Aerial permanent magnet starting generator initial inter-turn short circuit fault diagnosis method

Abstract

This invention discloses a method for diagnosing initial inter-turn short-circuit faults in aircraft permanent magnet generators. Specifically, it involves: establishing a mathematical model of the inter-turn short-circuit fault state based on a permanent magnet synchronous generator health model; determining the stator current of the fault phase as the input signal for sample interpolation and fault diagnosis; constructing a missing sample generation model based on a conditional Wasserstein generative adversarial network, and introducing gradient penalty, self-attention mechanism, and physical consistency constraints to achieve the interpolation of missing fault level samples, ensuring the integrity of the training dataset; and constructing a dual-domain feature fusion neural network of ordinary differential equations, achieving accurate identification of initial fine-grained faults through time-frequency dual-domain feature fusion and continuous dynamic modeling. The method of this invention generates missing samples that conform to the fault evolution law in terms of time-domain waveform and frequency-domain features. The diagnostic model trained based on the completed sample set achieves a diagnostic accuracy of 98.9% in the initial inter-turn short-circuit multi-classification task.

Description

Technical Field

[0001] This invention belongs to the field of AC motor control technology, specifically relating to a method for diagnosing initial inter-turn short circuit faults in aircraft permanent magnet generators. Background Technology

[0002] In the field of aerospace engineering, aircraft predictive health management (PHM) has become a crucial technology system for ensuring flight safety, improving operational efficiency, and reducing lifecycle costs. PHM can generally be divided into on-board PHM and post-flight PHM. Post-flight PHM relies on flight data to conduct comprehensive health assessments and initial fault identification of critical components. Permanent magnet starters / generators, as a vital component of multi-electric aircraft, directly impact the aircraft's power supply quality and operational reliability. These motors operate under complex flight environments characterized by high temperatures and vibrations, making them susceptible to various electrical or mechanical faults, with stator winding turn-to-turn short circuits being a typical example. If such faults are not addressed promptly, the short-circuit branch current will cause localized overheating, further accelerating insulation degradation and fault propagation, potentially leading to serious faults such as phase-to-phase short circuits or phase loss.

[0003] In the initial stage, when the short-circuit turns ratio is 1% to 5%, the fault severity is relatively mild, and the system can usually maintain basic operation. The impact of the fault on motor performance and control stability gradually increases, but it generally does not immediately lead to functional failure. If the degree of inter-turn short-circuit fault can be detected and accurately identified in this stage, corresponding maintenance, protection, or fault-tolerant control measures can be taken before the fault spreads further, thereby effectively suppressing the fault propagation, preventing it from developing into a serious short-circuit fault, and improving the operational safety and reliability of the permanent magnet starter / generator system.

[0004] Data-driven methods, which avoid complex mechanism modeling and possess the advantage of autonomously learning fault representations from historical data, have become an important research direction in inter-turn short-circuit fault diagnosis. For initial inter-turn short-circuit faults with a short-circuit turn ratio of 1% to 5%, their fault characteristics are usually weak, and the characteristic boundaries between adjacent fault degrees are not obvious. Traditional methods mostly rely on artificial features such as current amplitude and harmonic content. Limited by the ability to represent single features and discrete classification modeling methods, it is difficult to characterize the continuous evolution of fault characteristics with changes in the short-circuit turn ratio, thus limiting the precise identification of the initial inter-turn short-circuit fault level.

[0005] However, achieving the aforementioned fine-grained fault identification requires a complete training sample covering different fault severity levels. In actual operation, when generators operate at different speeds and fault severity varies, the fault samples collected at different speeds often only cover a portion of the initial fault severity, making it difficult to form a complete fault sample set, resulting in the loss of some fault level samples. An incomplete sample set directly weakens the network model's ability to learn the feature distribution of different fault severity levels, reducing the model's generalization performance and fault severity identification accuracy. Therefore, how to obtain complete and effective fault samples under conditions of missing fault level samples, and accurately identify the fine-grained differences in initial inter-turn short-circuit faults, has become a critical problem that urgently needs to be solved in data-driven initial inter-turn short-circuit fault diagnosis. Summary of the Invention

[0006] The purpose of this invention is to provide a method for diagnosing initial inter-turn short circuit faults in aircraft permanent magnet generators, which solves the problem of low accuracy in identifying initial inter-turn short circuit faults in existing methods.

[0007] The technical solution adopted in this invention is a method for diagnosing initial inter-turn short circuit faults in aircraft permanent magnet generators, specifically as follows: Step 1: Based on the health model of permanent magnet synchronous generator, establish a mathematical model under the inter-turn short-circuit fault state, and determine the stator current of the fault phase as the input signal for sample interpolation and fault diagnosis. Step 2: Construct a missing sample generation model based on a conditional Wasserstein generative adversarial network, and introduce gradient penalty, self-attention mechanism and physical consistency constraint to imput missing fault level samples and form a dataset. Step 3: Construct a dual-domain feature fusion neural network of ordinary differential equations. Through time-frequency dual-domain feature fusion, continuous dynamics modeling, and time attention convergence, the initial fine-grained fault degree can be identified.

[0008] The invention is further characterized in that, Step 1 specifically involves: Step 1.1, set up a permanent magnet synchronous generator a When an inter-turn short-circuit fault occurs in the stator winding of a phase, the faulty phase winding can be considered as a healthy part. a 1. Fault Section a 2, and through the short-circuit resistor R f Forming a closed loop; where, R f Represents the remaining insulation resistance; the short-circuit branch current is denoted as... i f ; Neglecting the effects of core magnetic saturation and other nonlinear factors, the electrical equations for the faulty branch are shown in equation (1): (1) in, (2) (3) (4) (5) (6) (7) In the formula, V abcf For the three phases abc and the faulty branch f The voltage; η The proportion of the number of turns of the short-circuited coil to the total number of turns of the coil in that phase; L Each phase of the stator is self-sensing; M Mutual induction; R s Phase resistance; i abcf For the three phases abc and the faulty branch f The current; λ PM,abcf for abc Three-phase and faulty branch f Permanent magnet linkage; λ PM,a , λ PM,b , λ PM,c , λ PM,f They are respectively abc Three-phase and faulty branch permanent magnet flux linkage; V 0,abcf for abc Zero-sequence voltage matrix of three phases and faulty branch f; V 0 represents the zero-sequence voltage component of each of the three phases; Step 1.2: Starting from the phase voltage equations in equation (2), and considering the neutral point voltage... u By influencing the effect of 0, the expression for the short-circuit branch current is derived. (8) In the formula, (9) in, u 0 represents the neutral point voltage; Step 1.3, when the permanent magnet synchronous generator is under an inter-turn short-circuit fault, the direct-axis and quadrature-axis voltage equations in the rotating coordinate system are as shown in equation (10): (10) In the formula, u' d Direct-axis voltage of a permanent magnet synchronous generator under inter-turn short-circuit fault conditions; u' q Quadrature-axis voltage of a permanent magnet synchronous generator under inter-turn short-circuit fault conditions; i' d and i' q The measured currents of the d-axis and q-axis of the permanent magnet synchronous generator under the inter-turn short-circuit fault condition are shown in Equation (11); (11) right dq The axis current is transformed to obtain... abc Phase stator current; (12) Therefore, the fault a The phase stator current serves as a fault feature sample for the missing sample interpolation model and the fault diagnosis model.

[0009] Step 2 specifically involves: Step 2.1, based on the fault phase stator current signal determined in Step 1.3, let the acquired fault phase stator current sample be: (13) Where C represents the number of input current signal channels, N Indicates the sampling window length; the fault severity condition label is denoted as... y , is used to represent the short-circuit turns ratio; where, y =0 indicates the normal state. y =0.01, 0.02, 0.03, 0.04, and 0.05 represent the initial inter-turn short-circuit fault severity corresponding to short-circuit turns ratios of 1% to 5% respectively; The collected set of fault severity labels is represented as follows: (14) The set of missing fault degree labels that need to be imputed is represented as follows: (15) in, y ava This represents the set of collectable fault levels. y mis This represents the set of missing fault levels that require sample imputation; A dataset consisting of collectable samples is represented as: (16) in, Dava Indicates the collectable sample set. x i Indicates the first i One current sample, y i ava This indicates the collectable fault level label corresponding to the sample. n ava Indicates the number of samples that can be collected; Step 2.2, given the collectable sample set D ava Based on the fault degree condition label, a missing sample generation model based on conditional Wasserstein generative adversarial network is constructed. Step 2.3: After constructing the CWGAN model, a gradient penalty term is introduced into the loss function to construct the CWGAN-GP model; Step 2.4: Based on the obtained missing sample generation model CWGAN-GP, a self-attention mechanism is introduced into the generator and evaluator to construct the missing sample imputation model CWGAN-GP-SA. Let the intermediate features of the network be represented as: (twenty one) in, L Indicates the length of the feature sequence. d Indicates feature dimension; The query matrix, key matrix, and value matrix are obtained through linear mapping; (twenty two) in, W Q , W K , W V The learnable parameter matrix; The self-attention weight matrix can be represented as: (twenty three) in, d k Indicates the dimension of the key vector; The output of self-attention is: (twenty four) The enhanced feature representation is obtained by using residual connections: (25) in, These are learnable weight coefficients; Step 2.5: After completing the CWGAN-GP-SA model, a physical consistency constraint is further introduced into the generator loss function; The generator's total loss function is defined as: (26) in, Indicates the constraint weight coefficient. Represents physical consistency constraints; Step 2.6: Using the collectable sample set formed in Step 2.1 and the generator total loss function determined in Step 2.5, train the CWGAN-GP-SA model; during the training process, alternately update the evaluator and generator parameters; the evaluator learns the distribution difference between real samples and generated samples under a given fault level by minimizing the evaluator loss function; the generator generates current samples that are close to the real sample distribution and satisfy the physical consistency constraint by minimizing the generator total loss function. After the model training is completed, the missing fault level label defined in step 2.1 is input into the generator to generate current samples under the corresponding missing fault level. (28) in, Represents a random noise vector. This indicates that a fault severity label is missing. This represents a sample of missing fault level current from the generator output. The generated missing fault level sample set is represented as follows: (29) in, This represents the generated set of missing fault level samples. Indicates the number of samples generated; When the fault severity label set is missing y mis When ={0.01, 0.03, 0.04}, it means that... y =0.01、 y =0.03 and y Input 0.04 into the trained generator to generate initial inter-turn short-circuit fault current samples corresponding to short-circuit turns ratios of 1%, 3%, and 4%. Step 2.7: Merge the generated missing fault level sample set with the collectable sample set to obtain the complete initial inter-turn short-circuit fault sample set. ; (30).

[0010] In step 2.2, the missing sample generation model includes a generator and an evaluator; generator With random noise vector and fault severity condition labels As input, the output is the generated current sample corresponding to the fault level; (17) in, p z Represents a random noise distribution. This represents a sample of the current output from the generator. evaluator Using current samples and their corresponding fault severity condition labels as input, and outputting real-valued scores, the optimization objective of CWGAN under the given constraints is expressed as: (18) in, This represents the joint distribution of actual current samples and fault severity labels. Represents a random noise distribution. The distribution of condition labels indicates the degree of failure.

[0011] In step 2.3, the loss function of the evaluator is: (19) in, Represents the gradient penalty coefficient. Indicates that it is composed of real samples With generated samples Samples obtained through random interpolation; The generator's adversarial loss function is: (20).

[0012] In step 2.5, the physical consistency constraint is expressed as follows: (27) in, Represents temporal characteristic constraints. Represents frequency domain characteristic constraints. Indicates consistency constraints for the degree of failure. and These are the corresponding weighting coefficients.

[0013] Step 3 specifically involves: Step 3.1: The complete initial inter-turn short-circuit fault sample set obtained in Step 2 is randomly divided into training set, validation set and test set according to fault category label; the initial inter-turn short-circuit multi-classification task takes normal state and 1% to 5% short-circuit turn ratio fault state as classification objects, and the number of fault categories is M=6. The first in the complete fault sample set iThe current sequence samples are divided into T overlapping time windows arranged in chronological order according to a fixed window length and sliding step size; (31) Among them, the The first current sequence sample The current segment within the time window satisfies (32) in, Indicates the first i Current sequence samples The current segment in the t-th time window, C Indicates the number of current channels; Using single-phase current as the model input, i.e. Set the number of time windows to T, and the length of each window to [value missing]. There are 1024 sampling points, and the window sliding step is 1024 sampling points; For each time window obtained by division, the normalization process is performed according to the mean and standard deviation of the training set samples to obtain a normalized time window sequence, which is then used as the input of DDF-ODENet; Step 3.2: For the standardized time window sequence, construct time-domain branches and frequency-domain branches to extract current waveform features and spectral harmonic features respectively, so as to characterize the weak changes in the current signal of the initial inter-turn short circuit fault. Time-domain branch directly uses the current window As input, a one-dimensional convolutional network is used to extract local amplitude changes, periodic morphology changes, and subtle distortion features caused by faults in the current waveform. The process is represented as follows: (33) in, This represents a nonlinear feature extraction mapping implemented by a time-domain branch. For learnable parameters, Indicates the first The first sample Temporal feature vectors corresponding to each time window; After obtaining the time-domain features, a frequency-domain branch is constructed; first, a Fourier transform is performed on each current window, and its amplitude spectrum is taken. (34) in, Indicates Fourier transform, Indicates the amplitude value. Indicates the first The first sample The frequency domain amplitude spectrum corresponding to each time window; The effective positive frequency components in the amplitude spectrum are retained, while the DC component is removed. Logarithmic transformation and in-sample normalization are performed on the amplitude spectrum to obtain the preprocessed frequency domain input features. The preprocessed frequency domain input features are then input into the frequency domain feature extraction network. (35) in, This represents a nonlinear feature extraction mapping implemented by frequency domain branching. For learnable parameters, Indicates the first The first sample Frequency domain feature vectors corresponding to each time window; After obtaining the time-domain features and frequency-domain features respectively, the time-domain features and frequency-domain features under the t-th time window are concatenated to obtain the dual-domain concatenated features: (36) in, Indicates the first i The dual-domain concatenated features corresponding to the t-th time window of each sample. This indicates a concatenation operation along the feature dimension; By using fully connected mapping, layer normalization, and nonlinear activation functions, the concatenated dual-domain features are projected onto a unified latent space. (37) in, This represents the two-domain fusion mapping function. For learnable parameters, Indicates the first The first sample The fused feature vectors corresponding to each time window; For the first i Repeat the above time-domain branching extraction, frequency-domain branching extraction, feature splicing and fusion mapping process for all time windows of a current sequence sample to obtain its dual-domain fusion feature sequence. (38) in, Indicates the number of time windows. This represents the dual-domain fusion feature sequence corresponding to the i-th current sequence sample; Step 3.3: After obtaining the dual-domain fusion feature sequence, the Neural ODE module is introduced to represent the hidden state changes as follows: (39) in, Indicates time The hidden state below, This represents a dynamic evolution function parameterized by a neural network. These are learnable parameters; By numerically integrating the above ordinary differential equations, the continuous evolution of the fused feature sequence in the latent space is obtained. (40) in, This represents the characteristics of time-series faults after Neural ODE modeling; After obtaining the continuous evolution features of the Neural ODE, a temporal attention mechanism is introduced; for the fused features of the t-th time window, the attention weight is expressed as: (41) in, The first value calculated by the attention network represents the value of the second value. The importance score of each time window This represents the corresponding attention weight; The fusion features of all time windows are weighted and summed according to the attention weights; (42) in, This represents the global features after temporal attention convergence; The Neural ODE output features are concatenated with the temporal attention features to obtain the final feature representation used for classification: (43) Will Inputting the data into a fully connected classifier yields the classification output: (44) in, This represents a fully connected classifier. This represents the classification output corresponding to the i-th sample; Step 3.4: Compare the predicted probability of the fault category obtained after processing the above classification output with the Softmax function with the true fault category label, construct a classification loss function, and optimize the network parameters of DDF-ODENet through this loss function; The classification loss function is calculated using training set samples, and the DDF-ODENet network parameters are updated. After each round of training, the model recognition performance is evaluated using the validation set, and the model parameters corresponding to the highest accuracy on the validation set are saved. After training, the saved optimal model is tested using the test set to obtain the final fault diagnosis accuracy, confusion matrix, and classification evaluation results.

[0014] The classification loss function is expressed as: (45) in, Indicates batch size. Indicates the number of fault categories. Indicates the first The sample at the th The true label distribution on the class, This indicates that the model predicts the sample belongs to the first... The probability of a class.

[0015] The beneficial effects of this invention are: The method of this invention can generate missing fault level samples based on CWGAN-GP-SA when there are missing samples of the fault level. The generated samples are consistent with the simulated fault samples in terms of time domain waveform and frequency domain characteristics, and the average point-by-point difference in amplitude between the two is only 1.81%. After completing the sample set, a diagnostic model is further trained based on DDF-ODENet. This model can integrate time domain, frequency domain and continuous dynamic evolution characteristics, and achieve a diagnostic accuracy of 98.9% for the initial inter-turn short circuit multi-classification task. Attached Figure Description

[0016] Figure 1 This is an overall structural block diagram of the method for diagnosing initial inter-turn short circuit faults in aviation permanent magnet generators according to the present invention. Figure 2 This is a schematic diagram of an inter-turn short circuit. Figure 3 This is a comparison diagram of the three-phase current waveforms before and after an inter-turn short circuit occurs; Figure 4 This is a schematic diagram of the CWGAN-GP-SA principle in the initial inter-turn short circuit fault diagnosis method for aviation permanent magnet generators of the present invention; Figure 5 This is a time-domain comparison diagram of the current waveforms generated by CWGAN-GP-SA with a short-circuit turns ratio of 1%, CGAN with a short-circuit turns ratio of 1%, and the simulated short-circuit turns ratio of 1%. Figure 6 This is a comparison chart of the percentage error at each point between the CGAN-generated short-circuit turns ratio current and the simulated short-circuit turns ratio current. Figure 7 This is a comparison chart of the percentage error at each point between the short-circuit turns ratio current generated by the CWGAN-GP-SA invention with the simulated short-circuit turns ratio current with a 1% short-circuit current. Figure 8 This is a comparison chart of the third harmonic content under various fault levels for CWGAN-GP-SA complete 1%, 3%, and 4% samples, CGAN complete 1%, 3%, and 4% samples, and complete simulation samples of the present invention. Figure 9 This is a schematic diagram of the DDF-ODENet in the initial inter-turn short circuit fault diagnosis method for aviation permanent magnet generators of the present invention; Figure 10 This is a confusion matrix diagram of the DDF-ODENet diagnostic results of the present invention; Figure 11 This is a trend chart of the diagnostic loss function of the DDF-ODENet of this invention; Figure 12 This is a trend chart of the diagnostic accuracy of the DDF-ODENet system of this invention; Figure 13 This is a trend chart showing the diagnostic accuracy of various methods after designing an ablation experiment using the DDF-ODENet method of this invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings.

[0018] Example 1 This invention provides a method for diagnosing initial inter-turn short-circuit faults in aircraft permanent magnet generators, such as... Figure 1 As shown, please follow these steps: Step 1: Based on the health model of the permanent magnet synchronous generator, establish a mathematical model under the inter-turn short-circuit fault state, and determine the stator current of the fault phase as the input signal for subsequent sample interpolation and fault diagnosis, starting from the influence of the fault branch current on the stator current; specifically: Step 1.1, set up a permanent magnet synchronous generator a When an inter-turn short-circuit fault occurs in the stator winding of a phase, the faulty phase winding can be considered as a healthy part. a 1. Fault Section a 2, and through the short-circuit resistor R f This forms a closed loop. R f Indicates the remaining insulation resistance; when R f When it is large enough, the motor is in good condition; when R f When the value is 0, it indicates that the faulty section is completely short-circuited. The short-circuit branch current is denoted as... i f ;like Figure 2 As shown; Neglecting the effects of core magnetic saturation and other nonlinear factors, the electrical equations for the faulty branch are shown in equation (1): (1) in, (2) (3) (4) (5) (6) (7) In the formula, f The subscript indicates the period after the failure; V abcf For the three phases abc and the faulty branch f The voltage; η The proportion of the number of turns of the short-circuited coil to the total number of turns of the coil in that phase; L Each phase of the stator is self-sensing; M Mutual induction; R s Phase resistance; i abcf For the three phases abc and the faulty branch f The current; λ PM,abcf for abc Three-phase and faulty branch f Permanent magnet linkage; λ PM,a , λ PM,b , λ PM,c , λ PM,f They are respectively abc Three-phase and faulty branch permanent magnet flux linkage; V 0,abcf for abc Zero-sequence voltage matrix of three phases and faulty branch f; V 0 represents the zero-sequence voltage component of each of the three phases.

[0019] Step 1.2: Starting from the phase voltage equations in equation (2), and considering the neutral point voltage... u The effect of 0 allows us to derive the expression for the short-circuit branch current; (8) In the formula, (9) in, u 0 is the neutral point voltage; from equations (8) and (9), it can be seen that the short-circuit branch current is... i f Based on short-circuit turns ratio η Short-circuit resistance R f Together with the neutral point voltage, it determines the electrical variables of the motor, thus directly reflecting the impact of the degree of inter-turn short-circuit fault on the motor's electrical variables.

[0020] Step 1.3, when the permanent magnet synchronous generator is under an inter-turn short-circuit fault, the direct-axis and quadrature-axis voltage equations in the rotating coordinate system are as shown in equation (10): (10) In the formula, u' d Direct-axis voltage of a permanent magnet synchronous generator under inter-turn short-circuit fault conditions; u' q Quadrature-axis voltage of a permanent magnet synchronous generator under inter-turn short-circuit fault conditions; i' d and i' q The measured currents of the d-axis and q-axis of the permanent magnet synchronous generator under the inter-turn short-circuit fault condition are shown in Equation (11); (11) As can be seen from the above analysis, the initial inter-turn short circuit fault will introduce additional fault branch current on the stator side, thereby changing the voltage and current distribution relationship of the motor.

[0021] right dq The axis current is transformed to obtain... abc Phase stator current; (12) Therefore, the fault a The phase stator current serves as a fault feature sample for subsequent missing sample interpolation models and fault diagnosis models.

[0022] Step 2 addresses the issue of missing initial fault sample data by constructing a missing sample generation model based on a conditional Wasserstein generative adversarial network. Gradient penalty, self-attention mechanism, and physical consistency constraint are then introduced sequentially to achieve high-quality imputation of missing fault level samples, thereby forming a complete training dataset. Figure 4 As shown; specifically: Step 2.1, based on the fault phase stator current signal determined in Step 1.3, let the acquired fault phase stator current sample be: (13) Where C represents the number of input current signal channels, N Indicates the sampling window length. The fault severity condition label is denoted as... y , is used to represent the short-circuit turns ratio. Among them, y =0 indicates the normal state. y =0.01, 0.02, 0.03, 0.04, and 0.05 represent the initial inter-turn short-circuit fault severity corresponding to short-circuit turns ratios of 1% to 5%.

[0023] The collected set of fault severity labels is represented as follows: (14) The set of missing fault degree labels that need to be imputed is represented as follows: (15) in, y ava This represents the set of collectable fault levels. y mis This represents the set of missing fault levels that require sample imputation. By distinguishing between collectable fault levels and missing fault levels, the training objects and imputation targets for the subsequent generative model can be clearly defined.

[0024] A dataset consisting of collectable samples is represented as: (16) in, D ava Indicates the collectable sample set. x i Indicates the first i One current sample, y i ava This indicates the collectable fault level label corresponding to the sample. n ava Indicates the number of samples that can be collected.

[0025] Step 2.2, given the collectable sample set D ava Based on fault severity condition labels, a missing sample generation model based on a conditional Wasserstein generative adversarial network is constructed. The model includes a generator and an evaluator. The fault severity condition labels are input into the model as continuous conditional variables, enabling the generative model to learn the correspondence between current samples and short-circuit turns ratio, thus possessing the ability to generate samples according to a specified missing fault level.

[0026] generator With random noise vector and fault severity condition labels As input, the output is the generated current sample corresponding to the fault level; (17) in, p z To represent random noise distribution, we take the standard normal distribution. This represents the current sample output by the generator.

[0027] evaluator Using current samples and their corresponding fault severity condition labels as input, the output is a real-valued score, which characterizes how closely the input sample approximates the distribution of the true sample under a given fault severity condition. Under these constraints, the optimization objective of CWGAN is expressed as: (18) in, This represents the joint distribution of actual current samples and fault severity labels. Represents a random noise distribution. The distribution of condition labels indicates the degree of failure.

[0028] Step 2.3: After constructing the CWGAN model, to improve the stability of the Wasserstein distance estimation and constrain the continuity of the evaluator function, a gradient penalty term is introduced into the loss function of the evaluator, thus constructing the CWGAN-GP model. The loss function of the evaluator is defined as: (19) in, Represents the gradient penalty coefficient. Indicates that it is composed of real samples With generated samples Samples obtained through random interpolation. The evaluator is constrained by a gradient penalty term to ensure it meets the continuity requirement for Wasserstein distance estimation, thereby mitigating training instability and gradient anomaly issues.

[0029] The adversarial loss function of the generator is defined as: (20) The generator minimizes the adversarial loss, making the generated samples approximate the real sample distribution under a given fault level.

[0030] Step 2.4: Based on the obtained CWGAN-GP model, a self-attention mechanism is introduced into the generator and evaluator to construct the CWGAN-GP-SA missing sample imputation model. The self-attention mechanism is used to model the correlation between different time positions in the current sequence, obtaining an enhanced feature representation, thereby enhancing the model's ability to model the global correlation of long current sequences. Let the intermediate feature representation of the network be: (twenty one) in, L Indicates the length of the feature sequence. d Representing feature dimensions. The query matrix, key matrix, and value matrix are obtained through linear mapping; (twenty two) in, W Q , WK , W V The self-attention weight matrix can be represented as: (The matrix contains learnable parameters.) (twenty three) in, d k This represents the dimension of the key vector. The self-attention output is: (twenty four) To avoid destroying the original feature representation after introducing a self-attention module, this invention uses residual connections to obtain the enhanced feature representation: (25) in, These are learnable weight coefficients. By introducing a self-attention mechanism, the generator can better maintain waveform consistency across multiple cycle scales when generating current samples, while the evaluator can judge from a global perspective whether the input sample conforms to the true current distribution characteristics under a given fault level.

[0031] Step 2.5: After completing the CWGAN-GP-SA modeling, a physical consistency constraint is further introduced into the generator loss function to ensure that the generated samples simultaneously satisfy the time-domain variation law, frequency-domain variation law, and fault severity condition correspondence of the initial inter-turn short-circuit fault. The generator total loss function is defined as: (26) in, Indicates the constraint weight coefficient. This represents a physical consistency constraint.

[0032] The physical consistency constraint is represented as follows: (27) in, Represents temporal characteristic constraints. Represents frequency domain characteristic constraints. Indicates consistency constraints for the degree of failure. and These are the corresponding weighting coefficients.

[0033] Time-domain feature constraints limit the amplitude, root mean square (RMS) value, and peak-to-peak value of the generated samples, ensuring they conform to the changing trends of the current time-domain characteristics under the corresponding fault severity. Frequency-domain feature constraints limit the fundamental amplitude, main harmonic components, and harmonic energy distribution of the generated samples, ensuring they conform to the spectral variation patterns of initial inter-turn short-circuit faults. Fault severity condition constraints constrain the correspondence between the generated samples and the input fault severity condition labels. Through these three types of constraints, the generated samples not only closely approximate the real samples in statistical distribution but also conform to the inter-turn short-circuit fault mechanism in terms of the correspondence between time domain, frequency domain, and fault severity.

[0034] Step 2.6: Using the collectable sample set formed in Step 2.1 and the generator total loss function determined in Step 2.5, the CWGAN-GP-SA model is trained. During training, the evaluator and generator parameters are updated alternately. The evaluator learns the distribution difference between real samples and generated samples under a given fault level by minimizing the evaluator loss function; the generator generates current samples that approximate the real sample distribution and satisfy physical consistency constraints by minimizing the generator total loss function.

[0035] After the model training is completed, the missing fault level label defined in step 2.1 is input into the generator to generate current samples under the corresponding missing fault level. (28) in, Represents a random noise vector. This indicates that a fault severity label is missing. This indicates the missing fault level current sample output by the generator.

[0036] The generated missing fault level sample set is represented as follows: (29) in, This represents the generated set of missing fault level samples. This indicates the number of samples generated.

[0037] When the fault severity label set is missing y mis When ={0.01, 0.03, 0.04}, it means that... y =0.01、 y =0.03 and y Inputting 0.04 into the trained generator will generate initial inter-turn short-circuit fault current samples corresponding to short-circuit turns ratios of 1%, 3%, and 4%.

[0038] Step 2.7: Merge the generated missing fault level sample set with the collectable sample set to obtain the complete initial inter-turn short-circuit fault sample set. ; (30) At this point, step 2 has completed the sample imputation for the missing fault set. The complete fault sample set includes normal state and initial inter-turn short-circuit fault samples corresponding to a short-circuit turn ratio of 1% to 5%, which are used as training samples for the fault diagnosis classification model.

[0039] Step 3 addresses the problem of weak initial inter-turn short-circuit fault characteristics and difficulty in distinguishing adjacent fault levels. A dual-domain feature fusion neural network is constructed. Through time-frequency dual-domain feature fusion, continuous dynamics modeling, and time attention convergence, accurate identification of the initial fine-grained fault severity is achieved. Figure 9 As shown; specifically: Step 3.1: The complete initial inter-turn short-circuit fault sample set obtained in Step 2 is randomly divided into training set, validation set, and test set according to the fault category label. The sample size ratio of the training set, validation set, and test set is 70:15:15.

[0040] Specifically, 70% of the samples are first randomly selected from the complete fault sample set according to category stratification as the training set, and the remaining 30% are used as a temporary sample set. Then, the temporary sample set is randomly divided into a validation set and a test set according to category stratification, with each set accounting for 15% of the complete fault sample set. Through the above stratification method, the sample proportions of each fault category corresponding to the normal state and the 1% to 5% short-circuit turns ratio are kept consistent in the training set, validation set, and test set.

[0041] The initial inter-turn short circuit multi-classification task uses normal state and fault state with a short-circuit turn ratio of 1% to 5% as the classification objects, and the number of fault categories is M=6.

[0042] To utilize the dynamic changes of the current signal in different time segments, the first fault sample in the complete fault sample set... i The current sequence samples are divided into T overlapping time windows arranged in chronological order according to a fixed window length and sliding step size; (31) Among them, the The first current sequence sample The current segment within the time window satisfies (32) in, Indicates the first i Current sequence samples The current segment in the t-th time window, C Indicates the number of current channels. L This indicates the sampling length of a single time window.

[0043] This invention uses single-phase current as the model input, that is Set the number of time windows to T, and the length of each window to [value missing]. The window has 1024 sampling points and a sliding step size. Because the window length is greater than the sliding step size, adjacent time windows will share some sampling points. Compared with non-overlapping partitioning, this method can reduce the information fragmentation caused by window segmentation and better preserve the continuous variation characteristics of the current signal between adjacent time segments.

[0044] For each time window obtained by division, the normalization process is performed according to the mean and standard deviation of the training set samples to obtain a normalized time window sequence, which is used as the input of DDF-ODENet to reduce the impact of the difference in amplitude scale of different samples on model training.

[0045] Step 3.2: For the standardized time window sequence, construct time-domain branches and frequency-domain branches to extract current waveform features and spectral harmonic features respectively, so as to characterize the slight changes in the current signal of the initial inter-turn short-circuit fault.

[0046] Time-domain branch directly uses the current window As input, a one-dimensional convolutional network is used to extract local amplitude changes, periodic morphology changes, and subtle distortion features caused by faults in the current waveform. The process is represented as follows: (33) in, This represents a nonlinear feature extraction mapping implemented by a time-domain branch. For learnable parameters, Indicates the first The first sample The temporal feature vector corresponding to each time window.

[0047] This invention employs a one-dimensional convolutional residual structure in the temporal branch to enhance the network's ability to extract local waveform distortion features. Simultaneously, a channel attention module is introduced into the multi-channel features after convolutional mapping, enabling the model to adaptively highlight effective feature channels related to the fault state.

[0048] Since relying solely on time-domain waveforms is insufficient to fully characterize harmonic variations caused by initial inter-turn short-circuit faults, a frequency-domain branch is further constructed after obtaining the time-domain characteristics. First, a Fourier transform is performed on each current window, and its amplitude spectrum is obtained. (34) in, Indicates Fourier transform, Indicates the amplitude value. Indicates the first The first sample The frequency domain amplitude spectrum corresponding to each time window.

[0049] Since the current time-domain sampling sequence is a real number sequence, its Fourier transform spectrum has conjugate symmetry. Therefore, the effective positive frequency components in the amplitude spectrum are retained, while the DC component is removed. Subsequently, the amplitude spectrum is subjected to logarithmic transformation and intra-sample normalization to obtain the preprocessed frequency domain input features. The preprocessed frequency domain input features are then input into the frequency domain feature extraction network. (35) in, This represents a nonlinear feature extraction mapping implemented by frequency domain branching. For learnable parameters, Indicates the first The first sample Frequency domain feature vectors corresponding to each time window.

[0050] The frequency domain branch is mainly used to extract the fundamental amplitude changes, main harmonic energy distributions, and spectral structure differences caused by the initial inter-turn short-circuit fault. Since the time domain branch has already preserved the timing morphology information of the current waveform, the frequency domain branch focuses on the harmonic energy distribution in the amplitude spectrum, thus complementing the time domain characteristics.

[0051] After obtaining the time-domain and frequency-domain features respectively, in order to fully utilize the complementary information of the two in waveform morphology and spectral structure, the time-domain features and frequency-domain features under the t-th time window are concatenated to obtain the dual-domain concatenated features: (36) in, Indicates the first i The dual-domain concatenated features corresponding to the t-th time window of each sample. This indicates a concatenation operation along the feature dimension. This operation combines time-domain waveform distortion features and frequency-domain harmonic structure features into a unified dual-domain representation, providing input for subsequent fusion mapping.

[0052] Since the time-domain features and frequency-domain features have different sources, their feature scales and distributions may differ. Therefore, by using fully connected mapping, layer normalization, and nonlinear activation functions, the spliced ​​dual-domain features are projected onto a unified latent space to ensure that subsequent continuous dynamic modeling can be performed at the same feature scale. (37) in, This represents the two-domain fusion mapping function. For learnable parameters, Indicates the first The first sample The fused feature vectors correspond to each time window. Through this mapping, time-domain features and frequency-domain features from different sources and at different scales are unified into the same latent space, thus forming a dual-domain fault feature representation.

[0053] For the first i Repeat the above time-domain branching extraction, frequency-domain branching extraction, feature splicing and fusion mapping process for all time windows of a current sequence sample to obtain its dual-domain fusion feature sequence. (38) in, Indicates the number of time windows. This represents the dual-domain fused feature sequence corresponding to the i-th current sequence sample. This sequence contains both time-domain waveform distortion information and frequency-domain harmonic structure information within each time window, while also preserving the dynamic relationship between different time windows. It serves as the input for the subsequent Neural ODE module, used to further model the continuous evolution process of the initial inter-turn short-circuit fault characteristics in the latent space.

[0054] Step 3.3: After obtaining the dual-domain fused feature sequence, considering the gradual evolution characteristics of the initial inter-turn short-circuit fault, the fault features between adjacent time windows are not completely independent but have a continuous changing relationship. To characterize the continuous evolution process of the fused features in the latent space, a Neural ODE module is introduced to represent the latent state changes as: (39) in, Indicates time The hidden state below, This represents a dynamic evolution function parameterized by a neural network. These are learnable parameters.

[0055] By numerically integrating the above ordinary differential equations, the continuous evolution of the fused feature sequence in the latent space can be obtained. (40) in, This represents the characteristics of time-series faults after Neural ODE modeling.

[0056] The above ordinary differential equations are solved using the fourth-order Runge-Kutta method, and the dual-domain fusion feature information of each time window is absorbed by the gated observation fusion mechanism, so as to take into account both the continuous evolution trend and the observation information of the current window.

[0057] After obtaining the continuous evolution features of Neural ODE, considering that the contributions of different time windows to the fault diagnosis results are not entirely the same, a temporal attention mechanism is introduced to highlight the key time windows. For the fused features of the t-th time window, its attention weight is expressed as: (41) in, The first value calculated by the attention network represents the value of the second value. The importance score of each time window This represents the corresponding attention weight.

[0058] The fusion features of all time windows are weighted and summed according to the attention weights; (42) in, This represents the global features resulting from the convergence of temporal attention.

[0059] After completing continuous evolution modeling and key window weighting, the Neural ODE output features are concatenated with temporal attention features to obtain the final feature representation for classification: (43) Will Inputting the data into a fully connected classifier yields the classification output: (44) in, This represents a fully connected classifier. This represents the classification output corresponding to the i-th sample. Softmax processing is applied to the classification output to obtain the predicted probability of each fault category, and the corresponding fault category is determined based on the predicted probability.

[0060] Step 3.4: Compare the predicted fault category probabilities obtained after processing the classification output using the Softmax function with the true fault category labels to construct a classification loss function. This loss function is then used to optimize the network parameters of DDF-ODENet. The training set is used to optimize the network parameters of DDF-ODENet, the validation set is used to evaluate model performance and determine the optimal model parameters during training, and the test set is used to independently evaluate the fault diagnosis performance of the optimal model after training is complete.

[0061] For a batch of samples, the classification loss function is expressed as: (45) in, Indicates batch size. Indicates the number of fault categories. Indicates the first The sample at the th The true label distribution on the class, This indicates that the model predicts the sample belongs to the first... The probability of a class.

[0062] During training, the classification loss function is calculated using training set samples and the DDF-ODENet network parameters are updated. After each round of training, the model recognition performance is evaluated using the validation set, and the model parameters corresponding to the highest accuracy on the validation set are saved. After training, the saved optimal model is tested using the test set to obtain the final fault diagnosis accuracy, confusion matrix, and classification evaluation results.

[0063] With the above structure, DDF-ODENet extracts current waveform distortion features through time-domain branching, extracts harmonic structure features through frequency-domain branching, and uses Neural ODE to model the continuous evolution process of the dual-domain fused feature sequence, thereby realizing the identification of fine-grained differences in the initial inter-turn short-circuit fault.

[0064] Example 2 The method of this invention was verified by ablation experiments. The DDF-ODENet of this invention was compared with ODENet with only time-domain branch + ODENet, only frequency-domain branch + ODENet, and only ODENet. The comparison and verification only changed the branch part, while keeping the other hyperparameter designs the same. The training results are shown in Table 1: Table 1. Ablation experimental results of the model of the present invention and the comparative model.

[0065] When there are missing samples for a fault level, the method of this invention can generate missing fault level samples that conform to the fault evolution law based on CWGAN-GP-SA, so that the generated samples are consistent with the simulated fault samples in terms of time domain waveform and frequency domain characteristics, and the average point-by-point difference in amplitude between the two is only 1.81%. After completing the sample set, a diagnostic model is further trained based on DDF-ODENet, which enables the model to integrate time domain, frequency domain and continuous dynamic evolution characteristics, and achieve a diagnostic accuracy of 98.9% for the initial inter-turn short circuit multi-classification task.

[0066] Example 3 Figure 3 This shows the three-phase stator current waveforms of a permanent magnet synchronous generator before and after an inter-turn short-circuit fault in phase a at t=1s. Figure 3It can be seen that before the fault occurred, the amplitudes of the three-phase stator currents were basically the same, with a phase difference of about 120°, and the current waveforms maintained good symmetry. When t=1s, an inter-turn short-circuit fault occurred in phase a, and a short-circuit branch current was formed in the faulty winding of phase a, which increased the amplitude of the phase a stator current. At the same time, due to the electromagnetic coupling relationship between the three-phase windings, the currents of phases b and c were also affected, and the system changed from a three-phase symmetrical state before the fault to a three-phase asymmetrical state after the fault.

[0067] Example 4 Figures 5-7 The comparison results of current samples obtained by different generation methods and simulated current samples are given when the short-circuit turns ratio is 1%. Figure 5 A time-domain comparison diagram of the simulated current waveform, the current waveform generated by CGAN, and the current waveform generated by the CWGAN-GP-SA of this invention; by Figure 5 It can be seen that after training using the initial inter-turn short-circuit fault current samples that can be collected, the generative model can generate current samples under the corresponding short-circuit turns ratio according to the specified missing fault level label. Taking the 1% short-circuit turns ratio fault sample as an example, the current waveform generated by CWGAN-GP-SA can maintain a periodic change trend and amplitude range similar to the simulated current waveform, indicating that the generative model has the ability to interpolate missing fault level samples.

[0068] Further comparison of the magnified local areas reveals that, compared to the current waveform generated by CGAN, the current waveform generated by the CWGAN-GP-SA model in this invention is closer to the simulated current waveform, with a higher degree of waveform overlap. This result indicates that the CWGAN-GP-SA model used in this invention can more fully learn the time-domain distribution characteristics of the initial inter-turn short-circuit fault current samples, and the generated missing fault level samples are closer to the true values. This model can be used to complete the fault sample set and provide data support for subsequent fault diagnosis model training.

[0069] Combination Figure 6 and Figure 7 The point-by-point error results show that the current generated by CGAN deviates significantly from the simulated current at some sampling points. However, the present invention CWGAN-GP-SA introduces Wasserstein distance, gradient penalty, and self-attention mechanism on the basis of conditional generative adversarial network, which can improve the training stability of the generative model and enhance the modeling ability of global correlation and local waveform changes of current sequence. The point-by-point error of the generated current is smaller overall and the error fluctuation range is lower, indicating that its generated samples have a better approximation effect in terms of time domain amplitude and local waveform details.

[0070] Example 5 Figure 8The figure shows a comparison of the third harmonic amplitudes of simulated values, CGAN-generated values, and the CWGAN-GP-SA-generated values ​​of this invention under different short-circuit turns ratios. FFT analysis of the generated samples reveals that the third harmonic amplitude of the CWGAN-GP-SA-generated values ​​is closer to the simulated values ​​than the CGAN-generated values. This indicates that the model can better learn the frequency domain characteristics of fault current changes, thereby improving the effectiveness of interpolating missing fault samples.

[0071] Example 6 Figures 10-12 The diagram shows the diagnostic results and training process curves of the DDF-ODENet fault diagnosis model of this invention. Figure 10 The confusion matrix of the diagnostic results shows that the identification results of each fault category are mainly concentrated on the diagonal, indicating that the model can accurately distinguish the degree of initial inter-turn short circuit faults corresponding to the normal state and the short-circuit turns ratio of 1% to 5%. Figure 11 The loss function trends for the training and validation sets show that as the number of iterations increases, the loss value decreases rapidly and gradually stabilizes, indicating that the model training process has good convergence. Figure 12 The results show the accuracy trends of the training and validation sets. The accuracy increases rapidly with the number of iterations and remains at a high level, indicating that the model has good fault identification capabilities. These results demonstrate that the DDF-ODENet of this invention can effectively fuse the time-domain, frequency-domain, and continuous dynamic characteristics of fault current, achieving accurate diagnosis of the severity of initial inter-turn short-circuit faults.

[0072] Figure 13 This is a comparison chart of the diagnostic accuracy trends of the DDF-ODENet of this invention and various ablation models. Compared with only time-domain branch + ODENet, only frequency-domain branch + ODENet, and only ODENet, the diagnostic accuracy of the DDF-ODENet of this invention is higher, indicating that the fusion of time-domain features, frequency-domain features, and Neural ODE continuous dynamic modeling can more fully characterize the initial inter-turn short-circuit fault characteristics, thereby improving the identification effect of different short-circuit turn ratio fault degrees.

Claims

1. A method for diagnosing initial inter-turn short circuit faults in aircraft permanent magnet generators, characterized in that, Specifically: Step 1: Based on the health model of permanent magnet synchronous generator, establish a mathematical model under the inter-turn short-circuit fault state, and determine the stator current of the fault phase as the input signal for sample interpolation and fault diagnosis. Step 2: Construct a missing sample generation model based on a conditional Wasserstein generative adversarial network, and introduce gradient penalty, self-attention mechanism and physical consistency constraint to imput missing fault level samples and form a dataset. Step 3: Construct a dual-domain feature fusion neural network of ordinary differential equations. Through time-frequency dual-domain feature fusion, continuous dynamics modeling, and time attention convergence, the initial fine-grained fault degree can be identified.

2. The method for diagnosing initial inter-turn short circuit faults in aircraft permanent magnet generators as described in claim 1, characterized in that, In step 1, specifically: Step 1.1, set up a permanent magnet synchronous generator a When an inter-turn short-circuit fault occurs in the stator winding of a phase, the faulty phase winding can be considered as a healthy part. a 1. Fault Section a 2, and through the short-circuit resistor R f Forming a closed loop; where, R f Represents the remaining insulation resistance; the short-circuit branch current is denoted as... i f ; Neglecting the effects of core magnetic saturation and other nonlinear factors, the electrical equations for the faulty branch are shown in equation (1): （1） in, （2） （3） （4） （5） （6） （7） In the formula, V abcf For the three phases abc and the faulty branch f The voltage; η The proportion of the number of turns of the short-circuited coil to the total number of turns of the coil in that phase; L Each phase of the stator is self-sensing; M Mutual induction; R s Phase resistance; i abcf For the three phases abc and the faulty branch f The current; λ PM,abcf for abc Three-phase and faulty branch f Permanent magnet linkage; λ PM,a , λ PM,b , λ PM,c , λ PM,f They are respectively abc Three-phase and faulty branch permanent magnet flux linkage; V 0,abcf for abc Zero-sequence voltage matrix of three phases and faulty branch f; V 0 represents the zero-sequence voltage component of each of the three phases; Step 1.2: Starting from the phase voltage equations in equation (2), and considering the neutral point voltage... u By influencing the effect of 0, the expression for the short-circuit branch current is derived. （8） In the formula, （9） in, u 0 represents the neutral point voltage; Step 1.3, when the permanent magnet synchronous generator is under an inter-turn short-circuit fault, the direct-axis and quadrature-axis voltage equations in the rotating coordinate system are as shown in equation (10): （10） In the formula, u' d Direct-axis voltage of a permanent magnet synchronous generator under inter-turn short-circuit fault conditions; u' q Quadrature-axis voltage of a permanent magnet synchronous generator under inter-turn short-circuit fault conditions; i' d and i' q The measured currents of the d-axis and q-axis of the permanent magnet synchronous generator under the inter-turn short-circuit fault condition are shown in Equation (11); （11） right dq The axis current is transformed in coordinates to obtain... abc Phase stator current; （12） Therefore, the fault a The phase stator current serves as a fault feature sample for the missing sample interpolation model and the fault diagnosis model.

3. The method for diagnosing initial inter-turn short circuit faults in aircraft permanent magnet generators as described in claim 2, characterized in that, Step 2 specifically involves: Step 2.1, based on the fault phase stator current signal determined in Step 1.3, let the acquired fault phase stator current sample be: （13） Where C represents the number of input current signal channels, N Indicates the sampling window length; the fault severity condition label is denoted as... y , is used to represent the short-circuit turns ratio; where, y =0 indicates the normal state. y =0.01, 0.02, 0.03, 0.04, and 0.05 represent the initial inter-turn short-circuit fault severity corresponding to short-circuit turns ratios of 1% to 5% respectively; The collected set of fault severity labels is represented as follows: （14） The set of missing fault degree labels that need to be imputed is represented as follows: （15） in, y ava This represents the set of collectable fault levels. y mis This represents the set of missing fault levels that require sample imputation; A dataset consisting of collectable samples is represented as: （16） in, D ava Indicates the collectable sample set. x i Indicates the first i One current sample, y i ava This indicates the collectable fault level label corresponding to the sample. n ava Indicates the number of samples that can be collected; Step 2.2, given the collectable sample set D ava Based on the fault degree condition label, a missing sample generation model based on conditional Wasserstein generative adversarial network is constructed. Step 2.3: After constructing the CWGAN model, a gradient penalty term is introduced into the loss function to construct the CWGAN-GP model; Step 2.4: Based on the obtained missing sample generation model CWGAN-GP, a self-attention mechanism is introduced into the generator and evaluator to construct the missing sample imputation model CWGAN-GP-SA. Let the intermediate features of the network be represented as: （21） in, L Indicates the length of the feature sequence. d Indicates feature dimension; The query matrix, key matrix, and value matrix are obtained through linear mapping; （22） in, W Q , W K , W V The learnable parameter matrix; The self-attention weight matrix can be represented as: （23） in, d k Indicates the dimension of the key vector; The output of self-attention is: （24） The enhanced feature representation is obtained by using residual connections: （25） in, These are learnable weight coefficients; Step 2.5: After completing the CWGAN-GP-SA model, a physical consistency constraint is further introduced into the generator loss function; The generator's total loss function is defined as: （26） in, Indicates the constraint weight coefficient. Represents physical consistency constraints; Step 2.6: Using the collectable sample set formed in Step 2.1 and the generator total loss function determined in Step 2.5, train the CWGAN-GP-SA model; during the training process, alternately update the evaluator and generator parameters; the evaluator learns the distribution difference between real samples and generated samples under a given fault level by minimizing the evaluator loss function; the generator generates current samples that are close to the real sample distribution and satisfy the physical consistency constraint by minimizing the generator total loss function. After the model training is completed, the missing fault level label defined in step 2.1 is input into the generator to generate current samples under the corresponding missing fault level. （28） in, Represents a random noise vector. Indicates the absence of a fault severity label. This represents a sample of missing fault level current from the generator output. The generated missing fault level sample set is represented as follows: （29） in, This represents the generated set of missing fault level samples. Indicates the number of samples generated; When the fault severity label set is missing y mis When ={0.01, 0.03, 0.04}, it means that... y =0.01、 y =0.03 and y Input 0.04 into the trained generator to generate initial inter-turn short-circuit fault current samples corresponding to short-circuit turns ratios of 1%, 3%, and 4%. Step 2.7: Merge the generated missing fault level sample set with the collectable sample set to obtain the complete initial inter-turn short-circuit fault sample set. ; （30）。 4. The method for diagnosing initial inter-turn short circuit faults in aircraft permanent magnet generators as described in claim 3, characterized in that, In step 2.2, the missing sample generation model includes a generator and an evaluator; generator With random noise vector and fault severity condition labels As input, the output is the generated current sample corresponding to the fault level; （17） in, p z Represents a random noise distribution. This represents a sample of the current output from the generator. evaluator Using current samples and their corresponding fault severity condition labels as input, and outputting real-valued scores, the optimization objective of CWGAN under the given constraints is expressed as: （18） in, This represents the joint distribution of actual current samples and fault severity labels. Represents a random noise distribution. The distribution of condition labels indicates the degree of failure.

5. The method for diagnosing initial inter-turn short circuit faults in aircraft permanent magnet generators as described in claim 4, characterized in that, In step 2.3, the loss function of the evaluator is: （19） in, Represents the gradient penalty coefficient. Indicates that it is composed of real samples With generated samples Samples obtained through random interpolation; The generator's adversarial loss function is: （20）。 6. The method for diagnosing initial inter-turn short circuit faults in aircraft permanent magnet generators as described in claim 5, characterized in that, In step 2.5, the physical consistency constraint is expressed as follows: （27） in, Represents temporal characteristic constraints, Represents frequency domain characteristic constraints. Indicates consistency constraints for the degree of failure. and These are the corresponding weighting coefficients.

7. The method for diagnosing initial inter-turn short circuit faults in aircraft permanent magnet generators as described in claim 6, characterized in that, Step 3 specifically involves: Step 3.1: The complete initial inter-turn short-circuit fault sample set obtained in Step 2 is randomly divided into training set, validation set and test set according to fault category label; the initial inter-turn short-circuit multi-classification task takes normal state and 1% to 5% short-circuit turn ratio fault state as classification objects, and the number of fault categories is M=6. The first in the complete fault sample set i The current sequence samples are divided into T overlapping time windows arranged in chronological order according to a fixed window length and sliding step size; （31） Among them, the The first current sequence sample The current segment within the time window satisfies （32） in, Indicates the first i Current sequence samples The current segment in the t-th time window, C Indicates the number of current channels; Using single-phase current as the model input, i.e. Set the number of time windows to T, and the length of each window to [value missing]. There are 1024 sampling points, and the window sliding step is 1024 sampling points; For each time window obtained by division, the normalization process is performed according to the mean and standard deviation of the training set samples to obtain a normalized time window sequence, which is then used as the input of DDF-ODENet; Step 3.2: For the standardized time window sequence, construct time-domain branches and frequency-domain branches to extract current waveform features and spectral harmonic features respectively, so as to characterize the weak changes in the current signal of the initial inter-turn short circuit fault. Time-domain branch directly uses the current window As input, a one-dimensional convolutional network is used to extract local amplitude changes, periodic morphology changes, and subtle distortion features caused by faults in the current waveform. The process is represented as follows: （33） in, This represents a nonlinear feature extraction mapping implemented by a time-domain branch. For learnable parameters, Indicates the first The first sample Temporal feature vectors corresponding to each time window; After obtaining the time-domain features, a frequency-domain branch is constructed; first, a Fourier transform is performed on each current window, and its amplitude spectrum is taken. （34） in, Indicates Fourier transform, Indicates the amplitude value. Indicates the first The first sample The frequency domain amplitude spectrum corresponding to each time window; The effective positive frequency components in the amplitude spectrum are retained, while the DC component is removed. Logarithmic transformation and in-sample normalization are performed on the amplitude spectrum to obtain the preprocessed frequency domain input features. The preprocessed frequency domain input features are then input into the frequency domain feature extraction network. （35） in, This represents a nonlinear feature extraction mapping implemented by frequency domain branching. For learnable parameters, Indicates the first The first sample Frequency domain feature vectors corresponding to each time window; After obtaining the time-domain features and frequency-domain features respectively, the time-domain features and frequency-domain features under the t-th time window are concatenated to obtain the dual-domain concatenated features: （36） in, Indicates the first i The dual-domain concatenated features corresponding to the t-th time window of each sample. This indicates a concatenation operation along the feature dimension; By using fully connected mapping, layer normalization, and nonlinear activation functions, the concatenated dual-domain features are projected onto a unified latent space. （37） in, This represents the two-domain fusion mapping function. For learnable parameters, Indicates the first The first sample The fused feature vectors corresponding to each time window; For the i Repeat the above time-domain branching extraction, frequency-domain branching extraction, feature splicing and fusion mapping process for all time windows of a current sequence sample to obtain its dual-domain fusion feature sequence. （38） in, Indicates the number of time windows. This represents the dual-domain fusion feature sequence corresponding to the i-th current sequence sample; Step 3.3: After obtaining the dual-domain fusion feature sequence, the Neural ODE module is introduced to represent the hidden state changes as follows: （39） in, Indicates time The hidden state below, This represents a dynamic evolution function parameterized by a neural network. These are learnable parameters; By numerically integrating the above ordinary differential equations, the continuous evolution of the fused feature sequence in the latent space is obtained. （40） in, This represents the characteristics of time-series faults after Neural ODE modeling; After obtaining the continuous evolution features of Neural ODE, a temporal attention mechanism is introduced; for the fused features of the t-th time window, the attention weight is expressed as: （41） in, The first value calculated by the attention network represents the value of the second value. The importance score of each time window This represents the corresponding attention weight; The fusion features of all time windows are weighted and summed according to the attention weights; （42） in, This represents the global features after temporal attention convergence; The Neural ODE output features are concatenated with the temporal attention features to obtain the final feature representation used for classification: （43） Will Input the fully connected classifier and obtain the classification output: （44） in, This represents a fully connected classifier. This represents the classification output corresponding to the i-th sample; Step 3.4: Compare the predicted probability of the fault category obtained after processing the above classification output with the Softmax function with the true fault category label, construct a classification loss function, and optimize the network parameters of DDF-ODENet through this loss function; The classification loss function is calculated using training set samples, and the DDF-ODENet network parameters are updated. After each round of training, the model recognition performance is evaluated using the validation set, and the model parameters corresponding to the highest accuracy on the validation set are saved. After training, the saved optimal model is tested using the test set to obtain the final fault diagnosis accuracy, confusion matrix, and classification evaluation results.

8. The method for diagnosing initial inter-turn short circuit faults in aircraft permanent magnet generators as described in claim 7, characterized in that, The classification loss function is expressed as: （45） in, Indicates batch size. Indicates the number of fault categories. Indicates the first The sample at the th The true label distribution on the class, This indicates that the model predicts the sample belongs to the first... The probability of a class.

Images