A method for constructing an eccentricity quantification model of a motor rotor and application thereof

Abstract

The application belongs to the field of quantitative characterization of eccentric fault of rotating machinery, and particularly relates to a method for constructing a motor rotor eccentric quantitative characterization model and application thereof, comprising: establishing a motor mathematical model, generating motor data under different motor eccentricities, and obtaining a model labeled sample set; sampling to construct an actual unlabeled sample set of an actual motor; training an approximate characterization model on the labeled sample set; and performing adversarial training on the approximate characterization model and a data discriminator using the unlabeled sample set and motor data corresponding to the model characterization of the motor in a normal state in the labeled sample set to correct the approximate characterization model; wherein, when constructing a static eccentric quantitative characterization model, the motor data is the back-EMF fundamental wave amplitude of each phase winding, and when constructing a dynamic eccentric quantitative characterization model, the motor data is the back-EMF harmonic amplitude of each phase winding. The method can realize quantitative characterization of motor rotor eccentricity by using a neural network even when there is no data under eccentric fault conditions.

Description

Technical Field

[0001] This invention belongs to the field of quantitative characterization of eccentricity faults in rotating machinery, and more specifically, relates to a method for constructing a quantitative characterization model of motor rotor eccentricity and its application. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) offer advantages such as reliability, high efficiency, high torque-to-current ratio, and high power density, making them widely used in electric vehicles, wind power generation, transportation, and propulsion systems. Therefore, reliable operation of PMSMs is essential, which can be achieved through timely maintenance and fault diagnosis. Due to manufacturing errors, almost all motors exhibit rotor eccentricity. This eccentricity induces harmonics in the motor's current and voltage, leading to torque pulsation and additional losses. Furthermore, motor eccentricity generates unbalanced magnetic pull, exacerbating eccentricity faults and potentially causing rotor rubbing and damage in severe cases. Therefore, timely prediction of motor eccentricity is crucial for mitigating its negative impacts.

[0003] Since motor eccentricity is difficult to measure directly, it is necessary to predict the degree of eccentricity based on relevant electromagnetic physical quantities of the motor. However, due to the complexity of the motor's electromechanical system, it is difficult to obtain an analytical expression of motor eccentricity directly from relevant physical quantities. Neural networks, as a good function approximator, can effectively solve this problem. As a data-driven model, neural networks require a large amount of labeled data for training to obtain good representations. However, for the problem of eccentricity prediction, the actual eccentricity of the motor is difficult to obtain directly, lacking eccentricity labels for training the neural network, and lacking a large amount of actual data under large motor eccentricity conditions. Under these circumstances, the obtained neural network model has poor generalization ability and cannot effectively predict the degree of motor eccentricity. Summary of the Invention

[0004] In view of the shortcomings and improvement needs of the existing technology, the present invention provides a method for constructing a quantitative characterization model of motor rotor eccentricity and its application. Its purpose is to improve the accuracy of characterizing motor rotor eccentricity in the context of the lack of motor data under the condition of eccentricity fault.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for constructing a quantitative characterization model of motor rotor eccentricity is provided, comprising:

[0006] A mathematical model of the motor is established, and motor data under different motor eccentricities is generated to obtain a labeled sample set of the model. Motor data under normal conditions of the actual motor are sampled to construct an actual unlabeled sample set of the motor. An approximate representation model is obtained by training on the labeled sample set of the model.

[0007] The actual unlabeled sample set and the motor data corresponding to the normal condition of the motor in the model labeled sample set are used to perform adversarial training on the approximate representation model and the data discriminator to correct the approximate representation model and obtain the eccentric quantitative representation model. The output of the feature extractor in the approximate representation model is used as the input of the data discriminator.

[0008] Specifically, when a quantitative characterization model for static eccentricity is pre-constructed to characterize static eccentricity, the motor data is the fundamental amplitude of the back EMF of each phase winding; when a quantitative characterization model for dynamic eccentricity is pre-constructed to characterize dynamic eccentricity, the motor data is the harmonic amplitude of the back EMF of each phase winding.

[0009] Furthermore, if the motor data is the no-load induced electromotive force, then the mathematical model of the motor is:

[0010]

[0011] In the formula, N is the number of winding turns, l is the motor depth, and ω r Let θ be the rotor's mechanical angular velocity. w+ and θ w- These represent the spatial positions of the positive and negative terminals of the winding, respectively; R is the stator inner diameter; r is the rotor outer diameter; θ e For the eccentric angle, O s and O r These are the geometric centers of the stator and rotor, respectively, and δ0 is the air gap length of the motor under normal conditions. denoted as eccentricity, h as the height of the permanent magnet, B0 as the amplitude of the air gap magnetic flux density under normal conditions, p as the number of rotor pole pairs, ɑ as the rotor magnetic pole position, and B(·) as the air gap magnetic flux density.

[0012] Furthermore, when constructing a quantitative eccentricity characterization model to characterize dynamic eccentricity, the motor data is the no-load back EMF harmonic amplitude of any phase winding in each phase.

[0013] Furthermore, during adversarial training, when fixing the parameters of the data discriminator and updating the parameters of the approximate representation model, the target loss function loss3 is:

[0014] loss3=λ1·loss1-λ2·loss2;

[0015] In the formula, loss1 represents the representation loss of the approximate representation model, loss2 represents the motor data source discrimination loss of the data discriminator, and λ1 and λ2 are weighting coefficients with positive values.

[0016] Furthermore, both the approximate representation model and the data discriminator are composed of fully connected networks.

[0017] The present invention also provides a method for quantitative characterization of motor rotor eccentricity, which uses the eccentricity quantitative characterization model constructed by the method described above for constructing a motor rotor eccentricity quantitative characterization model to perform quantitative characterization of eccentricity.

[0018] The present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed by a processor, it controls the device where the storage medium is located to execute a method for constructing a quantitative characterization model of motor rotor eccentricity as described above and / or a method for quantitative characterization of motor rotor eccentricity as described above.

[0019] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0020] (1) This invention models electromechanical equipment based on physical laws, and uses modeling data to train the network to learn physical laws. It can obtain a neural network fault prediction model that is close to the actual physical laws. Then, adversarial training combined with actual data is used to correct the neural network parameters, so as to realize a representation model that can more accurately represent the eccentricity quantitatively. This solves the problem that it is difficult to accurately represent the eccentricity of the motor rotor using neural networks when there is a lack of data in the case of eccentricity fault and when there are eccentricity labels.

[0021] (2) This invention provides corresponding sample set construction principles for different types of eccentricity. When constructing a quantitative characterization model for static eccentricity, the motor data is the fundamental amplitude of the back EMF of each phase winding. When constructing a quantitative characterization model for dynamic eccentricity, the motor data is the harmonic amplitude of the back EMF of each phase winding. The quantitative characterization model for eccentricity constructed in this invention distinguishes between static and dynamic eccentricity characterization models, thereby improving the characterization accuracy of the model.

[0022] (3) Since it is impossible to obtain motor data under abnormal conditions during actual sampling, the method proposed in this invention can be used to perform adversarial correction by using motor data under normal conditions in the model sample set and the actual sample during specific adversarial correction, and can also obtain a high-precision quantitative characterization model. Attached Figure Description

[0023] Figure 1 A flowchart illustrating a method for constructing a quantitative characterization model of motor rotor eccentricity provided in an embodiment of the present invention;

[0024] Figure 2 This is the network structure diagram required for constructing a quantitative characterization model of motor rotor eccentricity provided in the embodiments of the present invention;

[0025] Figure 3A flowchart of a quantitative characterization model for training motor rotor eccentricity provided in an embodiment of the present invention;

[0026] Figure 4 A comparison diagram of the characterization results obtained by the method of the present invention under static eccentricity and the actual eccentricity provided in an embodiment of the present invention;

[0027] Figure 5 A comparison diagram of the characterization results obtained by the method of the present invention under dynamic eccentricity and the actual eccentricity, provided for an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0029] Example 1

[0030] A method for constructing a quantitative characterization model of motor rotor eccentricity, such as... Figure 1 As shown, it includes:

[0031] A mathematical model of the motor is established, and motor data under different motor eccentricities is generated to obtain a labeled sample set of the model. Motor data under normal conditions of the actual motor are sampled to construct an actual unlabeled sample set of the motor. An approximate representation model is obtained by training on the labeled sample set of the model.

[0032] Using actual unlabeled sample sets and model labeled sample sets, motor data corresponding to the normal state of the motor is represented by the model. The approximate representation model and the data discriminator are subjected to adversarial training to correct the approximate representation model and obtain the eccentric quantitative representation model. The output of the feature extractor in the approximate representation model is used as the input of the data discriminator.

[0033] When a quantitative characterization model for static eccentricity is pre-constructed to characterize static eccentricity, the motor data is the fundamental amplitude of the back EMF of each phase winding. When a quantitative characterization model for dynamic eccentricity is pre-constructed to characterize dynamic eccentricity, the motor data is the harmonic amplitude of the back EMF of each phase winding.

[0034] It should be noted that the normal condition for a motor is an eccentricity of less than 10%. When constructing the actual unlabeled sample set of motors, the presence or absence of harmonics can be determined through the collected actual motor data. The presence of harmonics indicates that the motor has dynamic eccentricity, and in this case, the back EMF harmonic amplitude can be used as motor data for network training.

[0035] The different motor data for different types of eccentricity are explained below:

[0036] Because of the uneven air gap in the motor, the magnitude of the back EMF will change. In static eccentricity, since the rotor's eccentric position is fixed, its effect on the back EMF is time-invariant, therefore no harmonic components are generated, only the fundamental component exists. However, when dynamic eccentricity occurs, the rotor's eccentric position changes with time, and its effect on the back EMF is time-varying, thus introducing harmonic components into the back EMF. This will be further explained in the subsequent mathematical model.

[0037] This method models the motor, generates motor data under different eccentricities and under normal conditions using the model, trains a physical adversarial network, and achieves a quantitative representation of motor rotor eccentricity. This solves the problem of accurately representing motor rotor eccentricity in the absence of data on eccentricity fault conditions and the difficulty in training neural networks using only data from normal conditions. Specifically:

[0038] (1) Regarding the construction of data samples

[0039] A mathematical model of the motor is established using known physical information, and motor data under different motor eccentricities is generated. Corresponding rotor eccentricity label Constructing the sample set of the generative model Where X 1h ∈X1 represents motor data with an eccentricity of less than 10% in the generated data. An eccentricity of less than 10% indicates that the motor is in normal condition.

[0040] By sampling the induced electromotive force (EMF) data of an actual motor using sensors, the induced EMF data of the motor under normal conditions is obtained, forming a sample set of the induced EMF of the actual motor under normal conditions.

[0041] (2) About network models

[0042] like Figure 2 As shown, the neural network model consists of three parts: a feature extractor (network 1), an eccentricity characterizer (network 2), and a data discriminator (network 3). Each part is composed of fully connected networks, with network 1 forming fully connected relationships with networks 2 and 3, respectively.

[0043] (3) Regarding network training

[0044] like Figure 3 As shown, using the generated dataset Network 1 and Network 2 are trained. The trained Network 1 and Network 2 can achieve an approximate representation (i.e., approximate diagnosis) of the motor rotor eccentricity. The loss function used for training Network 1 and Network 2 is the mean square error function. in For network pairs 1 and 2 The predicted value.

[0045] Furthermore, using actual samples and model sample set Adversarial training is performed on the feature extractor (Network 1), eccentricity characterizer (Network 2), and data discriminator (Network 3) to correct the approximate model. The adversarially trained network can accurately characterize the eccentricity of the motor rotor.

[0046] The adversarial training process is as follows: S1, with network 1 and network 2 fixed, and the model data X... 1h And actual data X 2h Input into network 1 and mark it. k represents X 1h The number of samples in the sample, minimizing the loss function The parameters of network 3 are updated, where S1 represents the outputs of Network 1 and Network 3. S2, Fixing Network 3, update the parameters of Network 1 and Network 2 by minimizing the loss function loss3 = λ1·loss1 - λ2·loss2. Here, λ1 and λ2 are weight coefficients, which are positive numbers. S3, Repeat steps S1 and S2 until the training iteration threshold is reached, completing network training.

[0047] Preferably, the above motor data is the no-load induced electromotive force, then the above motor mathematical model is:

[0048]

[0049] In the formula, N is the number of winding turns, l is the motor depth, and ω r Let θ be the rotor's mechanical angular velocity. w+ and θ w- These represent the spatial positions of the positive and negative terminals of the winding, respectively; R is the stator inner diameter; r is the rotor outer diameter; θ e For the eccentric angle, O s and O r These are the geometric centers of the stator and rotor, respectively, and δ0 is the air gap length of the motor under normal conditions. denoted as eccentricity, h as the height of the permanent magnet, B0 as the amplitude of the air gap magnetic flux density under normal conditions, p as the number of rotor pole pairs, ɑ as the rotor magnetic pole position, and B(·) as the air gap magnetic flux density.

[0050] The process of constructing the above mathematical model of the motor is as follows:

[0051] When the rotor is eccentric, the air gap length at point θ on the circumference of the motor is:

[0052]

[0053] Where R is the stator inner diameter, r is the rotor outer diameter, and θ e For the eccentric angle, O s and O r These are the geometric centers of the stator and rotor, respectively, and δ0 is the air gap length without eccentricity. Eccentricity.

[0054] When the rotor is eccentric, the air gap magnetic flux density at point θ on the circumference of the motor is:

[0055]

[0056] Where h is the height of the permanent magnet, B0 is the air gap magnetic flux density amplitude under normal conditions, p is the number of rotor pole pairs, and ɑ is the rotor magnetic pole position.

[0057] According to the law of electromagnetic induction and the direction of the induced electromotive force, the no-load induced electromotive force in the winding is:

[0058]

[0059] Where N is the number of winding turns, l is the motor depth, and ω r Let θ be the rotor's mechanical angular velocity. w+ and θ w- These represent the spatial positions of the positive and negative terminals of the winding, respectively.

[0060] The different motor data corresponding to different types of eccentricity are further explained below:

[0061] When static eccentricity occurs, the eccentric position is fixed, and the rotor moves at an angular velocity ω. r The no-load induced electromotive force of the above winding can also be written as:

[0062]

[0063] Where α0 is the initial position of the rotor magnetic pole (the initial value of α), and t is time.

[0064] The corresponding magnitude expression can then be written as:

[0065]

[0066] When dynamic eccentricity occurs, the rotor's geometric center rotates around the stator's geometric center, with an eccentricity angle θ. e With angular velocity ω r The change in the no-load induced electromotive force of the winding can be written as:

[0067]

[0068] Where, θ e0 The initial position of the eccentric angle (θ) eIf the initial value is 0, perform a Maclaurin expansion on it and keep the first two terms, then (3) can be written as:

[0069]

[0070] The magnitude of the 1±1 / p harmonic component can be expressed as:

[0071]

[0072]

[0073] As can be seen from equation (1), when the motor experiences static eccentricity, the back EMF of the winding does not contain harmonic components, but only contains components with a frequency of pω. r The fundamental component, as can be seen from equation (2), shows that the magnitude of the back EMF of each winding changes and is related to the degree of eccentricity. Furthermore, due to the spatial position of each winding (θ... w+ and θ w- The back EMF amplitude of each winding is different under the same eccentricity. It can be seen from equation (4) that when the motor is dynamically eccentric, the 1±1 / p harmonic component will appear in the winding back EMF. It can be seen from equations (5) and (5) that the magnitude of the harmonic component is related to the degree of eccentricity. Due to the relative positions of the positive and negative poles of each phase winding (θ) w+ -θ w- Since the harmonic components of each winding are the same, the harmonic components of each winding are the same. Therefore, it is preferable that only the spectrum of one phase winding is needed to characterize the dynamic eccentricity.

[0074] Preferably, during adversarial training, when fixing the parameters of the data discriminator and updating the parameters of the approximate representation model, the target loss function loss3 is:

[0075] loss3=λ1·loss1-λ2·loss2;

[0076] In the formula, loss1 represents the representation loss of the approximate representation model, loss2 represents the motor data source discrimination loss of the data discriminator, and λ1 and λ2 are weighting coefficients with positive values.

[0077] Minimizing this loss function makes the features extracted by the feature extractor from the actual data and the model data similar, thus enabling the characterizer to characterize the eccentricity of the actual data based on the extracted features.

[0078] Preferably, both the approximate representation model and the data discriminator described above are composed of fully connected networks.

[0079] To better illustrate the present invention, the following examples are provided:

[0080] A specific implementation example using Ansys finite element simulation is provided. The established mathematical model data is designated as X1, and the finite element simulation data is designated as X2. The high-fidelity data only includes the back EMF data under normal motor conditions. For the quantitative characterization of static eccentricity, the input data of the neural network is the back EMF amplitude of each phase winding; for the quantitative characterization of dynamic eccentricity, the input data of the neural network is the spectrum of the back EMF of phase 1 winding.

[0081] The specific implementation method of the model construction method is as follows:

[0082] S1. Obtain motor back EMF data under different eccentricities from the motor mathematical model and simulation respectively;

[0083] S2. Construct a neural network. In network 1, each layer has 20 neurons and 20 neurons, and the activation function is tanh. In network 2, each layer has 10 neurons and 1 neuron, and the activation functions are tanh and sigmoid, respectively. In network 3, each layer has 20 neurons and 1 neuron, and the activation functions are tanh and sigmoid, respectively.

[0084] S3. Fix the parameters of network 3, train network 1 and network 2 using model data, use the mean square error function of predicted eccentricity and actual eccentricity as the loss function loss1, and use the Adam algorithm to optimize network 1 and network 2 to obtain a low-fidelity network.

[0085] S4. Fix the parameters of networks 1 and 2, input the model data and simulation data into the neural network, minimize loss2 to train network 3 to distinguish between model data and simulation data;

[0086] S5. Fix the parameters of network 3, input the model data and simulation data into the neural network to minimize loss 3 and train networks 1 and 2;

[0087] S6. Repeat S4 and S5 until the number of training iterations is reached, then stop training and obtain the trained neural network.

[0088] S7. Input the back EMF data obtained from the simulation under different eccentricities into the trained neural network to achieve the prediction of eccentricity.

[0089] like Figure 4 , Figure 5 This represents the quantitative characterization results of eccentricity. Figure 4 This is a comparison chart of the characterization results of the method in this embodiment under static eccentricity and the actual eccentricity. Figure 5The figure shows a comparison between the characterization results of the method in this embodiment and the actual eccentricity under dynamic eccentricity. It can be seen that the characterization results of rotor eccentricity using the method of this invention are close to the actual eccentricity. Furthermore, the figure also compares the method with that using only normal finite metadata to train the characterizer (i.e., method 2 shown in the figure). It can be seen that without using this invention, the ideal characterization effect cannot be obtained, proving the effectiveness of this invention.

[0090] Example 2

[0091] A method for quantitative characterization of motor rotor eccentricity is provided, which uses the eccentricity quantitative characterization model constructed by the method described in Embodiment 1 above to perform quantitative characterization of eccentricity.

[0092] The relevant technical solutions are the same as in Embodiment 1, and will not be repeated here.

[0093] Example 3

[0094] A computer-readable storage medium includes a stored computer program, wherein when the computer program is run by a processor, it controls the device where the storage medium is located to execute a method for constructing a quantitative characterization model of motor rotor eccentricity as described in Embodiment 1 above and / or a method for quantitative characterization of motor rotor eccentricity as described in Embodiment 2 above.

[0095] The relevant technical solutions are the same as in Embodiment 1, and will not be repeated here.

[0096] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a quantitative characterization model of motor rotor eccentricity, characterized in that, include: Establish a mathematical model of the motor, generate motor data under different motor eccentricities, and obtain a labeled sample set of the model. Sample motor data under normal conditions of actual motors to construct an actual unlabeled sample set of motors; An approximate representation model is obtained by training the model on a set of labeled samples. The actual unlabeled sample set and the motor data corresponding to the normal condition of the motor in the model labeled sample set are used to perform adversarial training on the approximate representation model and the data discriminator to correct the approximate representation model and obtain the eccentric quantitative representation model. The output of the feature extractor in the approximate representation model is used as the input of the data discriminator. Specifically, when a quantitative characterization model for static eccentricity is pre-constructed to characterize static eccentricity, the motor data is the fundamental amplitude of the back EMF of each phase winding; when a quantitative characterization model for dynamic eccentricity is pre-constructed to characterize dynamic eccentricity, the motor data is the harmonic amplitude of the back EMF of each phase winding. The motor data is the no-load induced electromotive force, then the mathematical model of the motor is: In the formula, N is the number of winding turns, l is the motor depth, and ω r Let θ be the rotor's mechanical angular velocity. w+ and θ w- These represent the spatial positions of the positive and negative terminals of the winding, respectively; R is the stator inner diameter; r is the rotor outer diameter; θ e For the eccentric angle, O s and O r These are the geometric centers of the stator and rotor, respectively, and δ0 is the air gap length of the motor under normal conditions. denoted as eccentricity, h as the height of the permanent magnet, B0 as the amplitude of the air gap magnetic flux density under normal conditions, p as the number of rotor pole pairs, ɑ as the rotor magnetic pole position, and B(·) as the air gap magnetic flux density.

2. The construction method according to claim 1, characterized in that, When constructing a quantitative eccentricity characterization model to characterize dynamic eccentricity, the motor data is the no-load back EMF harmonic amplitude of any phase winding in each phase.

3. The construction method according to claim 1, characterized in that, During adversarial training, when the parameters of the data discriminator are fixed and the parameters of the approximate representation model are updated, the target loss function loss3 is: loss3 = λ1 · loss1 - λ2 · loss2; In the formula, loss1 represents the representation loss of the approximate representation model, loss2 represents the motor data source discrimination loss of the data discriminator, and λ1 and λ2 are weighting coefficients with positive values.

4. The construction method according to claim 1, characterized in that, Both the approximate representation model and the data discriminator are composed of fully connected networks.

5. A method for quantitatively characterizing rotor eccentricity of an electric motor, characterized in that, The eccentricity quantitative characterization model constructed using the method for constructing a quantitative characterization model of motor rotor eccentricity as described in any one of claims 1 to 4 is used for quantitative characterization of eccentricity.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed by a processor, it controls the device where the storage medium is located to perform a method for constructing a quantitative characterization model of motor rotor eccentricity as described in any one of claims 1 to 4 and / or a method for quantitative characterization of motor rotor eccentricity as described in claim 5.

Images