Robust fault diagnosis method for rotating machinery under unknown working condition

Abstract

The invention provides a robust fault diagnosis method for a rotating machine under an unknown working condition, and the method comprises the steps: constructing an operation and maintenance database through the monitoring data of the rotating machines with the same or similar models; constructing a data set required by model training by using the data in the operation and maintenance database; training a conditional diffusion model generation model and a known working condition fault diagnosis model by using the constructed data set; then, using the conditional diffusion model to generate various working state data when the rotating machine works under unknown working conditions; using the known working condition fault diagnosis model and data in an operation and maintenance database to generate a confrontation sample of the known working condition; thirdly, training a fault diagnosis model by using the generated unknown working condition data, the generated adversarial sample data and data in an operation and maintenance database; and finally, performing fault diagnosis on the working state of the rotating machine under the unknown working condition by using the trained fault diagnosis model.

Description

Technical Field

[0001] The invention relates to the technical field of intelligent diagnosis of rotating machinery, and in particular to a robust fault diagnosis method of rotating machinery under unknown working conditions. Background Art

[0002] Rotating machinery, including gearboxes, bearings, etc., is a key component in industrial production, and its health status has a profound impact on the continuity of the production line and product quality. However, fluctuations in environmental factors and operating conditions often lead to complex and changeable working conditions of rotating machinery, which in turn affects its performance and service life. Therefore, real-time monitoring of the working status of rotating machinery and accurate fault diagnosis are crucial to promoting preventive maintenance and improving equipment reliability.

[0003] In recent years, learning-based methods have been widely used in the field of fault diagnosis. For example, document CN113312719A, "A method for fault diagnosis of rotating machinery based on cross entropy of class imbalance weights", solves the problem of class imbalance in samples to a certain extent by first constructing and training a sparse autoencoder model and then using the model to generate samples. However, it can only be used when there are data samples under the working condition. It cannot be used when the working condition requiring fault diagnosis is different from other working conditions. For another example, document CN 117909800A, "A method for fault diagnosis of rotating machinery equipment based on an interpretable implicit diffusion model", first preprocesses the fault vibration signal data, then trains an interpretable implicit diffusion model, and then, based on the existing fault vibration signal and fault category, augments and expands the vibration data set under the original fault vibration signal sample through the implicit diffusion model, and finally trains a fault diagnosis model based on a neural network. This method generates samples to generate fault categories with less data in known working conditions, thereby improving the performance of the model in fault diagnosis. However, this method does not pay attention to the characteristic relationship between the data of the working conditions. In the absence of data on a certain working condition, it is impossible to perform fault diagnosis tasks on rotating machinery under that working condition.

[0004] It can be seen that the existing rotating machinery fault detection methods usually need to rely on a large amount of data under known working conditions. Only through in-depth mining and analysis of the working condition data can the accurate identification of rotating machinery faults under known working conditions be achieved.

[0005] However, in practical applications, the training data of the fault diagnosis model is usually limited, and only the fault data of the rotating machinery under common working conditions can be collected. At this time, when the fault diagnosis model trained with the fault data under common working conditions is used to identify unknown working condition data, the recognition accuracy will decrease; when the working conditions vary greatly, the model may even be unusable. In addition, when the model is trained too few times or the data used for training is too little, the model may be underfitting; when the model is trained too many times or the data used for training is too much, the model may be overfitting. Summary of the invention

[0006] Therefore, in view of the defects and deficiencies in the prior art, the present invention provides a robust fault diagnosis method for rotating machinery under unknown working conditions, including: extracting features from vibration signal data in the operation and maintenance database, encoding working condition information features and working state features, and obtaining data samples; using the data samples to train the conditional automatic diffusion model to obtain the conditional automatic diffusion model; for any unknown working condition sample, using the conditional automatic diffusion model to generate vibration data samples of various working states of the rotating machinery under unknown working conditions; using the data samples to train the fault diagnosis model to obtain a known working condition fault diagnosis model; using the FGSM method to generate adversarial data samples of known working conditions based on the known working condition fault diagnosis model; using the vibration signal samples in the operation and maintenance database, the generated adversarial data samples and the generated vibration signal samples to train the fault diagnosis model to obtain the unknown working condition fault diagnosis model of the rotating machinery. The method is conducive to accurately and effectively diagnosing the working state of the unknown working condition of the rotating machinery.

[0007] The specific technical solutions are as follows:

[0008] A robust fault diagnosis method for rotating machinery under unknown working conditions: first, an operation and maintenance database is constructed using monitoring data of rotating machinery of the same or similar models; the data in the operation and maintenance database are used to construct a data set required for model training; and the constructed data set is used to train a conditional diffusion model generation model and a known working condition fault diagnosis model; then, the conditional diffusion model is used to generate various working state data of the rotating machinery when working under unknown working conditions; and the known working condition fault diagnosis model and the data in the operation and maintenance database are used to generate adversarial samples of known working conditions; then, the generated unknown working condition data, the generated adversarial sample data and the data in the operation and maintenance database are used to train the fault diagnosis model; finally, the trained fault diagnosis model is used to perform fault diagnosis on the working state of the rotating machinery under unknown working conditions.

[0009] The rotating machinery includes gear boxes, bearings, etc.; the detection data includes the operating condition information of the rotating machinery, the historical vibration signals of the rotating machinery and the corresponding working status.

[0010] Furthermore, the specific method of training the fault diagnosis model through the data in the operation and maintenance database is as follows:

[0011] Take out some data from the operation and maintenance database, including at least: vibration signal data M collected by sensors when the rotating machinery is in S working states including normal and S-1 faults under K different working conditions K,S =[m 0 ,m 1 ,...,m p-1 ], where p represents the number of sampling points of the vibration signal; the working condition information corresponding to the signal M is encoded using label coding to obtain the working condition information feature code C;

[0012] Then, the working state of the signal m is encoded using one-hot encoding to obtain the working state feature encoding Y = [q 0 ,q 1 ,...,q S-1 ];

[0013] Assume the working state is s, then

[0014] Then use the fast Fourier transform k=0,1,2,...,p-1 for signal M K,S Extract features to obtain the signal feature vector X; combine the i-th working condition information feature code, working state feature code and signal feature vector to construct a training data sample, denoted as D = [X i ,Y i ,C i ], where X i ,Y i ,C i They respectively represent the frequency domain signal feature vector of the rotating machinery of the i-th training data, the working state feature code of the rotating machinery, and the working condition information feature code of the rotating machinery.

[0015] Furthermore, when encoding information with numerical values ​​such as the operating speed of the rotating machinery, the numerical value is directly placed in the code C, and when encoding information without numerical values ​​such as whether there is a load, 0 or 1 is directly used for encoding.

[0016] Furthermore, the conditional diffusion model includes: an encoder module Encoder(), a decoder module Decoder(), a working condition information feature encoding module Clip() and a diffusion model module Diffuison();

[0017] The encoder module extracts features from the input vector and converts the input vector into a feature vector in a latent representation space; the calculation formula of the encoder module is: i,0=Encoder(X i ), where X i The frequency domain signal feature vector of the rotating machinery representing the i-th data, Z i,0 The feature vector of the latent representation space after the transformation of the i-th data without noise;

[0018] The decoder module is used to convert a feature vector in the latent representation space into a frequency domain feature vector in the frequency domain space, and the calculation formula is: in, The feature vector representing the latent representation space that needs to be decoded, The feature vector representing the decoded frequency domain space;

[0019] The working condition information feature encoding module is used to convert the input working condition information vector into a high-dimensional vector and input it into the diffusion model module; the calculation formula of the working condition information feature encoding module is: i,high = Clip(Y i ,C i ), where Y i Indicates the working status characteristic code of the i-th data, C i Indicates the characteristic code of the working condition information of the i-th data, Y i,high Represents the high-dimensional spatial feature vector of the working condition information of the i-th data;

[0020] The diffusion model module includes the noise adding process Addnoise() and the noise removing process Denoise(): Among them, the noise addition process simulates the data degradation process by gradually adding noise to the feature data in the potential representation space. The noise addition process calculation formula is: Z i,T =Addnoise(Z i,0 ,Y i,high ,T), T represents the number of times the noise adding process is repeated, Z i,T Represents the feature vector of the potential representation space of the i-th data after adding noise T times; the denoising process gradually denoises the degraded data to restore the data before adding noise. The denoising process calculation formula is: T represents the number of times the denoising process is repeated which is the same as the number of times the denoising process is repeated.

[0021] Furthermore, the conditional diffusion model is trained using the data in the operation and maintenance database. Specifically, when training the conditional diffusion model, the loss function is first used. Calculate the loss of the encoder module and the decoder module; use the stochastic gradient descent method to update the parameters of the conditional diffusion model where w 1,t , w 1,t+1They represent the weight parameters of the conditional diffusion model when the model is updated t times and t+1 times, η 1 is the learning rate of the conditional diffusion model weight update, is the loss function L with respect to w 1,t Gradient; after updating b1 times, it is considered that the encoder module and decoder module have been trained and converged; then the diffusion model module and the working condition information feature encoding module of the model are trained; the loss function is used Calculate the loss, where represents the expected value, ε t represents the noise vector when adding noise for the tth time, Z i,t represents the latent space feature vector obtained after denoising for the tth time, t represents the number of current denoising times, εθ(Z i,t ,t,Y i,high ) represents the noise predicted by the model during the denoising process; the parameters of the model are updated using the stochastic gradient descent method; after the update times b2, it is considered that the encoder module and the decoder module have been trained and converged, and the conditional diffusion model training is completed.

[0022] Furthermore, the conditional diffusion model is used to generate rotating machinery vibration signal samples under different working conditions as follows: the working state feature code Y ε and the working condition information feature code C of the unknown working condition u n Input into the working condition information feature encoding module of the conditional diffusion model, and randomly generate a Gaussian noise Z ε Input into the diffusion model module, use the denoising process of the diffusion model to generate new feature samples of the latent space according to the output of the working condition information feature encoding module; use the decoder module of the model to decode the generated latent space feature samples into feature samples of the frequency domain space; repeat the above process until the number of generated samples reaches the preset value r, completing the generation of unknown working condition data, and record the generated data as After completing the data generation work, use the data D in the operation and maintenance database and generate data Training of fault diagnosis model.

[0023] Furthermore, an artificial neural network model is constructed as a known working condition fault diagnosis model, and the forward propagation formula of the model is: where f(),X,w 2 They respectively represent the mapping function of the known working condition fault diagnosis model, the input vector of the known working condition fault diagnosis model, and the weight parameter of the known working condition fault diagnosis model; Represents the prediction vector of the fault diagnosis model; the data D in the operation and maintenance database is used as the training data of the known working condition fault diagnosis model, and the cross entropy function is used The prediction loss of the model is calculated as the loss function of the fault diagnosis model;

[0024] Use the stochastic gradient descent method with momentum to update the parameters of the fault diagnosis model for known working conditions where v t represents the accumulation of historical gradients in the updating process of the fault diagnosis model under known working conditions, a represents the momentum coefficient, and w 2,t represents the weight parameter when the fault diagnosis model under known working conditions is updated t times, η 2 is the learning rate of the fault diagnosis model under known working conditions, is the loss function L crossentropy About w 2,t Gradient; when the weight parameters of the known working condition fault diagnosis model are updated b times in the model training, it is considered that the known working condition fault diagnosis model has converged, that is, the known working condition fault diagnosis model has completed training.

[0025] Furthermore, the gradient-based adversarial sample generation algorithm FGSM is used to generate adversarial samples. All samples in the database are sequentially generated through the FGSM algorithm to generate adversarial samples. The generated adversarial sample data is recorded as D′.

[0026] For the FGSM algorithm, the calculation formula is: Where X′, X represent the input vector of the known working condition model and the adversarial sample generated corresponding to the input vector, respectively. Represents the gradient of the cross entropy loss function of the known working condition diagnosis model with respect to the model input vector, ∈ represents the size of the perturbation of the generated adversarial sample, and usually requires debugging to find a suitable value. The sign() function returns the sign of the gradient (positive or negative), which is used to determine the direction of the perturbation.

[0027] Furthermore, based on the unknown working condition fault diagnosis model that is the same as the known working condition fault diagnosis model, the data D in the operation and maintenance database, the generated adversarial sample data D′ and the generated data of the generative model are used. The unknown working condition fault diagnosis model is trained as the training data of the fault diagnosis model to obtain the final fault diagnosis model.

[0028] Furthermore, after the fault diagnosis model training is completed, the sensor data collected when the rotating machinery is in an unknown working condition u is input into the fault diagnosis model to obtain the fault diagnosis result of the fault diagnosis model for the rotating machinery under the unknown working condition u.

[0029] And, a non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: during operation, the robust fault diagnosis method for unknown working conditions of a rotating machinery as described above is used.

[0030] Compared with the prior art, the present invention and its preferred solution rely on the known working condition data of the rotating machinery to realize the generation of signals of its various working states under unknown working conditions, providing important data support for the training of the fault diagnosis model of unknown working conditions. And by performing a white-box attack on the fault diagnosis model of known working conditions, the model is trained using the generated adversarial samples to obtain a more robust model. Through this method, operation and maintenance personnel can adopt accurate operation and maintenance strategies based on the fault diagnosis results of the rotating machinery, which promotes the progress of intelligent operation and maintenance of rotating machinery. In the process of evaluating the working state of the rotating machinery under unknown working conditions, the present invention does not need to rely on the operating data of the rotating machinery under unknown working conditions, which enables the present invention to solve to a certain extent the problem that the data of various working states of the rotating machinery under various working conditions are difficult to collect, and the model is difficult to adapt to various working conditions. It has at least the following beneficial effects:

[0031] 1. The conditional diffusion model technology used can deeply analyze and understand the sensor signal characteristics of rotating machinery under different working conditions and working states in the operation and maintenance database. Through comparison and learning, the conditional diffusion model can generate vibration simulation signals of rotating machinery in different working states under unknown working conditions, providing valuable supplementary data for the training of fault diagnosis models. And the present invention uses adversarial samples to train the model, which can make the model more robust. From the perspective of engineering practice, the present invention expands the fault diagnosis model's processing capabilities for unknown working conditions, which greatly improves the accuracy of rotating machinery working state pattern recognition. This extension enables operation and maintenance personnel to more comprehensively assess the health of the equipment, thereby preventing potential problems in advance, and helping to promote the sustainable development of rotating machinery operation and maintenance technology.

[0032] 2. In practical applications, another major advantage of the present invention is its efficient use of data. By making full use of the information in the operation and maintenance database, vibration signals under unknown working conditions can be effectively generated, thereby reducing the fault diagnosis model's dependence on actual measurement data; by adding adversarial samples to the data for training the fault diagnosis model, the overfitting phenomenon of the model is reduced, thereby making the model more robust. For equipment with potential economic loss risks, this means that while reducing the cost of collecting actual fault samples, the accuracy and efficiency of diagnosis can still be guaranteed, reducing maintenance costs and improving equipment availability.

[0033] 3. Compared with the fault diagnosis model based on transfer learning, the present invention has significant advantages in the migration adaptability of the model. It does not require additional acquisition or measurement of data of unknown working conditions, directly improves the versatility and practicality of the model, simplifies the operation process, and improves the flexibility of the technology. This enables operation and maintenance personnel to apply new technologies to actual work more quickly, reduces the complexity of technology implementation, and thus improves the efficiency of overall operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0035] Figure 1 It is a framework diagram of the overall process of the embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the features and advantages of this patent more obvious and easy to understand, the following embodiments are specifically described in detail with the accompanying drawings as follows:

[0037] The main process of the method of the present invention is as follows Figure 1 As shown, the following steps are included:

[0038] 1) Take out some data from the operation and maintenance database. These data are: under K different working conditions, the vibration signal data M collected by the sensor when the rotating machinery is in S working states (normal, fault 1...fault S-1) K,S =[m 0 ,m 1 ,...,m p-1 ], where p represents the number of sampling points of the vibration signal. The working condition information corresponding to the signal M is encoded using label encoding to obtain the working condition information feature code C. For example, when encoding information with numerical values ​​such as the working speed of the rotating machinery, the numerical value can be directly placed in the code C. When encoding information without numerical values ​​such as whether there is a load, 0 or 1 is directly used for encoding. For specific encodings, please refer to the following table;

[0039] Working condition information name Specific working condition information Working condition information feature coding Working speed 3000r / min [3000, c2, c3, .....] 5000r / min [5000, c2, c3, .....] Load conditions High load [c1,1,c3,.....] Low load [c1,0,c3,.....] wait

[0040] Use one-hot encoding to encode the working state of signal m to obtain the working state feature code Y = [q 0 ,q 1 ,...,q S-1 ], for example, if the working state is s, then Using Fast Fourier Transform k=0,1,2,...,p-1 for signal M K,S Feature extraction is performed to obtain the signal feature vector X. The i-th working condition information feature code, working state feature code and signal feature vector are combined to construct a training data sample, which is recorded as D = [X i ,Y i ,C i ], where X i ,Y i ,C i They respectively represent the frequency domain signal feature vector of the rotating machinery of the i-th training data, the working state feature code of the rotating machinery, and the working condition information feature code of the rotating machinery.

[0041] 2) Construction of the conditional diffusion model

[0042] The conditional diffusion model is divided into: encoder module Encoder(), decoder module Decoder(), working condition information feature encoding module Clip() and diffusion model module Diffuison(). The encoder module enables the model to extract features from the input vector and convert the input vector into a feature vector in a potential representation space. The calculation formula of the encoder module is: Z i,0 =Encoder(X i ), where X i The frequency domain signal feature vector of the rotating machinery representing the i-th data, Z i,0 Represents the feature vector of the latent representation space after the i-th data is converted without noise. The decoder module can convert a feature vector in the latent representation space into a frequency domain feature vector in the frequency domain space. The calculation formula of the decoder is: Among them, Z i,0 The feature vector representing the latent representation space that needs to be decoded, Represents the feature vector of the decoded frequency domain space. The working condition information feature encoding module can convert the input working condition information vector into a high-dimensional vector and input it into the diffusion model module of the model, so that the diffusion model module can use the working condition information as a condition. The calculation formula of the working condition information feature encoding module is: i,high = Clip(Y i ,C i ), where Y i Indicates the working status characteristic code of the i-th data, C i Indicates the characteristic code of the working condition information of the i-th data, Y i,high Represents the high-dimensional spatial feature vector of the working condition information of the i-th data. The diffusion model module includes the noise adding process Addnoise() and the noise removing process Denoise(), that is, The noise adding process simulates the data degradation process by gradually adding noise to the feature data in the potential representation space. The noise adding process is calculated as: i,T =Addnoise(Z i,0 ,Y i,high ,T),Z i,0 where Y represents the feature vector of the potential representation space of the i-th data without noise, i,high represents the high-dimensional spatial feature vector of the working condition information of the i-th data, T represents the number of times the noise addition process is repeated, which is a hyperparameter (can be 400), Z i,TRepresents the feature vector of the potential representation space of the i-th data after adding noise T times. The denoising process restores the data before adding noise by gradually denoising the degraded data. The calculation formula of the denoising process is: Where Z i,T represents the feature vector of the potential representation space of the i-th data after adding noise T times, Y i,high represents the high-dimensional spatial feature vector of the working condition information of the i-th data, T represents the number of times the denoising process is repeated, and its value is the same as the number of times the denoising process is repeated. The feature vector of the potential representation space of the i-th data obtained after denoising T times. At the same time, in the process of model denoising and denoising, the output of the working condition information feature encoding module is used as a parameter, and the conditional information is integrated with the diffusion process, so that the model can learn the characteristics between different working conditions, and then generate samples that are more in line with the characteristics of unknown working conditions.

[0043] 3) Use the data in the operation and maintenance database to train the conditional diffusion model.

[0044] When training the conditional diffusion model, first use the loss function Calculate the loss of the encoder module and the decoder module, where X i represents the i-th data sample, Represents the reconstructed sample of the i-th data output by the decoder. Use the stochastic gradient descent method to update the parameters of the conditional diffusion model where w 1,t , w 1,t+1 They represent the weight parameters of the conditional diffusion model when the model is updated t times and t+1 times, η 1 is the learning rate of the conditional diffusion model weight update, is the loss function L with respect to w 1,t After updating a certain number of times b (usually 100), the encoder module and decoder module are considered to have converged. Then the diffusion model module and the working condition information feature encoding module of the model are trained. Use the loss function Calculate the loss, where represents the expected value, ε t represents the noise vector when adding noise for the tth time, Z i,t represents the latent space feature vector obtained after denoising for the tth time, t represents the number of times of current denoising, and Y i,high Represents the high-dimensional spatial feature vector of the working condition information of the i-th data, εθ(Z i,t ,t,Y i,high) represents the noise predicted by the model during the denoising process. The parameters of the model are updated using the stochastic gradient descent method. After a certain number of updates b (usually 100), it is considered that the encoder module and the decoder module have been trained and converged, and the conditional diffusion model training is completed.

[0045] 4) Use the conditional diffusion model to generate rotating machinery vibration signal samples under different working conditions. Encode the working state feature Y ε and the working condition information feature code C of the unknown working condition u n Input into the model's operating condition information feature encoding module and randomly generate a Gaussian noise Z ε Input into the diffusion model module, use the denoising process of the diffusion model to generate new feature samples of the latent space according to the output of the working condition information feature encoding module. Use the decoder module of the model to decode the generated latent space feature samples into feature samples in the frequency domain space. Repeat the above process until the number of generated samples reaches the preset value r, completing the generation of unknown working condition data, and record the generated data as After completing the data generation work, use the data D in the operation and maintenance database and generate data Training of fault diagnosis model.

[0046] 5) An artificial neural network model is constructed as a known working condition fault diagnosis model. The forward propagation formula of the model is: in They represent the input vector of the known working condition fault diagnosis model, the mapping function of the known working condition fault diagnosis model, the input vector of the known working condition fault diagnosis model, and the weight parameter of the known working condition fault diagnosis model. The data D in the operation and maintenance database is used as the training data of the known working condition fault diagnosis model, and the cross entropy function is used. The prediction loss of the model is calculated as the loss function of the fault diagnosis model, where Y represents the prediction vector of the fault diagnosis model and the working state vector corresponding to the sample.

[0047] 6) Train the known working condition fault diagnosis model. Use the stochastic gradient descent method with momentum to update the parameters of the known working condition fault diagnosis model. where v t represents the accumulation of historical gradients in the updating process of the fault diagnosis model under known working conditions, a represents the momentum coefficient, which is usually 1, and w 2,t represents the weight parameter when the fault diagnosis model under known working conditions is updated t times, η 2 is the learning rate of the fault diagnosis model under known working conditions, is the loss function L crossentropy About w 2,tWhen the weight parameters of the known working condition fault diagnosis model are updated a certain number of times b (usually 100) during model training, the known working condition fault diagnosis model is considered to have converged, that is, the known working condition fault diagnosis model has completed training.

[0048] 7) Generate adversarial samples. Use the FGSM algorithm to generate adversarial samples, and its calculation formula is: Where X′, X represent the input vector of the known working condition model and the adversarial sample generated corresponding to the input vector. It represents the gradient of the cross entropy loss function of the known working condition diagnosis model with respect to the model input vector, ∈ represents the size of the disturbance of the generated adversarial sample, which usually needs to be debugged to find a suitable value, and the sign() function returns the sign of the gradient (positive or negative) to determine the direction of the disturbance. All samples in the database are sequentially generated through the FGSM algorithm to generate adversarial samples, and the generated adversarial sample data is recorded as D′.

[0049] 8) Construct an artificial neural network model as an unknown working condition fault diagnosis model, and the forward propagation formula of the model is: in Respectively represent the input vector of the fault diagnosis model, the mapping function of the fault diagnosis model, the input vector of the fault diagnosis model, and the weight parameter of the fault diagnosis model. As training data for the fault diagnosis model, the cross entropy function is used The prediction loss of the model is calculated as the loss function of the fault diagnosis model, where They represent the prediction vector of the fault diagnosis model and the working state vector corresponding to the sample respectively. The parameters of the unknown working condition fault diagnosis model are updated using the stochastic gradient descent method with momentum until the unknown working condition fault diagnosis model training is completed.

[0050] 9) After the fault diagnosis model is trained, the vibration signal data collected when the rotating machinery is in an unknown working condition u is input into the fault diagnosis model to obtain the fault diagnosis result of the fault diagnosis model for the rotating machinery under the unknown working condition u.

[0051] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0052] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0053] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0054] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0055] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

[0056] This patent is not limited to the above-mentioned optimal implementation mode. Anyone can derive other forms of robust fault diagnosis methods under unknown working conditions of rotating machinery under the inspiration of this patent. All equal changes and modifications made according to the scope of the patent application of the present invention should be covered by this patent.

Claims

1. A robust fault diagnosis method for rotating machinery under unknown working conditions, characterized in that: First, an operation and maintenance database is constructed with monitoring data of rotating machinery of the same or similar models; the data in the operation and maintenance database are used to construct the data set required for model training; and the constructed data set is used to train the conditional diffusion model and the known working condition fault diagnosis model; then, the conditional diffusion model is used to generate various working state data of the rotating machinery when working under unknown working conditions; and the known working condition fault diagnosis model and the data in the operation and maintenance database are used to generate adversarial samples of known working conditions; then, the generated unknown working condition data, the generated adversarial sample data and the data in the operation and maintenance database are used to train the fault diagnosis model; finally, the trained fault diagnosis model is used to perform fault diagnosis on the working state of the rotating machinery under unknown working conditions.

2. The method for robust fault diagnosis of a rotating machinery under unknown working conditions according to claim 1, characterized in that: The specific method of training the fault diagnosis model using data in the operation and maintenance database is as follows: Take out some data from the operation and maintenance database, including at least: vibration signal data M collected by sensors when the rotating machinery is in S working states including normal and S-1 faults under K different working conditions K,S =[m0,m1,...,m p-1 ], where p represents the number of sampling points of the vibration signal; the working condition information corresponding to the signal M is encoded using label coding to obtain the working condition information feature code C; Then, the working state of the signal m is encoded using one-hot encoding to obtain the working state feature encoding Y = [q0, q1, ..., q S-1 ]; Assume the working state is s, then Then use the fast Fourier transform For signal M K,S Extract features to obtain the signal feature vector X; combine the i-th working condition information feature code, working state feature code and signal feature vector to construct a training data sample, denoted as D = [X i ,Y i ,C i ], where X i ,Y i ,C i They respectively represent the frequency domain signal feature vector of the rotating machinery of the i-th training data, the working state feature code of the rotating machinery, and the working condition information feature code of the rotating machinery.

3. The method for robust fault diagnosis of a rotating machinery under unknown working conditions according to claim 2, characterized in that: When encoding information with numerical values, such as the operating speed of a rotating machine, the numerical value is directly placed in the code C. When encoding information without numerical values, such as whether there is a load, 0 or 1 is directly used for encoding.

4. The method for robust fault diagnosis of a rotating machinery under unknown working conditions according to claim 2, characterized in that: The conditional diffusion model includes: an encoder module Encoder(), a decoder module Decoder(), a working condition information feature encoding module Clip() and a diffusion model module Diffuison(); The encoder module extracts features from the input vector and converts the input vector into a feature vector in a latent representation space; the calculation formula of the encoder module is: i,0 =Encoder(X i ), where X i The frequency domain signal feature vector of the rotating machinery representing the i-th data, Z i,0 The feature vector of the latent representation space after the transformation of the i-th data without noise; The decoder module is used to convert a feature vector in the latent representation space into a frequency domain feature vector in the frequency domain space, and the calculation formula is: in, The feature vector representing the latent representation space that needs to be decoded, The feature vector representing the decoded frequency domain space; The working condition information feature encoding module is used to convert the input working condition information vector into a high-dimensional vector and input it into the diffusion model module; the calculation formula of the working condition information feature encoding module is: i,high = Clip(Y i ,C i ), where Y i Indicates the working status characteristic code of the i-th data, C i Indicates the characteristic code of the working condition information of the i-th data, Y i,high Represents the high-dimensional spatial feature vector of the working condition information of the i-th data; The diffusion model module includes the noise adding process Addnoise() and the noise removing process Denoise(): Among them, the noise addition process simulates the data degradation process by gradually adding noise to the feature data in the potential representation space. The noise addition process calculation formula is: Z i,T =Addnoise(Z i,0 ,Y i,high ,T), T represents the number of times the noise adding process is repeated, Z i,T Represents the feature vector of the potential representation space of the i-th data after adding noise T times; the denoising process gradually denoises the degraded data to restore the data before adding noise. The denoising process calculation formula is: T represents the number of times the denoising process is repeated which is the same as the number of times the denoising process is repeated.

5. The robust fault diagnosis method for rotating machinery under unknown working conditions according to claim 4 is characterized in that: The training of the conditional diffusion model using the data in the operation and maintenance database is as follows: When training the conditional diffusion model, first use the loss function Calculate the loss of the encoder module and the decoder module; use the stochastic gradient descent method to update the parameters of the conditional diffusion model where w 1,t , w 1,t+1 They represent the weight parameters of the conditional diffusion model when the model is updated t times and t+1 times, respectively. η1 is the learning rate of the conditional diffusion model weight update. is the loss function L with respect to w 1,t Gradient; after updating b1 times, it is considered that the encoder module and the decoder module have been trained and converged; Then train the diffusion model module and the working condition information feature encoding module of the model; use the loss function Calculate the loss, where represents the expected value, ε t represents the noise vector when adding noise for the tth time, Z i,t represents the latent space feature vector obtained after denoising for the tth time, t represents the number of current denoising times, εθ(Z i,t ,t,Y i,high ) represents the noise predicted by the model during the denoising process; the parameters of the model are updated using the stochastic gradient descent method; after the update times b2, it is considered that the encoder module and the decoder module have been trained and converged, and the conditional diffusion model training is completed.

6. A robust fault diagnosis method for rotating machinery under unknown working conditions according to claim 5, characterized in that: The conditional diffusion model is used to generate rotating machinery vibration signal samples under different working conditions. Specifically, the working state feature code Y ε and the working condition information feature code C of the unknown working condition u u Input into the working condition information feature encoding module of the conditional diffusion model, and randomly generate a Gaussian noise Z ε Input into the diffusion model module, use the denoising process of the diffusion model to generate new feature samples of the latent space according to the output of the working condition information feature encoding module; use the decoder module of the model to decode the generated latent space feature samples into feature samples of the frequency domain space; repeat the above process until the number of generated samples reaches the preset value r, completing the generation of unknown working condition data, and record the generated data as After completing the data generation work, use the data D in the operation and maintenance database and generate data Training of fault diagnosis model.

7. The robust fault diagnosis method for rotating machinery under unknown working conditions according to claim 6, characterized in that: An artificial neural network model is constructed as a known working condition fault diagnosis model. The forward propagation formula of the model is: Wherein f(), X, w2 represent the mapping function of the known working condition fault diagnosis model, the input vector of the known working condition fault diagnosis model, and the weight parameter of the known working condition fault diagnosis model respectively; Represents the prediction vector of the fault diagnosis model; the data D in the operation and maintenance database is used as the training data of the known working condition fault diagnosis model, and the cross entropy function is used The prediction loss of the model is calculated as the loss function of the fault diagnosis model; Use the stochastic gradient descent method with momentum to update the parameters of the fault diagnosis model for known working conditions where v t represents the accumulation of historical gradients in the updating process of the fault diagnosis model under known working conditions, a represents the momentum coefficient, and w 2,t represents the weight parameter when the known working condition fault diagnosis model is updated t times, η2 is the learning rate of the known working condition fault diagnosis model, is the loss function L crossentropy About w 2,t Gradient; when the weight parameters of the known working condition fault diagnosis model are updated b times in the model training, it is considered that the known working condition fault diagnosis model has converged, that is, the known working condition fault diagnosis model has completed training.

8. The robust fault diagnosis method for rotating machinery under unknown working conditions according to claim 7, characterized in that: The adversarial samples are generated using the gradient-based adversarial sample generation algorithm FGSM. All samples in the database are sequentially generated through the FGSM algorithm to generate adversarial samples. The generated adversarial sample data is recorded as D'.

9. The method for robust fault diagnosis of rotating machinery under unknown working conditions according to claim 8, characterized in that: Based on the unknown working condition fault diagnosis model that is the same as the known working condition fault diagnosis model, the data D in the operation and maintenance database, the generated adversarial sample data D′ and the generated data of the generative model are used. The unknown working condition fault diagnosis model is trained as the training data of the fault diagnosis model to obtain the final fault diagnosis model.

10. The robust fault diagnosis method for rotating machinery under unknown working conditions according to claim 9, characterized in that: After the fault diagnosis model is trained, the sensor data collected when the rotating machinery is in an unknown working condition u is input into the fault diagnosis model to obtain the fault diagnosis result of the fault diagnosis model for the rotating machinery under the unknown working condition u.

Images