Source field black box model-based mechanical equipment migration fault diagnosis method and system

Abstract

The invention provides a mechanical equipment migration fault diagnosis method and system based on a source field black box model, and relates to the technical field of mechanical fault diagnos.The method comprises the steps that acceleration signals of mechanical equipment in various states are obtained and preprocessed, and a preprocessed data set is obtained; constructing a source field black box model and training to obtain a trained source field black box model; constructing a target domain model; obtaining the KL divergence loss of the output of the source domain black box model and the output of the target domain model; obtaining the consistency loss of the target domain model by adopting an interpolation consistency regularization method; obtaining nuclear norm loss of the target domain model; obtaining supervised loss with an adversarial regularizer of the target domain model; optimizing the target domain model; and performing fault diagnosis on the mechanical equipment based on the optimized target domain model. The source domain knowledge is extracted and transferred to the target domain model by using the output of the source domain black box model, and source domain data and model parameters are not needed.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and in particular to a mechanical equipment migration fault diagnosis method and system based on a source domain black box model. Background Art

[0002] With the rapid development of modern industrial processes, the safe and stable operation of mechanical equipment has become increasingly important. In recent years, various monitoring and diagnostic technologies have emerged one after another, promoting the discovery of equipment weaknesses and the development of operation and maintenance technologies. Among them, intelligent diagnosis and intelligent operation and maintenance based on deep learning, as well as industrial big data analysis, have developed rapidly. In addition, this technology breaks the conventional diagnostic method guided by expert knowledge in the past, forming a more intelligent end-to-end data-driven diagnostic model, which has broader application prospects.

[0003] Although deep learning-based intelligent diagnosis technology has made remarkable progress, training deep learning-based diagnostic models usually requires massive amounts of labeled data. However, due to the difficulty of disassembling and assembling some equipment or the existence of safety restrictions, it is extremely expensive or even impractical to collect enough labeled data for each diagnostic scenario or task. In addition, limited by the independent and identically distributed assumption of the intelligent diagnostic model, under the influence of changes in working conditions and equipment status, data collected at different times present different probability distributions, making it difficult to directly apply previously trained diagnostic models across domains. Therefore, unsupervised domain adaptation technology has attracted more and more attention. This technology can effectively solve the problem of domain probability distribution differences and solve the diagnostic task of unlabeled data in the target domain through the diagnostic knowledge of labeled source domain data.

[0004] However, no matter what domain adaptation strategy is adopted, labeled source domain data is always required. In addition, due to the high-frequency sampling of monitoring data, the transmission and storage costs of massive source data are high, especially for portable or edge diagnosis needs. To address this problem, passive adaptive diagnosis methods are proposed, which can access well-trained source domain models but cannot obtain labeled source domain data. Specifically, during the learning process, the model parameters are open and can be retrained to adapt to the target domain. Although this method has achieved satisfactory results, there is still a certain risk of exposing the details of the source domain model. Summary of the invention

[0005] The purpose of the present invention is to provide a mechanical equipment migration fault diagnosis method and system based on a source domain black box model, which only requires the output of the source domain black box model to extract and transfer the source domain knowledge to the target domain model without requiring source domain data and model parameters. At the same time, it prevents the model from falling into local optimality and obtains a more accurate target domain model.

[0006] A mechanical equipment migration fault diagnosis method based on a source domain black box model, comprising:

[0007] S1, acquiring acceleration signals generated by stiffness fluctuations caused by the moving contact of mechanical equipment parts under various states, preprocessing the acceleration signals to obtain a preprocessed data set, and dividing the preprocessed data set to obtain a source domain data set and a target domain data set;

[0008] S2, constructing a source domain black box model, based on the source domain dataset, training the source domain black box model in combination with a smoothed cross entropy loss function to obtain a trained source domain black box model;

[0009] S3, inputting any source domain data in the source domain data set into the trained source domain black box model to obtain an output acceleration signal of the source domain black box model;

[0010] S4, build the target domain model;

[0011] S5, inputting any target domain data in the target domain data set into the target domain model to obtain an output acceleration signal of the target domain model;

[0012] S6, obtaining the KL divergence loss between the acceleration signal output by the source domain black box model and the acceleration signal output by the target domain model; the KL divergence loss expression is:

[0013]

[0014] Where:

[0015] D kl (p s ||p t ) represents the KL divergence loss, p s represents the output acceleration signal of the black box model in the source domain, p t represents the output acceleration signal of the target domain model, D kl (b s ||b t ) represents the KL divergence of the prediction probability of the cth category of the source domain black box model and the target domain model, represents the KL divergence of the predicted probabilities of the source domain black box model and the target domain model excluding non-c classes, is the predicted probability of the cth class of the source domain black box model, represents the prediction probability of the source domain black box model excluding non-c classes, is the predicted probability of the c-th sample of the target domain model, represents the predicted probability of the target domain model excluding non-c categories, K represents the total number of fault categories, represents the predicted probability of the i-th category of the source domain black box model, represents the predicted probability of the i-th category of the target domain model, λ is the first hyperparameter, and the c-th category is the fault category with the highest predicted probability;

[0016] S7, based on the target domain data set, using an interpolation consistency regularization method to obtain the consistency loss of the target domain model;

[0017] The consistency loss expression is:

[0018]

[0019] Where: X t represents the target domain dataset, L mix represents the consistency loss, represents the qth target area acceleration signal data in the target area data set, represents the jth target area acceleration signal data in the target area dataset, The input of the target domain model is The output when The input of the target domain model is Output when Mix β (a, b) = β·a+(1-β)·b, β∈Beat(δ,δ) means that β is sampled from the Beta distribution, δ is the parameter value,

[0020] for and The KL divergence between

[0021] S8, obtaining the nuclear norm loss of the target domain model based on the acceleration signal output by the target domain model;

[0022] S9, obtaining a supervised loss with an adversarial regularizer of the target domain model based on the acceleration signal output by the target domain model;

[0023] S10, optimizing the target domain model based on the KL divergence loss, the consistency loss, the nuclear norm loss, and the supervised loss with the adversarial regularizer;

[0024] S11, repeat S5-S10 to obtain the final target domain model;

[0025] S12, based on the final target domain model, inputting the real-time acceleration signal of the mechanical equipment into the final target domain model to perform fault diagnosis on the mechanical equipment.

[0026] Optionally, the smoothed cross entropy loss function expression is:

[0027]

[0028] Where: is the source domain dataset, is the label corresponding to the source domain dataset, represents the black box model of the source domain, represents the output of the source domain black box model when the input is x, represents the predicted probability of the kth class of the source domain black box model, is a smooth label, and When y = k, q k =1, otherwise q k =0, α is the smoothing parameter, and u(k)=1 / K is the uniform distribution of labels.

[0029] Optionally, the nuclear norm loss expression is:

[0030]

[0031] Where: L nm is the nuclear norm loss, B is the length of the qth target domain data in the target domain dataset, || || * represents the nuclear norm.

[0032] Optionally, the supervised loss expression with adversarial regularizer is:

[0033]

[0034] Where: L ce is the supervised loss function with adversarial regularizer; y e For easy sample X e Pseudo-label of θ t are the parameters of the target domain model, θ Ω is the parameter of Ω, Ω is the adversarial domain adaptation discriminator, For easy sample sets, is the difficult sample set, L sa is the supervised loss with adversarial regularizer, f t is the target domain model, f t (X h ) indicates that the input of the target domain model is X h The output when f t (X e ) indicates that the input of the target domain model is X e Output when .

[0035] Optionally, obtaining the self-entropy of each target domain data in the target domain data set;

[0036] The average value of each of the self-entropies is obtained, and the target domain data corresponding to the self-entropies less than the average value is used as the easy sample set, and the target domain data corresponding to the self-entropies greater than or equal to the average value is used as the difficult sample set.

[0037] Optionally, the expression of the self-entropy is:

[0038]

[0039] Where: express The self-entropy, Indicates that the target domain model number input is The predicted probability of the kth class is .

[0040] The present invention also provides a mechanical equipment migration fault diagnosis system based on a source domain black box model, which comprises:

[0041] A data acquisition module, used to acquire acceleration signals of mechanical equipment in various states, preprocess the acceleration signals to obtain a preprocessed data set, and divide the preprocessed data set into a source domain data set and a target domain data set;

[0042] The first model building module is used to build a source domain black box model, and train the source domain black box model based on the source domain dataset in combination with a smoothed cross entropy loss function to obtain the trained source domain black box model;

[0043] A model training module, used for inputting any source domain data in the source domain data set into the trained source domain black box model to obtain the source domain black box model output;

[0044] The second model building module is used to build a target domain model;

[0045] A data input module, used to input any target domain data in the target domain data set into the target domain model to obtain a target domain model output;

[0046] The KL divergence loss module is used to obtain the KL divergence loss between the output of the source domain black box model and the output of the target domain model; the KL divergence loss expression is:

[0047]

[0048] Where: D kl (p s ||p t) represents the KL divergence loss, p s represents the output of the source domain black box model, p t represents the target domain model output, D kl (b s ||b t ) represents the KL divergence of the prediction probability of the cth category of the source domain black box model and the target domain model, represents the KL divergence of the predicted probabilities of the source domain black box model and the target domain model excluding the cth class. is the predicted probability of the cth class of the source domain black box model, represents the predicted probability of other classes excluding class c by the source domain black box model, is the predicted probability of the c-th sample of the target domain model, represents the prediction probability of other classes excluding the cth class in the target domain model, K represents the total number of fault categories, represents the predicted probability of the i-th category of the source domain black box model, represents the prediction probability of the i-th category of the target domain model, λ is the first hyperparameter, and the c-th category is the fault category with the highest probability of accurate prediction;

[0049] A consistency loss module, used to obtain the consistency loss of the target domain model by using an interpolation consistency regularization method based on the target domain data set;

[0050] The consistency loss expression is:

[0051]

[0052] Where: X t represents the target domain dataset, L mix represents the consistency loss, represents the qth target domain data in the target domain dataset, represents the jth target domain data in the target domain dataset, The input of the target domain model is The output when The input of the target domain model is Output when Mix β (a, b) = β·a+(1-β)·b, β∈Beat(δ,δ) means that β is sampled from the Beta distribution, δ is the parameter value, for and The KL divergence between

[0053] A nuclear norm loss module, used for obtaining the nuclear norm loss of the target domain model based on the target domain model output;

[0054] A supervised loss with an adversarial regularizer, for obtaining a supervised loss with an adversarial regularizer of the target domain model based on the target domain model output;

[0055] A model optimization module, used for optimizing the target domain model based on the KL divergence loss, the consistency loss, the nuclear norm loss and the supervised loss with adversarial regularizer;

[0056] A repeated execution module, used for repeatedly executing the data input module to the model optimization module to obtain a final target domain model;

[0057] A fault diagnosis module is used to perform fault diagnosis on mechanical equipment based on the final target domain model.

[0058] The effects of the present invention are as follows:

[0059] The mechanical equipment migration fault diagnosis method method of the present invention is based on the source domain black box model. The constructed target domain model does not require training set data, labels and parameters of the source model. It only relies on the predicted output obtained from the training model to construct the target domain model for unlabeled test data. On the one hand, it greatly ensures the privacy and security of the data. On the other hand, since there is no need to obtain the parameters of the source domain black box model, it greatly expands the applicability and convenience.

[0060] The present invention proposes a mechanical equipment migration fault diagnosis method based on a source domain black box model. The target domain model is optimized by obtaining the KL divergence loss of the source domain black box model output and the target domain model output. In order to avoid the local optimum of the model, the target domain model is optimized by combining consistency loss, nuclear norm loss and supervised loss with an adversarial regularizer, thereby improving the diagnostic performance of the target domain model.

[0061] The present invention provides a mechanical equipment migration fault diagnosis method based on the source domain black box model, establishes a learning method based on easy samples to guide difficult samples, and greatly improves the diagnostic performance of the target domain model. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a flow chart of the mechanical equipment migration fault diagnosis method based on the source domain black box model of the present invention;

[0063] Figure 2 It is a structural schematic diagram of the source domain black box model of the present invention;

[0064] Figure 3 It is a schematic diagram of detailed structural parameters of the diagnostic network of the source domain black box model of the present invention;

[0065] Figure 4It is a structural schematic diagram of the rolling bearing experimental device of the present invention;

[0066] Figure 5 This is a schematic diagram of the structure of the nuclear power gearbox experimental device of the present invention.

[0067] In the figure: 1. drive motor; 2. coupling; 3. planetary gearbox; 4. parallel shaft gearbox; 5. speed torque sensor; 6. rolling bearing to be tested; 7. acceleration sensor; 8. magnetic powder brake.

[0068] 11. Driving motor; 12. Coupling; 13. Planetary gearbox under test; 14. Acceleration sensor; 15. Support bearing; 16. Bevel gear; 17. Gear pump. DETAILED DESCRIPTION

[0069] Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[0070] Figure 1 FIG. 1 is a flow chart of a mechanical equipment migration fault diagnosis method based on a source domain black box model of the present invention. Figure 1 As shown, the present invention provides a mechanical equipment migration fault diagnosis method based on a source domain black box model, especially for state detection and fault diagnosis of bearing mechanical equipment (such as motors, gear boxes and other rotating mechanical equipment), which specifically includes the following steps:

[0071] S1, obtain the acceleration signal, i.e., vibration signal, generated by stiffness fluctuation caused by the moving contact of mechanical equipment parts under various states, and preprocess the acceleration signal to obtain a preprocessed data set, and divide the preprocessed data set into a source domain data set and a target domain data set; specifically, the acceleration signal of the mechanical equipment under various health states is collected by an acceleration sensor, and several data sets are obtained by window sliding. For several data sets, normal standardization preprocessing is used to normalize the data amplitude to the interval [0,1] to ensure data consistency, and obtain the preprocessed data set.

[0072] S2, build a source domain black box model, based on the source domain data set, combined with the smooth cross entropy loss function to train the source domain black box model to obtain a trained source domain black box model. The input data and parameters of the trained source domain black box model become unavailable. The smooth cross entropy loss function is expressed as:

[0073]

[0074] Where: is the source domain dataset, is the label corresponding to the source domain dataset, represents the black box model of the source domain, represents the output of the source domain black box model when the input is x, represents the predicted probability of the kth class of the source domain black box model, is a smooth label, and When y = k, q k =1, otherwise q k =0, α is the smoothing parameter, α=0.1, u(k)=1 / K is the uniform distribution of labels.

[0075] S3, input any source domain data in the source domain dataset into the trained source domain black box model to obtain the source domain black box model output.

[0076] S4, build the target domain model.

[0077] S5, input any target domain data in the target domain data set into the target domain model to obtain the target domain model output.

[0078] S6, obtain the KL divergence loss between the source domain black box model output and the target domain model output. The KL divergence loss expression is:

[0079]

[0080] Where: D kl (p s ||p t ) represents the KL divergence loss, p s represents the output of the source domain black box model, p t represents the target domain model output, D kl (b s ||b t ) represents the KL divergence of the prediction probability of the cth category of the source domain black box model and the target domain model, represents the KL divergence of the predicted probabilities of the source domain black box model and the target domain model excluding the cth class. is the predicted probability of the cth class of the source domain black box model, represents the predicted probability of other classes excluding class c by the source domain black box model, is the predicted probability of the c-th sample of the target domain model, represents the prediction probability of other classes excluding the cth class in the target domain model, K represents the total number of fault categories, represents the predicted probability of the i-th category of the source domain black box model, represents the prediction probability of the i-th category of the target domain model, λ is the first hyperparameter, and the c-th category is the fault category with the highest probability of accurate prediction.

[0081] The prediction probability of the source domain black box model output and the target domain model output excluding the cth class is expressed as Then the predicted probability of the non-target class of the i-th (i≠c) class is Therefore, the following expression can be obtained:

[0082]

[0083] This gives the KL divergence loss expression.

[0084] S7, based on the target domain dataset, uses the interpolation consistency regularization method to obtain the consistency loss of the target domain model. The consistency loss expression is:

[0085]

[0086] Where: X t represents the target domain dataset, L mix represents the consistency loss, represents the qth target domain data in the target domain dataset, represents the jth target domain data in the target domain dataset, The input of the target domain model is The output when The input of the target domain model is Output when Mix β (a, b) = β·a+(1-β)·b, β∈Beat(δ,δ) means β is sampled from Beta distribution, δ is the parameter value, δ = 0.3, for and The KL divergence between .

[0087] The consistency loss is implemented based on MixUp.

[0088] S8, obtain the nuclear norm loss of the target domain model based on the output of the target domain model. The nuclear norm loss expression is:

[0089]

[0090] Where: L nm is the nuclear norm loss, B is the length of the qth target domain data in the target domain dataset, || || * represents the nuclear norm.

[0091] S9, obtain the supervised loss of the target domain model with adversarial regularizer based on the output of the target domain model. The supervised loss expression with adversarial regularizer is:

[0092]

[0093] Where: Lce is the supervised loss function with adversarial regularizer; y e For easy sample X e Pseudo-label of θ t are the parameters of the target domain model, θ Ω is the parameter of Ω, Ω is the adversarial domain adaptation discriminator, For easy sample sets, is the difficult sample set, L sa is the supervised loss with adversarial regularizer, f t is the target domain model, f t (X h ) indicates that the input of the target domain model is X h The output when f t (X e ) indicates that the input of the target domain model is X e Output when .

[0094] Preferably, for the target domain model that is prone to fall into the local optimal parameter space guided by easy samples, an easy sample guided difficult sample adaptive learning strategy is designed, and the self-entropy of each target domain data in the target domain dataset is first obtained.

[0095] Furthermore, the average self-entropy is used as the division standard to obtain the average value of each entropy, and the target domain data corresponding to the self-entropy less than the average value is used as the easy sample set, and the target domain data corresponding to the self-entropy greater than or equal to the average value is used as the difficult sample set.

[0096] The expression of self-entropy is:

[0097]

[0098] Where: express The self-entropy, Indicates that the target domain model number input is The predicted probability of the kth class is .

[0099] By introducing the adversarial domain adaptation discriminator Ω, the distance between easy samples and difficult samples in the target samples is shortened in the high-dimensional space.

[0100] S10, optimizes the target domain model based on KL divergence loss, consistency loss, nuclear norm loss and supervised loss with adversarial regularizer.

[0101] S11, repeat S5-S10 to obtain the final target domain model.

[0102] S12, perform fault diagnosis on mechanical equipment based on the final target domain model.

[0103] Specifically, mechanical equipment such as gear boxes, motors and bearings are suitable for fault diagnosis using the method of the present invention.

[0104] The present invention also provides a mechanical equipment migration fault diagnosis system based on a source domain black box model, which comprises:

[0105] The data acquisition module is used to acquire the acceleration signal of the mechanical equipment under various states, preprocess the acceleration signal to obtain a preprocessed data set, and divide the preprocessed data set into a source domain data set and a target domain data set.

[0106] The first model building module is used to build a source domain black box model. Based on the source domain dataset, the source domain black box model is trained in combination with a smoothed cross entropy loss function to obtain a trained source domain black box model.

[0107] The model training module is used to input any source domain data in the source domain data set into the trained source domain black box model to obtain the source domain black box model output.

[0108] The second model building module is used to build a target domain model.

[0109] The data input module is used to input any target domain data in the target domain data set into the target domain model to obtain the target domain model output.

[0110] The KL divergence loss module is used to obtain the KL divergence loss between the source domain black box model output and the target domain model output. The KL divergence loss expression is:

[0111]

[0112] Where: D kl (p s ||p t ) represents the KL divergence loss, p s represents the output of the source domain black box model, p t represents the target domain model output, D kl (b s ||b t ) represents the KL divergence of the prediction probability of the cth category of the source domain black box model and the target domain model, represents the KL divergence of the predicted probabilities of the source domain black box model and the target domain model excluding the cth class. is the predicted probability of the cth class of the source domain black box model, represents the predicted probability of other classes excluding class c by the source domain black box model, is the predicted probability of the c-th sample of the target domain model, represents the prediction probability of other classes excluding the cth class in the target domain model, K represents the total number of fault categories, represents the predicted probability of the i-th category of the source domain black box model, represents the prediction probability of the i-th category of the target domain model, λ is the first hyperparameter, and the c-th category is the fault category with the highest probability of accurate prediction.

[0113] The consistency loss module is used to obtain the consistency loss of the target domain model based on the target domain dataset using the interpolation consistency regularization method. The consistency loss expression is:

[0114]

[0115] Where: X t represents the target domain dataset, L mix represents the consistency loss, represents the qth target domain data in the target domain dataset, represents the jth target domain data in the target domain dataset, The input of the target domain model is The output when The input of the target domain model is Output when Mix β (a, b) = β·a+(1-β)·b, β∈Beat(δ,δ) means that β is sampled from the Beta distribution, δ is the parameter value, for and The KL divergence between .

[0116] The nuclear norm loss module is used to obtain the nuclear norm loss of the target domain model based on the target domain model output.

[0117] Supervised loss with adversarial regularizer, used to obtain the supervised loss with adversarial regularizer of the target domain model based on the target domain model output.

[0118] The model optimization module is used to optimize the target domain model based on KL divergence loss, consistency loss, nuclear norm loss, and supervised loss with adversarial regularizer.

[0119] The repeated execution module is used to repeatedly execute the data input module to the model optimization module to obtain the final target domain model.

[0120] The fault diagnosis module is used to diagnose faults of mechanical equipment based on the final target domain model.

[0121] The present invention also provides a rolling bearing migration fault diagnosis method based on a source domain black box model, which comprises:

[0122] S10, acquiring the acceleration signal of the rolling bearing under various states, and preprocessing the acceleration signal to obtain a preprocessed data set, and dividing the preprocessed data set to obtain a source domain data set and a target domain data set. Specifically, the acceleration signal of the mechanical equipment under various health states is collected by an acceleration sensor, and several data sets are obtained by window sliding. For several data sets, normal standardization preprocessing is used to normalize the data amplitude to the [0,1] interval to ensure data consistency, and obtain a preprocessed data set. The rolling bearing is any one of a deep groove ball bearing, a cylindrical roller bearing and a thrust bearing. The rolling bearing includes a bearing outer ring, a bearing inner ring, rolling elements and a cage. The rolling bearing is arranged on a bearing seat, and the acceleration sensor measures the acceleration signal of the bearing seat. The inner ring of the bearing cooperates with the shaft and rotates with the shaft; the outer ring of the bearing cooperates with the bearing seat to play a supporting role; the rolling elements are evenly distributed between the inner ring of the bearing and the outer ring of the bearing through the cage.

[0123] S20, construct a source domain black box model, and train the source domain black box model based on the source domain data set and the smooth cross entropy loss function to obtain a trained source domain black box model. The input data and parameters of the trained source domain black box model become unavailable. The smooth cross entropy loss function is expressed as:

[0124]

[0125] Where: is the source domain dataset, is the label corresponding to the source domain dataset, represents the black box model of the source domain, represents the output of the source domain black box model when the input is x, represents the predicted probability of the kth class of the source domain black box model, is a smooth label, and When y = k, q k =1, otherwise q k =0, α is the smoothing parameter, α=0.1, u(k)=1 / K is the uniform distribution of labels.

[0126] S30, inputting any source domain data in the source domain data set into the trained source domain black box model to obtain the source domain black box model output.

[0127] S40, build the target domain model.

[0128] S50, inputting any target domain data in the target domain data set into the target domain model to obtain a target domain model output.

[0129] S60, obtaining the KL divergence loss between the source domain black box model output and the target domain model output. The KL divergence loss expression is:

[0130]

[0131] Where:

[0132] D kl (p s ||p t ) represents the KL divergence loss, p s represents the output of the source domain black box model, p t represents the target domain model output, D kl (b s ||b t ) represents the KL divergence of the prediction probability of the cth category of the source domain black box model and the target domain model, represents the KL divergence of the predicted probabilities of the source domain black box model and the target domain model excluding the cth class. is the predicted probability of the cth class of the source domain black box model, represents the predicted probability of other classes excluding class c by the source domain black box model, is the predicted probability of the c-th sample of the target domain model, represents the prediction probability of other classes excluding the cth class in the target domain model, K represents the total number of fault categories, represents the predicted probability of the i-th category of the source domain black box model, represents the prediction probability of the i-th category of the target domain model, λ is the first hyperparameter, and the c-th category is the fault category with the highest probability of accurate prediction.

[0133] The prediction probability of the source domain black box model output and the target domain model output excluding the cth class is expressed as Then the predicted probability of the non-target class of the i-th (i≠c) class is Therefore, the following expression can be obtained:

[0134]

[0135] This gives the KL divergence loss expression.

[0136] S70, based on the target domain data set, the interpolation consistency regularization method is used to obtain the consistency loss of the target domain model. The consistency loss expression is:

[0137]

[0138] Where: X t represents the target domain dataset, L mixrepresents the consistency loss, represents the qth target domain data in the target domain dataset, represents the jth target domain data in the target domain dataset, The input of the target domain model is The output when The input of the target domain model is Output when Mix β (a, b) = β·a+(1-β)·b, β∈Beat(δ,δ) means β is sampled from Beta distribution, δ is the parameter value, δ = 0.3, for and The KL divergence between .

[0139] The consistency loss is implemented based on MixUp.

[0140] S80, obtaining the nuclear norm loss of the target domain model based on the target domain model output. The nuclear norm loss expression is:

[0141]

[0142] Where: L nm is the nuclear norm loss, B is the length of the qth target domain data in the target domain dataset, || || * represents the nuclear norm.

[0143] S90, obtain the supervised loss of the target domain model with the adversarial regularizer based on the output of the target domain model. The supervised loss expression with the adversarial regularizer is:

[0144]

[0145] Where: L ce is the supervised loss function with adversarial regularizer; y e For easy sample X e Pseudo-label of θ t are the parameters of the target domain model, θ Ω is the parameter of Ω, Ω is the adversarial domain adaptation discriminator, For easy sample sets, is the difficult sample set, L sa is the supervised loss with adversarial regularizer, f t is the target domain model, f t (X h ) indicates that the input of the target domain model is X h The output when f t (X e ) indicates that the input of the target domain model is Xe Output when .

[0146] Preferably, for the target domain model that is prone to fall into the local optimal parameter space guided by easy samples, an easy sample guided difficult sample adaptive learning strategy is designed, and the self-entropy of each target domain data in the target domain dataset is first obtained.

[0147] Furthermore, the average self-entropy is used as the division standard to obtain the average value of each entropy, and the target domain data corresponding to the self-entropy less than the average value is used as the easy sample set, and the target domain data corresponding to the self-entropy greater than or equal to the average value is used as the difficult sample set.

[0148] The expression of self-entropy is:

[0149]

[0150] Where: express The self-entropy, Indicates that the target domain model number input is The predicted probability of the kth class is .

[0151] By introducing the adversarial domain adaptation discriminator Ω, the distance between easy samples and difficult samples in the target samples is shortened in the high-dimensional space.

[0152] S100 optimizes the target domain model based on KL divergence loss, consistency loss, nuclear norm loss, and supervised loss with adversarial regularizer.

[0153] S110, repeat S50-S100 to obtain the final target domain model.

[0154] S120, performing fault diagnosis on the rolling bearing based on the final target domain model.

[0155] Specifically, a rolling bearing is taken as an example for further description. The sampling frequency of the acceleration signal of the rolling bearing is 10 kHz.

[0156] The traditional method that can obtain source domain data and source model parameters and the diagnostic method without passive domain data but can obtain source model parameters are taken as references, and the specific comparison methods are as follows: 1) directly testing the target data based on the source model trained only on the source domain labeled data (SourceOnly); 2) the traditional domain adaptation fault diagnosis method based on maximum mean difference (Maximum mean discrepancy, MAD); 3) the traditional domain adaptation migration fault diagnosis method based on correlation alignment (Correlation alignment-based adaptation diagnosis, CAD); 4) the migration fault diagnosis method based on adversarial domain adaptation (Adversarial learning-based adaptation diagnosis, ALD); 5) the domain adaptation fault diagnosis method of passive data (Source hypothesis transfer, SHOT); 6) the improved domain adaptation fault diagnosis method of passive data (Source free adaptation diagnosi s, SFAD); 7) the comparative fault diagnosis method (Decoupled self-distillation in KTDA, KTDArd) constructed based on the traditional knowledge distillation method instead of the knowledge decoupled self-distillation method of the present invention. Among them, for intelligent diagnosis methods under transmission field adaptation such as MAD, CAD, ALD, 70% of the training data and test data are used for training the source model and the target model, and the remaining target data is used for the final model performance verification test. For SHOT, KTDArd and the method SBAD proposed by the present invention, the source data cannot be obtained, and the model can only be trained with the target data based on the source model, and 70% of the target data is used for model training and the remaining data is used for model verification testing.

[0157] Specifically applied to rolling bearing fault diagnosis, the present invention uses transmission test bench bearing data for verification. The schematic diagram of the test bench is shown in Figure 4 The source domain black box model structure used is as follows: Figure 2 The specific process and parameters are as shown in Figure 3 shown.

[0158] The test bench includes a drive motor 1, a coupling 2, a speed torque sensor 5, a planetary gearbox 3, a parallel shaft gearbox 4, a rolling bearing to be tested 6, a magnetic powder brake 8 and an acceleration sensor 7. The power of the test bench is provided by the drive motor 1. The present invention uses vibration data with a sampling frequency of 10kHz for analysis. The rolling bearing to be tested includes seven health states, including normal, inner ring crack, inner ring wear, outer ring crack, outer ring wear, inner and outer ring cracks, and cage failure. The test conditions include W11: speed 1020pm load 0.7Nm; W12: speed 1020rpm load 1.2Nm; W13: speed 1020rpm load 2Nm; W21: speed 780rpm load 0.7Nm; W22: speed 780rpm load 1.2Nm. Through the above 5 test conditions, 14 groups of migration diagnostic tasks were established, of which 6 groups of diagnostic tasks were speed changes, and the other 8 groups of diagnostic tasks were load changes.

[0159] Based on the constructed acceleration signal dataset, after end-to-end training and testing of multiple methods, the final diagnostic classification accuracy results are shown in Table 1. "W11→W12" means migration diagnosis from source domain condition W11 to target domain condition W12, and the rest are similar. Each diagnostic method under each diagnostic task is tested 5 times to take the average test accuracy and standard deviation to more comprehensively verify the effectiveness of the present invention. The expression method of √ and × in Table 1 describes whether the tested diagnostic method meets the setting of passive data and passive model parameters. It can be seen from the table that: 1) Even if the source data and source model parameters cannot be obtained, the average diagnostic accuracy (98.71%) of the SBAD method of the present invention is still the best; 2) Since the SourceOnly method cannot know the target data distribution, its average diagnostic accuracy as a conventional diagnostic method is only 83.53%; 3) Although the ALD, SHOT and SFAD methods are better than SBAD in some diagnostic tasks, the ALD, SHOT and SFAD methods can obtain more prior knowledge, namely source data or source model parameters; 4) After replacing the decoupled self-distillation method, the diagnostic performance of the KTDArd method decreases, verifying the effectiveness of the target sample feature mining in the present invention; 5) The SBAD proposed in the present invention performs better in multiple groups of diagnostic tasks, and the minimum diagnostic accuracy also exceeds 93.08%. Therefore, although the source data and source model parameters cannot be obtained during the training of the target diagnostic model, the SBAD proposed in the present invention still stands out among various mainstream diagnostic methods.

[0160]

[0161]

[0162] Specifically applied to gearbox fault diagnosis, another embodiment of the present invention uses nuclear power gearbox (scaled) test bench test data for verification. The schematic diagram of the test bench is shown in FIG. Figure 4 The source domain black box model structure used is as follows: Figure 2 The specific process and parameters are as shown in Figure 3 shown.

[0163] The test bench includes a drive motor 1, a coupling 2, a planetary gearbox to be tested 3, a support bearing 5, a bevel gear 6, a gear pump 7 and an acceleration sensor 4, and the power of the test bench is provided by the drive motor 1. The present invention uses vibration data with a sampling frequency of 12.8kHz for analysis. The planetary gearbox to be tested includes eight health states, including normal, sun gear peeling, sun gear wear, sun gear crack, planetary gear pitting, planetary gear wear, planetary gear crack, and planetary gear severe crack. The test conditions include T11: speed 746pm load is 25% of the rated load; T12: speed 746rpm load is 50% of the rated load; T13: speed 746rpm load is 75% of the rated load; T21: speed 895rpm load is 25% of the rated load; T22: speed 895rpm load is 50% of the rated load; T23: speed 895rpm load is 75% of the rated load. Through the above 6 test conditions, 12 groups of migration diagnostic tasks with changes in speed and load are established.

[0164] Based on the constructed acceleration signal data set, after end-to-end training and testing of multiple methods, the final diagnostic classification accuracy results are shown in Table 1. "T11→T22" means migration diagnosis from source domain condition T11 to target domain condition T22, and the rest are similar. In Table 2, the expression methods of √ and × are used to describe whether the tested diagnostic method meets the setting of passive data and passive model parameters. It can be seen from the table that: 1) Except for the Sourceonly method, the average diagnostic accuracy of other methods exceeds 94%; 2) The average diagnostic accuracy of the KTDA method reaches 99.82; 3) The traditional ALD method that requires source domain data, model and target domain unlabeled data has an average diagnostic accuracy of 99.33%, but there is still a gap of 0.49% compared with the SBAD method proposed in this invention; 4) Although the average diagnostic accuracy of the SFAD method is the best, it needs to meet the prerequisite of obtaining the source domain diagnostic model parameters, which is difficult to achieve in some actual industrial scenarios. Therefore, the SBAD method proposed in the present invention has excellent results in terms of diagnostic performance and applicability, stands out among various mainstream diagnostic methods, and can handle more mechanical equipment diagnostic tasks under the constraints of passive field data and source model parameters.

[0165]

[0166]

[0167] The embodiments described above are only descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A mechanical equipment migration fault diagnosis method based on a source domain black box model, characterized in that: It includes: S1, acquiring acceleration signals generated by stiffness fluctuations caused by the moving contact of mechanical equipment parts in various states, preprocessing the acceleration signals to obtain a preprocessed data set, and dividing the preprocessed data set to obtain a source domain data set and a target domain data set; S2, constructing a source domain black box model, based on the source domain dataset, training the source domain black box model in combination with a smoothed cross entropy loss function to obtain a trained source domain black box model; S3, inputting any source domain data in the source domain data set into the trained source domain black box model to obtain an output acceleration signal of the source domain black box model; S4, build the target domain model; S5, inputting any target domain data in the target domain data set into the target domain model to obtain an output acceleration signal of the target domain model; S6, obtaining the KL divergence loss between the acceleration signal output by the source domain black box model and the acceleration signal output by the target domain model; the KL divergence loss expression is: Where: D kl (p s ||p t ) represents the KL divergence loss, p s represents the output acceleration signal of the black box model in the source domain, p t represents the output acceleration signal of the target domain model, D kl (b s ||b t ) represents the KL divergence of the prediction probability of the cth category of the source domain black box model and the target domain model, represents the KL divergence of the predicted probabilities of the source domain black box model and the target domain model excluding non-c classes, is the predicted probability of the cth class of the source domain black box model, represents the prediction probability of the source domain black box model excluding non-c classes, is the predicted probability of the c-th sample of the target domain model, represents the predicted probability of the target domain model excluding non-c categories, K represents the total number of fault categories, represents the predicted probability of the i-th category of the source domain black box model, represents the prediction probability of the i-th category of the target domain model, λ is the first hyperparameter, and the c-th category is the fault category with the highest probability of accurate prediction; S7, based on the target domain data set, using an interpolation consistency regularization method to obtain the consistency loss of the target domain model; The consistency loss expression is: Where: X t represents the target domain dataset, L mix represents the consistency loss, represents the qth target area acceleration signal data in the target area data set, represents the jth target area acceleration signal data in the target area dataset, The input of the target domain model is The output when The input of the target domain model is Output when Mix β (a, b) = β·a+(1-β)·b, β∈Beat(δ,δ) means that β is sampled from the Beta distribution, δ is the parameter value, for and The KL divergence between S8, obtaining the nuclear norm loss of the target domain model based on the acceleration signal output by the target domain model; S9, obtaining a supervised loss with an adversarial regularizer of the target domain model based on the acceleration signal output by the target domain model; S10, optimizing the target domain model based on the KL divergence loss, the consistency loss, the nuclear norm loss, and the supervised loss with the adversarial regularizer; S11, repeat S5-S10 to obtain the final target domain model; S12, based on the final target domain model, inputting the real-time acceleration signal of the mechanical equipment into the final target domain model to perform fault diagnosis on the mechanical equipment.

2. The mechanical equipment migration fault diagnosis method based on the source domain black box model according to claim 1 is characterized in that: The smoothed cross entropy loss function expression is: Where: X s is the source domain dataset, Y s is the label corresponding to the source domain dataset, represents the black box model of the source domain, represents the output of the source domain black box model when the input is x, represents the predicted probability of the kth class of the source domain black box model, is a smooth label, and When y = k, q k =1, otherwise q k =0, α is the smoothing parameter, and u(k)=1 / K is the uniform distribution of labels.

3. The mechanical equipment migration fault diagnosis method based on the source domain black box model according to claim 1 is characterized in that: The nuclear norm loss expression is: Where: L nm is the nuclear norm loss, B is the length of the qth target domain data in the target domain dataset, || || * represents the nuclear norm.

4. The mechanical equipment migration fault diagnosis method based on the source domain black box model according to claim 1 is characterized in that: The supervised loss expression with adversarial regularizer is: Where: L ce is the supervised loss function with adversarial regularizer; y e For easy sample X e Pseudo-label of θ t are the parameters of the target domain model, θ Ω is the parameter of Ω, Ω is the adversarial domain adaptation discriminator, For easy sample sets, is the difficult sample set, L sa is the supervised loss with adversarial regularizer, f t is the target domain model, f t (X h ) indicates that the input of the target domain model is X h The output when f t (X e ) indicates that the input of the target domain model is X e Output when .

5. The mechanical equipment migration fault diagnosis method based on the source domain black box model according to claim 4 is characterized in that: Obtaining the self-entropy of each target domain data in the target domain data set; The average value of each self-entropy is obtained, and the target domain data corresponding to the self-entropy less than the average value is used as the easy sample set, and the target domain data corresponding to the self-entropy greater than or equal to the average value is used as the difficult sample set.

6. The mechanical equipment migration fault diagnosis method based on the source domain black box model according to claim 5 is characterized in that: The expression of the self-entropy is: Where: express The self-entropy, Indicates that the target domain model number input is The predicted probability of the kth class when .

7. A mechanical equipment migration fault diagnosis system based on source domain black box model, characterized in that: It includes: A data acquisition module, used to acquire acceleration signals of mechanical equipment in various states, preprocess the acceleration signals to obtain a preprocessed data set, and divide the preprocessed data set into a source domain data set and a target domain data set; The first model building module is used to build a source domain black box model, and train the source domain black box model based on the source domain dataset in combination with a smoothed cross entropy loss function to obtain the trained source domain black box model; A model training module, used for inputting any source domain data in the source domain data set into the trained source domain black box model to obtain the source domain black box model output; The second model building module is used to build a target domain model; A data input module, used to input any target domain data in the target domain data set into the target domain model to obtain a target domain model output; The KL divergence loss module is used to obtain the KL divergence loss between the output of the source domain black box model and the output of the target domain model; the KL divergence loss expression is: Where: D kl (p s ||p t ) represents the KL divergence loss, p s represents the output of the source domain black box model, p t represents the target domain model output, D kl (b s ||b t ) represents the KL divergence of the prediction probability of the cth category of the source domain black box model and the target domain model, represents the KL divergence of the predicted probabilities of the source domain black box model and the target domain model excluding the cth class. is the predicted probability of the cth class of the source domain black box model, represents the predicted probability of other classes excluding class c by the source domain black box model, is the predicted probability of the c-th sample of the target domain model, represents the prediction probability of other classes excluding the cth class in the target domain model, K represents the total number of fault categories, represents the predicted probability of the i-th category of the source domain black box model, represents the prediction probability of the i-th category of the target domain model, λ is the first hyperparameter, and the c-th category is the fault category with the highest probability of accurate prediction; A consistency loss module, used to obtain the consistency loss of the target domain model by using an interpolation consistency regularization method based on the target domain data set; The consistency loss expression is: Where: X t represents the target domain dataset, L mix represents the consistency loss, represents the qth target domain data in the target domain dataset, represents the jth target domain data in the target domain dataset, The input of the target domain model is The output when The input of the target domain model is Output when Mix β (a, b) = β·a+(1-β)·b, β∈Beat(δ,δ) means that β is sampled from the Beta distribution, δ is the parameter value, for and The KL divergence between A nuclear norm loss module, used for obtaining the nuclear norm loss of the target domain model based on the target domain model output; A supervised loss with an adversarial regularizer, for obtaining a supervised loss with an adversarial regularizer of the target domain model based on the target domain model output; A model optimization module, used for optimizing the target domain model based on the KL divergence loss, the consistency loss, the nuclear norm loss and the supervised loss with adversarial regularizer; A repeated execution module, used for repeatedly executing the data input module to the model optimization module to obtain a final target domain model; A fault diagnosis module is used to perform fault diagnosis on mechanical equipment based on the final target domain model.

Images