the output after lossy compression; f (x) is the input of the subsequent network.
6. The method for diagnosing the fault of the three-phase asynchronous motor according to claim 1, wherein the digital capsule layer is specifically:
the inputs to the digital capsule layer are:
Figure FDA0003398106660000021
Figure FDA0003398106660000022
wherein u isiIs the output of the previous capsule layer; wijIs a weight matrix;
Figure FDA0003398106660000023
is a prediction vector; c. CijIs the coupling coefficient; sjIs an intermediate vector;
sjobtaining an output vector through a squaring activation function of the digital capsule layer, specifically:
Figure FDA0003398106660000024
7. the method as claimed in claim 6, wherein the digital capsule layer searches for the optimal coupling coefficient c through a dynamic routing algorithmij(ii) a The coupling coefficient is formed by a prediction vector
Figure FDA0003398106660000025
And the output vector vjUpdating, namely:
Figure FDA0003398106660000026
Figure FDA0003398106660000027
8. the method for diagnosing the fault of the three-phase asynchronous motor according to claim 6, wherein the loss function of the digital capsule layer is as follows:
Lc＝Tcmax(0,m+-||vc||)2+λ(1-Tc)max(0,||vc||-m-)2
wherein, TcFor a true tag, T when the input sample class coincides with ccTaking 1, otherwise, taking 0; | v | (V)c| | is the modular length of the vector, i.e., the probability of the fault type; m is+And m-Respectively, an upper boundary and a lower boundary, i.e. when | | | vc||＞m+Or vc||＜m-The loss function is 0; and lambda is a ratio parameter and is used for adjusting the ratio of the two terms.
9. A method for diagnosing faults of a three-phase asynchronous motor according to claim 1, characterized in that step 3 uses an Adam optimizer to optimize the total loss.
10. A computer-readable medium, wherein the computer-readable medium stores therein the method for diagnosing the fault of the three-phase asynchronous motor according to any one of claims 1 to 9.
CN202111487994.0A 2021-12-08 2021-12-08 Fault diagnosis method for three-phase asynchronous motor and computer readable medium Pending CN114386453A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111487994.0A CN114386453A (en) 2021-12-08 2021-12-08 Fault diagnosis method for three-phase asynchronous motor and computer readable medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111487994.0A CN114386453A (en) 2021-12-08 2021-12-08 Fault diagnosis method for three-phase asynchronous motor and computer readable medium

Publications (1)

Publication Number Publication Date
CN114386453A true CN114386453A (en) 2022-04-22

Family

ID=81195935

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111487994.0A Pending CN114386453A (en) 2021-12-08 2021-12-08 Fault diagnosis method for three-phase asynchronous motor and computer readable medium

Country Status (1)

Country Link
CN (1) CN114386453A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115114848A (en) * 2022-06-13 2022-09-27 成都星云智联科技有限公司 Three-phase asynchronous motor fault diagnosis method and system based on hybrid CNN-LSTM
CN116026586A (en) * 2023-01-31 2023-04-28 东华大学 Harmonic reducer delivery qualification judging method and device

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115114848A (en) * 2022-06-13 2022-09-27 成都星云智联科技有限公司 Three-phase asynchronous motor fault diagnosis method and system based on hybrid CNN-LSTM
CN115114848B (en) * 2022-06-13 2023-12-26 成都星云智联科技有限公司 Three-phase asynchronous motor fault diagnosis method and system based on hybrid CNN-LSTM
CN116026586A (en) * 2023-01-31 2023-04-28 东华大学 Harmonic reducer delivery qualification judging method and device

Similar Documents

Publication Publication Date Title
Chen et al. Intelligent fault diagnosis for rotary machinery using transferable convolutional neural network
Zhao et al. Improved generative adversarial network for vibration-based fault diagnosis with imbalanced data
Zhu et al. Stacked pruning sparse denoising autoencoder based intelligent fault diagnosis of rolling bearings
CN106682688A (en) Pile-up noise reduction own coding network bearing fault diagnosis method based on particle swarm optimization
CN114386453A (en) Fault diagnosis method for three-phase asynchronous motor and computer readable medium
CN112254964A (en) Rolling bearing fault diagnosis method based on rapid multi-scale convolution neural network
CN111046916A (en) Motor fault diagnosis method and system based on void convolution capsule network
CN110929847A (en) Converter transformer fault diagnosis method based on deep convolutional neural network
CN113988126B (en) Rolling bearing fault diagnosis method based on few-label data feature migration
CN110657984B (en) Planetary gearbox fault diagnosis method based on reinforced capsule network
CN115290326A (en) Rolling bearing fault intelligent diagnosis method
CN112132102A (en) Intelligent fault diagnosis method combining deep neural network with artificial bee colony optimization
CN116702076A (en) Small sample migration learning fault diagnosis method, system, computer and storage medium based on CNN feature fusion
CN115392333A (en) Equipment fault diagnosis method based on improved end-to-end ResNet-BilSTM dual-channel model
CN114441173B (en) Rolling bearing fault diagnosis method based on improved depth residual error shrinkage network
CN117192358A (en) Permanent magnet synchronous motor turn-to-turn short circuit fault diagnosis method based on convolutional neural network
CN115099278A (en) Equipment state diagnosis method based on wavelet packet time-frequency diagram and ResNet migration network
CN113092044B (en) Rotary machine fault diagnosis method based on weighted level visible graph
CN113409213B (en) Method and system for enhancing noise reduction of time-frequency diagram of fault signal of plunger pump
CN115600134A (en) Bearing transfer learning fault diagnosis method based on domain dynamic impedance self-adaption
CN115829029A (en) Channel attention-based self-distillation implementation method
CN114254669A (en) Method for constructing time sequence period ResNet network model in rolling bearing fault diagnosis
CN114186593A (en) Bearing fault identification method, device, equipment and storage medium under noise condition
Yan et al. Fault Diagnosis of Rolling Bearing with Small Samples Based on Wavelet Packet Theory and Random Forest
CN112116172A (en) Criminal phase prediction method based on probability map model

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