Demagnetization degree detection method of permanent magnet synchronous motor based on multi-source adaptive SDP

Abstract

The application discloses a kind of permanent magnet synchronous motor demagnetization degree detection methods based on multi-source adaptive SDP, comprising the following steps: permanent magnet synchronous motor is modeled, and obtains the multi-source signal data of different demagnetization degrees;Data expansion is carried out using sliding window technology;Based on parameter adaptive symmetrical point mode SDP technology, time lag parameter and angle amplification factor in SDP are adaptively selected;Design the image generation method of multi-source signal fusion, mine potential information in different signals, convert one-dimensional single variable data into two-dimensional SDP image data;The generated image is divided into training set, validation set and test set, training set and validation set are used for convolutional neural network iterative training, test set is input into the trained model to detect demagnetization degree.The application significantly improves the accuracy and reliability of permanent magnet synchronous motor demagnetization degree detection, and the effectiveness of the application is verified through comparative experiment.

Description

TECHNICAL FIELD

[0001] The application relates to the field of motor fault detection, in particular to a permanent magnet synchronous motor demagnetization degree detection method based on multi-source adaptive SDP. BACKGROUND

[0002] The permanent magnet synchronous motor has the advantages of high efficiency, high power factor and simple structure and has been widely applied in industrial automation, new energy vehicles and new energy power generation. In the actual operation of the motor, mild demagnetization faults will inevitably occur, and when the demagnetization fault reaches a certain degree, corresponding measures need to be taken, so the demagnetization degree detection of the permanent magnet synchronous motor is very important.

[0003] At present, the methods for diagnosing the demagnetization degree of the permanent magnet synchronous motor are divided into three categories based on parameter identification, signal processing and artificial intelligence. In the prior art, various adaptive control methods are adopted to identify the parameters of the demagnetization of the permanent magnet synchronous motor; in order to extract the signal features of the demagnetization of the permanent magnet synchronous motor, signal processing methods such as fundamental wave extraction, modal decomposition and fractal counting are introduced; in order to improve the accuracy of the demagnetization fault diagnosis of the permanent magnet synchronous motor, fuzzy processing and BP neural network and other artificial intelligence methods are used.

[0004] In the operation of the permanent magnet synchronous motor, different degrees of demagnetization have different effects on various signals, and the collected signals may be affected by noise and other uncertain factors, and the signal features are not clear. In the signal processing of the prior art of the permanent magnet synchronous motor, the time sequence data of a single signal source is mostly used, and the diversity and richness of the data are insufficient to fully mine the potential information in the data, and the fault features cannot be fully reflected, thereby affecting the accuracy and reliability of the demagnetization degree detection. SUMMARY

[0005] In order to solve the above technical problems, the application provides a permanent magnet synchronous motor demagnetization degree detection method based on multi-source adaptive SDP, which has simple algorithm and high detection precision.

[0006] The technical scheme for solving the above technical problems is: a permanent magnet synchronous motor demagnetization degree detection method based on multi-source adaptive SDP, comprising the following steps:

[0007] S1: modeling the permanent magnet synchronous motor, and obtaining multi-source signal data of different demagnetization degrees based on the mechanism and simulation experiment of the permanent magnet synchronous motor;

[0008] S2: data expansion by using sliding window technology;

[0009] S3: adaptive selection of time lag parameters and angle amplification factors in the SDP based on parameter adaptive symmetric point mode SDP technology;

[0010] S4: calculating the demagnetization degree of the permanent magnet synchronous motor based on the selected time lag parameters and angle amplification factors.

[0010] S4: design a multi-source signal fusion image generation method to mine potential information in different signals and convert one-dimensional single variable data into two-dimensional SDP image data;

[0011] S5: divide the generated image into a training set, a validation set, and a test set, the training set and the validation set are used for iterative training of the convolutional neural network, and the test set is input into the trained model for demagnetization degree detection.

[0012] The above-mentioned permanent magnet synchronous motor demagnetization degree detection method based on multi-source adaptive SDP, in step S1, the simulation analysis results of the permanent magnet synchronous motor under different demagnetization degrees are two-dimensionally modeled by using RMxprt for the demagnetization fault of the permanent magnet synchronous motor. First, the resistance, inductance, stator, rotor, winding and permanent magnet material of the permanent magnet synchronous motor are set in the RMxprt engineering project, the stator current I, electromagnetic torque T and back EMF V are used as diagnostic signals, and four groups of data are obtained through simulation, i.e. the stable region data under four types of demagnetization conditions of 0%, 30%, 60% and 90%, each group of data includes three-phase stator current, back EMF and electromagnetic torque data.

[0013] The above-mentioned permanent magnet synchronous motor demagnetization degree detection method based on multi-source adaptive SDP, in step S2, the sliding window in the sliding window technology is defined by a left pointer and a right pointer, and the left pointer and the right pointer together determine the range of a window and slide the window on the data at a set step size; in each sliding process, the size of the window remains unchanged, and the content of the window is expanded or contracted by moving the left pointer and the right pointer;

[0014] The specific process of the sliding window technology is as follows: set the array nums, the window size L and the step size step, first initialize the left boundary and the right boundary of the window as the starting position of the array and the position of L-1, i.e. left pointer left=0, right pointer right=L-1, then start sliding the window with step size step; in each sliding process, first check whether the right pointer is out of bounds, if not, move the right pointer to the right by step, and at the same time move the left pointer to the right by step to keep the window size unchanged, then remove the old elements at the left boundary of the window and add new elements entering the window, and update the data in the window after each sliding; if it is out of bounds, stop the window sliding.

[0015] The above-mentioned permanent magnet synchronous motor demagnetization degree detection method based on multi-source adaptive SDP, in step S3, the core principle of the symmetry point mode SDP technology is to convert one-dimensional time series signals into symmetric images in polar coordinates, and this conversion makes different signals exhibit different symmetric patterns in the SDP image, and the conversion calculation method is:

[0016] ;

[0017] ;

[0018] ;

[0019] in, for The radius of the moment point in polar coordinates, that is, the distance between the point and the center of the polar coordinates; 、 They are The angle of rotation of the moment point along the initial line counterclockwise and clockwise; for The amplitude of the signal at that moment; for The amplitude of the signal at that moment; is the minimum amplitude of the signal; is the maximum amplitude of the signal; is the time lag parameter; is the rotation angle of the mirror symmetry plane; is the angle magnification factor.

[0020] In the above-mentioned method for detecting the demagnetization degree of a permanent magnet synchronous motor based on multi-source adaptive SDP, in step S3, the autocorrelation function theory in parameter adaptation is:

[0021]

[0022] in, is the autocorrelation coefficient, for The amplitude of the signal at that moment; for The amplitude of the signal at that moment; is the signal mean; is the total duration; is the order of the time delay.

[0023] In the above-mentioned method for detecting the demagnetization degree of a permanent magnet synchronous motor based on multi-source adaptive SDP, the core idea of ​​the time lag parameter adaptive mechanism in step S3 is to dynamically adapt to different signal types through the dual strategies of autocorrelation peak detection and zero-crossing detection. The specific process is as follows:

[0024] First, through the autocorrelation coefficient, intercept The effective interval of the quantized signal and its delay If there is similarity between the order versions, the main peak position is the corresponding signal period;

[0025] Then, dynamic peak detection and period extraction are performed. In order to improve the detection robustness in a noisy environment, a dynamic threshold is designed. , expressed as:

[0026]

[0027] in, 、 are the mean and standard deviation of the first 100 autocorrelation coefficients respectively; assuming that the autocorrelation coefficients obey the normal distribution, more than The value of is a significant outlier, so only when the autocorrelation coefficient is greater than Only peaks with

[0028] Finally, the main peak position is determined by detecting the autocorrelation coefficient, and then the fundamental frequency period is determined, so as to obtain The calculation of the periodic realization time lag parameter is expressed as:

[0029]

[0030] in, is the time lag parameter; is the period; in order to meet the stability requirements of the phase space reconstruction theorem, set .

[0031] In the above-mentioned method for detecting the demagnetization degree of a permanent magnet synchronous motor based on multi-source adaptive SDP, in step S3, the adaptive optimization of the angle magnification factor is based on the joint modeling of statistical characteristics and perceptual sensitivity:

[0032] First, the original signal is quantile normalized, which is expressed as:

[0033]

[0034] in, is the input signal amplitude, 、 They are the 5% and 95% quantile values ​​of the signal data respectively. The purpose of the above formula is to use piecewise function to The value is constrained to be within the boundary value, and the 5% and 95% quantiles are used to truncate the signal amplitude range to filter out extreme outliers. In addition, zero denominator protection is added, and a constant value of 0.5 is output when the signal amplitude does not fluctuate;

[0035] Then calculate the standard deviation of the normalized signal amplitude, the expression is:

[0036]

[0037] in, is the standard deviation of the signal amplitude, is the signal mean, and finally the exponential sensitivity model is used to The exponential sensitivity model expression in the angle range mapped to the experience division is:

[0038]

[0039] wherein, is an angle amplification factor, is a signal standard deviation.

[0040] The above permanent magnet synchronous motor demagnetization degree detection method based on multi-source adaptive SDP, in step S4, the stator current, electromagnetic torque and back EMF data are fused into an SDP image for demagnetization degree detection;

[0041] In terms of data quantity, the original data is expanded by using the sliding window technique, wherein the window size is set to 1000 and the step size is set to 150, and the data is expanded by sliding window value operation;

[0042] In terms of image generation, four groups of data sets are generated by parameter adaptive SDP technology for the signals of the permanent magnet synchronous motor under four different demagnetization degrees of 0%, 30%, 60% and 90%, each group of data set fuses single-phase stator current, electromagnetic torque and single-phase back EMF; the three channels of each image are composed of one phase of stator current, corresponding back EMF and electromagnetic torque, and the image is intuitively divided into three regions, and different signals exhibit different symmetric patterns in the SDP image.

[0043] The above permanent magnet synchronous motor demagnetization degree detection method based on multi-source adaptive SDP, in step S5, in terms of data set division, the total amount of data is kept unchanged, the total number of images and the set division ratio are adjusted, and the training set, the validation set and the test set are divided by the random sampling method;

[0044] The learning rate is set to 0.045 and is reduced by 0.001 every 10 iterations; the solver is selected as the stochastic gradient descent momentum SGDM; the convolution layer is configured as a two-layer structure, the filter size of the first convolution layer is 3x3, the filter number is 8, the step size is set to 1, and the padding value is 2; the filter size of the second convolution layer is also 3x3, the filter number is increased to 16, the step size remains 1, and the padding value is still 2; the maximum pooling layer is set to two layers, and the pooling size of each layer is 3x3, the step size is 2, and the padding value is 0.

[0045] The beneficial effects of the present invention are as follows: First, a permanent magnet synchronous motor is modeled to obtain multi-source signals with varying degrees of demagnetization; a sliding window technique is then used for data expansion; Finally, an image generation method based on multi-source signal fusion, using parameter-adaptive symmetrical point pattern (SDP) technology, is designed to mine the potential information in the different signals and convert the one-dimensional single-variable data into two-dimensional image data; Finally, the generated image is input into a convolutional neural network for demagnetization degree detection. This invention significantly improves the accuracy and reliability of permanent magnet synchronous motor demagnetization degree detection, and its effectiveness is verified through comparative experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is the overall flow chart of the present invention.

[0047] Figure 2 This is the back EMF waveform of a 5kW motor under normal conditions.

[0048] Figure 3 This is the back EMF waveform of a 5kW motor under 30% demagnetization fault.

[0049] Figure 4 This is the back EMF waveform of a 5kW motor under 60% demagnetization fault.

[0050] Figure 5 This is the back EMF waveform of a 5kW motor under 90% demagnetization fault.

[0051] Figure 6 This is the stator current waveform of a 5kW motor under normal conditions.

[0052] Figure 7 This is the stator current waveform of a 5kW motor with a 30% demagnetization fault.

[0053] Figure 8 This is the stator current waveform of a 5kW motor with a 60% demagnetization fault.

[0054] Figure 9 This is the stator current waveform of a 5kW motor with a 90% demagnetization fault.

[0055] Figure 10 Schematic diagram of the average torque of a 5kW motor under different demagnetization degrees.

[0056] Figure 11 The schematic diagram of SDP conversion.

[0057] Figure 12 Schematic diagram of a two-dimensional SDP image.

[0058] Figure 13 This is the confusion matrix diagram after training classification.

[0059] Figure 14 This is the confusion matrix diagram after manually selecting parameters.

[0060] Figure 15 This is a confusion matrix diagram for one-dimensional data. DETAILED DESCRIPTION

[0061] The present invention will be further described below with reference to the accompanying drawings and examples.

[0062] like Figure 1 As shown, a method for detecting the demagnetization degree of a permanent magnet synchronous motor based on a multi-source adaptive SDP includes the following steps:

[0063] S1: Model the permanent magnet synchronous motor and obtain multi-source signal data with different demagnetization degrees based on the mechanism and simulation experiments of the permanent magnet synchronous motor. The multi-source signals include stator current, electromagnetic torque, and back EMF.

[0064] Based on the simulation analysis results of the permanent magnet synchronous motor under different demagnetization degrees, in order to quickly and accurately build a model of the permanent magnet synchronous motor, RMxprt was used to perform two-dimensional modeling of the demagnetization fault of the permanent magnet synchronous motor. First, the core parameters of the permanent magnet synchronous motor, such as resistance, inductance, stator, rotor, winding, and permanent magnet material, were set in the RMxprt project. Then, the basic parameters required for the motor were input. The main parameters are shown in Table 1.

[0065]

[0066] Figures 2-5 The back EMF of the motor is shown at different levels of uniform demagnetization, with analysis performed after the permanent magnet synchronous motor is operating smoothly. Taking phase A as an example, when the permanent magnet is well excited, the back EMF amplitude reaches approximately 222V. As the degree of demagnetization increases, the back EMF amplitude decreases, reaching approximately 183V when the permanent magnet is demagnetized to 90%.

[0067] Figures 6-9 The figure shows the stator current of the motor at different levels of demagnetization. Taking phase A as an example, when the permanent magnets are well-excited, the stator current amplitude is approximately 2.86 A. As the degree of demagnetization increases, the stator current amplitude continues to increase. When the permanent magnets are demagnetized to 90%, the stator current amplitude increases to approximately 4 A.

[0068] Figure 10 is the average torque under different demagnetization degrees. It can be seen that as the degree of uniform demagnetization fault increases, the average torque decreases.

[0069] Stator current I, electromagnetic torque T and back EMF V are used as diagnostic signals. Four sets of data are obtained through simulation, namely stable region data under four types of demagnetization conditions: 0%, 30%, 60% and 90%. Each set of data contains 23,346 points of three-phase stator current, back EMF and electromagnetic torque data respectively.

[0070] S2: Use sliding window technology for data expansion.

[0071] The sliding window in sliding window technology is defined by a left pointer and a right pointer. The left and right pointers together determine the range of the window and slide the window over the data at a set step size. During each slide, the window size remains unchanged, and the window content is expanded or contracted by moving the left and right pointers.

[0072] The specific process of the sliding window technique is as follows: set the array nums, the window size L, and the step size step. First, initialize the left and right edges of the window to the start position of the array and position L-1, respectively, that is, the left pointer left = 0 and the right pointer right = L-1. Then, begin sliding the window with a step size of step. During each sliding process, first check whether the right pointer has crossed the boundary. If it has not crossed the boundary, move the right pointer to the right by step positions, and simultaneously move the left pointer to the right by step positions to maintain the window size. Then, remove the old elements at the left edge of the window and add the new elements that have entered the window. After each sliding process, the data in the window is updated. If the boundary is crossed, the window sliding stops. Due to the limited amount of simulation data due to the periodicity of the data waveform, the window size is 1000 and the step size is 150.

[0073] S3: Based on the parameter-adaptive symmetrical point mode SDP technology, the time lag parameter and angle magnification factor in the SDP are adaptively selected.

[0074] The core principle of Symmetrical Point Pattern (SDP) technology lies in converting a one-dimensional time series signal into a symmetrical image in polar coordinates. This conversion causes different signals to exhibit distinct symmetrical patterns in the SDP image. SDP images are simple and intuitive, and they more clearly reflect signal characteristics than traditional time-domain and frequency-domain images.

[0075] The basic principle of SDP image transformation method is as follows Figure 11 As shown, Figure 11 In the figure, areas I to III are the three channels of the image. When rotating clockwise, area I is (-60°) 300° to 60°, area II is 60° to 180°, and area III is 180° to 300°. This allows multiple signals to be fused into one image. The conversion calculation method is:

[0076] ;

[0077] ;

[0078] ;

[0079] wherein, is the radius of the point in polar coordinates, i.e. the distance between the point and the center of the polar coordinates; , are the angles of the point at the time instant along the initial line counterclockwise and clockwise, which determine the position of the point in polar coordinates; is the amplitude of the signal at the time instant; is the amplitude of the signal at the time instant; is the minimum amplitude of the signal; is the maximum amplitude of the signal; is the time lag parameter, used to adjust the dispersion or aggregation degree of the petal, which can reflect the small differences between time series in detail; is the rotation angle of the mirror symmetry plane; is the angle amplification factor, which is less than , used to adjust the opening angle of the petal in the symmetric point pattern. Taking , , the three regions are converted separately.

[0080] The autocorrelation function theory in parameter self-adaptation is:

[0081]

[0082] wherein, is the autocorrelation coefficient, is the amplitude of the signal at the time instant; is the amplitude of the signal at the time instant; is the mean value of the signal; is the total duration; is the order of time delay, used to detect periodicity. The numerator part of the formula aims to calculate the similarity of the signal and its delayed order version after de-meaning. If the signal after delaying order is highly similar to the original signal, the numerator will accumulate a large number of positive products, and the result tends to 1. If the signal after delaying is in the opposite relationship with the original signal, the numerator result is negative. If it is a completely random signal, the numerator value tends to 0. The denominator part is to calculate the variance of the original signal, which can directly compare the autocorrelation of signals with different dimensions and different amplitudes through normalization. ​​

[0083] The core idea of ​​the time lag parameter adaptation mechanism is to dynamically adapt to different signal types through the dual strategies of autocorrelation peak detection and zero-crossing detection. The specific process is as follows:

[0084] First, through the autocorrelation coefficient, intercept The effective interval of the quantized signal and its delay If there is similarity between the order versions, the main peak position is the corresponding signal period;

[0085] Then, dynamic peak detection and period extraction are performed. In order to improve the detection robustness in a noisy environment, a dynamic threshold is designed. , expressed as:

[0086]

[0087] in, 、 are the mean and standard deviation of the first 100 autocorrelation coefficients respectively; the autocorrelation coefficient of the signal may gradually decay as the delay increases, so intercepting the first 100 delay points can effectively avoid long-tail interference. At the same time, this dynamic threshold is set to 0.2 as the lower limit, which can avoid missing weak cycles even if the signal autocorrelation is extremely low. Assuming that the autocorrelation coefficient follows a normal distribution, it exceeds The value of is a significant outlier, so only when the autocorrelation coefficient is greater than Only peaks with

[0088] Finally, the main peak position is determined by detecting the autocorrelation coefficient, and then the fundamental frequency period is determined, so as to obtain The calculation of the periodic realization time lag parameter is expressed as:

[0089]

[0090] in, is the time lag parameter; is the period; in order to meet the stability requirements of the phase space reconstruction theorem, set .because The choice of must satisfy Takens embedding theorem, the core of which is to avoid overlapping or over-discrete trajectories in the reconstructed phase space, and when When , the phase space trajectory is optimally expanded in the principal axis direction, which can capture the dynamic characteristics and avoid the coordinate redundancy caused by high-frequency sampling.

[0091] The adaptive optimization of the angle magnification factor is based on the joint modeling of statistical characteristics and perceptual sensitivity:

[0092] First, the original signal is quantile normalized, which is expressed as:

[0093]

[0094] wherein, is the input signal amplitude, , are the 5% and 95% quantile values of the signal data respectively, the purpose of the above formula is to constrain the values beyond the boundary values into the boundary values by piecewise function, and filter extreme outliers by truncating the signal amplitude range by 5% and 95% quantile, and the zero denominator protection is added, and a constant value 0.5 is output when there is no fluctuation in the signal amplitude;

[0095] The normalized signal amplitude standard deviation is then calculated, and the expression is:

[0096]

[0097] wherein, is the signal amplitude standard deviation, is the signal mean, and finally the is mapped into the experience divided angle range by the exponential sensitivity model, and the expression of the exponential sensitivity model is:

[0098]

[0099] wherein, is the angle amplification factor, is the signal standard deviation. The effect of this method is to dynamically balance the detail visibility and noise suppression demand, and the quantile normalization filters part of the interference of outliers, and the exponential sensitivity model makes the angle amplification factor automatically increase to enhance the weak fault characteristics in low noise, and decrease to avoid display overload in strong noise.

[0100] S4: Design an image generation method of multi-source signal fusion, and mine potential information in different signals to convert one-dimensional single variable data into two-dimensional SDP image data.

[0101] In order to fully mine the potential information in different data and improve the accuracy of demagnetization degree detection, the stator current, electromagnetic torque and back EMF data are fused into an SDP image for demagnetization degree detection, so as to more comprehensively reflect the fault characteristics, thereby improving the accuracy and reliability of detection, and part of the image is as shown in Figure 12 .

[0102] In terms of data quantity, the original data is expanded by using the sliding window technology, wherein the window size is set to 1000 and the step size is set to 150, the data is expanded by sliding window value operation, and the total amount of data is 23346;

[0103] ​In terms of image generation, the signals of the permanent magnet synchronous motor at four different demagnetization levels of 0%, 30%, 60%, and 90% are used to generate four sets of data sets through parameter adaptive SDP technology. Each data set integrates three types of signals: single-phase stator current, electromagnetic torque, and single-phase back EMF. The three channels of each image are composed of one phase of the stator current, the corresponding back EMF, and the electromagnetic torque. The image is intuitively divided into three areas, and different signals show different symmetrical patterns in the SDP image.

[0104] S5: The generated images are divided into training set, validation set and test set. The training set and validation set are used for iterative training of the convolutional neural network, and the test set is input into the trained model for demagnetization degree detection.

[0105] In terms of data set division, the total amount of data remains unchanged, and the total number of images and the set division ratio are adjusted. Each data set contains 149 SDP images. Through random sampling, 92 images are selected as training sets, 30 images as validation sets, and the remaining 27 images as test sets. The total number of training sets, validation sets, and test sets of the four data sets is 368, 120, and 108, respectively.

[0106] The learning rate is set to 0.045 and decreases by 0.001 every 10 iterations; the solver uses stochastic gradient descent momentum SGDM; the convolution layer is configured as a two-layer structure, the filter size of the first convolution layer is 3×3, the number of filters is 8, the step size is set to 1, and the padding value is 2. The filter size of the second convolution layer is also 3×3, the number of filters is increased to 16, the step size is kept at 1, and the padding value is still 2; the maximum pooling layer is set to two layers, the pooling size of each pooling layer is 3×3, the step size is 2, and the padding value is 0. The confusion matrix after training classification is shown in the figure below. Figure 13 shown.

[0107] In order to verify the effectiveness of the adaptive parameters, the angle magnification factor and time lag parameters that are more suitable for the 5kW permanent magnet synchronous motor are selected by artificial experience through enumeration method. and time lag parameters Increase from 30 to 70, each time increasing by 10. Through manual comparison, select =50°, =50.

[0108] The other aspects of data processing and convolutional neural network parameter settings remain unchanged, and the convolutional neural network is put into training. The confusion matrix after training classification is as follows Figure 14 shown.

[0109] To verify the effectiveness of the method for converting one-dimensional time series data into two-dimensional multi-source images, a comparative experiment was conducted using one-dimensional convolutional neural network classification of one-dimensional time series data. Based on data from a 5 kW motor, one-dimensional time series data of stator current, electromagnetic torque, and back EMF were collected. To maintain consistency across the comparative experiments, the total data volume after sliding window processing was kept to 149,000 points, with 1,000 points per segment. Therefore, the four demagnetization fault data sets (0%, 30%, 60%, and 90%) each contained 447 segments. 347 segments were randomly selected as the training set, 70 segments as the validation set, and 30 segments as the test set, resulting in a total of 1,388 segments for the training set, 280 segments for the validation set, and 120 segments for the test set. Two convolutional layers were used. The first layer had a filter length of 5, a number of 8, a stride of 1, and a padding of 2. The second layer had a filter length of 16, with all other parameters remaining the same as the first layer. Two max pooling layers were used, each with a length of 3, a stride of 2, and a padding of 0. The confusion matrix of one-dimensional data is as follows Figure 15 shown.

[0110] Through the above experiments, the accuracy of the three methods is compared on the test set. The test set accuracy is shown in Table 2.

[0111]

[0112] As shown in Table 2, the accuracy of the present invention is 99.07%, the accuracy of the parameters selected by manual experience is 97.72%, and the accuracy of directly using one-dimensional time series data is 70.83%. This verifies the effectiveness of the method of converting one-dimensional time series data into two-dimensional multi-source images and the adaptive parameters.

Claims

1. A method for detecting the demagnetization degree of a permanent magnet synchronous motor based on multi-source adaptive SDP, characterized in that: The following steps are involved: S1: Model the permanent magnet synchronous motor and obtain multi-source signal data with different demagnetization degrees based on the mechanism and simulation experiments of the permanent magnet synchronous motor; S2: Data expansion using sliding window technology; S3: symmetrical point mode SDP technology based on parameter adaptation, adaptively selecting the time lag parameter and angle magnification factor in SDP; In step S3, the autocorrelation function theory in parameter adaptation is: ； in, is the autocorrelation coefficient, for The amplitude of the signal at that moment; for The amplitude of the signal at that moment; is the signal mean; is the total duration; is the order of time delay; S4: Design an image generation method that fuses multiple source signals, mines the potential information in different signals, and converts one-dimensional univariate data into two-dimensional SDP image data; In step S4, the stator current, electromagnetic torque, and back EMF data are integrated into an SDP image to perform demagnetization degree detection; In terms of data quantity, the original data is expanded by using the sliding window technology, where the window size is set to 1000 and the step size is set to 150, and the data is expanded by sliding window value operation; For image generation, the parameter-adaptive SDP technology generates four datasets of signals from a permanent magnet synchronous motor at four different demagnetization levels: 0%, 30%, 60%, and 90%. Each dataset combines three signals: single-phase stator current, electromagnetic torque, and single-phase back EMF. Each image has three channels, each consisting of one phase of the stator current, the corresponding back EMF, and the electromagnetic torque. The image is intuitively divided into three regions, and different signals exhibit distinct symmetrical patterns in the SDP image. S5: The generated images are divided into training set, validation set and test set. The training set and validation set are used for iterative training of the convolutional neural network, and the test set is input into the trained model for demagnetization degree detection.

2. The method for detecting the demagnetization degree of a permanent magnet synchronous motor based on multi-source adaptive SDP according to claim 1 is characterized in that: In step S1, RMxprt is used to perform two-dimensional modeling of the demagnetization fault of the permanent magnet synchronous motor based on the simulation analysis results under different demagnetization degrees. First, the resistance, inductance, stator, rotor, winding and permanent magnet material of the permanent magnet synchronous motor are set in the RMxprt project, and the stator current I, electromagnetic torque T and back electromotive force V are used as diagnostic signals. Four sets of data are obtained through simulation, namely stable region data under four types of demagnetization conditions: 0%, 30%, 60% and 90%, and each set of data includes three-phase stator current, back electromotive force and electromagnetic torque data.

3. The method for detecting the demagnetization degree of a permanent magnet synchronous motor based on multi-source adaptive SDP according to claim 1, characterized in that: In step S2, the sliding window in the sliding window technique is defined by a left pointer and a right pointer, which together determine the range of a window and slide the window on the data with a set step size; During each sliding process, the size of the window remains unchanged, and the content of the window is expanded or contracted by moving the left and right pointers; The specific process of the sliding window technology is as follows: set the array nums, window size K and step size step, first initialize the left and right boundaries of the window to the starting position and K-1 position of the array respectively, that is, the left pointer left=0, the right pointer right=K-1, and then start sliding the window with a step size step; in each sliding process, first check whether the right pointer is out of bounds. If not, move the right pointer to the right by step positions, and at the same time move the left pointer to the right by step positions to keep the window size unchanged, then remove the old elements at the left boundary of the window and add the new elements entering the window. The data in the window can be updated after each sliding; if it is out of bounds, stop sliding the window.

4. The method for detecting the demagnetization degree of a permanent magnet synchronous motor based on multi-source adaptive SDP according to claim 1, characterized in that: In step S3, the core principle of the symmetrical point pattern (SDP) technology is to convert the one-dimensional time series signal into a symmetrical image in polar coordinates. This conversion makes different signals show different symmetrical patterns in the SDP image. The conversion calculation method is: ； ； ； in, for The radius of the moment point in polar coordinates, that is, the distance between the point and the center of the polar coordinates; 、 They are The angle of rotation of the moment point along the initial line counterclockwise and clockwise; for The amplitude of the signal at that moment; for The amplitude of the signal at that moment; is the minimum amplitude of the signal; is the maximum amplitude of the signal; is the time lag parameter; is the rotation angle of the mirror symmetry plane; is the angle magnification factor.

5. The method for detecting the demagnetization degree of a permanent magnet synchronous motor based on multi-source adaptive SDP according to claim 4 is characterized in that: In step S3, the core idea of ​​the time lag parameter adaptive mechanism is to dynamically adapt to different signal types through the dual strategies of autocorrelation peak detection and zero-crossing detection. The specific process is as follows: First, through the autocorrelation coefficient, intercept The effective interval of the quantized signal and its delay If there is similarity between the order versions, the main peak position is the corresponding signal period; Then, dynamic peak detection and period extraction are performed. In order to improve the detection robustness in a noisy environment, a dynamic threshold is designed. , expressed as: ； in, 、 are the mean and standard deviation of the first 100 autocorrelation coefficients respectively; assuming that the autocorrelation coefficients obey the normal distribution, more than The value of is a significant outlier, so only when the autocorrelation coefficient is greater than Only peaks with Finally, the main peak position is determined by detecting the autocorrelation coefficient, and then the fundamental frequency period is determined, so as to obtain The calculation of the periodic realization time lag parameter is expressed as: ； in, is the time lag parameter; is the period; in order to meet the stability requirements of the phase space reconstruction theorem, set .

6. The method for detecting the demagnetization degree of a permanent magnet synchronous motor based on multi-source adaptive SDP according to claim 5, characterized in that: In step S3, the adaptive optimization of the angle magnification factor is based on the joint modeling of statistical characteristics and perceptual sensitivity: First, the original signal is quantile normalized, which is expressed as: ； in, is the input signal amplitude, 、 They are the 5% and 95% quantile values ​​of the signal data respectively. The purpose of the above formula is to use piecewise function to The value is constrained to be within the boundary value, and the 5% and 95% quantiles are used to truncate the signal amplitude range to filter out extreme outliers. In addition, zero denominator protection is added, and a constant value of 0.5 is output when the signal amplitude does not fluctuate; Then calculate the standard deviation of the normalized signal amplitude, the expression is: ； in, is the standard deviation of the signal amplitude, is the signal mean, and finally the exponential sensitivity model is used to Mapped to the angle range of empirical division, the exponential sensitivity model expression is: ； in, is the angle magnification factor, is the signal standard deviation.

7. The method for detecting the demagnetization degree of a permanent magnet synchronous motor based on multi-source adaptive SDP according to claim 1, characterized in that: In step S5, the data set is divided into a training set, a validation set, and a test set by randomly sampling the data set while keeping the total amount of data unchanged and adjusting the total number of images and the set division ratio. The learning rate is set to 0.045 and decreases by 0.001 every 10 iterations; the solver uses stochastic gradient descent momentum SGDM; the convolution layer is configured as a two-layer structure, the filter size of the first convolution layer is 3×3, the number of filters is 8, the step size is set to 1, and the padding value is 2, the filter size of the second convolution layer is also 3×3, the number of filters is increased to 16, the step size remains 1, and the padding value is still 2; the maximum pooling layer is set to two layers, and the pooling size of each pooling layer is 3×3, the step size is 2, and the padding value is 0.

Images