CN115688553A - Method for diagnosing turn-to-turn short circuit fault of unbalanced sample of permanent magnet synchronous motor - Google Patents
Method for diagnosing turn-to-turn short circuit fault of unbalanced sample of permanent magnet synchronous motor Download PDFInfo
- Publication number
- CN115688553A CN115688553A CN202210988049.7A CN202210988049A CN115688553A CN 115688553 A CN115688553 A CN 115688553A CN 202210988049 A CN202210988049 A CN 202210988049A CN 115688553 A CN115688553 A CN 115688553A
- Authority
- CN
- China
- Prior art keywords
- turn
- sample
- network
- fault
- short circuit
- 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
Images
Landscapes
- Control Of Electric Motors In General (AREA)
Abstract
The invention discloses a method for diagnosing turn-to-turn short circuit faults of unbalanced samples of a permanent magnet synchronous motor. The method comprises the following steps: 1) Generating a simulation stator current signal by using simulation software, and pre-training a depth residual shrinkage network; 2) Collecting three-phase stator current signals of the motor with different degrees of turn-to-turn short circuit, and establishing a real data set; 3) Putting the real data set into a generated countermeasure network for sample expansion to generate a balanced data set; 4) Optimizing and modifying the pre-training depth residual error network based on a sparse representation theory; 5) And inputting the data set into a trained deep residual shrinkage network by a transfer learning method, and performing noise reduction processing on the sample by using a shrinkage structure in the deep residual shrinkage network to realize fault classification of turn-to-turn short circuits.
Description
Technical Field
The invention belongs to the technical field of state detection and fault diagnosis of permanent magnet synchronous motors, and relates to a method for diagnosing turn-to-turn short circuit faults of unbalanced samples of a permanent magnet synchronous motor.
Background
The permanent magnet synchronous motor provides excitation by the permanent magnet, so that the structure of the motor is simpler, and the processing and assembling cost is reduced. The motor has the advantages of high efficiency, low noise, small volume, light weight, small loss and the like, so that the motor is widely applied to the fields of various automatic processing, electric vehicles and the like. The permanent magnet synchronous motor has a complex structure and is operated in a severe environment, so that a fault is inevitable. The stator turn-to-turn short circuit phenomenon is a typical fault of the permanent magnet synchronous motor. In the early stage of turn-to-turn short-circuited stator winding fault, the motor still can run, but if the fault is not processed in time, the fault degree can be further aggravated, and in the serious case, the motor is damaged, the normal work of the motor is also influenced, and an external system of the motor is damaged. Therefore, the fast and accurate turn-to-turn short circuit fault diagnosis method for the synchronous motor is important for guaranteeing safe and reliable operation of the motor, improving efficiency and the like.
Most deep learning diagnosis methods are performed based on big data, but the frequency of faults in actual situations is very low, so that short-circuit fault samples of the motor are too few, collected fault signals are mostly of a fault type, and a balanced data set is difficult to construct, which seriously affects the diagnosis effect. Meanwhile, the environment of the motor is more than that of a high-noise environment, and noise is generated in the working process of the motor, generally, for small and medium-sized motors, electromagnetic noise is a main source of overall noise and is caused by electromagnetic force acting on the surfaces of a stator and a rotor. The data of the motor is collected, and then the influence of noise is removed.
Disclosure of Invention
The invention aims to provide a method for diagnosing turn-to-turn short circuit fault of an unbalanced sample of a permanent magnet synchronous motor, so as to overcome the defects in the background technology.
The technical solution for realizing the purpose of the invention is as follows:
a method for diagnosing turn-to-turn short circuit faults of unbalanced samples of a permanent magnet synchronous motor comprises the following steps:
step 2, building a deep residual error shrinkage network, and training the deep residual error shrinkage network by using the simulation data set in the step 1;
step 3, collecting real turn-to-turn short circuit fault data of the permanent magnet synchronous motor through a current sensor, and establishing a real sample set;
step 4, building a generating countermeasure network, carrying out sample expansion on the real data set, combining the expanded sample with the real data sample, and building an experimental data set;
and 5, optimizing and modifying the pre-training depth residual error shrinkage network based on a sparse representation theory, and training the modified depth residual error shrinkage network on an experimental data set by a transfer learning method until the model is fitted.
And 6: and carrying out threshold analysis on each sample by using a shrinkage structure in the deep residual shrinkage network, thereby realizing noise reduction treatment and finally carrying out fault classification.
Compared with the prior art, the invention has the remarkable advantages that:
(1) The method uses the deep residual shrinkage network, has better characteristic extraction effect by adding the multilayer convolution layer compared with the traditional convolution network, and can effectively classify the motor faults by carrying out noise reduction treatment on the samples by analyzing the soft threshold value of each sample in the shrinkage structure.
(2) The invention uses the transfer learning method, greatly reduces the dependency of the deep residual shrinkage network on a large amount of fault data, can realize the diagnosis of the fault through a small amount of fault data, and realizes higher fault diagnosis accuracy.
(3) The method uses the generation countermeasure network, can reduce the influence of the unbalanced sample on the deep residual shrinkage network diagnosis by generating the data sample, and realizes the high-precision fault diagnosis.
(4) The invention uses the sparse theory, improves the noise resistance and overfitting resistance of the depth residual shrinkage network model, and improves the real-time performance of the system.
Drawings
FIG. 1 is a flowchart of a method for diagnosing an unbalanced sample turn-to-turn short circuit fault of a permanent magnet synchronous motor;
FIG. 2 is a parameter diagram of a deep residual shrinkage network according to the present invention;
FIG. 3 is a schematic diagram of a deep residual shrinkage network structure according to the present invention;
FIG. 4 is a schematic diagram of a deep residual shrinkage network structure according to the present invention;
FIG. 5 is a flowchart of training a deep residual shrinkage neural network based on deep migration learning according to the present invention.
Detailed Description
Specific embodiments of the present invention are described below in conjunction with the accompanying drawings in order for those skilled in the art to better understand the present invention. It should be noted that in the following description, a detailed description of known functions and designs will be omitted when it may obscure the main contents of the present invention. Referring to fig. 1, fig. 2, fig. 3, fig. 4, and fig. 5, the method for diagnosing the turn-to-turn short circuit fault of the unbalanced sample of the permanent magnet synchronous motor specifically includes the following steps:
1) Constructing a normal model and an inter-turn short circuit fault model according to a permanent magnet synchronous motor voltage equation and a mechanical motion equation: a modeling equation of a normal model of the permanent magnet synchronous motor:
mechanical equation of motion:
voltage equation:
wherein J is moment of inertia, T e For electromagnetic torque, T L As load torque, ω m Representing the motor speed, t is time, B is the motor damping coefficient, i a 、i b 、i c Representing the stator winding current of the machine, e a 、e b 、e c The three-phase winding flux linkage of the motor is represented, R represents stator winding resistance, L represents stator leakage inductance, and M represents stator mutual inductance.
The modeling equation of the turn-to-turn short circuit fault of the permanent magnet synchronous motor is as follows:
voltage equation:
mechanical equation of motion:
wherein, V a 、V b 、V c Representing the three-phase stator voltage, M, of a PMSM ab 、M ac Indicating motor isMutual inductance between A-phase stator winding and B, C phase winding in normal operation, R represents stator winding resistance, R represents f Represents the resistance of the fault section, mu represents the degree of fault (ratio of short-circuited section to total length of stator), i a 、i b 、i c Representing stator winding current, i f Indicating a fault section current, e a 、e b 、e c Representing a three-phase winding flux linkage, e f Representing a fault part flux linkage, T is time, L represents stator leakage inductance, M represents stator mutual inductance, J represents rotational inertia, and T e For electromagnetic torque, T L As load torque, ω m And B represents the motor rotating speed, and is a motor damping coefficient.
After a normal motor model and an inter-turn short circuit fault model are built, simulation data under the normal condition of the motor and under the conditions of 5% inter-turn short circuit, 10% inter-turn short circuit and 15% inter-turn short circuit are collected respectively. And each group is one thousand groups of stator three-phase current signal data with two period lengths to establish a simulation data set.
2) The method comprises the steps of constructing a deep residual shrinkage neural network model, selecting four groups of network blocks to construct a network, enabling a first network block to comprise three residual blocks (nine layers of convolution layers in total), enabling a second network block to comprise four residual blocks (twelve layers of convolution layers in total), enabling a third network block to comprise six residual blocks (eighteen layers of convolution layers in total), enabling a fourth network block to comprise three residual blocks (nine layers of convolution layers in total), enabling each residual block to comprise an identity mapping, returning to the previous module when the model is poor in operation effect, enabling the residual blocks to further comprise a shrinkage structure, enabling soft threshold analysis to be conducted on each input sample through an attention system, achieving noise reduction, setting network parameters such as learning rate, step length, convolution kernel size and filling size of the deep residual shrinkage network model, and selecting a cross entropy loss function.
3) The simulation data set is input to train the neural network, the cross entropy loss function is used for representing the error between the network output label and the expected target label, and the weight is updated layer by adopting a gradient descent method in the reverse propagation process, so that the loss function is reduced to the minimum. Wherein the cross entropy function is:
wherein p (x) i ) For the true data label probability distribution, q (x) i ) To fit the label probability distribution, m is the total number of samples, i represents the ith sample, loss represents the Loss function, H (p (x) i ),q(x i ) Represents a cross entropy loss function.
4) The method comprises the steps of collecting real normal data of a motor and 200 groups of turn-to-turn short circuit fault data of three degrees of 5%, 10% and 15% by using a current sensor to serve as a real data training set, wherein each group of data is a stator current noise-containing signal with two period lengths.
5) Constructing and generating a confrontation network model:
the result graph for generating the confrontation network model in the invention is shown in fig. 2, and the construction process specifically comprises the following steps:
(1) Two convolutional layers are selected to replace a full connection layer to construct a generator, so that generated data are more stable, and the discriminator is replaced by the full connection layer.
(2) The method comprises the steps of inputting random signals into a generator for signal expansion, inputting turn-to-turn short circuit fault data of different degrees into a discriminator to serve as a basis for discrimination, selecting an Adam optimization algorithm, continuously modifying weights among nodes of each layer and bias of each layer by using back propagation to achieve the effect of reducing errors, continuously updating parameters in repeated iteration to obtain a state under the lowest error, namely stopping iteration when the error value is the lowest error, and taking an optimization objective function for generating a countermeasure network as a criterion for the lowest error.
The final optimization objective function for generating the countermeasure network is as follows:
wherein: wherein
Indicating that the value of the parameter D to be optimized for the discriminator is maximum firstAnd then minimizing the value of the parameter G to be optimized for the generator. X-P data(x) Statistical distribution p representing x fit to the real data data I.e. x belongs to the real data. Z-P Z(Z) Indicating that z corresponds to the encoded statistical distribution Pz, which is a random number sampled from the encoded statistical distribution. D (x) is a probability value that the discriminator discriminates the real sample x as true, D (G (z)) is a probability value that the discriminator discriminates the generated sample G (z) as real, and the network satisfies the following condition:
1. finally, judging whether the true accuracy of the data is fifty percent by a discriminator of the trained model;
2. the false fault data generated by the generator is consistent with the fault sample distribution in the data set.
(3) The experiment randomly samples 60 groups of normal samples and 60 groups of fault samples with certain turn-to-turn short circuit from a real data training set every time, and the iteration is carried out for 60 times. When D (G (z)) =1, the generated model output data and the fault data reach the optimal state, and the network parameters at this time are recorded as the network model of the final data expansion.
(4) And respectively performing data expansion on the turn-to-turn short circuit fault data of different degrees by using the trained generation countermeasure network, extracting samples in the real data set according to different proportions to establish an unbalanced data set, and combining the data samples generated by the generation countermeasure network with the unbalanced data set to obtain a balanced data set. And putting the two types of data sets into a one-dimensional convolutional neural network at the same time, and analyzing whether the generated sample can solve the problem of sample imbalance. The sample set composition and the diagnosis results after model processing are as follows:
| sample set | A | II | III | Fourthly | Five are | Six ingredients |
| Accuracy of diagnosis | 82％ | 71％ | 77％ | 88％ | 87.7％ | 88.1％ |
From the experimental results, the method aims at the problems that in the actual situation, the collected fault data is insufficient and the samples are unbalanced, generates the countermeasure network and generates the samples, so that the influence of the unbalanced samples on the diagnosis result can be reduced, and the method has wide application prospect in fault diagnosis of the motor and other fault samples under the condition of unbalance.
6) Keeping the network parameters of the convolutional layer in the pre-trained deep residual shrinkage network unchanged, and adding a global average pooling layer in the convolutional layer and the fully-connected layer, thereby solving the problem of overlarge parameter quantity, and inputting a real data training setForward propagation is used for network training, and backward propagation loss function is added based on minimization l 2 And (3) the sparse representation of the norm improves the anti-noise capability of the network, and the network is trained until the loss function is smaller than a target value, so that the network model is completed. The addition is based on 2 The loss function after sparse representation of the norm is:
wherein j (w, b) is a neural network loss function, w is a network weight parameter, b is a network bias parameter,
expressed as the neural network loss function minimum, p (x) i ) For the true data label probability distribution, q (x) i ) To fit the label probability distribution, m represents the number of input signals, i represents the ith sample, λ is the sparse coefficient,
is the vector parameter W 2 And (4) norm.
Respectively constructing balanced and unbalanced data sets of collected normal and 5%, 10% and 15% turn-to-turn short circuit fault data of the permanent magnet synchronous motor, wherein the data sets are specifically composed as follows; the depth residual shrinkage network trained on the data sets by using a conventional method and the depth residual shrinkage network model in the method of the invention are used for diagnosing and classifying the real data sets, and the accuracy of the classification result is as follows:
according to the experimental process and the experimental results, aiming at the problem of unbalanced fault samples under the real condition, the invention expands the sample data by generating the confrontation network model to generate a balanced sample set, applies the deep residual shrinkage network after the simulation pre-training to the real data set by a transfer learning method, avoids the problem of insufficient model training caused by too little real fault data, and simultaneously introduces the sparse representation theory to carry out anti-noise treatment. The invention adopts the deep residual shrinkage network, and can autonomously realize the noise reduction treatment of the fault sample by using the shrinkage structure in the network, thereby further performing noise reduction optimization on the data set and finally realizing fault classification. The method can realize high-precision fault diagnosis, has wide application prospect in the field of fault diagnosis of motors and other simulatable machines, and provides a new method for fault diagnosis technology based on deep learning.
Claims (7)
1. A method for diagnosing turn-to-turn short circuit fault of an unbalanced sample of a permanent magnet synchronous motor is characterized by comprising the following steps of:
step 1, building a permanent magnet synchronous motor normal model and an inter-turn short circuit fault model through simulation software, adjusting the fault degree of the inter-turn short circuit fault model, respectively collecting stator current signals of different fault degrees of the permanent magnet synchronous motor normal model and the inter-turn short circuit model, and building a simulation data set;
step 2, building a deep residual error shrinkage network, and training the deep residual error shrinkage network by using the simulation data set in the step 1;
step 3, collecting real fault data of the turn-to-turn short circuit of the permanent magnet synchronous motor through a current sensor, and establishing a real sample set;
step 4, building a generating countermeasure network, carrying out sample expansion on the real data set, combining the expanded sample with the real data sample, and building an experimental data set;
and 5, optimizing and modifying the pre-training depth residual error network based on a sparse representation theory, and training the modified depth residual error shrinkage network on an experimental data set by a transfer learning method until the model is fitted.
Step 6: and carrying out threshold analysis on each sample by using a shrinkage structure in the deep residual shrinkage network, thereby realizing noise reduction treatment and finally carrying out fault classification.
2. The method for diagnosing the turn-to-turn short circuit fault of the unbalanced sample of the permanent magnet synchronous motor according to claim 1, wherein the simulation feature data based pre-training deep residual shrinkage network in the step 2 is pre-trained by utilizing a simulation data set through establishing the network, so that the corresponding features of the turn-to-turn short circuit fault of the permanent magnet synchronous motor in a stator current time domain signal are understood, and the noise reduction processing, the correct recognition and the fault degree classification of the data are completed.
3. The method for diagnosing the turn-to-turn short circuit fault of the unbalanced sample of the permanent magnet synchronous motor according to claim 1, wherein the real fault data in the step 3 are original noise-containing stator current time-domain signals of the permanent magnet synchronous motor, which are collected by a current sensor and not preprocessed.
4. The method for diagnosing the turn-to-turn short circuit fault of the unbalanced sample of the permanent magnet synchronous motor according to claim 1, wherein the objective function of the generative countermeasure network in the step 4 is as follows:
wherein: wherein
The method comprises the steps of firstly maximizing the value of a parameter D to be optimized of a discriminator and then minimizing the value of a parameter G to be optimized of a generator; X-P data(x) Statistical distribution p representing x fits to the real data data I.e. x belongs to real data; Z-P Z(Z) Indicating that z corresponds to the encoded statistical distribution Pz, z being a random number sampled from the encoded statistical distribution; logD (x) is the mapping function that discriminates the model, log (1-D (G (z)) is the mapping function that generates the model D (x) represents the true sample that the discriminator will use to representx is the probability value for discriminating the true sample, and D (G (z)) is the probability value for discriminating the generated sample G (z) as the true sample by the discriminator;
and the network satisfies the following conditions:
(1) Finally, a discriminator of the trained model judges whether the data is true or not, and the accuracy is fifty percent;
(2) The pseudo-fault data generated by the generator is consistent with the distribution of fault samples in the data set.
5. The method for diagnosing the turn-to-turn short circuit fault of the unbalanced sample of the permanent magnet synchronous motor according to claim 1, wherein the step 4 of optimizing and transforming the network based on the sparse representation theory is to add a minimization l to a loss function of a model according to the sparse representation principle in a training composition 2 Sparse representation of norms.
6. The method for diagnosing the turn-to-turn short circuit fault of the unbalanced sample of the permanent magnet synchronous motor according to claim 1, wherein in the method based on the transfer learning, the reconstructed model is trained on a real sample fault data set until the model is fitted, the model is loaded on the real sample data set by the method based on the transfer learning, parameters of a network layer of the model except a full connection layer are kept unchanged, and the training is continued until a loss function is smaller than a target value.
7. The method for diagnosing the turn-to-turn short circuit fault of the unbalanced sample of the permanent magnet synchronous motor according to claim 1, wherein the deep residual shrinkage network in the step 6 is provided with an attention mechanism and a soft thresholding module, wherein the attention mechanism is used for finding local useful information by scanning global information and enhancing the found local information to inhibit other useless redundant information; the soft thresholding is to set a threshold value, set the characteristic value with the absolute value lower than a certain threshold value to zero, and adjust other characteristics to zero, namely shrink; therefore, different thresholds are set for noise-containing samples of different levels, noise reduction processing of the samples is achieved, and finally fault classification is achieved.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210988049.7A CN115688553A (en) | 2022-08-17 | 2022-08-17 | Method for diagnosing turn-to-turn short circuit fault of unbalanced sample of permanent magnet synchronous motor |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210988049.7A CN115688553A (en) | 2022-08-17 | 2022-08-17 | Method for diagnosing turn-to-turn short circuit fault of unbalanced sample of permanent magnet synchronous motor |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115688553A true CN115688553A (en) | 2023-02-03 |
Family
ID=85061357
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210988049.7A Pending CN115688553A (en) | 2022-08-17 | 2022-08-17 | Method for diagnosing turn-to-turn short circuit fault of unbalanced sample of permanent magnet synchronous motor |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115688553A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116626491A (en) * | 2023-07-26 | 2023-08-22 | 江苏大学 | Five-phase permanent magnet motor turn-to-turn short circuit fault diagnosis method |
-
2022
- 2022-08-17 CN CN202210988049.7A patent/CN115688553A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116626491A (en) * | 2023-07-26 | 2023-08-22 | 江苏大学 | Five-phase permanent magnet motor turn-to-turn short circuit fault diagnosis method |
| CN116626491B (en) * | 2023-07-26 | 2023-11-03 | 江苏大学 | Five-phase permanent magnet motor turn-to-turn short circuit fault diagnosis method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112926728B (en) | Small sample turn-to-turn short circuit fault diagnosis method for permanent magnet synchronous motor | |
| CN112052796A (en) | Permanent magnet synchronous motor fault diagnosis method based on deep learning | |
| CN106682688A (en) | Pile-up noise reduction own coding network bearing fault diagnosis method based on particle swarm optimization | |
| CN115688553A (en) | Method for diagnosing turn-to-turn short circuit fault of unbalanced sample of permanent magnet synchronous motor | |
| CN116010900A (en) | Multi-scale feature fusion gearbox fault diagnosis method based on self-attention mechanism | |
| CN115345277A (en) | PMSM turn-to-turn short circuit diagnosis method based on time sequence generation countermeasure network | |
| CN115545070A (en) | Intelligent diagnosis method for unbalance-like bearing based on comprehensive balance network | |
| CN116071352A (en) | Method for generating surface defect image of electric power safety tool | |
| CN111766513B (en) | Capsule network-based variable-working-condition multi-fault diagnosis method for three-phase induction motor | |
| Radhi et al. | Faults diagnosis in stator windings of high speed solid rotor induction motors using fuzzy neural network | |
| Fu et al. | Development of deep convolutional neural network with adaptive batch normalization algorithm for bearing fault diagnosis | |
| CN116796597A (en) | Multi-objective optimization method suitable for multi-parameter coupling permanent magnet vernier rim propulsion motor | |
| CN113311364B (en) | Permanent magnet synchronous motor inverter open-circuit fault diagnosis method based on multi-core SVM | |
| CN116541980A (en) | Rotor punching size parameter optimization method | |
| Li et al. | GAN-DRSN based Inter-turn Short Circuit Fault Diagnosis of PMSM | |
| CN111859574B (en) | Synchronous reluctance motor rotor optimization design method for reducing torque pulsation | |
| CN113742638B (en) | Kurtosis-based STLBO motor bearing fault diagnosis method based on FastICA and approximation solution domain | |
| Zhao et al. | Fault diagnosis of asynchronous induction motor based on BP neural network | |
| CN113093052B (en) | Method for judging turn number of stator winding fault of turn-to-turn short circuit motor | |
| Shi et al. | Online detection method for inter-turn short-circuit fault of permanent magnet synchronous motor based on deep learning | |
| Fang et al. | deep transfer learning in inter-turn short circuit fault diagnosis of PMSM | |
| Ren et al. | Application of RBF neural network optimized based on K-means cluster algorithm in fault diagnosis | |
| Zhang et al. | fault diagnosis of servo drive system of CNC machine based on deep learning | |
| Dong et al. | A novel generative method for machine fault diagnosis | |
| CN116626491B (en) | Five-phase permanent magnet motor turn-to-turn short circuit fault diagnosis method |
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 |