A digital-analog linkage rotor system feature migration fault diagnosis method and system

Abstract

The application relates to a digital-analog linkage rotor system feature migration fault diagnosis method and system, which comprises the following steps: introducing a misalignment fault and an imbalance fault into a constructed rotor system dynamics model to establish a rotor system fault dynamics differential equation and solve to obtain a fault displacement simulation signal; taking the fault displacement simulation signal as the input of a generator in a constructed gradient penalty condition deep convolution generative adversarial network, inputting an obtained initial generated signal and a real signal and fault label information into a discriminator to obtain a generated signal combined with mechanism characteristics and actual mechanical characteristics; and constructing a cross-condition-domain adaptive fault diagnosis model based on a feature migration learning theory, taking data of the generated signal combined with mechanism characteristics and actual mechanical characteristics as a source domain, taking to-be-tested other-condition data as a target domain, training the cross-condition-domain adaptive fault diagnosis model to perform cross-condition fault diagnosis on the rotor system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, in particular to a digital-analog linkage gas turbine rotor system feature migration fault diagnosis method and system. BACKGROUND

[0002] The gas turbine is a kind of power equipment applied in aviation, energy and other fields, and the rotor system is the core component of the gas turbine. Due to the complexity and variability of the operating environment and working conditions, the rotor system is prone to imbalance, misalignment and other faults, and if the faults are not found in time, the working performance and safe operation of the gas turbine will be directly affected. The mechanism model-based and data-driven methods have been relatively mature and successfully applied in rotor system fault diagnosis, but each has the problem of being difficult to apply to actual scenarios, and the digital-analog linkage method of mechanism and data fusion can make up for the shortcomings of each other.

[0003] In recent years, the digital-analog linkage fault diagnosis method has been studied in many fields at home and abroad. In the existing literature, principal component analysis and fuzzy C-means clustering (PCA-FCM) algorithm are used to classify PTA historical working conditions, a combination of BP neural network and partial least squares (PLS) multi-working condition model is established, and the material consumption index and quality index of PTA production process are predicted. In order to solve the energy consumption problem in the steelmaking-continuous casting process in the steel production, an improved hybrid model is proposed combined with the mechanism model, which improves the prediction accuracy, and the RBF neural network is used to directly learn the relevant process data of the mechanism model, the network is trained, and the material consumption and recovery are predicted. In order to realize the fine control of the energy flow network, the input-output model of the converter based on mechanism and data-driven is constructed. According to the actual production data, the relevant parameters of the converter smelting are obtained by using mathematical statistics and regression method. The final temperature of the molten steel is predicted by using neural network, and it is compared with the target molten steel temperature used by the mechanism model, so as to correct the mechanism model and improve the accuracy of the model. The data-driven fault diagnosis method faces the problem that fault data is difficult to obtain in different working conditions of complex equipment, and the data-driven method usually does not consider the mechanism and internal structure of the fault system, so it is difficult to accurately diagnose the fault. The existing digital-analog linkage research has been applied in many industrial scenes, but there is no perfect method in the rotor system fault diagnosis. SUMMARY

[0004] In view of the above problems, the purpose of the present application is to provide a digital-analog linkage gas turbine rotor system feature migration fault diagnosis method and system, which is based on conditional deep convolution generation adversarial network (CDCGAN) and sub-field transfer learning to construct a rotor system fault diagnosis model, and can solve the problem of fault diagnosis of the gas turbine rotor system under the complex background of few fault samples or uneven distribution of the rotor system and unlabeled test data.

[0005] To achieve the above object, the application adopts the following technical scheme: a rotor system feature migration fault diagnosis method of digital-analog linkage, comprising: based on a rotor system fault dynamics model, introducing misalignment faults and imbalance faults into the constructed rotor system dynamics model to establish a rotor system fault dynamics differential equation, and solving the rotor system fault dynamics differential equation to obtain a fault displacement simulation signal; taking the fault displacement simulation signal as the input of the generator in the constructed gradient penalty conditional deep convolution generative adversarial network, inputting the obtained initial generated signal together with the real signal and fault label information into the discriminator to obtain a generated signal integrated with mechanism characteristics and actual mechanical characteristics; constructing a cross-condition domain adaptive fault diagnosis model based on feature migration learning theory, taking the data of the generated signal integrated with mechanism characteristics and actual mechanical characteristics as the source domain, taking the to-be-tested other condition data as the target domain, training the cross-condition domain adaptive fault diagnosis model, and performing cross-condition fault diagnosis on the rotor system by the trained cross-condition domain adaptive fault diagnosis model.

[0006] Further, based on the rotor system fault dynamics model, the misalignment faults and imbalance faults are introduced into the constructed rotor system dynamics model, comprising: the rotor system fault dynamics model comprises a rotor shaft imbalance model for describing the imbalance, a rolling bearing model for describing the rolling bearing, a coupling misalignment fault model for describing the misalignment, and a fault model for describing the remaining faults; a rotor system lumped mass model is established based on the lumped mass degree of freedom method, and a rotor system fault dynamics differential equation is established based on the rotor system fault model.

[0007] Further, solving the rotor system fault dynamics differential equation comprises: the rotor system dynamics differential equation with misalignment faults is solved by the Runge-Kutta method to obtain a time displacement signal that can accurately show the motion characteristic change of the rotor system within a set parameter range as the fault displacement simulation signal under the fault state of the rotor system.

[0008] Further, inputting the obtained initial generated signal together with the real signal and fault label information into the discriminator to obtain a generated signal integrated with mechanism characteristics and actual mechanical characteristics, comprising: converting the real signal into a two-dimensional time-frequency signal by combining the deconvolution short-time fractional Fourier transform; inputting the generated signal and the converted two-dimensional time-frequency signal of the real signal together with the fault label information into the discriminator, punishing the norm of the parameters of the gradient penalty conditional deep convolution generative adversarial network to make the update of the parameters more stable, determining the objective function of the gradient penalty conditional deep convolution generative adversarial network, and then obtaining the generated signal integrated with mechanism characteristics and actual mechanical characteristics.

[0009] Further, the norm of the parameters of the gradient penalty conditional deep convolutional generative adversarial network is punished, including: introducing an orthogonal gradient penalty term, the gradient penalty term is the square or absolute value of the norm of the network parameters, and the norm of the network parameters is punished to make the update of the network parameters more stable.

[0010] Further, the objective function of the gradient penalty conditional deep convolutional generative adversarial network is:

[0011]

[0012] In the formula, V(.) represents a value function, represents the probability that the discriminator D(.) obtains real data;

[0013] represents the probability that the discriminator D(.) judges the sample of the generator G(z) to be true.

[0014] Further, a cross-condition domain adaptive fault diagnosis model is established based on feature transfer learning theory, including: a feature extraction layer, an adaptive layer and a fully connected layer; ResNet-50 network is used as the feature extraction layer for feature extraction, and a local maximum mean difference adaptive layer is introduced to align the sub-field distribution; a fully connected layer is added after the ResNet-50 global average pooling layer, which is set as the adaptive layer, and the local maximum mean difference is used for sub-field domain adaptation to obtain the cross-condition domain adaptive fault diagnosis model; wherein, the ResNet-50 network includes 49 convolution layers and 1 fully connected layer.

[0015] A rotor system feature transfer fault diagnosis system of digital-analog linkage, comprising: a first processing module, based on a rotor system fault dynamics model, introducing misalignment faults and imbalance faults into the constructed rotor system dynamics model to establish a rotor system fault dynamics differential equation, and solving the rotor system fault dynamics differential equation to obtain a fault displacement simulation signal; a second processing module, taking the fault displacement simulation signal as the input of the generator in the constructed gradient penalty conditional deep convolutional generative adversarial network, inputting the obtained initial generated signal together with the real signal and fault label information into the discriminator to obtain a generated signal integrating mechanism characteristics and actual mechanical characteristics; a fault diagnosis module, constructing a cross-condition domain adaptive fault diagnosis model based on feature transfer learning theory, taking the data of the generated signal integrating mechanism characteristics and actual mechanical characteristics as the source domain, taking the to-be-tested other working condition data as the target domain, training the cross-condition domain adaptive fault diagnosis model, and performing cross-condition fault diagnosis on the rotor system by the trained cross-condition domain adaptive fault diagnosis model.

[0016] A computer-readable storage medium storing one or more programs, the one or more programs comprising instructions which, when executed by a computing device, cause the computing device to perform any of the above-described methods.

[0017] A computing device comprising: one or more processors, memory, and one or more programs, where the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the above-described methods.

[0018] The present application has the following advantages due to the above technical solutions:

[0019] 1、The improved Runge-Kutta method can effectively handle this nonlinear case, which can dynamically adjust the step size according to the needs to ensure the accuracy of the numerical solution, so as to give a relatively accurate numerical solution, and has good stability, which can effectively avoid the divergence or oscillation of the numerical solution.

[0020] 2、The short-time fractional Fourier transform method based on deconvolution is used for time-frequency conversion in the present application, which can make the time-frequency graph resolution higher, the energy distribution more concentrated, and the characteristics more obvious, which is more conducive to fault diagnosis.

[0021] 3、The gradient penalty algorithm is applied in the conditional deep convolution generative adversarial network to solve the problem of decreased generalization ability, gradient disappearance or gradient explosion of the model in training and testing, and the norm of the model parameters is punished to make the update of the model parameters more stable.

[0022] 4、The mechanism model and data-driven are combined in the fault diagnosis of the present application, so that the obtained samples combine the mechanism characteristics and conform to the actual mechanical characteristics, and the linkage of the two methods brings the improvement of the diagnostic performance, and realizes the diagnosis problem across working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is a rotor system feature migration fault diagnosis method flowchart of the digital-analog linkage in the embodiment of the present application;

[0024] Figure 2 is a simplified model diagram of a rotor system in the embodiment of the present application;

[0025] Figure 3 is an unbalance fault model diagram of a rotor system in the embodiment of the present application;

[0026] Figure 4 is a rolling bearing model diagram in the embodiment of the present application;

[0027] Figure 5is a misalignment fault schematic diagram in the embodiment of the application;

[0028] Figure 6 is a generator structure diagram of a conditional deep convolutional generative adversarial network in the embodiment of the application;

[0029] Figure 7 is a discriminator structure diagram of a conditional deep convolutional generative adversarial network in the embodiment of the application;

[0030] Figure 8 is a Resnet-50 model structure diagram in the embodiment of the application;

[0031] Figure 9 is a sub-field domain adaptive cross-condition fault diagnosis model structure diagram in the embodiment of the application;

[0032] Figure 10 is a fault diagnosis model diagram based on digital-analog linkage in the embodiment of the application;

[0033] Figure 11a is a rotational speed at 1670r / min normal state signal time domain waveform comparison diagram in the embodiment of the application;

[0034] Figure 11b is a rotational speed at 1670r / min normal state signal frequency domain waveform comparison diagram in the embodiment of the application;

[0035] Figure 12a is a rotational speed at 1670r / min misalignment fault time domain waveform comparison diagram in the embodiment of the application;

[0036] Figure 12b is a rotational speed at 1670r / min misalignment fault frequency domain waveform comparison diagram in the embodiment of the application;

[0037] Figure 13a is a rotational speed at 1670r / min imbalance fault time domain waveform comparison diagram in the embodiment of the application;

[0038] Figure 13b is a rotational speed at 1670r / min imbalance fault frequency domain waveform comparison diagram in the embodiment of the application;

[0039] Figure 14 is a training flowchart of a conditional deep convolutional generative adversarial network in the embodiment of the application;

[0040] Figure 15 is a 1B-A and 2B-A confusion matrix diagram in the embodiment of the application. DETAILED DESCRIPTION

[0041] The application will be described in detail below with reference to the accompanying drawings and embodiments.

[0042] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present application.

[0043] It is to be noted that the terms used herein are only intended to describe specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form, unless the context clearly indicates otherwise, and it should also be understood that the terms "comprise" and / or "include" as used in the specification indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0044] A rotor system feature migration fault diagnosis method based on digital-analog linkage is invented for the problem of few fault samples and cross-condition fault diagnosis of gas turbine rotor system. The method first uses fault simulation signals of the rotor system and real experimental data under the same working condition as input, and uses the conditional deep convolution generative adversarial network to combine mechanism characteristics and mechanical characteristics of the generated samples. Secondly, based on the theory of transfer learning, a cross-condition domain adaptive fault diagnosis model is established, which uses data combining mechanism characteristics and mechanical characteristics as the source domain and the to-be-tested other working condition data as the target domain to train the fault diagnosis model. Through experimental verification, the method can effectively improve the generalization of the model and realize cross-condition fault diagnosis of the rotor system.

[0045] In an embodiment of the present application, a rotor system feature migration fault diagnosis method based on digital-analog linkage is provided. As shown in Figure 1 the method includes the following steps:

[0046] 1) Based on the rotor system fault dynamics model, the misalignment fault and the imbalance fault are introduced into the constructed rotor system dynamics model to establish the rotor system fault dynamics differential equation, and the rotor system fault dynamics differential equation is solved to obtain the fault displacement simulation signal;

[0047] 2) The fault displacement simulation signal is used as the input of the generator in the constructed gradient penalty conditional deep convolution generative adversarial network (CDCGAN-ogp), and the obtained initial generated signal is input into the discriminator together with the real signal and the fault label information to obtain a generated signal combining mechanism characteristics and actual mechanical characteristics;

[0048] 3) Constructing a cross-condition domain adaptive fault diagnosis model based on the feature transfer learning theory, using the data of the generated signal combining the mechanism characteristics and the actual mechanical characteristics as the source domain, and using the data of the other conditions to be tested as the target domain, training the cross-condition domain adaptive fault diagnosis model, and performing cross-condition fault diagnosis on the rotor system by the trained cross-condition domain adaptive fault diagnosis model.

[0049] In use, the application obtains mechanism characteristic simulation data through rotor dynamics simulation, thereby creating conditions for data enhancement in the digital-mechanical linkage method. A rotor system dynamics differential equation is established based on the Lagrange equation by using the lumped mass method, unbalance, misalignment or other faults are introduced, the rotor fault dynamics differential equation set is solved by the 4 / 5 order Runge-Kutta algorithm, the vibration displacement signals of the lumped mass in the X and Y directions are obtained, and the vibration signals are obtained by integral operation; the rotor system fault simulation signals and the real experimental data under the same condition are used as inputs, and the generated samples combine mechanism characteristics and mechanical characteristics based on the conditional deep convolution generative adversarial network; secondly, a cross-condition domain adaptive fault diagnosis model is established based on the transfer learning theory, the data combining mechanism characteristics and mechanical characteristics are used as the source domain, and the data of the other conditions to be tested are used as the target domain to train the fault diagnosis model, the application has high recognition accuracy, strong interpretability and low sample dependence, and can effectively perform fault diagnosis on the rotor system.

[0050] In the above step 1), based on the rotor system fault dynamics model, the misalignment fault and the unbalance fault are introduced into the constructed rotor system dynamics model, including the following steps:

[0051] 1.1) The rotor system fault dynamics model includes an unbalance model, a rolling bearing model, a misalignment fault model and a fault model for describing the remaining faults;

[0052] Specifically, in this embodiment, as shown in Figure 2 , a rotor system simplified model and a fault test bench simplified model are established based on the actual structure and working characteristics of the steam turbine. The construction of the fault model includes the following steps:

[0053] 1.1.1) Constructing an unbalance model;

[0054] As shown in Figure 3 , the rotor system unbalance fault model, Figure 3 , where o is the center of mass, o' is the mass center of the unbalance mass, ω is the rotational speed, m is the unbalance mass, e is the eccentricity, is the initial phase, and the unbalance force can be represented as F = meω 2 .

[0055] Let x1, y1, x2, y2 represent the horizontal and vertical displacement of the center of rotation and the center of mass, respectively, and the displacement in each direction is:

[0056]

[0057] System kinetic energy function:

[0058]

[0059] System potential energy function:

[0060]

[0061] System dissipation energy function:

[0062]

[0063] Let L be the Lagrangian function, then L = T - V, and the generalized force Q j is expressed by the Lagrange equation as:

[0064]

[0065] where q j is the generalized coordinate.

[0066] Substituting equations (3) to (5) into the vibration differential equation of the geometric center O of the disc:

[0067]

[0068] Therefore, the unbalance force exerted by the unbalance mass on the rotor system is:

[0069]

[0070] Let the mass of the rotor be m, the rotational speed of the rotor be ω, the unbalance mass be m, and the eccentricity of the unbalance mass be e. When the unbalance mass rotates, it will generate a centrifugal force F c , whose size is: F c = meω 2 According to Newton's second law, the centrifugal force is equal to the mass multiplied by the acceleration, that is: F c = ma, combining the above two formulas, we can get: a = eω 2 / m, since the acceleration is the source of the exciting force, therefore the size of the exciting force is: F c = ma = eω 2 r, therefore, the size of the rotor unbalance exciting force is related to the unbalance mass, the rotational speed and the eccentricity, and the size of the exciting force can be reduced by reducing the unbalance mass or increasing the stiffness of the rotor.

[0071] 1.1.2) Constructing rolling bearing model;

[0072] The rotor system bearing generally uses rolling bearing, such as Figure 4 As shown, the rolling bearing is composed of bearing raceway, ball and cage. The force analysis is carried out on the rolling bearing, and the linear velocity of the outer ring is v out , the linear velocity of the inner ring is v in , the radius of the outer ring is R, the radius of the inner ring is r, the angular velocity of the outer ring is ω out , and the angular velocity of the inner ring is ω in , so v out = ω out R, v in = ω in r.

[0073] The linear velocity of the cage is:

[0074]

[0075] Since the outer ring is fixed, ω out = 0, so the angular velocity of the cage is:

[0076]

[0077] Let the number of balls be N b , then the angular velocity position of the jth ball is θ j ,

[0078]

[0079] j = 1, 2,... N b Let the displacement of the shaft center in the horizontal and vertical directions be x, y, and the bearing clearance be δ c , then the normal deformation of the jth ball is:

[0080] δ j = x cos θ j + y sin θ j - δ c (11)

[0081] The contact force is calculated by the nonlinear Hertz contact theory:

[0082] F j = C b (x cos θ j + y sin θ j - δ c ) 3 / 2 H(x cos θ j + y sin θ j - δ c ) (12)

[0083] where C b is the Hertz contact stiffness, H is the Heaviside function

[0084] F j The components in the horizontal and vertical directions are:

[0085]

[0086] The axial force generated by the rolling bearing is the sum of the forces on the balls:

[0087]

[0088] 1.1.3) Construction of misalignment model;

[0089] As Figure 5 is the schematic diagram of the coupling when the comprehensive failure occurs, O1 and O2 are the rotation centers of the left and right end couplings, O and O3 are the static and dynamic centers of the coupling shell respectively. ΔL is the installation distance of the left and right end couplings, the angle deviation is γ, ΔY is the misalignment amount when parallel misalignment occurs, and Δα is the misalignment amount when angular misalignment occurs. The comprehensive misalignment amount of the rotor system can be expressed as the comprehensive of the parallel misalignment amount and the angular misalignment amount:

[0090]

[0091] When the rotor system is running, point O3 makes a circular motion with the comprehensive misalignment amount ΔE as the diameter. Assuming that the coordinates of point O3 are O3(X, Y), the angular velocity of the rotating shaft is ω, and α is the independent variable, then the expressions of the dynamic centers x and y of the coupling shell are:

[0092]

[0093] Taking the derivative with respect to time t, the motion velocity is:

[0094]

[0095] The velocity of point O3 is:

[0096]

[0097] The running angular velocity of the coupling shell is equal to that of the rotating shaft, so:

[0098]

[0099] Therefore, the angular velocity of point O3 around the center of the circle is:

[0100]

[0101] The angular velocity of point O3 is twice the rotational angular velocity of the shaft, so when the rotor system is running, the eccentricity fault will generate centrifugal force due to the mass of the coupling shell, exciting the rotor to vibrate.

[0102] Differentiating again with respect to time t gives:

[0103]

[0104] The acceleration of point O3 is:

[0105]

[0106] Let the mass of the coupling be m c Therefore, under the comprehensive eccentricity fault, the eccentric excitation force P of the rotor can be represented as:

[0107] P = -2m c ΔEω 2 (23)

[0108] The components of P in the x and y directions are:

[0109]

[0110] 1.1.4) Based on the above model, the force analysis of the rotor system is carried out to obtain the differential equation of the eccentricity fault dynamics;

[0111]

[0112] where, is the force generated in the x and y directions when the system is eccentric, are the rigid support forces of the rolling bearings, are the unbalanced forces exerted by the unbalanced mass on the rotor system.

[0113] 1.2) Based on the lumped mass degree of freedom method, a lumped mass model of the rotor system is established, and based on the fault model of the rotor system, a differential equation of the fault dynamics of the rotor system is established.

[0114] In this embodiment, the lumped mass degree of freedom method is used, assuming that the mass of the rotor is uniformly distributed around the rotational axis and is represented as a series of mass elements, each mass element having a specific mass and position. Two discs are fixed on a massless, elastic shaft, and the mass of the system is equivalent to the mass concentrated at the center of each disc and the center of each bearing support. From the left bearing to the right bearing, there are 4 nodes, labeled as 1-4 nodes, each node having only 2 degrees of freedom in the x and y directions, and the system has a total of 8 degrees of freedom. The unbalance and eccentricity models of the shaft, the rolling bearing model, and the force analysis of the rotor system are combined to obtain the differential equation of the fault dynamics.

[0115] In the above step 1), the rotor system fault dynamics differential equation is solved, specifically, the Runge-Kutta method is used to solve the rotor system dynamics differential equation with misalignment fault, and the time displacement signal capable of accurately showing the motion characteristic change of the rotor system in the set parameter range is obtained as the fault displacement simulation signal of the rotor system in the fault state.

[0116] In this embodiment, the 4 / 5 order Runge-Kutta algorithm is used to solve the rotor fault dynamics differential equation set, and the X and Y direction vibration displacement signals of the concentrated mass are obtained, and the vibration signals are obtained by integral operation. Specifically, the steps of solving the rotor system fault dynamics differential equation by using the Runge-Kutta method are as follows:

[0117] (1) The quantitative and relationship involved in the system are described by the output condition;

[0118] (2) Variable substitution in mathematics is used;

[0119] (3) Based on the above two steps, a function file capable of calculating the derivative is written;

[0120] (4) The solver slover-ode45 function is used to solve the column vector x.

[0121] In the above step 2), the conditional deep convolution generative adversarial network is constructed. The improvement based on the deep convolution generative adversarial network (DCGAN) is to introduce the conditional variable into the input of the generator and the discriminator, so as to supervise the model to generate samples of a specific category. This improvement is essentially a conditional constraint DCGAN, namely Conditional DCGAN (CDCGAN). By introducing the conditional variable y, the model can be guided to generate samples of a specific category or with specific attributes.

[0122] Wherein, the input of the generator is the rotor system dynamics simulation signal and the fault label. In the generator, transpose convolution is needed to expand the image information of the input data. The dimension of the transpose convolution decreases in turn from 512 to 256, 256 to 128, 128 to 64, and finally to 3. The moving step of the convolution kernel in the first three layers is 2 along the input data, and the last layer is 1. All the padding is set to 1, as shown in Figure 6 The generator structure of the CDCGAN-OGP is shown.

[0123] In the discriminator structure, the input data is processed using convolutional layers, and the size of the convolution kernel gradually increases, which is 64, 128, 256, and 512, respectively. A dropout layer is added after each convolution operation to prevent overfitting. Subsequently, the feature map is flattened through a fully connected layer, and the prediction of whether the sample is from the actual data or the generated data is output, as well as the class label of the sample. The step size in the convolution operation is set to 1, and the padding is set to 1 to ensure that the feature map size remains unchanged. As shown in FIG. 8, the discriminator of the CDCGAN-OGP is shown. Figure 7

[0124] The loss function is a joint spatial sampling of the sample, which reduces the difference between the real and generated samples. The sampling result X in is shown in equation (26).

[0125] X in = eps×R real +(1-eps)×R fake (26)

[0126] where eps represents random sampling; R real represents real sample sampling; and R fake represents false sample sampling.

[0127] According to the spatial joint sampling, the discriminator output D(X in ) is obtained, and the gradient calculation formula is obtained by jointly deriving the two:

[0128]

[0129] where i represents the number of samples contained in the batch size.

[0130] Through the square difference processing of the gradient, the sample features are adapted, and the loss function calculation formula is obtained:

[0131]

[0132] where N grad represents the number of rows of the matrix gradient, k represents the gradient coefficient, k = 1, and λ represents the gradient penalty parameter.

[0133] loss eps Adjusting the gradient tends to k, and the loss function difference of the real and false samples of the discriminator is weighted and combined to obtain the final discriminator loss function calculation formula:

[0134]

[0135] Affected by the functional structure of the generator, the loss generated by the model is concentrated in the convolution processing part. According to the gradient descent method, the loss function calculation formula is:

[0136] ​

[0137] For the entire sample space, the regional mean is calculated as the loss of the single iteration process sample training for the false sample set region, and the generator loss function calculation formula is obtained:

[0138]

[0139] In the formula, β represents the generator gradient penalty parameter.

[0140] Specifically, in this embodiment, the gradient penalty condition deep convolutional generative adversarial network constructed includes a generator, a discriminator and a loss function, etc. The goal of the discriminator is to maximize the discrimination ability of the input data, that is, to correctly distinguish the real samples and the generated samples as much as possible; the goal of the generator is to generate samples that combine mechanism characteristics and actual mechanical characteristics and are infinitely close to the measured data, so the simulation vibration signal and the label are taken as the input of the generator, the generated samples and the real data are taken as the input of the discriminator to train and test the model, and finally the fault samples combining mechanism characteristics and mechanical characteristics are obtained.

[0141] In this embodiment, the obtained initial generated signal is input into the discriminator together with the real signal and the fault label information, and the generated signal combining mechanism characteristics and actual mechanical characteristics is obtained, including the following steps:

[0142] 2.1) The real signal is converted into a two-dimensional time-frequency signal by the short-time fractional Fourier transform (RLSTFRFT) combined with deconvolution;

[0143] The short-time fractional Fourier transform (RLSTFRFT) combined with deconvolution is a short-time fractional Fourier transform combined with the Richardson-Lucy (R-L) algorithm, which operates on images in the spatial domain and does not involve Fourier transform in the frequency domain or convolution in the frequency domain. The deconvolution iteration formula is:

[0144]

[0145] S represents the number of iterations, the initial value RLFRFT0=FRFT, and finally:

[0146] RLFRFT=FRFT[k(t)]exp(-i(cotγ / 2)o 2 ) (33)

[0147] The short-time fractional Fourier transform (RLSTFRFT) based on the deconvolution algorithm performs R-L algorithm operation after calculating the fractional Fourier transform in each short-time window, thereby completing the final operation.

[0148] The gradient penalty term (ogp) can be in the form of the square of the norm or absolute value of the model parameters, defined as:

[0149]

[0150] The training objective function of the discriminator D becomes:

[0151]

[0152] where, is the real data x r and the linear difference of the generated data x f , expressed as:

[0153]

[0154] The goal of the discriminator D is to minimize the error L(G, D) so that the real distribution p r and the generator distribution p g EM distance is maximized, and the GP penalty term is close to 1.

[0155] The training objective of the generator G is:

[0156]

[0157] In order to make the generator distribution p g and the real distribution p r EM distance is minimized, considering and the generator is irrelevant, so the generator training objective is:

[0158]

[0159] The orthogonal gradient penalty term is used to correct the quality of the loss function and clarify the directionality of the gradient. By introducing the orthogonal gradient penalty term, the training of the generative adversarial network can be more effectively constrained, and the stability of the model and the quality of the generated samples can be improved. The penalty term is defined as:

[0160]

[0161] After adding the penalty term, the orthogonal gradient penalty is defined as:

[0162]

[0163] 2.2) The generated signal is converted into a two-dimensional time-frequency signal together with the real signal and the fault label information, and is input into the discriminator. The norm of the parameter of the gradient penalty condition deep convolution generative adversarial network is punished (ogp) to make the update of the parameter more stable, and the objective function of the gradient penalty condition deep convolution generative adversarial network is determined, and then the generated signal combined with the mechanism characteristics and the actual mechanical characteristics is obtained.

[0164] In the above step 2.2), the norm of the parameter of the gradient penalty condition deep convolution generative adversarial network is punished, specifically: a gradient penalty term is introduced, which is the square or absolute value of the norm of the network parameter, and the norm of the network parameter is punished to make the update of the network parameter more stable.

[0165] In the above step 2.2), the objective function of the gradient penalty condition deep convolution generative adversarial network is:

[0166]

[0167] In the formula, V(.) represents the value function, represents the probability that the discriminator D(·) obtains the real data;

[0168] represents the probability that the discriminator D(·) judges the sample of the generator G(z) to be true.

[0169] In the above step 3), a cross-condition domain adaptive fault diagnosis model is constructed based on the feature transfer learning theory, including: a feature extraction layer, an adaptive layer and a fully connected layer, etc.

[0170] 3.1) ResNet-50 network is used as the feature extraction layer for feature extraction, and a local maximum mean difference (LMMD) adaptive layer is introduced to align the sub-field distribution;

[0171] The local maximum mean difference (Local Maximum Mean Discrepancy, LMMD) is a statistical quantity used to measure the difference between two distributions (usually source domain and target domain). It is an extension based on the maximum mean difference (Maximum Mean Discrepancy, MMD) for local structure alignment in domain adaptation (Domain Adaptation) problems. Domain adaptation refers to using a model trained on a source domain to improve performance on a target domain when the source and target domain distributions are inconsistent. The formula is as follows:

[0172]

[0173] 3.2) Adding a fully connected layer after the ResNet-50 global average pooling layer, setting it as an adaptive layer, using local maximum mean difference measurement for sub-field domain adaptation, obtaining a cross-condition domain adaptation fault diagnosis model;

[0174] Wherein, the ResNet-50 network comprises 49 convolutional layers and 1 fully connected layer.

[0175] In this embodiment, in the ResNet-50 network, the first part is used for convolution of the input object, so as to preliminarily extract the features of the input. The second, third, fourth and fifth parts contain residual blocks. By introducing the residual block, the ResNet50 effectively solves the problem of gradient disappearance, so that the network can learn the feature representation more deeply, thereby achieving good performance in image recognition and other tasks. The specific model structure of the adopted Resnet-50 is as shown in Figure 8 Conv1, Conv2D is a convolutional layer, the batch normalization layer is represented by BN, the maximum pooling layer is represented by MAXP, layer represents a residual block, BLK1 represents a bottleneck with a down-sampling step, and BLK2 represents a bottleneck without a down-sampling step. A local maximum mean difference (LMMD) adaptive layer is introduced to align the sub-field distribution. After the ResNet-50 global average pooling layer, a fully connected layer is added, which is set as an adaptive layer, and LMMD measurement is used for sub-field domain adaptation. The cross-condition fault diagnosis model is as shown in Figure 9 .

[0176] The fault diagnosis model based on digital-analog linkage is combined with the data enhancement model and the cross-condition domain adaptation transfer fault diagnosis model based on the mechanism model of the above-mentioned CDCGAN-OGP. First, a large amount of normal and fault simulation data is obtained by solving the dynamics model of the rotor system. The simulation data and fault labels are used as the input of the data enhancement model generator of the CDCGAN-OGP, and the real data is used as the input of the discriminator. Through the adversarial learning of the generator and the discriminator, the generated signal is obtained, which is fused with the fault mechanism and infinitely close to the real signal, so as to achieve the purpose of data enhancement. Then, the generated sample is used as the source domain data, and the measured unlabeled sample is used as the target domain to train the cross-condition transfer learning fault diagnosis model, so as to solve the rotor system fault diagnosis under actual cross-condition. The mechanism model and the transfer learning fault diagnosis model are as shown in Figure 10 .

[0177] In order to further illustrate the present application, the following embodiments are described in detail.

[0178] In this embodiment, the rolling bearings used in the misalignment and imbalance experiments are all SKF-620. The rotor system of a gas turbine rotor test bench was selected for verification. It is important to clarify that this research does not aim for precise consistency between simulation and experimental data. The goal is to obtain a signal through dynamic simulation that is similar to the experimental signal but more representative of the rotor's dynamic mechanisms. To achieve this goal, a certain degree of error in the magnitude of the acceleration amplitude is acceptable, as the focus should be on the shape and characteristics of the signal rather than a precise match of its absolute value.

[0179] The results are obtained by solving the fault dynamics differential equations using the Runge-Kutta method, as shown below. Figure 11a , Figure 11b Comparison of time-domain and frequency-domain waveforms of the rotor system under healthy conditions. Figure 12a to Figure 13b The paper presents time-domain and frequency-domain waveform comparisons of the actual and simulated signals under fault conditions of 0.2 mm misalignment and 2.8 g imbalance at a speed of 1670 r / min.

[0180] This embodiment uses five fault categories as shown in Table 1. Simulation signals are obtained by setting parameters according to the above fault categories. The simulation signals are also sampled using an overlap sampling method. Three operating conditions are set as shown in Table 2, with 200 samples for each fault category under each condition, for a total of 1000 samples.

[0181] Table 1 Fault Categories

[0182]

[0183] Table 2 Operating Parameters

[0184]

[0185] The generated samples are consistent with the simulation samples, with 200 samples for each fault category under each operating condition, for a total of 1000 samples. The generator gradient penalty parameter β = 1 × 10 -6 The discriminator gradient penalty parameter λ = 1. The Adam optimization algorithm is used to train the generator and discriminator models. The training process is as follows: Figure 14 As shown.

[0186] Simulation data and generated data were used as source domain datasets, respectively. The actual signal to be diagnosed was divided into samples using the same overlapping sampling method and used as the target domain. Using the generated signal, which includes the mechanism model and actual mechanical characteristics, as the source domain, and the actual fault signal as the target domain, a cross-condition fault diagnosis model was used for fault diagnosis. The experimental scheme is shown in Table 3. The number of model training iterations remained consistent. Stochastic gradient descent was used with a momentum of 0.8 and a regularization factor of 5e.-3 The learning rate is set to 0.0001, the iteration number is 200, and the batch size is 32.

[0187] Table 3 Experimental scheme

[0188] Test task Source domain Target domain Task 1a Working condition A generated data Working condition B / C Task 1b Working condition B generated data Working condition A / C Task 1c Working condition C generated data Working condition A / B

[0189] In step 3) of the embodiment, three experimental schemes are set, for example, the A working condition simulation signal is the source domain, and the B and C working conditions are actual fault signals as the target domain for experiment; the A working condition generated signal is the source domain, and the B and C working conditions are actual fault signals as the target domain for experiment. The accuracy of the experimental results is shown in Table 4. It can be seen that the accuracy of each transfer case in the task of using the CDCGAN-OGP data enhancement model to generate data as the source domain is more than 85%, and the average accuracy is more than 91%, which verifies the effectiveness of the CDCGAN-OGP data enhancement model combined with the cross-working condition fault diagnosis model.

[0190] Table 4 Transfer fault diagnosis results

[0191] Source domain \ target domain A B C Average 2A 0.9908 0.8802 0.8798 0.9169 2B 0.9171 0.9936 0.8522 0.9176 2C 0.9236 0.9946 0.8800 0.9327

[0192] The simulation data and generated data under working condition B are respectively taken as the source domain, and the actual measured data under working condition A is taken as the target domain for fault diagnosis. As shown in Table 5, it can be seen from the classification confusion matrix of each fault sample that the effect of using the generated data combined with the mechanism characteristics and the actual mechanical characteristics for transfer diagnosis classification is much better than that of directly using the simulation signal for transfer fault diagnosis. This shows that by reasonably fusing the characteristics of the mechanism and the actual working condition, the generalization ability and diagnostic accuracy of the model can be effectively improved. Figure 15

[0193] In an embodiment of the present application, a digital-analog interaction rotor system feature transfer fault diagnosis system is provided, which comprises:

[0194] The first processing module introduces the misalignment fault and the imbalance fault into the constructed rotor system dynamics model based on the rotor system fault dynamics model, to establish a rotor system fault dynamics differential equation, and solve the rotor system fault dynamics differential equation to obtain a fault displacement simulation signal;

[0195] The second processing module takes the fault displacement simulation signal as the input of the generator in the constructed gradient penalty condition deep convolutional generative adversarial network, and inputs the obtained initial generated signal together with the real signal and fault label information into the discriminator to obtain a generated signal combined with the mechanism characteristics and the actual mechanical characteristics;

[0196] ​The fault diagnosis module is configured to build a cross-condition-domain adaptive fault diagnosis model based on feature transfer learning theory, to combine mechanism characteristics and actual mechanical characteristics to generate signal data as a source domain, to combine other condition data to be tested as a target domain, to train the cross-condition-domain adaptive fault diagnosis model, and to perform cross-condition fault diagnosis on the rotor system by using the trained cross-condition-domain adaptive fault diagnosis model.

[0197] In the above embodiment, based on the rotor system fault dynamics model, the misalignment fault and the imbalance fault are introduced into the rotor system dynamics model, including:

[0198] The rotor system fault dynamics model includes a fault model for describing an imbalance model, a fault model for describing a rolling bearing model, a fault model for describing a misalignment fault model, and a fault model for describing other faults.

[0199] The rotor system lumped mass model is established based on the lumped mass degree of freedom method, and the rotor system fault dynamics differential equation is established based on the rotor system fault model.

[0200] In the above embodiment, the rotor system fault dynamics differential equation is solved, including:

[0201] The Runge-Kutta method is used to solve the rotor system dynamics differential equation with the misalignment fault, and a time displacement signal that can accurately show the motion characteristic changes of the rotor system within a set parameter range is obtained as a fault displacement simulation signal under the fault state of the rotor system.

[0202] In the above embodiment, the obtained initial generated signal is input into the discriminator together with the real signal and the fault label information to obtain a generated signal that combines mechanism characteristics and actual mechanical characteristics.

[0203] The real signal is converted into a two-dimensional time-frequency signal by combining the deconvolution short-time fractional Fourier transform;

[0204] The generated signal and the two-dimensional time-frequency signal converted from the real signal are input into the discriminator together with the fault label information, the norm of the parameters of the gradient penalty condition deep convolution generative adversarial network is penalized to make the update of the parameters more stable, the objective function of the gradient penalty condition deep convolution generative adversarial network is determined, and then the generated signal that combines mechanism characteristics and actual mechanical characteristics is obtained.

[0205] In the above embodiment, the norm of the parameters of the gradient penalty condition deep convolution generative adversarial network is penalized, including:

[0206] The orthogonal gradient penalty term is introduced, and the gradient penalty term is the square or absolute value of the norm of the network parameters, and the norm of the network parameters is penalized to make the update of the network parameters more stable.

[0207] In this embodiment, the objective function of the gradient penalty conditional deep convolutional generative adversarial network is:

[0208]

[0209] In the formula, V(.) represents a value function, represents the probability that the discriminator D(.) obtains real data;

[0210] represents the probability that the discriminator D(.) judges the sample of the generator G(z) to be real.

[0211] In the above embodiment, a cross-condition domain adaptive fault diagnosis model is established based on feature transfer learning theory, including a feature extraction layer, an adaptive layer and a fully connected layer.

[0212] A ResNet-50 network is used as the feature extraction layer for feature extraction, and a local maximum mean difference adaptive layer is introduced to align the sub-field distribution.

[0213] A fully connected layer is added after the ResNet-50 global average pooling layer, which is set as the adaptive layer, and the local maximum mean difference is used for sub-field domain adaptation to obtain the cross-condition domain adaptive fault diagnosis model.

[0214] The ResNet-50 network includes 49 convolutional layers and 1 fully connected layer.

[0215] The system provided in this embodiment is used to execute the above method embodiments, and the specific process and detailed content are referred to the above embodiments, which will not be repeated here.

[0216] In an embodiment of the present application, a computing device, which can be a terminal, can include a processor, a communications interface, a memory, a display screen and an input device. The processor, the communications interface and the memory can communicate with each other through a communication bus. The processor is configured to provide computing and control capabilities. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The computer program is executed by the processor to implement the method of any of the above embodiments. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communications interface is configured to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, a management network, NFC (Near Field Communication) or other technologies. The display screen can be a liquid crystal display screen or an electronic ink display screen. The input device can be a touch layer overlaid on the display screen, or a key, a trackball or a touchpad arranged on the housing of the computing device, or an external keyboard, a touchpad or a mouse, etc. The processor can invoke the logical instructions in the memory.

[0217] In addition, the logical instructions in the memory described above can be implemented in the form of a software functional unit and sold or used as a standalone product, which can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application, in essence, or the parts that contribute to the prior art, or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0218] In an embodiment of the present application, a computer program product is provided, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions that, when executed by a computer, enable the computer to perform the methods provided by the various embodiments described above.

[0219] In an embodiment of the present application, a non-transitory computer-readable storage medium is provided, which stores server instructions. The computer instructions enable a computer to perform the methods provided by the various embodiments described above.

[0220] The computer readable storage medium provided by the above embodiment has similar implementation principles and technical effects to the method embodiment, and thus will not be described here.

[0221] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices generate a device that implements the functions specified in the flowcharts and / or block diagrams. Figure One one or more flows and / or blocks Figure One an apparatus that implements the functions specified in one or more blocks or flows.

[0222] These computer program instructions can also be stored in a computer readable memory that can direct the computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce a product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure One one or more flows and / or blocks Figure One an apparatus that implements the functions specified in one or more blocks or flows.

[0223] These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to produce a computer implemented process, so that the instructions executed on the computer or other programmable device provide a process for implementing the functions specified in the flowcharts and / or block diagrams. Figure One one or more flows and / or blocks Figure One an apparatus that implements the functions specified in one or more blocks or flows.

[0224] It should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application. In addition, it should be noted that, although the present application is described in specific embodiments, each embodiment does not contain only one independent technical solution. The description of the specification is for the sake of clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that those skilled in the art can understand.

Claims

1. A method for diagnosing feature migration faults in a digital-analog linked rotor system, characterized in that, include: Based on the rotor system fault dynamics model, misalignment faults and imbalance faults are introduced into the constructed rotor system dynamics model to establish the rotor system fault dynamics differential equations, and solve the rotor system fault dynamics differential equations to obtain the fault displacement simulation signal. The fault displacement simulation signal is used as the input to the generator in the constructed gradient-penalized conditional deep convolutional generative adversarial network. The initial generated signal, the real signal, and the fault label information are input into the discriminator to obtain a generated signal that combines the mechanism characteristics and the actual mechanical characteristics. A cross-operating-condition adaptive fault diagnosis model is constructed based on feature transfer learning theory. The data of the generated signal, which combines the mechanism characteristics and the actual mechanical characteristics, is used as the source domain, and the data of other operating conditions to be tested is used as the target domain. The cross-operating-condition adaptive fault diagnosis model is trained, and the trained cross-operating-condition adaptive fault diagnosis model is used to diagnose cross-operating-condition faults in the rotor system. A cross-operating-condition adaptive fault diagnosis model based on feature transfer learning theory is constructed, including: a feature extraction layer, an adaptive layer, and a fully connected layer; We use ResNet-50 as the feature extraction layer to extract features, and introduce a local maximum mean difference adaptive layer to align the sub-neighborhood distribution. A fully connected layer is added after the ResNet-50 global average pooling layer and set as an adaptive layer. The local maximum mean difference metric is used to adapt the sub-domain, resulting in a cross-condition adaptive fault diagnosis model. The ResNet-50 network consists of 49 convolutional layers and 1 fully connected layer.

2. The method for diagnosing feature migration faults in a digital-analog linked rotor system as described in claim 1, characterized in that, Based on the rotor system fault dynamics model, misalignment faults and imbalance faults are introduced into the constructed rotor system dynamics model, including: The rotor system fault dynamics model includes a model to describe shaft imbalance, a model to describe rolling bearings, a model to describe coupling misalignment faults, and a fault model to describe other faults. A lumped mass model of the rotor system is established based on the lumped mass degree of freedom method, and the differential equation of rotor system fault dynamics is established based on the rotor system fault model.

3. The method for diagnosing feature migration faults in a rotor system with digital-analog linkage as described in claim 1, characterized in that, Solving the differential equations of fault dynamics for the rotor system includes: The Runge-Kutta method is used to solve the dynamic differential equations of a rotor system with misalignment faults, and the time displacement signal that can accurately show the changes in the motion characteristics of the rotor system within the set parameter range is obtained as the fault displacement simulation signal of the rotor system under fault conditions.

4. The method for diagnosing feature migration faults in a rotor system with digital-analog linkage as described in claim 1, characterized in that, The initial generated signal, the real signal, and the fault tag information are input together into the discriminator to obtain a generated signal that combines mechanistic characteristics and actual mechanical characteristics, including: The real signal is converted into a two-dimensional time-frequency signal by combining a short-time fractional Fourier transform with deconvolution; The generated signal and the two-dimensional time-frequency signal converted from the real signal, along with the fault label information, are input into the discriminator. By penalizing the norm of the parameters of the gradient-penalized deep convolutional generative adversarial network, the parameter updates are made more stable. The objective function of the gradient-penalized deep convolutional generative adversarial network is determined, thereby obtaining a generated signal that combines mechanistic characteristics and actual mechanical characteristics.

5. The method for diagnosing feature migration faults in a rotor system with digital-analog linkage as described in claim 4, characterized in that, Penalizing the norm of the parameters of the gradient-penalized deep convolutional generative adversarial network includes: An orthogonal gradient penalty term is introduced, which is the square or absolute value of the norm of the network parameters. By penalizing the norm of the network parameters, the updates of the network parameters are made more stable.

6. The method for diagnosing feature migration faults in a rotor system with digital-analog linkage as described in claim 4, characterized in that, The objective function of a gradient-penalized conditional deep convolutional generative adversarial network is: ； In the formula, Represents the value function. Discriminator The probability of obtaining real data; Discriminator Decision generator The probability that the sample is true.

7. A feature migration fault diagnosis system for a rotor system with digital-analog linkage, characterized in that, include: The first processing module, based on the rotor system fault dynamics model, introduces misalignment faults and imbalance faults into the constructed rotor system dynamics model to establish the rotor system fault dynamics differential equations, and solves the rotor system fault dynamics differential equations to obtain the fault displacement simulation signal. The second processing module takes the fault displacement simulation signal as the input to the generator in the constructed gradient-penalized conditional deep convolutional generative adversarial network, and inputs the obtained initial generated signal, real signal and fault label information into the discriminator to obtain a generated signal that combines mechanism characteristics and actual mechanical characteristics. The fault diagnosis module constructs a cross-operating-condition adaptive fault diagnosis model based on feature transfer learning theory. It uses the generated signal data that combines the mechanism characteristics and actual mechanical characteristics as the source domain and the data of other operating conditions to be tested as the target domain to train the cross-operating-condition adaptive fault diagnosis model. The trained cross-operating-condition adaptive fault diagnosis model is used to perform cross-operating-condition fault diagnosis on the rotor system. A cross-operating-condition adaptive fault diagnosis model based on feature transfer learning theory is constructed, including: a feature extraction layer, an adaptive layer, and a fully connected layer; We use ResNet-50 as the feature extraction layer to extract features, and introduce a local maximum mean difference adaptive layer to align the sub-neighborhood distribution. A fully connected layer is added after the ResNet-50 global average pooling layer and set as an adaptive layer. The local maximum mean difference metric is used to adapt the sub-domain, resulting in a cross-condition adaptive fault diagnosis model. The ResNet-50 network consists of 49 convolutional layers and 1 fully connected layer.

8. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described in claims 1 to 6.

9. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods described in claims 1 to 6.

Images