Demagnetization Fault Diagnosis of Permanent Magnet Direct Drive Generator Based on Improved Residual Neural Network

The permanent magnet direct drive wind turbine is the core equipment of the wind turbine. Barrier analysis ensures its safe and reliable operation. The demagnetization failure of the permanent magnet motor will occur because the wind energy equipment works under bad conditions. Taking the permanent magnet direct drive wind turbine as the research object, this paper uses ANSYS simulation software to build a two-dimensional electromagnetic field model of the generator to simulate the uniform demagnetization fault of the permanent magnet direct drive generator to obtain the back electromotive force waveform under different operating conditions and convert the back electromotive force waveform into images as input data for fault diagnosis of the improved residual neural network. Two kinds of materials are used to verify the effect of the diagnosis model, and the simulation results show that both materials can be used for demagnetizing fault diagnosis, which proves the effectiveness of the fault diagnosis strategy.


Introduction
At present, most of the permanent magnet materials are rubidium iron boron.Although it has high magnetic energy and coercivity, its working temperature is low and demagnetization is easy to occur at high temperatures, which affects the performance of the motor and causes great harm to the motor.To make the permanent magnet direct drive wind turbine more safe and reliable operation, it is necessary to study and analyze the permanent magnet materials of the motor and replace the materials with better performance to improve the performance of the motor.
In the late 1980s, as a new type of permanent magnet material, two-phase nano-composite permanent magnet material gradually came into people's view.It mainly consists of the hard magnetic phase and soft magnetic phase [1] , which is a composite material with single magnetic properties formed through magnetic coupling between the soft magnetic phase and hard magnetic phase within the nano range.The new material, which integrates high saturation magnetization of the soft magnetic phase and high coercivity of the hard magnetic phase, has an unprecedented theoretical magnetic energy product.It is the only kind of megajoule permanent magnet material expected at present, and may also be the fourth generation of rare earth permanent magnet material.In 1988, Coehoorn et al. obtained this new amorphous thin strip by using the melt rapid quenching method [2] and obtained isotropic nano permanent magnetic powder through crystallization treatment, which greatly improved the remanence performance.Moreover, this composite material can work in high temperature environments, and the motor is not prone to demagnetizing failure [3,4] .This further increases the magnetic and magnetic materials research workers' attention and actively invested in the research.
In this paper, the demagnetization failure of the permanent magnet direct drive wind generator is taken as the research object.Rubidium iron boron and biphase nanomaterials are selected as the permanent magnet material of permanent magnet direct drive wind generators respectively.The finite element method is used for uniform demagnetization of the permanent magnet direct drive wind generator model.An improved fault diagnosis method based on the combination of residual neural network and attention mechanism is proposed to provide theoretical guidance for the demagnetization fault diagnosis application of the permanent magnet direct drive generator.

Permanent magnet material selection
To prove the application significance of the diagnostic strategy proposed in this paper, two kinds of permanent magnet materials, rubidium iron-boron and biphase nano were selected to establish the finite element model of permanent magnet direct drive wind turbine.Biphase nanomaterials have the following characteristics: first, the rare earth content is relatively low, which will reduce the cost of raw materials; Second, remanence and magnetic energy products are relatively high; Third, in temperature, heat resistance and oxidation resistance have more prominent advantages, which is more suitable for high magnetic flux density and easy to be magnetized and the magnetic field is small occasions.To reduce the harmful properties of the demagnetization of permanent magnets, biphase nanomaterials are applied to permanent magnets.Figure 1 shows that the B-H hysteresis curve of biphase nanomaterials is compared with that of common rubidium fe-boron, and the magnetic saturation strength of biphase nanomaterials is about 2.3 T, while the maximum magnetic saturation strength of NdFeB is only 1.75 T. The higher saturation magnetic density of the biphase nano means stronger magnetic aggregation [5] .

Simulation analysis of the back electromotive force of demagnetizing fault under two materials
In this part, the finite element numerical simulation method of permanent magnet material is used, and 30% uniform demagnetization is carried out.The back electromotive force of a permanent magnet is studied under the same demagnetization condition, and it is found that the change of the back electromotive force energy is due to the difference in demagnetization.Then, the back electromotive force of the motor is simulated, and the uniform failure of the 30% demagnetization field is simulated [6]   .
As can be seen from Figure 2, the magnetic circuit is a symmetrical permanent magnet wind farm [7] .Under the same condition, the internal magnetic field of the permanent magnet motor will not change significantly, and the amplitude of the back electromotive force will decrease gradually with the increase of demagnetization intensity.Due to the demagnetization of the permanent magnet, the magnetic density of the gas gap is reduced.The dashed line in the figure represents the back electromotive force waveform of uniform demagnetization by biphasic nanomaterials [8,9] , while the solid line represents the back electromotive force waveform of the original material.It can be seen from the figure that the back electromotive force amplitude of biphasic nanomaterials during demagnetization is larger than that of the original material, and its current value is smaller, thus causing less harm to the motor [10] .Figure 2. Back electromotive force waveform of the permanent magnet under uniform demagnetization.

Image processing of wavelet energy spectrum
Due to the influence of the surrounding environment, it is difficult for signal decomposition to feed back all effective information.Therefore, the wavelet energy spectrum method is used to graphically process the back electromotive force waveforms under two different operating states obtained in 2. The contour line on the time-scale plane is used to represent the square of the amplitude of the wavelet conversion coefficient of the back electromotive force signal x (t) on the translation factor b and the scale factor a. The processed time-scale wavelet energy spectrum is called the time-scale wavelet energy spectrum, and its expression is: Then the total energy in the time-scale domain is: According to the simulation programming software and Formula (2), the corresponding wavelet energy spectrum calculation program is compiled, and its projected image on the XOY plane is taken as the output image of the residual neural network.The time-scale energy spectrum of the back electroEMF signal under normal operation is obtained below, and the results are shown in Figure 3.

An improved image recognition model based on residual network and attention mechanism
ResNet is a new kind of neural network, which uses some ways to skip some convolutional layers.It is widely used in natural language, computer vision, and other fields.In the past, many scholars used CNN to detect malicious code, but the network depth is too high, resulting in the disappearance of the gradient phenomenon, and with the increase of the number of network layers, the weight of each layer of the network does not change with the number of network layers.By introducing residual units into the network, ResNet can well solve the degradation problem caused by increasing the number of layers.
Residuals do not have any immediate benefits or a better description of a particular feature, but ResNet can characterize more patterns layer by layer to get the best results.In ResNet, each residual unit is connected by leaps, which does not require a lot of calculation but can speed up the training and improve the training effect.The residual network can often simplify the complexity of the algorithm, so ResNet50 is introduced to diagnose faults.

Demagnetizing fault diagnosis and analysis
The image converted by the wavelet energy spectrum is input into the improved residual neural network to identify the demagnetization images of two different materials.Figures 4 and 5 are obtained respectively.It can be seen from the two figures that when a demagnetization fault occurs, the possibility of occurrence is 1.The demagnetization fault diagnosis can be performed on both materials, and the harm caused by demagnetization fault can be reduced by biphase nanomaterials.Therefore, the performance of the motor is improved.On this basis, combined with the attention mechanism, the residual network is trained, In this paper, Mean Average Precision (mAP) index is used to evaluate the accuracy of each type of model, wherein the AP index represents the average accuracy of each type of model.For the ResNet algorithm, the need to identify whether the target in the prediction object has failed, and the experimental data loss curve obtained in this chapter is shown in Figure 6. Figure 6.Improved loss value curve In the diagram, the horizontal axis is the number of iterations trained, and the vertical axis is the error.When the number of training rounds is 75, the model tends to be stable.To optimize the parameters of the model, the number of rounds was increased to 200.According to the training method in this paper, the model with better parameters is selected for fault diagnosis.The average recall rate is 91.11% and the average accuracy is 92.75%, which can be used in the fault diagnosis of permanent magnet direct drive generators based on back electromotive force.
To verify the advanced nature of the improved residual neural network in this paper, the unimproved residual neural network is compared with the improved residual neural network.Based on the data set established in this paper, the above experiments are carried out on the unimproved residual neural network under the same computer environment and parameter settings.The loss curve is shown in Figure 7.

Figure 7. Loss value curve
The original residual network is trained according to the training method of the improved residual network mentioned above.In the figure, the abscissa is the number of training rounds, and the ordinate is the loss value.When the number of training rounds is 200, the model tends to be stable.The model with the best parameters is also selected for evaluation, its average recall rate is 91% and its average accuracy is 86.96%.Compared with the above improved residual network evaluation results, it is found that the average recall rate is increased by 0.13% and the average accuracy is increased by 5.79%, which verifies the advanced nature of the improved residual neural network in this paper.

Summary
In this paper, based on 30% uniform demagnetization failure of the permanent magnet direct drive generator, two different materials are used to simulate the back electrodynamic force waveform when demagnetization failure occurs respectively, and the obtained waveform is input into the improved residual neural network for fault identification.It can be seen from the simulation results that two different materials can be used for demagnetization fault diagnosis.Moreover, the use of biphase nanomaterials can increase the amplitude of the back electromotive force and decrease the current value when the demagnetization fault occurs in the permanent magnet direct drive generator, which can greatly reduce the harm of demagnetization degree.Therefore, the use of new materials of permanent magnet can improve the performance of the motor so that the permanent magnet direct drive generator can operate more safely and stably.
2. The wavelet energy spectrum can be divided into three types: time-scale energy spectrum, time-scale energy spectrum and time-scale energy spectrum.Based on the energy refinement characteristics of the timescale energy spectrum, the application and advantages of the wavelet energy spectrum in back electromotive force signal analysis are expounded from the energy point of view.In a certain frequency and time range, the distribution law of the back electromotive force signal under different working conditions is shown.The image data can be obtained to make full preparation for the residual neural network diagnosis.Below is the time-scale energy spectrum of the back electromotive force signal.

Figure 3 .
Figure 3. Wavelet energy spectrum image conversion

Figure 4 .
Figure 4. Raw material diagnosis Figure 5. New material diagnosis