Generator stator turn-to-turn short circuit fault diagnosis method, device, system and medium
By constructing a real data set and using dual-way wavelet convolution feature extraction and multi-sensor data fusion, the problem of insufficient data in generator stator interturn short circuit fault diagnosis is solved, the diagnostic accuracy and stability are improved, and the signal characteristics are enhanced. use.
Patent Information
- Application Number
- CN202411217187.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-02
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-09-02
AI Technical Summary
In the prior art, in the diagnosis of short-circuit between stator turns of generators, insufficient data leads to limited model diagnostic capabilities, and a single type of data cannot fully reflect the stator status, resulting in low diagnostic accuracy.
By constructing a real data set and using a dual-way wavelet convolution feature extraction module and a multi-sensor data feature fusion module, combining the generation model for data expansion and feature fusion, the diagnosis of generator stator interturn short circuit faults is realized.
It improves the accuracy and stability of generator stator-turn short circuit fault diagnosis, enhances the utilization of correlation, complementarity and redundancy of multi-sensor signals, and improves diagnostic capabilities.
Smart Images

Figure CN118731690B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of generator fault diagnosis, and in particular to a generator stator turn-to-turn short circuit fault diagnosis method, device, system and medium. Background Art
[0002] Permanent magnet synchronous motors are more widely used in industrial fields such as automobiles due to their excellent performance in precise control, high efficiency, high torque-to-current ratio and high power density. However, if they suddenly fail, it will cause huge economic losses. Therefore, it is very important to identify the fault early and avoid the adverse effects caused by the fault.
[0003] Stator turn-to-turn short circuit fault is one of the common faults of permanent magnet synchronous motors. It is caused by the degradation of stator winding insulation due to the combination of thermal, electrical and mechanical stresses. At present, the main motor fault diagnosis methods at home and abroad are mathematical model methods, signal analysis methods and artificial intelligence methods. In recent years, the stator turn-to-turn short circuit fault diagnosis method of permanent magnet synchronous motor based on data-driven and deep learning has received great attention. The generalization of data-driven methods depends on reliable and sufficient data sets. Under actual operating conditions, the motor is basically in a healthy working state, and usually only a very small number of fault signals can be collected. This makes the model that relies on a large amount of data for fault diagnosis face great challenges.
[0004] In the past few decades, diagnostic methods such as transfer learning, generative models, and oversampling have become the mainstream methods for dealing with small sample problems. However, in practical applications, it is difficult to obtain enough real data to train models, and generated data may result in poor data quality, limiting the diagnostic ability of the model.
[0005] In addition, the short-circuit fault between the stator turns of the generator will cause changes in multiple physical quantities. A single type of data cannot fully reflect the state of the stator. Through multi-sensor data feature fusion, the features of different types of data can be fused and used as a basis for more accurate and comprehensive fault judgment. Summary of the invention
[0006] In order to solve the above problems, the present invention provides a method, device, system and medium for diagnosing stator turn-to-turn short-circuit fault of a generator. On the basis of generating high-quality data expansion samples, a two-path wavelet method is used to extract data features of multiple types of sensors and then feature fusion and fault diagnosis are performed to achieve effective diagnosis of stator turn-to-turn short-circuit fault.
[0007] In order to achieve the above object, the present invention provides a generator stator turn-to-turn short circuit fault diagnosis method, device, system and medium, and the technical solution adopted is as follows:
[0008] According to a first technical solution of the present invention, a method for diagnosing a stator turn-to-turn short circuit fault of a generator is provided, the method comprising:
[0009] Constructing a real data set, wherein the real data set includes current signals and vibration signals under different operating states of the generator, wherein the different operating states of the generator include a normal operating state of the generator, a stator turn-to-turn short circuit fault operating state with a degree of 2%, a stator turn-to-turn short circuit fault operating state with a degree of 4%, a generator air gap eccentricity fault operating state with a degree of 3%, and a generator air gap eccentricity fault operating state with a degree of 6%;
[0010] Performing data expansion on the real data set;
[0011] A diagnostic model is constructed, the diagnostic model includes a dual-path wavelet convolution feature extraction module, a multi-sensor data feature fusion module and a fault identification module. The diagnostic model responds to the input current signal and vibration signal, extracts the first current feature, the second current feature and the first vibration feature and the second vibration feature through the dual-path wavelet convolution feature extraction module, and fuses the first current feature and the second current feature to obtain the current feature, and fuses the first vibration feature and the second vibration feature to obtain the vibration feature. The current feature and the vibration feature are fed into the multi-sensor data feature fusion module for feature fusion to obtain the fusion feature. The fault identification module uses the fusion feature as input and determines the generator operation state according to the input fusion feature. The diagnostic model is trained based on the expanded real data set to obtain the trained diagnostic model, and the generator stator turn-to-turn short circuit fault diagnosis is implemented based on the trained diagnostic model.
[0012] Furthermore, before constructing the real data set, the method further includes:
[0013] The variation characteristics of the current signal and the vibration signal when a stator inter-turn short circuit fault occurs are determined, and the real data set is calibrated based on the variation characteristics of the current signal and the vibration signal when a stator inter-turn short circuit fault occurs.
[0014] Furthermore, the change characteristics of the current signal and the vibration signal during the stator inter-turn short circuit fault are determined, specifically including:
[0015] The generator air gap flux density, phase current and electromagnetic force per unit area on the stator winding are determined according to the following formula:
[0016] ;
[0017] In the formula, , , They are respectively the air gap flux density, air gap magnetic potential and air gap permeability of the generator; is the circumferential angle of the air gap; t For time; is the phase current; a is the number of parallel winding branches; for Winding factor of subharmonics; L is the effective length of stator winding; p is the pole pair number; q is the number of slots per pole and per phase; N c is the number of turns of a single coil; is the rotation speed; R is the normal winding resistance; R s is the winding resistance; is the electromagnetic force per unit area on the stator winding; is the magnetic permeability of air;
[0018] When a stator inter-turn short circuit fault occurs, the current signal generates third harmonics. The amplitude of the current signal increases with the increase of the number of short-circuited turns, and the amplitude of the vibration signal also increases with the increase of the number of short-circuited turns.
[0019] Furthermore, data expansion is performed on the real data set, specifically including:
[0020] Constructing a generative model, wherein the generative model includes a generator and a discriminator;
[0021] The generator responds to the input random noise, maps the random noise into a feature space through a fully connected layer, a continuous multi-layer deconvolution module and a Tanh activation, and obtains a generated sample, wherein the generated sample is a vector with the same size as the real sample included in the real data set;
[0022] The discriminator takes the generated sample as input, extracts features through one-dimensional convolution, activates the activation function, and continuously performs multi-layer convolutional neural networks to obtain distribution features of the generated sample, and determines the authenticity of the generated sample based on the distance between the distribution features of the generated sample and the distribution features of the real sample;
[0023] The generative model is trained using the real data set, and samples are generated based on the trained generative model output, and the generated samples are added to the real data set to expand the real data set.
[0024] Furthermore, using the real data set to train the generative model specifically includes:
[0025] The real sample is input into the discriminator to obtain the real sample prediction result, and the loss function of the real data is calculated. The random noise is input into the generator, and the generated sample is fed into the discriminator to obtain the loss function of the generated sample, and the gradient penalty phase and the total loss function of the discriminator are calculated. After back-propagating the gradient, the discriminator parameters are updated.
[0026] Furthermore, the gradient penalty phase and the total loss function of the discriminator are calculated by the following formula:
[0027] ;
[0028] In the formula, GP is the gradient penalty phase; K To set a constant; G represents a generator, D represents the discriminator, x Represents real data, E represents mathematical expectation; , represents a uniform distribution from 0 to 1, Indicates the generation of data; for P norm; Indicates along the sampling point and The straight lines between are uniformly sampled; is the gradient operator; represents the gradient penalty factor; represents the objective function of WDCGAN-GP, Represents the output of the discriminator after discriminating the generated samples, D ( x ) represents the output of the discriminator after discriminating the real sample. Represents the generated data output by the generator.
[0029] Furthermore, the dual-path wavelet convolution feature extraction module includes a first branch, a second branch and a feature fusion layer, the first branch includes a first convolution layer, the second branch includes a second convolution layer, two consecutive third convolution layers and a fourth convolution layer connected in sequence, the first convolution layer and the fourth convolution layer are both connected to the feature fusion layer, the first convolution layer includes a first wavelet convolution, a first BatchNorm layer and a first maximum pooling layer connected in sequence; the second convolution layer includes a second wavelet convolution, a second BatchNorm layer and a second maximum pooling layer connected in sequence, the third convolution layer includes a third convolution, a third BatchNorm layer, a third ReLU layer and a third maximum pooling layer connected in sequence, and the fourth convolution layer includes a fourth convolution, a fourth BatchNorm layer, a fourth ReLU layer and a fourth maximum pooling layer connected in sequence.
[0030] According to a second technical solution of the present invention, a generator stator turn-to-turn short circuit fault diagnosis device is provided, the device comprising:
[0031] a data set construction module, configured to construct a real data set, wherein the real data set includes current signals and vibration signals under different operating states of the generator, wherein the different operating states of the generator include a normal operating state of the generator, an operating state of a stator turn-to-turn short circuit fault with a degree of 2%, an operating state of a stator turn-to-turn short circuit fault with a degree of 4%, an operating state of a generator air gap eccentricity fault with a degree of 3%, and an operating state of a generator air gap eccentricity fault with a degree of 6%;
[0032] A data set expansion module, configured to perform data expansion on the real data set;
[0033] A fault diagnosis module is configured to construct a diagnosis model, which includes a dual-path wavelet convolution feature extraction module, a multi-sensor data feature fusion module and a fault identification module. The diagnosis model responds to the input current signal and vibration signal, extracts the first current feature, the second current feature and the first vibration feature and the second vibration feature through the dual-path wavelet convolution feature extraction module, and fuses the first current feature and the second current feature to obtain the current feature, and fuses the first vibration feature and the second vibration feature to obtain the vibration feature. The current feature and the vibration feature are fed into the multi-sensor data feature fusion module for feature fusion to obtain the fusion feature. The fault identification module uses the fusion feature as input and determines the generator operation state according to the input fusion feature. Based on the expanded real data set, the diagnosis model is trained to obtain the trained diagnosis model, and the generator stator turn-to-turn short circuit fault diagnosis is implemented based on the trained diagnosis model.
[0034] According to a third technical solution of the present invention, a generator stator turn-to-turn short-circuit fault diagnosis system is provided, the system comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the method described above.
[0035] According to a fourth technical solution of the present invention, a non-transitory computer-readable storage medium storing instructions is provided. When the instructions are executed by a processor, the method described above is executed.
[0036] The generator stator turn-to-turn short circuit fault diagnosis method, device, system and medium according to each solution of the present invention have at least the following technical effects:
[0037] 1. Introducing multiple continuous convolutional layers and deconvolutional layers improves the network's ability to extract data features.
[0038] 2. The Wasserstein distance is used to measure the distance between the actual data distribution and the generated data distribution to ensure the convergence of the model.
[0039] 3. In order to avoid the limitation of weight range, the gradient penalty phase (GP) is introduced. On the basis of satisfying the continuity condition of the equation, the norm is optimized by restricting the gradient.
[0040] 4. Lpalace wavelet transform is used to obtain time-frequency features from the original vibration signal and current signal with spatiotemporal characteristics to maximize the feature extraction capability of subsequent convolutional neural network.
[0041] 5. Dual-way convolution can be used to mine the multi-scale and multi-temporal and spatial characteristics of data.
[0042] 6. Vibration and current sensor data are selected, and the time domain data is extracted using a dual-path wavelet convolutional network. The features obtained by fusing the current and vibration features are used as the basis for diagnosis, which effectively integrates the correlation, complementarity and redundancy between sensor signals and improves the accuracy of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In the drawings, which are not necessarily drawn to scale, the same reference numerals may describe similar parts in different views. The same reference numerals with letter suffixes or different letter suffixes may represent different instances of similar parts. The accompanying drawings generally illustrate various embodiments by way of example and not limitation, and together with the specification and claims, are used to illustrate the embodiments of the invention. When appropriate, the same reference numerals are used throughout the drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the present apparatus or method.
[0044] Figure 1A flow chart of a generator stator turn-to-turn short circuit fault diagnosis method according to an embodiment of the present invention is shown.
[0045] Figure 2 A flow chart of performing data expansion on a real data set according to an embodiment of the present invention is shown.
[0046] Figure 3 A structural principle diagram of a generator of a generation model according to an embodiment of the present invention is shown.
[0047] Figure 4 A structural principle diagram of a discriminator of a generative model according to an embodiment of the present invention is shown.
[0048] Figure 5 A training flowchart of a generative model according to an embodiment of the present invention is shown.
[0049] Figure 6 A schematic diagram of the network structure of a diagnosis model according to an embodiment of the present invention is shown.
[0050] Figure 7 A structural diagram of a generator stator turn-to-turn short circuit fault diagnosis device according to an embodiment of the present invention is shown.
[0051] Figure 8 A structural diagram of a generator stator turn-to-turn short circuit fault diagnosis device according to an embodiment of the present invention is shown when the device includes a theoretical analysis module. DETAILED DESCRIPTION
[0052] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and specific embodiments, but are not intended to limit the present invention. For the various steps described herein, if there is no necessity for a causal relationship between each other, the order in which they are described as examples herein should not be regarded as a limitation, and those skilled in the art should know that they can be adjusted in order, as long as the logic between them is not destroyed, resulting in the inability to implement the entire process.
[0053] The present invention provides a method for diagnosing a stator turn-to-turn short circuit fault of a generator. Figure 1 , is a flow chart of a method for diagnosing a generator stator turn-to-turn short circuit fault according to an embodiment of the present invention. The method comprises the following steps:
[0054] Step S1, constructing a real data set, wherein the real data set includes current signals and vibration signals under different operating states of the generator, wherein the different operating states of the generator include a normal operating state of the generator, a stator turn-to-turn short circuit fault operating state with a degree of 2%, a stator turn-to-turn short circuit fault operating state with a degree of 4%, a generator air gap eccentricity fault operating state with a degree of 3%, and a generator air gap eccentricity fault operating state with a degree of 6%.
[0055] Exemplarily, in this embodiment, 350 sample points are collected at a sampling frequency of 10 kHz in a ratio of 13:6:6:5:5, including the normal operation state of the generator, the operation state of 2% stator turn-to-turn short circuit fault, the operation state of 4% stator turn-to-turn short circuit fault, the operation state of 3% generator air gap eccentricity fault, and the operation state of 6% generator air gap eccentricity fault, and the real data is input into the generation model to expand the data set. The expanded data sets of the five working states are divided into training sets and test sets according to a ratio of 3:1. The training set and the test set can be used for training and testing the model in the diagnosis step, respectively. The training set is used for diagnostic model training in subsequent steps, and the test set is used to verify the generation quality of the generated model and the classification effect of the diagnostic model. At the same time, the division of the training set and the test set can also be performed after the real data set is expanded, and the division ratio of 3:1 mentioned here in this embodiment is only an example, not a limitation of the present invention.
[0056] In some embodiments, in step S1, before constructing the real data set, the method further includes: determining the change characteristics of the current signal and the vibration signal during the stator inter-turn short circuit fault, and calibrating the real data set based on the change characteristics of the current signal and the vibration signal during the stator inter-turn short circuit fault.
[0057] Exemplarily, this embodiment takes the inter-turn short circuit fault of the A-phase winding of the permanent magnet synchronous generator as an example, and derives the changes of the stator phase current and the stator vibration during the fault.
[0058] The generator air gap flux density, phase current and electromagnetic force per unit area on the stator winding are determined according to the following formula:
[0059] ;
[0060] In the formula, , , They are respectively the air gap flux density, air gap magnetic potential and air gap permeability of the generator; is the circumferential angle of the air gap; t For time; is the phase current; a is the number of parallel winding branches; for Winding factor of subharmonics; Lis the effective length of stator winding; p is the pole pair number; q is the number of slots per pole and per phase; N c is the number of turns of a single coil; is the rotation speed; R is the normal winding resistance; R s is the winding resistance; is the electromagnetic force per unit area on the stator winding; is the magnetic permeability of air;
[0061] When the A-phase winding turns are short-circuited, the number of equivalent winding turns of this phase is reduced, the three-phase winding is no longer symmetrical, and the magnetic potential of the three-phase winding can no longer offset each other, so the third harmonic will appear. And as the number of short-circuited turns increases and the fault severity increases, the amplitude of the third harmonic will increase. From the above formula, it can be seen that the phase current is proportional to the magnetic potential. Therefore, when the stator turns are short-circuited, the phase current will also generate the third harmonic, and the amplitude will increase with the deepening of the fault severity.
[0062] It can be seen from the electromagnetic force per unit area of the stator winding that when a turn-to-turn short circuit occurs in the generator, the electromagnetic force is proportional to the square of the air gap magnetic potential. Therefore, the electromagnetic force will generate sixth harmonics, which will cause the stator winding to vibrate, and the vibration amplitude will increase with the increase of the turn-to-turn short circuit.
[0063] Therefore, it can be determined that the changing characteristics of the current signal and vibration signal during a stator inter-turn short-circuit fault are as follows: during a stator inter-turn short-circuit fault, the current signal generates third harmonics, the amplitude of the current signal increases with the increase of the number of short-circuited turns, and the amplitude of the vibration signal also increases with the increase of the number of short-circuited turns.
[0064] The above derivation not only verifies that the generator stator inter-turn short-circuit fault diagnosis can be performed based on the current signal and the vibration signal, but also can be used to calibrate the real samples in the real data set based on the change characteristics of the current signal and the vibration signal during the determined stator inter-turn short-circuit fault, wherein the real sample includes the current signal and the vibration signal and their corresponding labels, and the label is used to indicate the engine running state of the real sample, such as the real sample corresponding to the stator inter-turn short-circuit fault running state, by judging whether the current signal and the vibration signal in the real sample meet the above change characteristics. If they meet, it indicates the correctness of the real sample. If they do not meet, it indicates that the real sample has an error, and the cause of the error may be sensor failure, data entry error, etc. The wrong real sample can be removed from the real data set to ensure the correctness of the real sample.
[0065] Step S2, performing data expansion on the real data set.
[0066] In an exemplary embodiment, a small sample data set is expanded by a generative method.
[0067] Specifically, if Figure 2 As shown, the process of data expansion for the real data set includes the following steps:
[0068] Step S201, constructing a generative model, wherein the generative model includes a generator and a discriminator.
[0069] In step S202, the generator responds to the input random noise, maps the random noise into a feature space through a fully connected layer, a continuous multi-layer deconvolution module, normalization, and Tanh activation, and obtains a generated sample, where the generated sample is a vector having the same size as the real sample included in the real data set.
[0070] In step S203, the discriminator takes the generated sample as input, extracts features through one-dimensional convolution, activates the activation function, and continuously performs multi-layer convolutional neural networks to obtain the distribution characteristics of the generated sample, and determines the authenticity of the generated sample based on the distance between the distribution characteristics of the generated sample and the distribution characteristics of the real sample.
[0071] Step S204: train the generative model using the real data set, generate samples based on the trained generative model output, and add the generated samples to the real data set to expand the real data set.
[0072] In an exemplary embodiment, see Figure 3 and Figure 4 , respectively, are the structural diagrams of the generator and the discriminator, Figure 3 ① in the figure represents a deconvolutional network layer in the generator. There are five deconvolutional network layers in the generator. The data processing process of each deconvolutional network layer includes 4×2×1 deconvolution, BatchNorm normalization and Tanh activation in sequence. Figure 4 ② in the figure represents a convolutional network layer in the discriminator. The discriminator has four convolutional network layers. The data processing process of each convolutional network layer includes 4×2×1 one-dimensional convolution, InastanceNorm normalization and LeakyReLU activation in sequence.
[0073] Specifically, Figure 3As shown in the figure, in the generator, 1×100 random noise is used as input, passes through the fully connected layer, and then passes through 5 consecutive layers of 4×2×1 deconvolution modules, BatchNorm standardization, and ReLU activation. The input raw data is mapped to a larger feature space, and the generator's representation ability is improved, which can capture more comprehensive information of the data. Finally, after activation by the Tanh activation function, the original noise input to the generator is converted into a vector with the same size as the real sample, that is, the generated sample is obtained, which includes the current signal and the vibration signal.
[0074] like Figure 4 As shown in the figure, in the discriminator, the input is the current / vibration signal in the generated sample, which is firstly subjected to a 4×2×1 one-dimensional convolution to extract features and activated by the LeakyReLU activation function, and then passed through 4 consecutive layers of 4×2×1 convolutional neural networks. Finally, a fully connected layer is used to map the previously learned distributed feature representation to the sample label space to determine whether the data is real or not.
[0075] The discriminator determines whether the data is real or not by judging the distance between the distributed features of the generated sample and the real sample, and judges the authenticity of the generated sample based on the distance and the set distance threshold. For example, the set distance threshold is [a, b]. If the distance d ∈[a, b], then the generated sample is determined to be real data and can be added to the real data set as a real sample to expand the real data set. Otherwise, it is determined that the generated sample cannot be used as a sample to expand the real data set.
[0076] In an exemplary embodiment, Figure 5 As shown in the figure, it is a training flow chart for training the generative model using the real data set. Before training, the initial learning rate, convolution kernel size, step size and other hyperparameters of the generative model are set, and Adam is selected as the optimization algorithm. Referring to the WGAN-GP model, the Wasserstein distance is introduced to measure the distance between the real sample distribution and the generated sample distribution. In order to avoid the limitation of the weight range, the gradient penalty phase (GP) is introduced, and its definition can be expressed as:
[0077] ;
[0078] Introduced in the generative model GP It not only satisfies the continuity condition of the equation, but also optimizes the norm by limiting the gradient. The optimization goal of the generative model is to make the data generated by the generator as similar as possible to the real sample while making the judgment result of the discriminator as accurate as possible. The total loss function of the discriminator can be expressed as:
[0079] ;
[0080] In the formula, GP is the gradient penalty phase; K To set a constant, usually 1; G represents a generator, D represents the discriminator, x Represents real data, E represents mathematical expectation; , represents a uniform distribution from 0 to 1, Indicates the generation of data; for P norm; Indicates along the sampling point and The straight lines between are uniformly sampled; is the gradient operator; represents the gradient penalty factor; represents the objective function of WDCGAN-GP, Represents the output of the discriminator after discriminating the generated samples, D ( x ) represents the output of the discriminator after discriminating the real sample. Represents the generated data output by the generator.
[0081] Combination Figure 5 , the training process of the generative model includes training the discriminator and training the generator. The specific process is as follows:
[0082] First, train the discriminator. Since the gradient of the model parameters will accumulate during the training process, in order to prevent the gradient in the current training process from being affected by the previous data, the gradient must be cleared first. Then input the real sample into the discriminator model, obtain the real sample judgment result, and calculate the loss function of the real sample. Next, input random noise into the generative model, obtain the output generated sample, and then input it into the discriminant model to obtain the loss function of the generated data, and calculate the gradient penalty phase GP and the total loss function of the discriminator. After back-propagating the gradient, update the discriminator parameters.
[0083] Finally, train the generator. At the beginning of the training, the gradient is cleared, random noise is input to generate data, and the discriminator is used to obtain the discrimination result. The loss between the real label and the generated data is calculated, and the discriminator parameters are updated through back propagation. Finally, the generated data can successfully deceive the discriminator, so that the discriminator can judge the generated data as real data on the basis of high accuracy judgment.
[0084] Based on the trained generative model, a large number of generated samples close to real samples can be generated. In specific implementation, according to the actual situation, a certain type of real samples with fewer samples in the real data can be generated in a targeted manner, so that the types with fewer samples in the real data set can be expanded, and finally high-quality sample data is generated after 5000 iterations, so that the sample size of each type of data is balanced. The sample size of each type of data mentioned here refers to the number of real samples of different categories in the real data set, and the real samples of different categories refer to the current signal and vibration signal under different operating conditions of the generator.
[0085] Step S3, constructing a diagnostic model, training the diagnostic model based on the expanded real data set to obtain a trained diagnostic model, and implementing generator stator turn-to-turn short circuit fault diagnosis based on the trained diagnostic model.
[0086] Among them, the diagnostic model includes a dual-path wavelet convolution feature extraction module, a multi-sensor data feature fusion module and a fault identification module. The diagnostic model responds to the input current signal and vibration signal, extracts the first current feature, the second current feature and the first vibration feature and the second vibration feature through the dual-path wavelet convolution feature extraction module, and fuses the first current feature and the second current feature to obtain the current feature, and fuses the first vibration feature and the second vibration feature to obtain the vibration feature. The current feature and the vibration feature are fed into the multi-sensor data feature fusion module for feature fusion to obtain the fusion feature. The fault identification module uses the fusion feature as input and judges the operating status of the generator based on the input fusion feature.
[0087] In the dual-path wavelet convolution feature extraction module, the expanded one-dimensional vibration signal and current signal are used as the input of the diagnostic model, and richer features are extracted after dual-path wavelet convolution and a series of convolution modules. Since the input of the diagnostic model of the present invention is a vibration signal and a current signal, the Laplace wavelet basis function is selected as the analysis signal. The number of parameters of the Laplace wavelet convolution kernel is significantly less than that of the ordinary convolution parameter. On the basis of effectively extracting features, the complexity and amount of calculation of the network are reduced. At the same time, Laplace wavelet convolution also has the characteristics of multi-scale feature extraction and edge retention, which can better capture some details and overall features. In addition, the feature extraction of the dual branches can extract feature information from multiple levels. The features extracted from each of the two branches can be fused to express the data features more richly, making the model more distinguishable.
[0088] Stator turn-to-turn short circuit faults can cause changes in multiple physical quantities. Using only single sensor data as the signal source will lead to inaccurate diagnostic results when subject to external interference or changes in environmental conditions or operating conditions. In order to improve the accuracy of diagnosis, this paper uses vibration sensors and current sensors to obtain vibration signals and current signals as the basis for diagnosis. The features of the two types of signals are extracted and fused as the basis for fault identification, which helps to conduct a more detailed analysis of the generator status and ensure the stability and accuracy of the diagnosis.
[0089] In an exemplary embodiment, the dual-path wavelet convolution feature extraction module includes a first branch, a second branch and a feature fusion layer, the first branch includes a first convolution layer, the second branch includes a second convolution layer, two consecutive third convolution layers, a fourth convolution layer and a fifth convolution layer connected in sequence, the first convolution layer and the fifth convolution layer are both connected to the feature fusion layer, the first convolution layer includes a first wavelet convolution, a first BatchNorm layer and a first maximum pooling layer connected in sequence; the second convolution layer includes a second wavelet convolution, a second BatchNorm layer and a second maximum pooling layer connected in sequence, the third convolution layer includes a third convolution, a third BatchNorm layer, a third ReLU layer and a third maximum pooling layer connected in sequence, and the fourth convolution layer includes a fourth convolution, a fourth BatchNorm layer, a fourth ReLU layer and a fourth maximum pooling layer connected in sequence.
[0090] It is easy to understand that there is one convolution layer in the first branch, and four consecutive convolution layers in the second branch. Except for the convolution layer in the first branch and the second convolution layer in the second branch (that is, the third convolution layer in the five consecutive convolution layers in the second branch), which do not contain activation functions, the other convolution layers include sequentially connected convolution, BatchNorm layer, ReLU layer and maximum pooling layer, and the convolution kernel sizes of each convolution layer are not consistent.
[0091] Specifically, if Figure 6 As shown, the expanded real data set is used as the input of the diagnostic model for feature extraction through two branches. In order to extract feature information at different levels of the data, the network structures of the two branches in this embodiment are not consistent. The network structure of the first branch is a single-layer 18×2 wavelet convolution. The data in the second branch first passes through a 10×2 wavelet convolution layer, and then passes through two consecutive 8×1×1 convolution layers, and finally passes through an 8×2 convolution layer. Except for the 18×2 wavelet convolution module in the first branch and the 10×2 wavelet convolution module in the second branch, which only contain convolution layers, BatchNorm layers, and 1×2 maximum pooling layers, other convolution modules include convolution layers, BatchNorm layers, ReLU activation functions, and 1×2 maximum pooling layers. Figure 6The 8×1×1 convolutional layer that appears twice in a row is represented by ③. The vibration features (current features) extracted from the two branches become fused features with unchanged data types after being processed by the feature fusion layer. The fused features are further extracted by the GRU layer and sent to the adaptive average pooling layer for processing. The extracted vibration features and current features are fused by the feature layer to become fused features that contain both types of information. Finally, the fused features are mapped to the output category through the fully connected layer to realize the identification of the generator operating status and achieve the purpose of accurate diagnosis of stator inter-turn short-circuit faults.
[0092] In summary, the embodiment of the present invention provides a method for diagnosing short-circuit faults between turns of a generator when fault data is insufficient. The data set is expanded by a method based on WDCGAN-GP (i.e., the method of the present invention introduces GP and generates a model by training a real training set, corresponding to step S2 above). Compared with GAN, DCGAN, WGAN-GP and other methods, the method of the present invention improves the quality of generated data on the basis of ensuring the stability of model training. A large amount of nearly real data can be generated based on a small amount of fault data, and the time-frequency characteristics of the original sample are retained to a great extent. The dual-path wavelet convolutional neural network can well extract the features of the data and perform feature fusion on the collected multi-sensor signals. The fused features can fully reflect the status and performance of the generator, and enhance the comprehensiveness of the diagnostic information.
[0093] The present invention also provides a generator stator turn-to-turn short circuit fault diagnosis device, see Figure 7 , is a structural diagram of a generator stator turn-to-turn short-circuit fault diagnosis device, the device comprising:
[0094] The data set construction module 701 is configured to construct a real data set, wherein the real data set includes current signals and vibration signals under different operating states of the generator, wherein the different operating states of the generator include a normal operating state of the generator, a stator turn-to-turn short circuit fault operating state with a degree of 2%, a stator turn-to-turn short circuit fault operating state with a degree of 4%, a generator air gap eccentricity fault operating state with a degree of 3%, and a generator air gap eccentricity fault operating state with a degree of 6%;
[0095] A data set expansion module 702 is configured to perform data expansion on the real data set;
[0096] The fault diagnosis module 703 is configured to construct a diagnosis model, which includes a dual-path wavelet convolution feature extraction module, a multi-sensor data feature fusion module and a fault identification module. The diagnosis model responds to the input current signal and vibration signal, extracts the first current feature, the second current feature and the first vibration feature and the second vibration feature through the dual-path wavelet convolution feature extraction module, and fuses the first current feature and the second current feature to obtain the current feature, and fuses the first vibration feature and the second vibration feature to obtain the vibration feature. The current feature and the vibration feature are fed into the multi-sensor data feature fusion module for feature fusion to obtain the fusion feature. The fault identification module uses the fusion feature as input and determines the generator operation state according to the input fusion feature. Based on the expanded real data set, the diagnosis model is trained to obtain the trained diagnosis model, and the generator stator turn-to-turn short circuit fault diagnosis is implemented based on the trained diagnosis model.
[0097] In some embodiments, Figure 8 As shown, the apparatus further includes a theoretical analysis module 704, and the theoretical analysis module 704 is configured to:
[0098] The variation characteristics of the current signal and the vibration signal when a stator inter-turn short circuit fault occurs are determined, and the real data set is calibrated based on the variation characteristics of the current signal and the vibration signal when a stator inter-turn short circuit fault occurs.
[0099] In some embodiments, the theoretical analysis module 704 is further configured to:
[0100] The generator air gap flux density, phase current and electromagnetic force per unit area on the stator winding are determined according to the following formula:
[0101] ;
[0102] In the formula, , , They are respectively the air gap flux density, air gap magnetic potential and air gap permeability of the generator; is the circumferential angle of the air gap; t For time; is the phase current; a is the number of parallel winding branches; for Winding factor of subharmonics; L is the effective length of stator winding; p is the pole pair number; q is the number of slots per pole and per phase; N c is the number of turns of a single coil; is the rotation speed; Ris the normal winding resistance; R s is the winding resistance; is the electromagnetic force per unit area on the stator winding; is the magnetic permeability of air;
[0103] When a stator inter-turn short circuit fault occurs, the current signal generates third harmonics, the amplitude of the current signal increases with the increase of the number of short-circuited turns, and the amplitude of the vibration signal increases with the increase of the number of short-circuited turns.
[0104] In some embodiments, the data set expansion module 702 is further configured to:
[0105] Constructing a generative model, wherein the generative model includes a generator and a discriminator;
[0106] The generator responds to the input random noise, maps the random noise into a feature space through a fully connected layer, a continuous multi-layer deconvolution module, normalization, and ReLU activation to obtain a generated sample, wherein the generated sample is a vector with the same size as the real sample included in the real data set;
[0107] The discriminator takes the generated sample as input, extracts features through one-dimensional convolution, activates the activation function, and continuously performs multi-layer convolutional neural networks to obtain distribution features of the generated sample, and determines the authenticity of the generated sample based on the distance between the distribution features of the generated sample and the distribution features of the real sample;
[0108] The generative model is trained using the real data set, and samples are generated based on the trained generative model output, and the generated samples are added to the real data set to expand the real data set.
[0109] In some embodiments, the data set expansion module 702 is further configured to:
[0110] Input the real sample into the discriminator, obtain the real sample discrimination result, and calculate the loss function of the real data, input the random noise into the generator, feed the obtained generated sample into the discriminator, obtain the loss function of the generated sample, and calculate the gradient penalty phase and the total loss function of the discriminator, and update the discriminator parameters after back-propagating the gradient; in some embodiments, the data set expansion module 702 is further configured as follows:
[0111] The gradient penalty phase and the total loss function of the discriminator are calculated by the following formula:
[0112] ;
[0113] In the formula, GP is the gradient penalty phase;K To set a constant; G represents a generator, D represents the discriminator, x Represents real data, E represents mathematical expectation; , represents a uniform distribution from 0 to 1, Indicates the generation of data; for P norm; Indicates along the sampling point and The straight lines between are uniformly sampled; is the gradient operator; represents the gradient penalty factor; represents the objective function of WDCGAN-GP, Represents the output of the discriminator after discriminating the generated samples, D ( x ) represents the output of the discriminator after discriminating the real sample. Represents the generated data output by the generator.
[0114] In some embodiments, the dual-path wavelet convolution feature extraction module includes a first branch, a second branch, and a feature fusion layer, the first branch includes a first convolution layer, the second branch includes a second convolution layer connected in sequence, two consecutive third convolution layers, and a fourth convolution layer, the first convolution layer and the fourth convolution layer are both connected to the feature fusion layer, the first convolution layer includes a first wavelet convolution, a first BatchNorm layer, and a first maximum pooling layer connected in sequence; the second convolution layer includes a second wavelet convolution, a second BatchNorm, and a second maximum pooling layer connected in sequence, the third convolution layer includes a third convolution, a third BatchNorm layer, a third ReLU layer, and a third maximum pooling layer connected in sequence, and the fourth convolution layer includes a fourth convolution, a fourth BatchNorm layer, a fourth ReLU layer, and a fourth maximum pooling layer connected in sequence. In the diagnostic flowchart, in the embodiment of the present invention, the third convolution that appears twice is represented by ③ to reduce the length of this embodiment.
[0115] The generator stator inter-turn short circuit fault diagnosis device mentioned in the embodiment of the present invention belongs to the same technical concept as the method described previously, and the technical effects they achieve are basically the same, which will not be repeated here.
[0116] An embodiment of the present invention further provides a generator stator turn-to-turn short circuit fault diagnosis system, the system comprising:
[0117] Memory for storing computer programs;
[0118] A processor is used to execute the computer program to implement the generator stator turn-to-turn short circuit fault diagnosis method according to any embodiment of the present invention.
[0119] An embodiment of the present invention further provides a non-transitory computer-readable medium storing instructions, and when the instructions are executed by a processor, the generator stator turn-to-turn short circuit fault diagnosis method according to any embodiment of the present invention is executed.
[0120] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been explained in detail with reference to the above embodiments, it should be understood by those skilled in the art that the technical solutions recorded in the above embodiments may be modified, or some of the technical supports therein may be replaced with equivalent features; and these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the invention embodiments, and should be included in the protection scope of the present invention.
Claims
1. A method for diagnosing a generator stator turn-to-turn short circuit fault, characterized in that: The method comprises: Constructing a real data set, wherein the real data set includes current signals and vibration signals under different operating states of the generator, wherein the different operating states of the generator include a normal operating state of the generator, a stator turn-to-turn short circuit fault operating state with a degree of 2%, a stator turn-to-turn short circuit fault operating state with a degree of 4%, a generator air gap eccentricity fault operating state with a degree of 3%, and a generator air gap eccentricity fault operating state with a degree of 6%; Performing data expansion on the real data set; Constructing a diagnostic model, the diagnostic model includes a dual-path wavelet convolution feature extraction module, a multi-sensor data feature fusion module and a fault identification module, the diagnostic model responds to the input current signal and vibration signal, extracts the first current feature, the second current feature and the first vibration feature and the second vibration feature through the dual-path wavelet convolution feature extraction module, and fuses the first current feature and the second current feature to obtain the current feature, and fuses the first vibration feature and the second vibration feature to obtain the vibration feature, the current feature and the vibration feature are fed into the multi-sensor data feature fusion module for feature fusion to obtain the fusion feature, the fault identification module uses the fusion feature as input, and judges the generator operation state according to the input fusion feature, trains the diagnostic model based on the expanded real data set to obtain the trained diagnostic model, and implements the generator stator turn-to-turn short circuit fault diagnosis based on the trained diagnostic model; Before constructing the real data set, the method further includes: Determine the change characteristics of the current signal and the vibration signal when the stator inter-turn short circuit fault occurs, and calibrate the real data set based on the change characteristics of the current signal and the vibration signal when the stator inter-turn short circuit fault occurs; Determine the changing characteristics of the current signal and vibration signal during stator turn-to-turn short circuit fault, including: The generator air gap flux density, phase current and electromagnetic force per unit area on the stator winding are determined according to the following formula: ; In the formula, , , They are respectively the air gap flux density, air gap magnetic potential and air gap permeability of the generator; is the circumferential angle of the air gap; t For time; is the phase current; a is the number of parallel winding branches; for Winding factor of subharmonics; L is the effective length of stator winding; p is the pole pair number; q is the number of slots per pole and per phase; N c is the number of turns of a single coil; is the rotation speed; R is the normal winding resistance; R s is the winding resistance; is the electromagnetic force per unit area on the stator winding; is the magnetic permeability of air; When the stator turns are short-circuited, the current signal generates the third harmonic, the amplitude of the current signal increases with the increase of the number of short-circuited turns, and the amplitude of the vibration signal increases with the increase of the number of short-circuited turns; The real data set is expanded, specifically including: Constructing a generative model, wherein the generative model includes a generator and a discriminator; The generator responds to the input random noise, maps the random noise into a feature space through a fully connected layer, a continuous multi-layer deconvolution module, normalization, and Tanh activation, and obtains a generated sample, wherein the generated sample is a vector with the same size as the real sample included in the real data set; The discriminator takes the generated sample as input, extracts features through one-dimensional convolution, activates the activation function, and continuously performs multi-layer convolutional neural networks to obtain distribution features of the generated sample, and determines the authenticity of the generated sample based on the distance between the distribution features of the generated sample and the distribution features of the real sample; Using the real data set to train the generative model, and based on the trained generative model output, generating samples, and adding the generated samples to the real data set to expand the real data set; Using the real data set to train the generative model specifically includes: Input the real sample into the discriminator, obtain the real sample judgment result, and calculate the loss function of the real data, input random noise into the generator, feed the generated sample to the discriminator, obtain the loss function of the generated sample, and calculate the gradient penalty phase and the total loss function of the discriminator, and update the discriminator parameters after back-propagating the gradient; The gradient penalty phase and the total loss function of the discriminator are calculated by the following formula: ; In the formula, GP is the gradient penalty phase; K To set a constant; G represents a generator, D represents the discriminator, x Represents real data, E represents mathematical expectation; , represents a uniform distribution from 0 to 1, Indicates the generation of data; for P norm; Indicates along the sampling point and The straight lines between are uniformly sampled; is the gradient operator; represents the gradient penalty factor; represents the objective function of WDCGAN-GP, Represents the output of the discriminator after discriminating the generated samples, D ( x ) represents the output of the discriminator after discriminating the real sample. Represents the generated data output by the generator.
2. The method according to claim 1, characterized in that The dual-path wavelet convolution feature extraction module includes a first branch, a second branch and a feature fusion layer, the first branch includes a first convolution layer, the second branch includes a second convolution layer, two consecutive third convolution layers and a fourth convolution layer connected in sequence, the first convolution layer and the fourth convolution layer are both connected to the feature fusion layer, the first convolution layer includes a first wavelet convolution, a first BatchNorm layer and a first maximum pooling layer connected in sequence; the second convolution layer includes a second wavelet convolution, a second BatchNorm layer and a second maximum pooling layer connected in sequence, the third convolution layer includes a third convolution, a third BatchNorm layer, a third ReLU layer and a third maximum pooling layer connected in sequence, and the fourth convolution layer includes a fourth wavelet convolution, a fourth BatchNorm layer, a fourth ReLU layer and a fourth maximum pooling layer connected in sequence.
3. A generator stator turn-to-turn short circuit fault diagnosis device, characterized in that: The device comprises: a data set construction module, configured to construct a real data set, wherein the real data set includes current signals and vibration signals under different operating states of the generator, wherein the different operating states of the generator include a normal operating state of the generator, an operating state of a stator turn-to-turn short circuit fault with a degree of 2%, an operating state of a stator turn-to-turn short circuit fault with a degree of 4%, an operating state of a generator air gap eccentricity fault with a degree of 3%, and an operating state of a generator air gap eccentricity fault with a degree of 6%; A data set expansion module, configured to perform data expansion on the real data set; A fault diagnosis module is configured to construct a diagnosis model, the diagnosis model includes a dual-path wavelet convolution feature extraction module, a multi-sensor data feature fusion module and a fault identification module, the diagnosis model responds to the input current signal and vibration signal, extracts the first current feature, the second current feature and the first vibration feature and the second vibration feature through the dual-path wavelet convolution feature extraction module, and fuses the first current feature and the second current feature to obtain the current feature, and fuses the first vibration feature and the second vibration feature to obtain the vibration feature, the current feature and the vibration feature are fed to the multi-sensor data feature fusion module for feature fusion to obtain the fusion feature, the fault identification module uses the fusion feature as input, and determines the generator operation state according to the input fusion feature, trains the diagnosis model based on the expanded real data set to obtain the trained diagnosis model, and implements the generator stator turn-to-turn short circuit fault diagnosis based on the trained diagnosis model; A theoretical analysis module is configured to determine the change characteristics of the current signal and the vibration signal when a stator inter-turn short circuit fault occurs, and calibrate the real data set based on the change characteristics of the current signal and the vibration signal when a stator inter-turn short circuit fault occurs; The theoretical analysis module is further configured as follows: The generator air gap flux density, phase current and electromagnetic force per unit area on the stator winding are determined according to the following formula: ; In the formula, , , They are respectively the air gap flux density, air gap magnetic potential and air gap permeance of the generator; is the circumferential angle of the air gap; t For time; is the phase current; a is the number of parallel winding branches; for Winding factor of subharmonics; L is the effective length of stator winding; p is the pole pair number; q is the number of slots per pole and per phase; N c is the number of turns of a single coil; is the rotation speed; R is the normal winding resistance; R s is the winding resistance; is the electromagnetic force per unit area on the stator winding; is the magnetic permeability of air; When the stator turns are short-circuited, the current signal generates the third harmonic, the amplitude of the current signal increases with the increase of the number of short-circuited turns, and the amplitude of the vibration signal increases with the increase of the number of short-circuited turns; The data set expansion module is further configured to: Constructing a generative model, wherein the generative model includes a generator and a discriminator; The generator responds to the input random noise, maps the random noise into a feature space through a fully connected layer, a continuous multi-layer deconvolution module, normalization, and ReLU activation to obtain a generated sample, wherein the generated sample is a vector with the same size as the real sample included in the real data set; The discriminator takes the generated sample as input, extracts features through one-dimensional convolution, activates the activation function, and continuously performs multi-layer convolutional neural networks to obtain distribution features of the generated sample, and determines the authenticity of the generated sample based on the distance between the distribution features of the generated sample and the distribution features of the real sample; Using the real data set to train the generative model, and based on the trained generative model output, generating samples, and adding the generated samples to the real data set to expand the real data set; Input the real sample into the discriminator, obtain the real sample discrimination result, and calculate the loss function of the real data, input random noise into the generator, feed the generated sample to the discriminator, obtain the loss function of the generated sample, and calculate the gradient penalty phase and the total loss function of the discriminator, and update the discriminator parameters after back-propagating the gradient; The gradient penalty phase and the total loss function of the discriminator are calculated by the following formula: ; In the formula, GP is the gradient penalty phase; K To set a constant; G represents a generator, D represents the discriminator, x Represents real data, E represents mathematical expectation; , represents a uniform distribution from 0 to 1, Indicates the generation of data; for P norm; Indicates along the sampling point and The straight lines between are uniformly sampled; is the gradient operator; represents the gradient penalty factor; represents the objective function of WDCGAN-GP, Represents the output of the discriminator after discriminating the generated samples, D ( x ) represents the output of the discriminator after discriminating the real sample. Represents the generated data output by the generator.
4. A generator stator turn-to-turn short circuit fault diagnosis system, characterized in that: The system comprises: Memory for storing computer programs; A processor, configured to execute the computer program to implement the method according to any one of claims 1 to 2. 5 . A non-transitory computer-readable storage medium storing instructions, which, when executed by a processor, executes the method according to claim 1 .
