External gear pump airborne fault diagnosis method and system based on multi-teacher knowledge

Abstract

The invention relates to the technical field of artificial intelligence and fault signal diagnosis, in particular to an external gear pump airborne fault diagnosis method and system based on multi-teacher knowledge. According to the invention, on a data set {a pressure signal, a vibration signal and a fault label}, a student module is guided to train through multi-teacher network distillation knowledge, and then the student module is used to predict the fault category of the pressure signal; the multi-teacher network comprises at least two of a physical teacher module, a time sequence teacher module and a space teacher module; in the model training process, the loss function adopts the sum of the loss function of each teacher module and the knowledge distillation loss. According to the method, the problem that in the prior art, airborne rotating part diagnosis faces limitation of computing resources and data sources is solved, and efficient and reliable fault monitoring can be provided under complex working conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and fault signal diagnosis, and in particular to a method and system for on-board fault diagnosis of externally meshing gear pumps based on multi-teacher knowledge. BACKGROUND

[0002] In the aircraft fuel system, gear pumps play a vital role. Gear pumps can maintain stable fuel supply under different working conditions and are the core components for maintaining stable fuel flow. Once a fault occurs, it may lead to a decline in engine performance, and even endanger flight safety. If the gear pump stops running, fuel supply may be interrupted, which may cause thrust decay or engine shutdown. Pump failure may also cause fuel leakage or pressure fluctuations. Therefore, rapid and accurate fault detection of gear pumps is crucial.

[0003] At present, most rotating machinery fault diagnosis methods rely only on a single data type (usually vibration signals), and such methods have serious single-source dependence. Recently, researchers have begun to use multi-modal data methods, which effectively compensate for the limitations of fault diagnosis techniques that rely on a single data type. Multi-modal data integrates information from different sources such as vibration, temperature, pressure, electrical signals, and acoustic emission.

[0004] Multi-modal methods can collect data from different sources simultaneously, but actual on-board environments often rely only on pressure monitoring. The limitations of a single data source can reduce the performance and flexibility of diagnosis models trained based on multi-modal input, resulting in a significant reduction in their effectiveness in real-world scenarios.

[0005] Reliable multi-modal models that maintain good performance under incomplete data conditions have become a major challenge in the field of fault diagnosis research. In recent years, knowledge distillation has become an effective tool for knowledge transfer. This technology is initially used to compress and transfer the knowledge in complex high-performance teacher models to more compact and efficient student models. By transferring the knowledge of teacher models trained on complete modal data to student models that can only obtain partial modalities, the distillation process can be applied to modal missing scenarios. Even with limited modal information, student networks can indirectly absorb rich information through the distillation process. Knowledge distillation has shown significant results in handling cross-modal tasks such as emotion recognition, but systematic research on industrial equipment state monitoring and fault diagnosis under modal missing scenarios in aerospace scenarios is still lacking. In addition, existing methods rely on a single teacher model, limiting knowledge transfer across diverse representation spaces; especially in the field of on-board gear pump diagnosis, computational resources are severely limited, and existing model structures are difficult to meet the requirements. SUMMARY

[0006] In order to overcome the problem of limited computing resources and data sources in the prior art of diagnosing the rotating parts on board, the application provides a multi-teacher knowledge-based external gear pump on-board fault diagnosis method, which can provide efficient and reliable fault monitoring under complex working conditions.

[0007] The multi-teacher knowledge-based external gear pump on-board fault diagnosis method provided by the application trains a student module under the guidance of knowledge distillation of a multi-teacher network on a data set {pressure signal, vibration signal; fault label}, and then uses the student module to predict the fault category of the pressure signal.

[0008] The multi-teacher network includes at least two of a physical teacher module, a time sequence teacher module and a space teacher module; the physical teacher module adds the feature coefficients obtained by encoding the pressure signal and the feature coefficients obtained by frequency domain extraction, and then diagnoses the fault category; the time sequence teacher module extracts the time sequence fusion features of the pressure signal and the vibration signal, and predicts the fault category based on the fusion feature sequence; and the space teacher module diagnoses the fault category based on the spatial feature distribution of the pressure signal and the vibration signal.

[0009] During the model training process, the loss function uses the sum of the loss functions of each teacher module and the knowledge distillation loss.

[0010] Preferably, the physical teacher module includes:

[0011] A physical guiding unit extracts frequency domain feature coefficients based on the frequency domain variation values of the pressure signal.

[0012] An encoder generates encoded feature coefficients of the pressure signal.

[0013] The frequency domain feature coefficients and the encoded feature coefficients obtain fault feature coefficients through dimension stacking.

[0014] A first classifier predicts the classification based on the fault feature coefficients.

[0015] Preferably, the physical teacher module further includes a signal reconstruction unit which reconstructs the pressure signal based on the fault feature coefficients obtained by stacking the frequency domain feature coefficients and the encoded feature coefficients; and the loss function of the physical teacher unit is the sum of the signal reconstruction loss and the classification loss.

[0016] The reconstructed pressure signal

[0017] wherein, represents the fault feature coefficients, P(t) is the predicted coefficient of the nth frequency component of the frequency domain transformation; t is the time, I and J are the number of low frequency components and the number of high frequency components after frequency domain variation, respectively; g n and ζ n respectively represent The position and phase angle of the corresponding frequency component in the frequency spectrum, ω is the meshing frequency, and c0 represents the offset term.

[0018] Preferably, the time sequence teacher module diagnoses the fault label based on the pressure signal and the vibration signal; the pressure signal and the vibration signal are first equally divided into N small segments and time sequence aligned; a multi-head attention mechanism is used to fuse the time sequence aligned signal segments to generate a fusion feature sequence with a length of N; and the fusion feature sequence is used to predict the fault category.

[0019] Preferably, the time sequence teacher module further sequentially predicts a next fusion feature of each subsequence based on the subsequence on the fusion feature sequence to generate a reconstructed feature sequence corresponding to the fusion feature sequence; the reconstructed feature sequence is used to predict the fault category; the reconstructed feature sequence lacks a first feature value relative to the fusion feature sequence; a loss function of the time sequence teacher module is a sum of a signal reconstruction loss and a classification loss; and the signal reconstruction loss uses a mean square error of the fusion feature sequence and the reconstructed feature sequence.

[0020] Preferably, a GPT network is used to generate the reconstructed feature sequence based on the fusion feature sequence.

[0021] Preferably, a knowledge distillation loss L KD is calculated according to the following formula:

[0022]

[0023] wherein M is a combination of teacher modules, is a tensor value of a corresponding category c output by a classifier of a teacher module m e M, σ m,c is a probability value of the corresponding category c in a category probability distribution; w m is a weight of the teacher module m, is a category cross-entropy loss of the teacher module m, and C is a total number of fault categories.

[0024] Preferably, the student module is sequentially connected by two convolution units, an adaptive average pooling layer, and a fourth classifier, and the convolution unit is sequentially connected by a one-dimensional convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer.

[0025] Preferably, the data set is constructed in the following manner: first, the vibration signal and the pressure signal of the gear pump are collected, and a resampling method based on Fourier transform is used to correct the time axis of the sampling signal; then, the corrected vibration signal and the corrected pressure signal are time aligned and windowed, and the mean value of the first observation window is used as a reference value, and the corrected vibration signal and the corrected pressure signal are respectively subtracted by the corresponding reference value to form a data set sample.

[0026] The application provides an external gear pump on-board fault diagnosis system based on multi-teacher knowledge.

[0027] The application has the advantages that:

[0028] (1) The external gear pump on-board fault diagnosis method based on multi-teacher knowledge introduced in the application introduces multiple teacher models, wherein a time sequence teacher is used for modeling a time sequence mode, a space teacher focuses on a space feature, and a physical teacher describes the working mechanism of a gear pump. Through distilling the knowledge of the three teacher models, a student model enhances the processing capacity for complex information. In view of the strict limitation of computing resources in on-board deployment, the application adopts a knowledge distillation strategy to migrate the diagnosis capacity of the teacher models to a smaller and more efficient student network. The method significantly reduces the model size and computing cost while maintaining the diagnosis accuracy.

[0029] (2) The application constructs a fault diagnosis framework capable of fusing multiple space knowledge, adapting to modal missing and meeting the deployment efficiency requirements, so as to realize the actual needs of vehicle-mounted fault diagnosis and cope with the challenges of modal missing and deployment limitation. The framework combines physical-based modeling and deep learning in structure and combines multi-teacher knowledge migration and lightweight student model training in algorithm. The core is to use complete modal data in a ground laboratory environment through multiple teacher models, and then migrate the knowledge to a lightweight student network, aiming to solve the fault diagnosis problem of limited modal and insufficient computing resources in a vehicle-mounted scene.

[0030] (3) The application has higher diagnosis accuracy and robustness under the conditions of data missing and resource limitation, and verifies the application potential in actual engineering scenes such as aerospace. The physical teacher module proposed in the application is particularly suitable for pressure pulsation signals, and the multi-teacher strategy based on physical driving has good universality. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 A structure diagram of an external gear pump on-board fault diagnosis model based on multi-teacher knowledge is provided in the application.

[0032] Figure 2 A flowchart of an external gear pump on-board fault diagnosis method based on multi-teacher knowledge is provided in the application.

[0033] Figure 3 A structure diagram of a physical teacher module is provided in the application.

[0034] Figure 4This is a visualization of the features of different models under working condition 1 in the embodiment. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0036] Reference Figure 1 , Figure 2 This embodiment proposes an onboard fault diagnosis model for external gear pumps based on multi-teacher knowledge. The model is trained by a multi-teacher model to obtain a student model, which then predicts the type of onboard fault of the external gear pump based on pressure signals.

[0037] The multi-teacher diagnostic model constructed in this embodiment includes a student module and at least two teacher modules selected from the physics teacher module, the space teacher module, and the temporal teacher module.

[0038] The physics teacher module performs a Fast Fourier Transform (FFT) on the measured pressure signal P(t) to convert it to the frequency domain. Then, it uses a physical information-guided masking operation to preserve the frequency components of interest, thus retaining the feature coefficients within the band of interest. Next, it uses the nonlinear mapping capability of the neural network to map the pressure signal onto the encoded features. Finally, it adds the feature coefficients obtained from the physical guidance mask and the encoder to achieve signal reconstruction as combined feature coefficients.

[0039] Reference Figure 3 The physics teacher module includes an encoder, a physics guidance unit, a signal reconstruction unit, and a first classifier.

[0040] The physical guidance unit is used to map the input pressure signal P(t) into characteristic coefficients {a}. i ,b j |1≤a≤I,1≤b≤J};I and J are empirical values, representing the number of low-frequency components and the number of high-frequency components after the frequency domain transformation of P(t), respectively; the mapping model is:

[0041]

[0042] f1(t) represents the low-frequency component of the frequency domain transform of P(t), f2(t) represents the high-frequency component of the frequency domain transform of P(t), and a i and α i b represents the amplitude and phase angle of the i-th low-frequency component, respectively. j c j dj and β j are the amplitude, amplitude modulation coefficient, frequency modulation coefficient and phase angle of the jth high frequency component respectively; a i and b j , c j , d j are the coefficients to be learned; ω is the meshing frequency, l is the number of gears, and t is the time length; J m denotes the Bessel function; B(t) is the high frequency component concentrated at jω, and C(t) is the high frequency component concentrated at (j-1)ω or (j+1)ω.

[0043] The physical guidance unit performs physical information guidance based on the frequency domain variation value of the pressure signal P(t), extracts the frequency domain feature coefficient, that is, the feature coefficient {a i ,b j |1≤a≤I,1≤b≤J};

[0044] The encoder encodes the pressure signal P(t) to obtain the encoded feature coefficient. The encoder can specifically use a ResNet50 network.

[0045] The frequency domain feature coefficient and the encoded feature coefficient are superimposed in dimension to obtain the fault feature coefficient, and the signal reconstruction unit reconstructs the pressure signal based on the fault feature coefficient; the first classifier predicts the classification based on the fault feature coefficient .

[0046] Let the reconstructed signal of P(t) be denoted as is the fault feature coefficient;

[0047]

[0048] wherein, is the prediction coefficient of the nth frequency component of the frequency domain transformation of P(t); is the prediction value of a i , and is the prediction value of b j ;

[0049]

[0050] wherein, F(·) denotes the Fourier transform; R(·) denotes the coefficient extraction; φ(·) denotes the encoder, which can specifically use a ResNet50 network; g n and ζ n respectively denote the position and phase angle of the corresponding frequency component in the frequency spectrum, ω is the meshing frequency, and c0 denotes the offset term.

[0051] The loss function of the physical teacher unit is , which is the signal reconstruction loss and the classification loss ; the signal reconstruction loss The signal reconstruction loss adopts the square of the two-norm of the original input signal P(t) and the reconstructed signal The classification loss adopts the cross-entropy loss of the predicted label output by the first classifier and the real label.

[0052] The time sequence teacher module comprises a pressure coding unit, a vibration coding unit, a multi-head attention unit, a GPT decoding unit and a second classifier.

[0053] The pressure coding unit equally divides the pressure signal P(t) into N small segments p n , P(t) = {p n |1≤n≤N};

[0054] The vibration coding unit equally divides the vibration signal V(t) into N small segments v n , V(t) = {v n |1≤n≤N};

[0055] The multi-head attention unit adopts a multi-head cross-attention mechanism to perform multi-modal feature extraction and fusion on each sample {p n ,v n} after time sequence alignment, to generate fusion features f n ; the second classifier classifies according to the fusion feature sequence {fn|1≤m≤N}.

[0056] The GPT decoding unit predicts the next feature f m-1 2≤m≤N based on the fusion feature sequence {f1,f2,…,f 2≤m≤N.

[0057] The loss function of the time sequence teacher module is the sum of the signal reconstruction loss and the classification loss ; the signal reconstruction loss adopts the mean square error of the fusion feature sequence {f2,f3,…,f m} and the reconstructed feature sequence The classification loss adopts the cross-entropy loss of the predicted label output by the second classifier and the real label.

[0058] The space teacher module is composed of a spatial transformation network (Spatial Transformer Networks) and a third classifier, the spatial transformation network processes the pressure signal and the vibration signal to extract spatial features, and the third classifier predicts classification based on the spatial features; the loss function of the space teacher module adopts the cross-entropy loss of the predicted label output by the third classifier and the real label.

[0059] The student module is sequentially connected by two convolution units, an adaptive average pooling layer (AdpAvgpoolld), and a fourth classifier; the convolution unit is sequentially connected by a one-dimensional convolution layer (Convld), a batch normalization layer (BatchNormld), an activation layer, and a maximum pooling layer (Maxpoolld); the activation layer can specifically adopt a Relu activation function. The input of the student module extracts the time domain features before and after the convolution unit, and the batch normalization layer and the maximum pooling layer are combined to realize feature compression and optimization; finally, the adaptive average pooling is used to obtain the final low-dimensional features for classification prediction. In this way, the effective representation learning ability is guaranteed, and the computing efficiency is also improved. The loss L S The cross-entropy loss of the predicted label output by the fourth classifier and the real label.

[0060] The first classifier, the second classifier, the third classifier, and the fourth classifier have the same structure, which first generates a category tensor value based on the input data, and then activates the tensor value vector to obtain a category probability distribution. In specific implementation, the tensor value vector activation can adopt a Softmax function.

[0061] Let M be the combination of the teacher modules in the multi-teacher diagnosis model, and the tensor value of the corresponding category c output by the teacher module m e M is denoted as The probability value of the corresponding category c in the category probability distribution is denoted as σ m,c ; let the weight of the teacher module m be denoted as w m , and the knowledge distillation loss be denoted as L KD ;

[0062]

[0063] Where C is the total number of fault categories; is the category cross-entropy loss of the teacher module m; y c is a binary number representing the category c in the real label, that is, if the real label is the category c, then y c = 1, otherwise y c = 0.

[0064] In this embodiment, the multi-teacher diagnosis model is first trained on the ground data set {pressure signal, vibration signal; fault label} to convergence, and then the student model is extracted as a student model for predicting the on-board fault category of the external mesh gear pump, that is, the student model predicts the fault category based on the air pressure signal.

[0065] Referring to Figure 2 , the training steps of the multi-teacher diagnosis model are as follows:

[0066] S1, initialize the multi-teacher diagnosis model;

[0067] S2, extract training samples from the ground data set to iterate the multi-teacher diagnosis model, and then extract validation samples to calculate the loss function L of the multi-teacher diagnosis model total , which is the sum of the loss function of all teacher modules and the knowledge distillation loss L KD .

[0068] S3, update the multi-teacher diagnosis model by back propagation of the loss function L total ;

[0069] S4, repeat steps S2 and S3 until the multi-teacher diagnosis model converges, and then extract the student module as the student model.

[0070] The following specific embodiments are used to verify the experimental models given by the present application, and the experimental models are student models trained by different multi-teacher diagnosis models.

[0071] In this embodiment, different working conditions are simulated through a fuel gear pump experimental platform, data samples {pressure signal, vibration signal} are collected through an external gear pump fault test device and various sensors, and fault labels are manually labeled to form a data set {pressure signal, vibration signal; fault label}. In this embodiment, the fixed length of the data sample is 5120.

[0072] During the experiment, the data acquisition channel is configured in ADC mode, the input voltage range is ±5V, the sampling frequency is 20.48kHz, and the maximum recording time is 60 seconds. Each fault condition is repeated three times, and each collection time is not less than 30 seconds. Considering that environmental noise and changes in operating state may affect the data collection process, the pressure pulsation signal is preprocessed before the experiment to ensure the reliability of the data. The preprocessing includes the following two steps:

[0073] Correction for speed fluctuation: a resampling method based on Fourier transform is used to calibrate and stabilize the signal, and the time axis is corrected;

[0074] Drift compensation: a correction method based on mean value is introduced to correct the drift in the pressure signal. Specifically, the mean value of the first batch of observation samples is taken as the reference, and the subsequent collected data is subtracted from the reference value as the sample. Through the above correction, the offset caused by drift is successfully eliminated, so that more consistent and more accurate measurement results are obtained.

[0075] In this embodiment, the data set covers eight different operating states, i.e. real labels, including one normal state, five single fault conditions and two compound fault conditions.

[0076] The eight real labels and three working conditions in this embodiment are shown in Tables 1 and 2.

[0077] Table 1: Fault categories

[0078]

[0079]

[0080] Table 2: 3 working conditions

[0081] Operating condition Gear rotational speed Gear operating frequency Operating condition 1 600 r / min 10 Hz Operating condition 2 1050 r / min 17.5 Hz Operating condition 3 1500 r / min 25 Hz

[0082] In this embodiment, the total data set is divided into a training set, a validation set and a test set in a ratio of 6:2:2, wherein the training set contains 576 samples, and the validation set and the test set each contain 192 samples. The learning rate is set to 0.001, and an early stopping mechanism is introduced to prevent overfitting. The Adam optimizer is used in the training.

[0083] In this embodiment, four experimental models and three comparative models are trained and tested. The models are trained on the training set and the validation set, and then the fault categories are predicted based on the individual pressure signals on the test set.

[0084] The four experimental models in this embodiment are as follows:

[0085] Experimental model 1: The multi-teacher diagnosis model only contains a time sequence teacher module and a space teacher module; during the training process, the calculation of the total loss function L total refers to formulas (10-1) to (13-1):

[0086]

[0087] Experimental model 2: The multi-teacher diagnosis model only contains a physical teacher module and a space teacher module; during the training process, the calculation of the total loss function L total refers to formulas (10-2) to (13-2):

[0088]

[0089]

[0090] Experimental model 3: The multi-teacher diagnosis model only contains a time sequence teacher module and a physical teacher module; during the training process, the calculation of the total loss function L total refers to formulas (10-3) to (13-3):

[0091]

[0092] Experimental model 4: The multi-teacher diagnosis model contains a physical teacher module, a time sequence teacher module and a space teacher module; during the training process, the calculation of the total loss function L totalThe calculation of the formula (10-4) to the formula (13-4) and the formula (14) is as follows:

[0093]

[0094] wherein y c is a binary number in the true label representing the class c, that is, the true label is the class c, then y c = 1, otherwise y c = 0; σ p,c and respectively are the tensor value and the probability of the class c output by the physics teacher module, is the cross-entropy loss of the physics teacher module; σ t,c and respectively are the tensor value and the probability of the class c output by the time teacher module, is the cross-entropy loss of the time teacher module; σ s,c and respectively are the tensor value and the probability of the class c output by the space teacher module, is the cross-entropy loss of the space teacher module.

[0095] Specifically:

[0096]

[0097]

[0098] In this embodiment, the experimental models 1-4 are all trained on the ground data set, and then the student module is extracted as the experimental model for testing on the air data set.

[0099] In this embodiment, the comparative model adopts the DAFT network, the MMTM network, the MedFuse and the independent student module.

[0100] In this embodiment, the test results of each model are shown in Table 3.

[0101] Table 3: Model test accuracy under each working condition

[0102] Model abbreviation Operating condition 1 Operating condition 2 Operating condition 3 DAFT (Comparative Model 1) 0.828 0.989 0.849 MMTM (Comparative Model 2) 0.878 0.919 0.732 MedFuse (Comparative Model 3) 0.708 0.776 0.789 S (Comparative Model 4) 0.750 0.682 0.833 NOphy (Experimental Model 1) 0.850 0.979 0.843 NOtem (Experimental Model 2) 0.850 0.974 0.875 NOspa (Experimental Model 3) 0.891 0.875 0.713 ALL (Experimental Model 4) 0.979 1 0.948

[0103] As can be seen from Table 3, in the configuration using only the student model S, the model fails to utilize the multi-modal information from the vibration signals, although it receives the same pressure signal input as the physical teacher, and thus the accuracy is generally low. In contrast, comparing experimental models 1-4 and comparative model 3, it can be seen that the student model guided by the knowledge distilled from the teacher network shows consistent and significant performance improvement across the three working conditions. These results show that the student model is difficult to extract effective multi-modal features alone, and without the support of distillation, it is easy to fail in complex environments - unable to meet the requirements of robustness and accuracy in practical applications.

[0104] When the spatial teacher module is removed, the student model can still benefit from the physical teacher and the temporal teacher to obtain multi-modal knowledge. However, in working condition 3, the accuracy decreases significantly. This highlights the unique role of the spatial teacher in capturing the local spatial structure in the vibration signals and the pressure signals, and its absence seriously affects the model's ability to identify complex fault patterns.

[0105] When comparing the complete model ALL with the configurations in which any one of the teacher networks is removed, consistent decreases in diagnostic accuracy are observed. This observation emphasizes the necessity of the proposed multi-teacher structure, which jointly utilizes physical, temporal, and spatial knowledge. By distilling the knowledge of the temporal, spatial, and physical teachers to the student model using only the pressure signals, the proposed framework significantly reduces the performance fluctuations under the three operating conditions, indicating that the method not only compensates for the absence of modalities, but also achieves stable and robust fault identification capabilities under various working conditions.

[0106] In this embodiment, t-SNE is used to reduce the dimensionality of the features extracted by the model for test samples and visualize the analysis, Figure 4 which shows the fault features learned under working condition 1. From Figure 4 It can be seen that in the test results of NO phy and NO tem, some fault features are overlapped; using only the student network S, the fault feature overlap problem is more serious, only six faults are partially distinguished, and the features are not complete. This shows that the features extracted by the model cannot completely restore the fault state, and the performance is insufficient without the support of distillation. However, the fault feature distinguishing ability of the student model and all experimental models is better than that of DAFT, MMTM, and MedFuse.

[0107] The ALL model can clearly distinguish between eight fault states and form clear clusters in the feature space, showing good classification performance.

[0108] Of course, the present application is not limited to the details of the above-described exemplary embodiments but comprises the same or similar structures which can be realized in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all respects as illustrative and not restrictive, the scope of the present application being defined by the appended claims rather than the above description, and it is intended that all changes which come within the meaning and range of equivalency of the claims are embraced therein. Any reference signs in the claims should not be construed as limiting the claims to the figures in which the reference signs are used.

[0109] Furthermore, it should be understood that although the description is made on the embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and the skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be combined appropriately to form other embodiments which can be understood by the skilled in the art.

[0110] The technologies, shapes, and structural parts not described in detail in the present application are well-known technologies.

Claims

1. A method for diagnosing airborne faults in an external gear pump based on multi-teacher knowledge, characterized in that, In the dataset {pressure signal, vibration signal; fault label}, student modules are trained by knowledge distillation through a multi-teacher network, and then the student modules are used to predict the fault category of the pressure signal. The multi-teacher network includes at least two of the following: a physical teacher module, a temporal teacher module, and a spatial teacher module. The physical teacher module adds the characteristic coefficients obtained by encoding the pressure signal and the characteristic coefficients obtained by frequency domain extraction, and then diagnoses the fault category. The temporal teacher module extracts the temporal fusion features of the pressure signal and the vibration signal, and predicts the fault category based on the fusion feature sequence. The spatial teacher module diagnoses the fault category based on the spatial feature distribution of the pressure signal and the vibration signal. During model training, the loss function is the sum of the loss functions of each teacher module and the knowledge distillation loss.

2. The airborne fault diagnosis method for external gear pumps based on multi-teacher knowledge as described in claim 1, characterized in that, The physics teacher module includes: The physical guidance unit extracts frequency domain feature coefficients based on the frequency domain change value of the pressure signal; Encoder, the coding feature coefficients that generate pressure signals; The fault feature coefficients are obtained by superimposing the frequency domain feature coefficients and the coding feature coefficients in terms of dimensions; The first classifier predicts classification based on fault feature coefficients.

3. The on-board fault diagnosis method for external gear pumps based on multi-teacher knowledge as described in claim 1, characterized in that, The physics teacher module also includes a signal reconstruction unit, which reconstructs the pressure signal based on the fault feature coefficients obtained by superimposing frequency domain feature coefficients and coding feature coefficients; the loss function of the physics teacher unit is the sum of signal reconstruction loss and classification loss; Reconstructing pressure signals in, Represents the fault characteristic coefficient. Let g be the prediction coefficient of the nth frequency component after frequency domain transformation of P(t); t is time, and I and J are the number of low-frequency components and the number of high-frequency components after frequency domain transformation, respectively; n and ζ n They represent The corresponding frequency component has its position and phase angle in the spectrum, where ω is the meshing frequency and c0 represents the offset term.

4. The airborne fault diagnosis method for external gear pumps based on multi-teacher knowledge as described in claim 1, characterized in that, The temporal teacher module diagnoses fault labels based on pressure and vibration signals. First, the pressure and vibration signals are divided into N equal segments and time-aligned. A multi-head attention mechanism is used to fuse the time-aligned signal segments to generate a fused feature sequence of length N. The fault category is predicted based on the fused feature sequence.

5. The on-board fault diagnosis method for external gear pumps based on multi-teacher knowledge as described in claim 3, characterized in that, The temporal teacher module also predicts the next fusion feature of each subsequence based on the subsequences on the fusion feature sequence, generating the reconstructed feature sequence corresponding to the fusion feature sequence; predicts the fault category based on the reconstructed feature sequence; the reconstructed feature sequence lacks the first feature value relative to the fusion feature sequence; the loss function of the temporal teacher module is the sum of the signal reconstruction loss and the classification loss; the signal reconstruction loss uses the mean square error of the fusion feature sequence and the reconstructed feature sequence.

6. The on-board fault diagnosis method for external gear pumps based on multi-teacher knowledge as described in claim 5, characterized in that, The GPT network is used to generate reconstructed feature sequences based on fused feature sequences.

7. The airborne fault diagnosis method for external gear pumps based on multi-teacher knowledge as described in claim 1, characterized in that, Knowledge distillation loss L KD The calculation formula is: Where M represents the combination of teacher modules, Let σ be the tensor value of the classifier output corresponding to category c for the teacher module m∈M. m,c w represents the probability value of category c in the category probability distribution. m Let m be the weight of the teacher module. Let C be the category cross-entropy loss of the teacher module m, and C be the total number of fault categories.

8. The on-board fault diagnosis method for external gear pumps based on multi-teacher knowledge as described in claim 1, characterized in that, The student module consists of two convolutional units, an adaptive average pooling layer, and a fourth classifier connected sequentially. The convolutional unit consists of a sequentially connected one-dimensional convolutional layer, a batch normalization layer, an activation layer, and a max pooling layer.

9. The on-board fault diagnosis method for external gear pumps based on multi-teacher knowledge as described in any one of claims 1-8, characterized in that, The dataset was constructed as follows: First, vibration and pressure signals of the gear pump were collected, and the time axis of the sampled signals was corrected using a resampling method based on Fourier transform. Then, the corrected vibration and pressure signals were time-aligned and windowed. The mean of the first observation window was used as the reference value, and the corrected vibration and pressure signals were subtracted from the corresponding reference values ​​to form the dataset samples.

10. An onboard fault diagnosis system for an external gear pump based on multi-teacher knowledge, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, the processor is connected to the memory, and the processor is used to execute the computer program to implement the on-board fault diagnosis method for external gear pumps based on multi-teacher knowledge as described in any one of claims 1-9.

Images