Demagnetization prediction method for cylindrical linear motor based on shallow feedforward neural network

The motor demagnetization prediction model is constructed through shallow feedforward neural network, which solves the structural limitations and prior knowledge dependence of motor demagnetization fault diagnosis technology, and realizes accurate prediction of the degree and distribution of motor demagnetization, and is suitable for extreme environments such as the moon. motor fault diagnosis.

CN118709584BActive Publication Date: 2024-12-03HUNAN UNIV
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411213241.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-30
Publication Date
2024-12-03
Estimated Expiration
2044-08-30

AI Technical Summary

Technical Problem

The existing motor demagnetization fault diagnosis technology is limited by the motor structure, relies on prior knowledge, and cannot accurately predict the degree of demagnetization, especially in extreme lunar environments, which are difficult to use normally.

Method used

Using a shallow feedforward neural network method, a model is constructed to predict the degree of demagnetization and demagnetization distribution by obtaining the motor's current demagnetization and temperature rise demagnetization state, and a backpropagation algorithm and extreme learning machine are used to train the nerves. network, realizes accurate prediction of the degree of demagnetization of permanent magnets.

Benefits of technology

No motor intrusive operation is required, and the prior knowledge requirements are reduced. It can accurately predict the degree and distribution of motor demagnetization. It is suitable for long-term motors, providing intuitive demagnetization characteristics, and improving the accuracy and practicality of diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118709584B_ABST
    Figure CN118709584B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for predicting demagnetization of a cylindrical linear motor based on a shallow feedforward neural network, comprising the following steps: S1, obtaining preset data of the motor in a current demagnetization state and a temperature rise demagnetization state, and obtaining a first data set and a second data set respectively; S2, constructing a first model and a second model based on the first data set in combination with a shallow feedforward neural network; constructing a third model and a fourth model based on the second data set in combination with a shallow feedforward neural network; the first model and the third model include a model for predicting the degree of demagnetization; the second model and the fourth model include a model for predicting the distribution of demagnetization; S3, predicting the degree of demagnetization and the distribution of demagnetization of a permanent magnet under a current demagnetization state according to the first model and the second model respectively; predicting the degree of demagnetization and the distribution of demagnetization of a permanent magnet under a temperature rise demagnetization state according to the third model and the fourth model respectively. The present invention can accurately predict the degree of demagnetization and the distribution of demagnetization of a motor, and has practical value.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault diagnosis, and in particular to a method for predicting demagnetization of a cylindrical linear motor based on a shallow feedforward neural network. Background Art

[0002] The cylindrical linear motor has the advantages of no radial electromagnetic force and small magnetic leakage, and can be used as the preferred power generation component of the lunar Stirling power generation system. Due to the extreme environment on the lunar surface and the need for the lunar power generation system to meet the nonlinear and impact diversified load requirements, the cylindrical linear motor as a power generation component is very prone to temperature rise and current demagnetization failure. Therefore, it is extremely important to study the demagnetization fault diagnosis technology of the cylindrical linear motor.

[0003] However, the existing demagnetization fault diagnosis technologies mainly include direct diagnosis method, mathematical model diagnosis method, signal injection diagnosis method, etc. The direct diagnosis method measures the electromagnetic parameters inside the motor through the built-in monitoring equipment, which can accurately diagnose the demagnetization fault, but it needs to be considered in advance when designing the motor, and is not applicable to the finished motor; the mathematical model diagnosis method constructs a flux observer, observes online and combines table lookup to diagnose the demagnetization fault, but can only comprehensively characterize the occurrence of the fault and cannot predict the degree of the fault; the signal injection method injects voltage and current signals of specific waveforms and frequencies into the motor, observes the response signal and then discovers the fault state, but it is more dependent on the prior knowledge of the diagnoser and has limitations. At the same time, the extreme environment and personnel restrictions on the lunar surface also make the existing methods unable to be used normally. With the advancement of human technology, neural network related technologies have begun to be applied in various fields. Shallow feedforward neural network refers to a feedforward neural network with simple structure and fewer hidden layers. Usually, it has only one or two hidden layers, which is suitable for function fitting or prediction problems. When the data features are simple, shallow feedforward neural networks can effectively and quickly solve the problem. For example, the most basic extreme learning machine is a typical shallow feedforward neural network. If a shallow feedforward neural network is applied to demagnetization fault diagnosis technology, it can effectively capture and model complex nonlinear relationships and further predict the degree of demagnetization.

[0004] Therefore, a new technical solution is urgently needed to solve the technical problems that the existing motor demagnetization fault diagnosis method is limited by the motor structure, relies on prior knowledge, and cannot accurately predict the degree of demagnetization. Summary of the invention

[0005] The present invention provides a cylindrical linear motor demagnetization prediction method based on a shallow feedforward neural network, which is used to solve the technical problems that the existing motor demagnetization fault diagnosis method is limited by the motor structure, relies on prior knowledge and cannot accurately predict the degree of demagnetization.

[0006] To achieve the above object, the present invention provides a method for predicting demagnetization of a cylindrical linear motor based on a shallow feedforward neural network, comprising the following steps:

[0007] S1. Obtain first data of the motor in a current demagnetization state to obtain a first data set; obtain second data of the motor in a temperature rise demagnetization state to obtain a second data set; the first data includes time domain and frequency domain signals of no-load voltage and permanent magnet demagnetization distribution; the second data includes temperature value, time domain and frequency domain signals of no-load voltage and permanent magnet demagnetization distribution.

[0008] S2. Constructing a first model and a second model based on the first data set in combination with a shallow feedforward neural network; constructing a third model and a fourth model based on the second data set in combination with a shallow feedforward neural network; the first model and the third model include models for predicting the degree of demagnetization; the second model and the fourth model include models for predicting the demagnetization distribution.

[0009] S3. Predict the demagnetization degree and demagnetization distribution of the permanent magnet under the current demagnetization state according to the first model and the second model respectively; predict the demagnetization degree and demagnetization distribution of the permanent magnet under the temperature rise demagnetization state according to the third model and the fourth model respectively.

[0010] Preferably, before S1, it includes:

[0011] A first parameter of the motor is obtained, and a finite element simulation model of the motor is constructed in simulation software according to the first parameter.

[0012] The motor includes a single-phase generator electromagnet structure cylindrical linear motor; the first parameter includes the topological structure and size parameters of the motor.

[0013] Preferably, obtaining first data of the motor in a current demagnetization state in S1 to obtain a first data set includes:

[0014] A1. Input the nonlinear demagnetization curve into the simulation software according to the actual demagnetization data of the permanent magnet of the motor; set the permanent magnet to nonlinear demagnetization.

[0015] A2. Input a preset number of cycles of reverse-phase sinusoidal demagnetization current into the motor winding to simulate current demagnetization; set the calculation demagnetization working point function of the simulation software.

[0016] A3. Taking the amplitude of the anti-phase sinusoidal demagnetization current as a variable, increasing the current amplitude from 0 at a preset amplitude interval, obtaining the time domain and frequency domain signals of the motor no-load voltage from the normal state of the permanent magnet to the complete demagnetization state of the permanent magnet, and the demagnetization coefficient of the preset point set on the permanent magnet, to obtain the first data set.

[0017] The value of the preset amplitude interval is adjusted in real time according to the demagnetization coefficient to ensure that the demagnetization coefficient of the preset point set on the permanent magnet is gradually 0; the preset point set includes a preset number of demagnetization coefficient sampling points with an interval less than or equal to 1mm; the demagnetization coefficient includes:

[0018] ;

[0019] in, represents the demagnetization coefficient; It represents the initial remanence of the permanent magnet; It indicates the residual magnetism of permanent magnet after demagnetization.

[0020] Preferably, S2 comprises:

[0021] Build the first model:

[0022] B1. Analyze the no-load voltage frequency domain signal according to the first data set, and extract the harmonic component of the no-load voltage as the first input of the fifth model; the fifth model includes a neural network based on a back propagation algorithm.

[0023] B2. Calculating the average value of the demagnetization coefficient of the preset point set on the permanent magnet according to the first data set to obtain a first average demagnetization coefficient as the first output of the fifth model; calculating the average value of the demagnetization coefficient of the preset point set on the permanent magnet includes:

[0024] ;

[0025] in, represents the average demagnetization coefficient; Indicates the permanent magnet i The demagnetization coefficient of the demagnetization coefficient sampling point is n Represents the total number of demagnetization coefficient sampling points on the permanent magnet.

[0026] B3. Set the number of network neurons and the learning rate combination, divide the first input and the first output, and obtain the first training set; train the fifth model according to the first training set, use the back propagation algorithm to update the network weights and biases, determine the parameters of the network, and obtain the first model.

[0027] Preferably, S2 further includes:

[0028] Build the second model:

[0029] C1. Reset the input and output of the fifth model, and use the first average demagnetization coefficient as the second input of the fifth model.

[0030] C2. Divide the permanent magnet into a preset number of sub-regions, and calculate the first sub-average demagnetization coefficient of each sub-region according to the first data set; calculate the ratio of the first sub-average demagnetization coefficient of each sub-region to the first average demagnetization coefficient to obtain a first demagnetization distribution coefficient, and use the first demagnetization distribution coefficient of each sub-region as the second output of the fifth model; the calculation of the demagnetization distribution coefficient includes:

[0031] ;

[0032] in, represents the sub-average demagnetization coefficient of a sub-region on the permanent magnet, represents the average demagnetization coefficient of permanent magnets; Represents the demagnetization distribution coefficient of the permanent magnet.

[0033] Dividing the permanent magnet into a preset number of sub-regions includes: dividing the permanent magnet into a preset number of sub-regions according to the size of the permanent magnet and the amount of training data required.

[0034] C3. Set the number of network neurons and the learning rate combination, divide the second input and the second output, and obtain a second training set; train the fifth model according to the second training set, use the back propagation algorithm to update the network weights and biases, determine the parameters of the network, and obtain the second model.

[0035] Preferably, S3 includes:

[0036] The actual harmonic components are extracted from the actual no-load voltage frequency domain signal of the motor, and the average demagnetization coefficient of the permanent magnet under the current demagnetization state is predicted through the first model according to the actual harmonic components.

[0037] The average demagnetization coefficient of the permanent magnet under the current demagnetization state is input into the second model to obtain the demagnetization distribution diagram of the permanent magnet under the current demagnetization state; the demagnetization distribution diagram includes the demagnetization degree and the demagnetization distribution.

[0038] Preferably, obtaining second data of the motor in a temperature rise demagnetization state in S1 to obtain the second data set includes:

[0039] D1. According to the relationship between intrinsic magnetic induction intensity and magnetic induction intensity, the nonlinear demagnetization curve is converted into the intrinsic demagnetization curve, the intrinsic demagnetization curve is input into the simulation software, and the temperature rise characteristic of the permanent magnet is activated; the permanent magnet is set to nonlinear demagnetization.

[0040] The relationship between intrinsic magnetic induction and magnetic induction includes:

[0041] ;

[0042] in, represents the intrinsic magnetic induction intensity; Indicates the magnetic induction intensity; represents the magnetization intensity, Represents the magnetic permeability of vacuum.

[0043] D2. Correct the remanence and intrinsic coercivity of the intrinsic demagnetization curve according to the temperature coefficient of the permanent magnet, including:

[0044] ;

[0045] ;

[0046] in, Indicates the initial temperature The remanence of the lower permanent magnet, Indicates temperature The corresponding remanent magnetism of the magnet; Indicates the initial temperature The intrinsic coercive force of the permanent magnet; Indicates the initial temperature The intrinsic coercive force of the permanent magnet; represents the magnetic temperature coefficient; Represents the temperature coefficient of coercivity.

[0047] D3. Taking temperature as a variable, the permanent magnet temperature is adjusted from a preset initial temperature to a temperature exceeding the maximum operating temperature of the permanent magnet at a preset temperature rise interval, and the time domain and frequency domain signals of the motor no-load voltage when the permanent magnet goes from a normal state to when the temperature upper limit is reached, as well as the demagnetization coefficient of a preset point set on the permanent magnet are obtained to obtain a second data set.

[0048] Preferably, S2 further comprises:

[0049] Build the third model:

[0050] E1. Selecting a temperature value as a third input of a sixth model according to the second data set; the sixth model includes an extreme learning machine.

[0051] E2. Calculate the average value of the demagnetization coefficient of the preset point set on the permanent magnet according to the second data set to obtain a second average demagnetization coefficient as the third output of the sixth model.

[0052] E3. Set the number of network neurons, divide the third input and the third output to obtain a third training set; train the sixth model according to the third training set, randomly set the input layer weights and hidden layer neuron thresholds based on the extreme learning machine principle, and obtain the output layer weights by calculating the pseudo-inverse matrix, and finally determine the parameters of the network to obtain the third model.

[0053] Preferably, S2 further comprises:

[0054] Construct the fourth model:

[0055] F1. Reset the input and output of the sixth model, and use the second average demagnetization coefficient as the fourth input of the sixth model.

[0056] F2. Divide the permanent magnet into a preset number of sub-regions, and calculate the second sub-average demagnetization coefficient of each sub-region of the permanent magnet according to the second data set; calculate the ratio of the second sub-average demagnetization coefficient of each sub-region to the second average demagnetization coefficient to obtain the second demagnetization distribution coefficient, and use the second demagnetization distribution coefficient of each sub-region as the fourth output of the sixth model.

[0057] F3. Set the number of network neurons, divide the fourth input and the fourth output to obtain a fourth training set; train the sixth model according to the fourth training set, randomly set the input layer weights and hidden layer neuron thresholds based on the extreme learning machine principle, and obtain the output layer weights by calculating the pseudo-inverse matrix, and finally determine the parameters of the network to obtain the fourth model.

[0058] Preferably, S3 further includes:

[0059] The average demagnetization coefficient of the permanent magnet under the temperature rise demagnetization state is predicted by the third model according to the actual temperature value of the motor.

[0060] The average demagnetization coefficient of the permanent magnet under the temperature rise demagnetization state is input into the fourth model to obtain the demagnetization distribution diagram of the permanent magnet under the temperature rise demagnetization state; the demagnetization distribution diagram includes the demagnetization degree and the demagnetization distribution.

[0061] The present invention has the following beneficial effects:

[0062] The cylindrical linear motor demagnetization prediction method based on shallow feedforward neural network of the present invention can realize accurate prediction of the demagnetization degree of permanent magnet by acquiring the no-load voltage spectrum characteristics and the motor operating temperature; compared with the traditional data-driven demagnetization diagnosis method, the demagnetization prediction model based on shallow feedforward neural network can effectively capture and model complex nonlinear relationships, and further realize the prediction of demagnetization degree. According to the predicted demagnetization degree of permanent magnet, the demagnetization distribution of the magnet can be further obtained, and the intuitive characterization of the demagnetization of the permanent magnet is realized. There is no need for invasive operation of the motor, which reduces the requirements for prior knowledge of the operator, is suitable for motors that need to run for a long time, has a simple implementation method, and has a certain guiding role in the diagnosis and prediction of motor demagnetization. The prediction method based on neural network of the present invention effectively avoids the defects of quasi-expert systems, reduces the requirements for prior knowledge of users in the demagnetization diagnosis process, can more accurately predict the demagnetization degree of the motor, and then gives the demagnetization distribution of the permanent magnet of the motor, which has practical value.

[0063] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0065] Figure 1 It is a flow chart of a preferred embodiment of the present invention.

[0066] Figure 2 It is a cross-sectional schematic diagram of a single-phase dual-stator moving magnet cylindrical linear motor according to a preferred embodiment of the present invention.

[0067] Figure 3 It is a schematic diagram of the current demagnetization process of a preferred embodiment of the present invention.

[0068] Figure 4 It is a schematic diagram of the surface and internal temperature distribution of a motor according to a preferred embodiment of the present invention.

[0069] Figure 5 Schematic diagram of the no-load voltage waveform of current demagnetization in a preferred embodiment of the present invention.

[0070] Figure 6 It is a schematic diagram of the frequency spectrum change corresponding to the no-load voltage waveform of the current demagnetization in the preferred embodiment of the present invention.

[0071] Figure 7 Schematic diagram of the model performance curve of the preferred embodiment of the present invention.

[0072] Figure 8 is a model error histogram of a preferred embodiment of the present invention.

[0073] Fig. 9 It is a schematic diagram of the permanent magnet area division of a preferred embodiment of the present invention.

[0074] Fig.10 It is a schematic diagram of the prediction results and errors of permanent magnet No. 1 in the preferred embodiment of the present invention.

[0075] Fig.11 It is a schematic diagram of the prediction results and errors of permanent magnet No. 3 in the preferred embodiment of the present invention.

[0076] Fig.12 It is a schematic diagram of the current demagnetization simulation result of magnet No. 1 of the preferred embodiment of the present invention and the predicted demagnetization distribution of the present invention.

[0077] Fig.13 It is a schematic diagram of the current demagnetization simulation result of magnet No. 3 of the preferred embodiment of the present invention and the predicted demagnetization distribution of the present invention.

[0078] Fig.14It is a simulation result of temperature rise demagnetization of magnet No. 1 in a preferred embodiment of the present invention and a schematic diagram of the predicted demagnetization distribution of the present invention.

[0079] Fig.15 It is a simulation result of temperature rise and demagnetization of magnet No. 3 in a preferred embodiment of the present invention and a schematic diagram of the predicted demagnetization distribution of the present invention. DETAILED DESCRIPTION

[0080] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0081] See also Figure 1 In a preferred embodiment of the present invention, a method for predicting demagnetization of a cylindrical linear motor based on a shallow feedforward neural network is provided, comprising the following steps:

[0082] Prior to S1, this included:

[0083] The first parameters of the motor are obtained, and a finite element simulation model of the motor is constructed in the simulation software according to the first parameters. The motor includes a single-phase generator magnetic structure cylindrical linear motor; the first parameters include the topological structure and size parameters of the motor.

[0084] In the preferred embodiment of the present invention, a single-phase dual-stator moving magnet cylindrical linear motor is taken as an example, and a motor simulation model is built by ANSYS. The motor has an inner and outer double stator structure, and the permanent magnets are arranged in NSSN. The key size parameters of the motor are shown in Table 1. Figure 2 . Figure 2 Among them, M1, M2, M3 and M4 represent the No. 1, No. 2, No. 3 and No. 4 permanent magnets of the motor respectively, AA represents the external stator, BB represents the stator winding, CC represents the mover, and the red or blue color of M1, M2, M3 and M4 represents the polarity of the permanent magnet.

[0085] ;

[0086] S1. Obtain first data of the motor in a current demagnetization state to obtain a first data set; obtain second data of the motor in a temperature rise demagnetization state to obtain a second data set; the first data includes time domain and frequency domain signals of no-load voltage and permanent magnet demagnetization distribution; the second data includes temperature value, time domain and frequency domain signals of no-load voltage and permanent magnet demagnetization distribution.

[0087] In S1, first data of the motor in the current demagnetization state is obtained, and the first data set includes:

[0088] A1. Input the nonlinear demagnetization curve into the simulation software according to the actual demagnetization data of the permanent magnet of the motor; set the permanent magnet to nonlinear demagnetization.

[0089] In the preferred embodiment of the present invention, the permanent magnet material is NdFe35SH, the remanence is 1.23T, and the coercive force is 935kA / m. The actual demagnetization data of the permanent magnet includes a series of data of magnetic induction intensity B and magnetic field intensity H obtained by demagnetization experiments on the permanent magnet, including the demagnetization curve of the change of magnetic induction intensity B with magnetic field intensity H, which is generally found online or provided by the manufacturer.

[0090] A2. Input a preset number of cycles of reverse-phase sinusoidal demagnetization current into the motor winding to simulate current demagnetization; set the calculation demagnetization working point function of the simulation software.

[0091] In a preferred embodiment of the present invention, the simulated current demagnetization characterizes the sudden demagnetization current in the motor caused by the nonlinear and impactful diversified loads of the lunar distribution network. As a component of the Stirling power generation system of the lunar scientific research station, the motor needs to meet the power generation needs of diversified loads with nonlinear and impact characteristics such as scientific experiment cabins, multi-functional lunar rovers, communication base stations, radars, and deep space probes in the lunar scientific research station, causing voltage and current fluctuations in the lunar distribution network, resulting in sudden currents in the motor windings, causing permanent magnet current demagnetization failures. Therefore, for current demagnetization, the demagnetization caused by the diversified loads of the lunar scientific research station distribution network is mainly considered. Taking into account the randomness of the sudden current, and the current demagnetization is only related to the demagnetization magnetic field strength. Therefore, a large-value current can be added to the winding for a period of time to characterize the sudden demagnetization current caused by the diversified loads of the power grid. In a preferred embodiment of the present invention, in order to ensure that the permanent magnet is fully demagnetized, a sinusoidal current that is in antiphase with the no-load operating voltage of the motor is added within one cycle of the motor operation, and the demagnetization working point calculation function provided in the ANSYS software is set. After one cycle, refer to Figure 3 , the permanent magnet completes demagnetization.

[0092] A3. Taking the amplitude of the anti-phase sinusoidal demagnetization current as a variable, increasing the current amplitude from 0 at a preset amplitude interval, obtaining the time domain and frequency domain signals of the motor no-load voltage from the normal state of the permanent magnet to the complete demagnetization state of the permanent magnet, and the demagnetization coefficient of the preset point set on the permanent magnet, to obtain the first data set.

[0093] The value of the preset amplitude interval is adjusted in real time according to the demagnetization coefficient to ensure that the demagnetization coefficient of the preset point set on the permanent magnet is gradually reduced to 0; that is, to ensure that the demagnetization coefficient is gradually reduced to 0 at the interval required by the experiment, and the value of the preset amplitude interval is adjusted in real time according to the interval required by the demagnetization coefficient.

[0094] The preset point set includes a preset number of demagnetization coefficient sampling points with an interval less than or equal to 1 mm; in a preferred embodiment of the present invention, the number and interval of the demagnetization coefficient sampling points are selected according to the size of the permanent magnet of the motor.

[0095] The demagnetization coefficient includes:

[0096] ;

[0097] in, represents the demagnetization coefficient; It represents the initial remanence of the permanent magnet; It indicates the residual magnetism of permanent magnet after demagnetization.

[0098] In the preferred embodiment of the present invention, the amplitude of the anti-phase sinusoidal demagnetization current is used as a variable for parameter analysis to obtain the time domain data of the no-load voltage at different demagnetization levels and the corresponding spectrum data, and the demagnetization coefficients of each point of the four magnets are obtained. , forming the first data set. In order to ensure that the demagnetization distribution of the permanent magnet can be accurately characterized, the demagnetization coefficient of a point is taken at every 1mm interval on the permanent magnet. As can be seen from Table 1, the lengths of the positive and negative pole magnets are different. In the preferred embodiment of the present invention, permanent magnets No. 1 and No. 3 are analyzed, and the number of demagnetization distribution points finally selected is 6900 for permanent magnet No. 1 and 7600 for permanent magnet No. 3.

[0099] In S1, second data of the motor in the temperature rise demagnetization state is obtained, and the second data set includes:

[0100] D1. According to the relationship between intrinsic magnetic induction intensity and magnetic induction intensity, the nonlinear demagnetization curve is converted into the intrinsic demagnetization curve, the intrinsic demagnetization curve is input into the simulation software, and the temperature rise characteristic of the permanent magnet is activated; the permanent magnet is set to nonlinear demagnetization.

[0101] The relationship between intrinsic magnetic induction and magnetic induction includes:

[0102] ;

[0103] in, represents the intrinsic magnetic induction intensity; B Indicates the magnetic induction intensity; represents the magnetization intensity, represents the vacuum magnetic permeability. H Changes in magnetic induction intensity B At the same time, the magnetic polarization intensity J is also changing, which can be described by the intrinsic demagnetization curve J and H It is a curve reflecting the intrinsic magnetic properties of permanent magnetic materials, in which the magnetic polarization intensity can also be J Intrinsic magnetic induction , while temperature rise demagnetization mainly considers the influence of temperature on the intrinsic demagnetization curve.

[0104] D2. Correct the remanence and intrinsic coercivity of the intrinsic demagnetization curve according to the temperature coefficient of the permanent magnet, including:

[0105] ;

[0106] ;

[0107] in, Indicates the initial temperature The remanence of the lower permanent magnet, Indicates temperature The corresponding remanent magnetism of the magnet; Indicates the initial temperature The intrinsic coercive force of the permanent magnet; Indicates the initial temperature The intrinsic coercive force of the permanent magnet.

[0108] In order to quantitatively reflect the influence of temperature on the performance of magnets, some temperature stability parameters related to the environment are defined, such as the residual magnetism temperature coefficient and coercivity temperature coefficient , which means that for every 1°C increase in temperature, the remanence of the permanent magnet and intrinsic coercivity The percentage value that will be reduced is determined by these two coefficients, which depend on the intrinsic magnetic properties of the material.

[0109] In a preferred embodiment of the present invention, the initial temperature Set to 20℃; NdFe35SH permanent magnet corresponds to 1.23T, , -0.001 and -0.0056 respectively.

[0110] D3. Considering the high internal temperature of the motor in the lunar research station, the temperature is used as a variable, and the temperature of the permanent magnet is adjusted from the preset initial temperature to a temperature exceeding the maximum operating temperature of the permanent magnet at a preset temperature rise interval, and the time domain and frequency domain signals of the motor no-load voltage when the permanent magnet is in a normal state to the temperature upper limit are obtained, as well as the demagnetization coefficient of the preset point set on the permanent magnet, to obtain the second data set. In a preferred embodiment of the present invention, the preset initial temperature is 20°C.

[0111] In a preferred embodiment of the present invention, the preset temperature rise interval remains unchanged, and the demagnetization coefficient changes very little at a lower temperature, so some data that do not meet the experimental requirements are appropriately deleted to finally obtain the second data set.

[0112] In the preferred embodiment of the present invention, the atmospheric density of the lunar near-moon space is only 1 / 10 of that of the Earth. 12 The thin atmosphere makes the lunar surface directly exposed to the sun, and the solar irradiance during the day is as high as 1358W / m 2, the limit temperature can reach 150℃. In this extreme environment, it is difficult for the motor to dissipate heat, resulting in a high internal operating temperature and prone to temperature rise and demagnetization. Due to the compact internal structure of the cylindrical linear motor and limited heat dissipation capacity, its internal temperature distribution is relatively uniform. Figure 4 As the temperature rises, the four magnets demagnetize synchronously. At the same time, the selected motor operation feature is the no-load voltage signal, so the temperature rise caused by the winding copper loss is not considered. Therefore, the temperature of the four magnets is considered to be the same, and the temperature is used as a variable for parameter analysis. The maximum operating temperature of NdFe35SH is 150°C, so the temperature range is selected from 20°C to 180°C, and the no-load voltage time domain data and corresponding spectrum data at different temperatures and the demagnetization coefficients of each point of the four magnets are obtained. , constituting the second data set.

[0113] S2. Constructing a first model and a second model based on the first data set in combination with a shallow feedforward neural network; constructing a third model and a fourth model based on the second data set in combination with a shallow feedforward neural network; the first model and the third model include models for predicting the degree of demagnetization; the second model and the fourth model include models for predicting the demagnetization distribution.

[0114] Building the first model includes:

[0115] B1. Analyze the no-load voltage frequency domain signal according to the first data set, and extract the harmonic component of the no-load voltage as the first input of the fifth model; the fifth model includes a neural network based on a back propagation algorithm.

[0116] In the preferred embodiment of the present invention, the motor current demagnetization simulation no-load voltage waveform changes refer to Figure 5 , and its corresponding spectrum changes can be found in Figure 6 .Depend on Figures 5 and 6 It can be seen that with the occurrence of demagnetization fault, the no-load voltage waveform is distorted, and the harmonic components of each order in its voltage spectrum also fluctuate with demagnetization. Among them, the fundamental wave, double frequency, and triple frequency components account for the largest proportion, and the change amplitude is obvious, which is closely related to the demagnetization fault, while the quadruple frequency and higher frequency spectrum harmonic components account for a very small proportion, so the voltage fundamental wave, double frequency, and triple frequency components are selected as the characteristics of the voltage signal and as the input of the fifth model.

[0117] B2. Calculating the average value of the demagnetization coefficient of the preset point set on the permanent magnet according to the first data set to obtain a first average demagnetization coefficient as the first output of the fifth model; calculating the average value of the demagnetization coefficient of the preset point set on the permanent magnet includes:

[0118] ;

[0119] in, represents the average demagnetization coefficient; Indicates the permanent magneti The demagnetization coefficient of the demagnetization coefficient sampling point is n Represents the total number of demagnetization coefficient sampling points on the permanent magnet.

[0120] In a preferred embodiment of the present invention, the average demagnetization coefficient of permanent magnets No. 1 and No. 3 during the current demagnetization process is selected As the output of the fifth model. The nonlinear and impactful diversified loads of the lunar research station cause the voltage and current of the lunar distribution network to fluctuate, and eventually a sudden current appears in the motor winding, causing current demagnetization. By simulating this process, the demagnetization coefficient of each point on the permanent magnet is In order to characterize the influence of current demagnetization on the permanent magnet as a whole, the average demagnetization coefficient is used in the present invention to evaluate the degree of uneven demagnetization of the permanent magnet.

[0121] In the preferred embodiment of the present invention, it is considered that when the actual motor is running, the four magnets are always affected by the demagnetization magnetic field at the same time, and magnets No. 1 and No. 2 are symmetrical with magnets No. 4 and No. 3. Therefore, under the influence of the demagnetization magnetic field caused by the sudden current, the demagnetization of the four magnets in the motor is also symmetrically distributed. In the subsequent analysis, only the uneven demagnetization of the four magnets is targeted, and permanent magnets No. 1 and No. 3 are selected as the analysis objects to characterize the overall demagnetization.

[0122] B3. Set the number of network neurons and the learning rate combination, divide the first input and the first output, and obtain the first training set; train the fifth model according to the first training set, use the back propagation algorithm to update the network weights and biases, determine the parameters of the network, and obtain the first model.

[0123] In a preferred embodiment of the present invention, dividing the input and output of the fifth model can avoid overfitting of the training results.

[0124] The fifth model includes a neural network based on a back-propagation algorithm, specifically including:

[0125] The neural network based on back propagation is a typical feedforward neural network, including input layer, hidden layer and output layer. The data of the input layer is passed to the output layer through the hidden layer. The neurons between layers are connected by weights. The gradient descent and back propagation algorithms are used to adjust the weights between each layer to minimize the error between the prediction and the actual output. This method can learn and represent complex nonlinear relationships and is suitable for processing various types of data and problems. At the same time, the network can automatically adjust weights and biases to adapt to different data sets, and the model has strong generalization ability.

[0126] In the preferred embodiment of the present invention, first, in order to ensure that the data is as complete as possible and can characterize the demagnetization fault characteristics of the motor to a large extent, a total of 200 sets of demagnetization data of magnet No. 1 and magnet No. 3 are obtained; then, according to the characteristics of the data, a single hidden layer feedforward neural network is selected, and then the Sigmoid function is selected as the activation function of the hidden layer. This function can map any input value to between 0 and 1. The function is smooth and continuous, can characterize complex nonlinear relationships, and is suitable for prediction problems; then, the mean squared error (MSE) is used as the network loss function to calculate the difference between the model prediction and the true value; finally, according to the calculated difference, the back propagation algorithm is used to calculate the gradient, and the weight and bias of the network are updated by gradient descent.

[0127] In order to determine the number of hidden layer neuron nodes and the learning rate, the grid search method (GS) is used to exhaustively search all possible parameter combinations, cross-validate each combination, and evaluate the model performance. At the same time, considering that the model training results vary greatly under different data set division methods, and the corresponding optimal models are also different. Therefore, a five-fold cross-validation is used to divide the data set into 5 parts, and one of them is selected as the test set each time without repetition, and the other four are used as training sets to train the model. Then, the mean square error (MSE) of the model on the test set is calculated, and the 5 MSEs are averaged to obtain the final average error. This method reduces the impact of training data on training results and avoids overfitting of training results. Finally, the number of neuron nodes is determined to be 20, the learning rate is 0.01, and the model training results are shown in Figures 7 and 8 , Figure 7 is the model performance curve, Figure 8 is the model error histogram.

[0128] Figure 7 This reflects the error transformation during the training process. Since no validation set data was set when training the model, Figure 7 There are only two curves in the figure, and they are the performance curves corresponding to the last training process after the five-fold cross-validation divides the data in sequence. From the model performance curve, it can be seen that the mean square error of the training set and the test set decreases almost synchronously, and the mean square error drops to 10 at the 6th iteration. -6 , the training is terminated. Therefore, it can be seen from the performance curve that the model training results are relatively ideal. Figure 8 The error distribution of 20 groups of samples in the test set reflects the prediction accuracy of the model for each data of the test samples. It is not difficult to see that in the last training process, the error range of the test samples is between -0.0042 and 0.0035, indicating a high prediction accuracy for the test samples. In summary, the first model of the preferred embodiment of the present invention can achieve accurate prediction of permanent magnet demagnetization.

[0129] Building the second model includes:

[0130] C1. Reset the input and output of the fifth model, and use the first average demagnetization coefficient as the second input of the fifth model.

[0131] C2. Divide the permanent magnet into a preset number of sub-regions, and calculate the first sub-average demagnetization coefficient of each sub-region according to the first data set; calculate the ratio of the first sub-average demagnetization coefficient of each sub-region to the first average demagnetization coefficient to obtain a first demagnetization distribution coefficient, and use the first demagnetization distribution coefficient of each sub-region as the second output of the fifth model; the calculation of the demagnetization distribution coefficient includes:

[0132] ;

[0133] in, represents the sub-average demagnetization coefficient of a sub-region on the permanent magnet, represents the average demagnetization coefficient of permanent magnets; Represents the demagnetization distribution coefficient of the permanent magnet.

[0134] To obtain the actual demagnetization coefficient of each point on the permanent magnet is equivalent to finding the correlation function between each point in a set of data and the corresponding average value, which can be a function of data deviation or proportional relationship. By solving this correlation function, the actual value of each point can be inferred based on the average value.

[0135] Dividing the permanent magnet into a preset number of sub-regions includes: dividing the permanent magnet into a preset number of sub-regions according to the size of the permanent magnet and the amount of training data required.

[0136] In the preferred embodiment of the present invention, the amount of data of the demagnetization coefficient of each point of the permanent magnet is very large. In order to reduce the amount of data, see Fig. 9 , divide the magnet into multiple small areas. Among them, permanent magnet No. 1 is divided into 230 areas, and permanent magnet No. 3 is divided into 190 areas. The demagnetization distribution coefficient of each area on the magnet is regarded as matrix data, and the rows and columns of the matrix correspond to the length and width respectively. The uneven characterization on the magnet is equivalent to finding The functional relationship with this matrix. As output, the specific method is to expand the matrix into a row of data, each column corresponds to a region of the permanent magnet , which is equivalent to one model output, and there are 200 sets of data in total.

[0137] C3. Set the number of network neurons and the learning rate combination, divide the second input and the second output, and obtain a second training set; train the fifth model according to the second training set, use the back propagation algorithm to update the network weights and biases, determine the parameters of the network, and obtain the second model.

[0138] Building the third model includes:

[0139] E1. Selecting a temperature value as a third input of a sixth model according to the second data set. The sixth model includes an extreme learning machine.

[0140] E2. Calculate the average value of the demagnetization coefficient of the preset point set on the permanent magnet according to the second data set to obtain a second average demagnetization coefficient as the third output of the sixth model.

[0141] In a preferred embodiment of the present invention, the average demagnetization coefficient of permanent magnets No. 1 and No. 3 during the temperature rise demagnetization process is selected As the model output. Due to the extreme radiation environment and internal losses of the moon, as well as the limited internal heat dissipation capacity of the motor due to its compact structure, the motor runs at a high temperature and is prone to temperature rise and demagnetization. Figure 4 During operation, the temperature inside the motor is distributed in a hierarchical manner, and the temperature near the inner stator and the mover is relatively uniform. Therefore, the temperature changes of the four magnets are the same, and the motor temperature rise demagnetization can be regarded as a symmetrical demagnetization change. Therefore, consistent with the above-mentioned current demagnetization analysis object, only magnets 1 and 3 are analyzed. However, due to the synchronous demagnetization of the magnets during the temperature rise demagnetization process, the change range of each harmonic content in the no-load voltage spectrum is small, so only temperature is selected as the model input, and the average demagnetization coefficient of permanent magnets 1 and 3 is used as the output. Table 2 is a characteristic table of the temperature rise demagnetization spectrum.

[0142] ;

[0143] E3. Set the number of network neurons, divide the third input and the third output to obtain a third training set; train the sixth model according to the third training set, randomly set the input layer weights and hidden layer neuron thresholds based on the extreme learning machine principle, and obtain the output layer weights by calculating the pseudo-inverse matrix, and finally determine the parameters of the network to obtain the third model.

[0144] The sixth model includes an extreme learning machine, specifically including:

[0145] The extreme learning machine is a feedforward neural network. Different from the above-mentioned neural network based on feedback propagation, the extreme learning machine first randomly sets the input weights and biases to map the input data to the neurons of the hidden layer; then, by solving the pseudo-inverse matrix, the output layer weights are obtained to finally determine the network structure. The extreme learning machine usually has only one hidden layer. As mentioned above, the weights and biases of its hidden layer are randomly initialized and do not need to be adjusted by the back-propagation algorithm, while the weights of the output layer are directly calculated by mathematical calculations. Therefore, compared with the neural network based on the back-propagation method, this method does not require iteration, which greatly improves the training speed of the model. At the same time, because of the randomization of input and output parameters, the extreme learning machine is relatively simple to implement, does not require complex network structure and parameter tuning, is insensitive to the initial parameter selection, and can simplify the model building process.

[0146] In the preferred embodiment of the present invention, since the extreme learning machine does not need to involve multiple iterations, the training time is greatly saved, and the network parameters can be set randomly or manually, so there is no need to determine the optimal network parameter combination in advance. Taking temperature as input data, the average demagnetization coefficient of magnet No. 1 and magnet No. 3 As output data, the sample data is preprocessed using the maximum and minimum normalization method to uniformly scale the data to the range of 0 to 1. The above temperature rise demagnetization data set contains a total of 130 sets of data with temperatures ranging from 20°C to 180°C. In order to further simplify the model building process, a simple verification method is used to divide the sample data into 80% training sets and 20% test sets, which are then brought into the model training. The final prediction results and errors of permanent magnets No. 1 and No. 3 are shown in Figure 10 to Figure 11 , Fig.10 Characterize magnet No. 1, Fig.11 Characterize magnet No. 3. Figure 10 to Figure 11 It can be seen that the average demagnetization coefficient predicted by the extreme learning machine is It is almost consistent with the true value. The mean square errors of the predicted results of the two permanent magnets are 0.0009 and 0.0013 respectively. The maximum errors do not exceed 0.004, and the fluctuation of the error curve is also small, which shows that the third model of the present invention has high accuracy and can achieve accurate prediction of permanent magnet demagnetization.

[0147] Constructing the fourth model includes:

[0148] F1. Reset the input and output of the sixth model, and use the second average demagnetization coefficient as the fourth input of the sixth model.

[0149] F2. Divide the permanent magnet into a preset number of sub-regions, and calculate the second sub-average demagnetization coefficient of each sub-region of the permanent magnet according to the second data set; calculate the ratio of the second sub-average demagnetization coefficient of each sub-region to the second average demagnetization coefficient to obtain the second demagnetization distribution coefficient, and use the second demagnetization distribution coefficient of each sub-region as the fourth output of the sixth model.

[0150] F3. Set the number of network neurons, divide the fourth input and the fourth output to obtain a fourth training set; train the sixth model according to the fourth training set, randomly set the input layer weights and hidden layer neuron thresholds based on the extreme learning machine principle, and obtain the output layer weights by calculating the pseudo-inverse matrix, and finally determine the parameters of the network to obtain the fourth model.

[0151] In a preferred embodiment of the present invention, temperature is used as input, and the demagnetization distribution coefficient of each point on permanent magnets No. 1 and No. 3 is used as output. The data set is divided using the same method as in constructing the second model, and the sixth model is trained to obtain the fourth model.

[0152] In the preferred embodiment of the present invention, compared with the traditional data-driven demagnetization diagnosis method, the demagnetization prediction model based on a shallow feedforward neural network can effectively capture and model complex nonlinear relationships, and further predict the degree of demagnetization. In addition, from the model building process of the preferred embodiment of the present invention, it can be seen that by obtaining the no-load voltage spectrum signal as the input feature, there is no need for manual preprocessing or definition of new features, which simplifies the feature extraction work. The demagnetization distribution coefficient obtained by simulation As output data, the demagnetization distribution prediction model obtained through training can intuitively characterize the demagnetization distribution of permanent magnets and achieve functions that traditional methods cannot achieve. The structure and parameters of the neural network model itself can be adjusted and optimized according to specific problems and data characteristics, so that the method of the present invention can adapt to the demagnetization prediction of different motors and has more practical value. At the same time, the trained model can quickly process large-scale data and has extremely fast calculation speed in the reasoning stage, which is suitable for real-time prediction and decision-making, and is helpful for monitoring the demagnetization state of the motor.

[0153] S3, predicting the demagnetization degree and demagnetization distribution of the permanent magnet under the current demagnetization state according to the first model and the second model respectively; predicting the demagnetization degree and demagnetization distribution of the permanent magnet under the temperature rise demagnetization state according to the third model and the fourth model respectively. S3 specifically includes:

[0154] The actual harmonic components are extracted from the actual no-load voltage frequency domain signal of the motor, and the average demagnetization coefficient of the permanent magnet under the current demagnetization state is predicted through the first model according to the actual harmonic components.

[0155] The average demagnetization coefficient of the permanent magnet under the current demagnetization state is input into the second model to obtain the demagnetization distribution diagram of the permanent magnet under the current demagnetization state; the demagnetization distribution diagram includes the demagnetization degree and the demagnetization distribution.

[0156] The prediction results of the second model can be found in Figure 12 to Figure 13 , Fig.12 The simulation results of magnet No. 1 and the predicted demagnetization distribution of the present invention are shown in FIG. Fig.13 The simulation results of magnet No. 3 and the predicted demagnetization distribution of the present invention are shown in Fig. Figure 12 to Figure 13 It can be seen that the demagnetization coefficients of each area on both permanent magnet No. 1 and No. 3 are The results are almost consistent with the simulation results. Therefore, the method of the present invention can better characterize the current non-uniform demagnetization.

[0157] The average demagnetization coefficient of the permanent magnet under the temperature rise demagnetization state is predicted by the third model according to the actual temperature value of the motor.

[0158] The average demagnetization coefficient of the permanent magnet under the temperature rise demagnetization state is input into the fourth model to obtain the demagnetization distribution diagram of the permanent magnet under the temperature rise demagnetization state; the demagnetization distribution diagram includes the demagnetization degree and the demagnetization distribution.

[0159] The prediction results of the fourth model can be found in Figure 14 to Figure 15 , Fig.14 The simulation results of magnet No. 1 and the predicted demagnetization distribution of the present invention are shown in FIG. Fig.15 The simulation results of magnet No. 3 and the predicted demagnetization distribution of the present invention are shown in Figure 2. From the distribution prediction results, it can be seen that the method of the present invention can better characterize the temperature rise non-uniform demagnetization.

[0160] In summary, the cylindrical linear motor demagnetization prediction method based on shallow feedforward neural network of the present invention can realize accurate prediction of the demagnetization degree of permanent magnet by acquiring the no-load voltage spectrum characteristics and the motor operating temperature; compared with the traditional data-driven demagnetization diagnosis method, the demagnetization prediction model based on shallow feedforward neural network can effectively capture and model complex nonlinear relationships, and further realize the prediction of demagnetization degree. According to the predicted demagnetization degree of permanent magnet, the demagnetization distribution of the magnet can be further obtained, and the intuitive characterization of the demagnetization of the permanent magnet is realized. There is no need for invasive operation of the motor, which reduces the requirements for prior knowledge of the operator, is suitable for motors that need to run for a long time, has a simple implementation method, and has a certain guiding role in the diagnosis and prediction of motor demagnetization. The prediction method based on neural network of the present invention effectively avoids the defects of expert systems, reduces the requirements for prior knowledge of users in the demagnetization diagnosis process, can more accurately predict the demagnetization degree of the motor, and then gives the demagnetization distribution of the permanent magnet of the motor, which is more practical.

[0161] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A demagnetization prediction method for cylindrical linear motor based on shallow feedforward neural network, It is characterized in that The following steps are involved: S1. Acquire first data of the motor in a current demagnetization state to obtain a first data set; acquire second data of the motor in a temperature rise demagnetization state to obtain a second data set; the first data includes time domain and frequency domain signals of no-load voltage and permanent magnet demagnetization distribution; the second data includes temperature value, time domain and frequency domain signals of no-load voltage and permanent magnet demagnetization distribution; S2, constructing a first model and a second model based on the first data set in combination with a shallow feedforward neural network; constructing a third model and a fourth model based on the second data set in combination with a shallow feedforward neural network; the first model and the third model include models for predicting the degree of demagnetization; the second model and the fourth model include models for predicting the distribution of demagnetization; Constructing the first model includes: B1. Analyze the frequency domain signal of the no-load voltage according to the first data set, and extract the harmonic component of the no-load voltage as the first input of the fifth model; the fifth model includes a neural network based on the back propagation algorithm; B2. Calculating the average value of the demagnetization coefficients of the preset point set on the permanent magnet according to the first data set to obtain a first average demagnetization coefficient as the first output of the fifth model; the calculating the average value of the demagnetization coefficients of the preset point set on the permanent magnet includes: ; in, represents the average demagnetization coefficient; Indicates the permanent magnet The demagnetization coefficient of the demagnetization coefficient sampling point is Represents the total number of demagnetization coefficient sampling points on the permanent magnet; B3, setting the number of network neurons and the learning rate combination, dividing the first input and the first output to obtain a first training set; training the fifth model according to the first training set, updating the network weights and biases using a back propagation algorithm, determining the parameters of the network, and obtaining the first model; Constructing the second model includes: C1. resetting the input and output of the fifth model, and using the first average demagnetization coefficient as the second input of the fifth model; C2. Divide the permanent magnet into a preset number of sub-regions, and calculate the first sub-average demagnetization coefficient of each sub-region according to the first data set; calculate the ratio of the first sub-average demagnetization coefficient of each sub-region to the first average demagnetization coefficient to obtain a first demagnetization distribution coefficient, and use the first demagnetization distribution coefficient of each sub-region as the second output of the fifth model; the calculation of the demagnetization distribution coefficient includes: ; in, represents the sub-average demagnetization coefficient of a sub-region on the permanent magnet, represents the average demagnetization coefficient of permanent magnets; It represents the demagnetization distribution coefficient of the permanent magnet; The dividing the permanent magnet into a preset number of sub-regions includes: dividing the permanent magnet into a preset number of sub-regions according to the size of the permanent magnet and the amount of training data required; C3, setting the number of network neurons and the learning rate combination, dividing the second input and the second output to obtain a second training set; training the fifth model according to the second training set, updating the network weights and biases using a back propagation algorithm, determining the parameters of the network, and obtaining the second model; Constructing the third model includes: E1. Selecting a temperature value as a third input of a sixth model according to the second data set; the sixth model includes an extreme learning machine; E2. Calculating the average value of the demagnetization coefficient of the preset point set on the permanent magnet according to the second data set to obtain a second average demagnetization coefficient as the third output of the sixth model; E3. Set the number of network neurons, divide the third input and the third output to obtain a third training set; train the sixth model according to the third training set, randomly set the input layer weights and hidden layer neuron thresholds based on the extreme learning machine principle, and obtain the output layer weights by obtaining the pseudo-inverse matrix, and finally determine the parameters of the network to obtain the third model; Constructing the fourth model includes: F1. Reset the input and output of the sixth model, and use the second average demagnetization coefficient as the fourth input of the sixth model; F2. Divide the permanent magnet into a preset number of sub-regions, and calculate the second sub-average demagnetization coefficient of each sub-region of the permanent magnet according to the second data set; calculate the ratio of the second sub-average demagnetization coefficient of each sub-region to the second average demagnetization coefficient to obtain a second demagnetization distribution coefficient, and use the second demagnetization distribution coefficient of each sub-region as the fourth output of the sixth model; F3, setting the number of network neurons, dividing the fourth input and the fourth output to obtain a fourth training set; training the sixth model according to the fourth training set, randomly setting the input layer weights and hidden layer neuron thresholds based on the extreme learning machine principle, and obtaining the output layer weights by obtaining a pseudo-inverse matrix, and finally determining the parameters of the network to obtain the fourth model; S3. Predict the demagnetization degree and demagnetization distribution of the permanent magnet under the current demagnetization state according to the first model and the second model respectively; predict the demagnetization degree and demagnetization distribution of the permanent magnet under the temperature rise demagnetization state according to the third model and the fourth model respectively.

2. The method for predicting demagnetization of a cylindrical linear motor based on a shallow feedforward neural network according to claim 1, It is characterized in that Before said S1 includes: Acquire a first parameter of the motor, and construct a finite element simulation model of the motor in simulation software according to the first parameter; The motor comprises a single-phase generator-magnetic cylindrical linear motor; the first parameters comprise the topological structure and size parameters of the motor.

3. The demagnetization prediction method for a cylindrical linear motor based on a shallow feedforward neural network according to claim 2, It is characterized in that The first data of the motor in the current demagnetization state is obtained in S1, and the first data set is obtained including: A1. Input a nonlinear demagnetization curve into the simulation software according to the actual demagnetization data of the permanent magnet of the motor; set the permanent magnet to nonlinear demagnetization; A2. Input a preset number of cycles of anti-phase sinusoidal demagnetization current into the motor winding to simulate current demagnetization; set the calculation demagnetization working point function of the simulation software; A3, taking the amplitude of the anti-phase sinusoidal demagnetization current as a variable, increasing the current amplitude from 0 at a preset amplitude interval, obtaining the time domain and frequency domain signals of the motor no-load voltage from the normal state of the permanent magnet to the complete demagnetization state of the permanent magnet, and the demagnetization coefficient of the preset point set on the permanent magnet, to obtain the first data set; The value of the preset amplitude interval is adjusted in real time according to the demagnetization coefficient to ensure that the demagnetization coefficient of the preset point set on the permanent magnet is gradually 0; the preset point set includes a preset number of demagnetization coefficient sampling points with an interval less than or equal to 1mm; the demagnetization coefficient includes: ; in, represents the demagnetization coefficient; It represents the initial remanence of the permanent magnet; It indicates the residual magnetism of permanent magnet after demagnetization.

4. The method for predicting demagnetization of a cylindrical linear motor based on a shallow feedforward neural network according to claim 3, It is characterized in that The S3 includes: Extracting actual harmonic components through the actual no-load voltage frequency domain signal of the motor, and predicting the average demagnetization coefficient of the permanent magnet under the current demagnetization state through the first model according to the actual harmonic components; The average demagnetization coefficient of the permanent magnet under the current demagnetization state is input into the second model to obtain a demagnetization distribution diagram of the permanent magnet under the current demagnetization state; the demagnetization distribution diagram includes a demagnetization degree and a demagnetization distribution.

5. The method for predicting demagnetization of a cylindrical linear motor based on a shallow feedforward neural network according to claim 4, It is characterized in that The second data of the motor in the temperature rise demagnetization state is obtained in S1, and the second data set is obtained including: D1. Convert the nonlinear demagnetization curve into an intrinsic demagnetization curve according to the relationship between the intrinsic magnetic induction intensity and the magnetic induction intensity, input the intrinsic demagnetization curve into the simulation software, activate the temperature rise characteristic of the permanent magnet; set the permanent magnet to nonlinear demagnetization; The relationship between the intrinsic magnetic induction intensity and the magnetic induction intensity includes: ; in, represents the intrinsic magnetic induction intensity; Indicates the magnetic induction intensity; represents the magnetization intensity, represents the magnetic permeability of vacuum; Indicates the magnetic field strength; D2. Correcting the remanence and intrinsic coercivity of the intrinsic demagnetization curve according to the temperature coefficient of the permanent magnet, including: ; ; in, Indicates the initial temperature The remanence of the lower permanent magnet, Indicates temperature The corresponding remanent magnetism of the magnet; Indicates the initial temperature The intrinsic coercive force of the permanent magnet; Indicates the initial temperature The intrinsic coercive force of the permanent magnet; represents the magnetic temperature coefficient; represents the coercivity temperature coefficient; D3. Taking temperature as a variable, adjusting the temperature of the permanent magnet from a preset initial temperature to a temperature exceeding the maximum operating temperature of the permanent magnet at a preset temperature rise interval, obtaining the time domain and frequency domain signals of the no-load voltage of the motor when the permanent magnet goes from a normal state to when the temperature upper limit is reached, and the demagnetization coefficient of a preset point set on the permanent magnet, to obtain the second data set.

6. The method for predicting demagnetization of a cylindrical linear motor based on a shallow feedforward neural network according to claim 5, It is characterized in that The S3 further includes: Predicting the average demagnetization coefficient of the permanent magnet under the temperature rise demagnetization state by the third model according to the actual temperature value of the motor; The average demagnetization coefficient of the permanent magnet under the temperature rise demagnetization state is input into the fourth model to obtain a demagnetization distribution diagram of the permanent magnet under the temperature rise demagnetization state; the demagnetization distribution diagram includes a demagnetization degree and a demagnetization distribution.

Citation Information

Patent Citations

  • PMSM demagnetization fault diagnosis method based on torque signal fuzzy intelligent learning

    CN112101457A

  • Method and system for predicting sintered SmCo magnetic performance based on neural network

    CN112259176A

  • Permanent magnet synchronous motor demagnetization fault diagnosis method and device based on information fusion

    CN112285554A

  • Permanent magnet motor demagnetization fault diagnosis method based on bispectrum analysis and CNN

    CN117788841A

  • Control device, control method, control system, learning device, and model data generation method

    JP6465236B1