A zero-sample mechanical fault diagnosis method for variable working conditions driven by mechanism data coupling

Abstract

The present invention relates to a mechanical parameter pattern recognition method, and more specifically, to a mechanism-data coupling-driven, variable-operating-condition, zero-sample mechanical fault diagnosis method. This method addresses the problem that, while domain generalization networks can perform fault diagnosis for new operating conditions without data, the model's diagnostic performance depends on the distribution of source domain data and fails to consider the operating condition information and data distribution of the new operating condition, resulting in insufficient generalization performance and poor diagnostic capabilities. This mechanism-data coupling-driven, variable-operating-condition, zero-sample mechanical fault diagnosis method proposes a fault sequence conversion model based on a quadratic polynomial function when establishing the operating condition conversion relationship. This model effectively converts real data from the old operating condition into converted data for the new operating condition, which is then used to fine-tune the parameters of a generative adversarial network to generate generated data similar to the real data for the new operating condition.

Description

Technical Field

[0001] The present invention relates to a mechanical parameter pattern recognition method, and in particular to a variable working condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling. Background Art

[0002] Bearing fault diagnosis plays an important role in the safe and stable operation of rotating machinery. With the rapid development of intelligent sensing and artificial intelligence technologies, deep learning-based fault diagnosis methods, such as deep belief networks, deep autoencoders, convolutional neural networks, and recurrent neural networks, have achieved good results. However, deep learning-based fault diagnosis methods require a large number of labeled samples for training, and the training set and test set must maintain the same distribution. The distribution differences caused by cross-operating conditions will lead to a decrease in the diagnostic performance of such methods. Fault diagnosis methods based on transfer learning, such as maximum mean difference, adversarial learning, and parameter fine-tuning, can align the marginal distribution or conditional distribution of different operating conditions, thereby improving diagnostic performance. However, in actual operating conditions, equipment under new operating conditions is less likely to fail, and the fault type and location must be determined immediately when a fault occurs. Fault diagnosis methods based on transfer learning cannot complete fault diagnosis for new operating conditions without data.

[0003] To address the problem of fault diagnosis for new operating conditions with no data, researchers have drawn on ideas from adversarial learning and transfer learning to propose a data-coupled fault diagnosis model based on adversarial learning, dynamic models, and model transfer mechanisms. This model solves the problem of fault diagnosis for new operating conditions with no data. The main idea is to use dynamic models to simulate data from different operating conditions to supplement missing data, use adversarial learning and real data to correct the simulation data, and use simulation data from different operating conditions to establish a working condition conversion model to convert real data. This results in generated data for new operating conditions with a distribution similar to the real data for the new operating condition, enabling data-free fault diagnosis for the new operating condition.

[0004] In their paper "A Hybrid Generalization Network for Intelligent Fault Diagnosis of Rotating Machinery Under Unseen Working Conditions" (IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, 2021, doi:10.1109 / TIM.2021.3088489), Hunter et al. proposed a method for rotating machinery fault diagnosis under unknown working conditions and no data based on metric learning and adversarial learning. The steps of this method are: first, vibration data of the planetary gearbox under the old working conditions are collected; second, noise is added to the vibration data of the old working conditions to form enhanced data, and the domain generalization network is trained using domain classification loss, domain adversarial loss, and triplet loss to enhance the performance of the domain generalization network; finally, the trained domain generalization classifier is used to perform fault diagnosis on real data of unknown working conditions, with good diagnostic performance.

[0005] The shortcomings of the above-mentioned rotating machinery fault diagnosis method are: due to the influence of the problem of no data for new working conditions, the existing data distribution alignment transfer learning method cannot be implemented when there is no data in the target domain. Although the domain generalization network can realize fault diagnosis for new working conditions without data, the model diagnostic performance depends on the distribution of source domain data, and does not consider the working condition information and data distribution of the new working conditions. The generalization performance is insufficient, resulting in poor diagnostic ability. Summary of the Invention

[0006] The purpose of the present invention is to solve the problem that although the domain generalization network can realize fault diagnosis of new working conditions without data, the model diagnostic performance depends on the distribution of source domain data, does not consider the working condition information and data distribution of the new working condition, and the generalization performance is insufficient, resulting in poor diagnostic ability. Instead, a variable working condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling is provided.

[0007] In order to solve the deficiencies of the above-mentioned prior art, the present invention provides the following technical solutions:

[0008] The variable-operating-condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling is unique in that it includes the following steps:

[0009] Step 1: Obtain the old working condition supervision simulation dataset D SS1 and new working condition supervision simulation data set D SS2 ;

[0010] Establish a nonlinear dynamic model of the rolling bearing; input the parameters of the old working condition and the new working condition into the nonlinear dynamic model of the rolling bearing for simulation, wherein the parameters include labels, and one or more of bearing type, speed, load, health status, and defect size; convert the corresponding vibration time domain signal obtained into vibration frequency domain signal as samples, and organize the samples into P old working condition supervision simulation data sets according to the labels and P new working condition supervision simulation data sets D SS1 The i-th sample and its label in m SS1 D SS1 Number of samples, D SS2 The i-th sample and its label in m SS2 D SS2 Sample number; P ≥ 2;

[0011] Step 2: D SS1 、D SS2 Form a set of simulation data sets, and perform steps 2.1 to 2.4 on each of the P sets of simulation data sets in sequence;

[0012] Step 2.1, build a fault sequence conversion model;

[0013] Set the frequency domain threshold T to filter the D obtained in step 1 SS1 、D SS2 , respectively get the filtered old working condition supervision simulation data set Filtering the new working condition supervision simulation data set Fitting using a quadratic polynomial function and Obtain the fault sequence conversion model Ψ(·); for The i-th sample and its label in , for The i-th sample and its label in ;

[0014] Step 2.2: Obtain new working condition supervision conversion data D SG2 :

[0015] Get I and D under the old working condition SS1 The real vibration frequency domain signals with the same parameters are used as samples to form the real data set of old working condition supervision D SR1 Input the fault sequence conversion model Ψ(·) obtained in step 2.1 to obtain the new working condition supervision conversion data I=m SS1 ; DSR1 The i-th sample and its label in m SR1 D SR1 Number of samples, D SC2 The i-th sample and its label in m SC2 D SC2 Number of samples;

[0016] Step 2.3: Build and train the old working condition generative adversarial network;

[0017] Construct the old working condition generation adversarial network and use D SR1 and D SS1 , update the parameters of the discriminator D1 and generator G1 of the old working condition generative adversarial network in turn, and finally obtain the generator G1″ and discriminator D1″;

[0018] Step 2.4: Build and train a new working condition generative adversarial network;

[0019] According to the generator G1″ and the discriminator D1″, the parameters of the generator G2 and the discriminator D2 of the new working condition generation adversarial network are set to construct the new working condition generation adversarial network; using D SC2 and D SS2 , update the parameters of the discriminator D2 and generator G2 of the new working condition generative adversarial network in turn, and finally obtain the generator G2″ and discriminator D2″;

[0020] Step 3: Obtain new working condition supervision data set D SG2 ;

[0021] Replace the three Ds in step 1 SS2 Input the corresponding generator G1″ obtained in step 2.3 to obtain the corresponding new working condition supervision and no fine-tuning generated dataset The three D' SG1 The new working condition supervision is merged to generate the total data set D without fine-tuning SG1 ; Replace the three D in step 1 SS2 Input the corresponding generator G2″ obtained in step 2.4 to obtain the corresponding new working condition supervision fine-tuning generation dataset The three D' SG2 The new working condition supervision fine-tuning is combined to generate the total data set D SG2 ;

[0022] D′ SG1 The i-th sample and its label in m SG1 D′ SG1 Number of samples, D′ SG2 The i-th sample and its label in m SG2 D′SG2 Number of samples;

[0023] Step 4: Obtain fault diagnosis results;

[0024] Collect the vibration time domain signal of the rolling bearing under the new working condition and convert it into the vibration frequency domain signal as a sample to form an unsupervised real data set for the new working condition Calculate D SR2 Each sample in D SG2 The distance between each sample in And sort; D SR2 The label of each sample in D SG2 The I samples with the smallest distance from it have the highest proportion of labels, and we can get D SR2 The prediction result of the label to which each sample belongs; D SR2 The jth sample in m SR2 D SR2 Number of samples, D SG2 The i-th sample in .

[0025] Furthermore, the step 2.3 is specifically as follows:

[0026] Step 2.3.1: Construct the old working condition generation adversarial network and set the number of iterations of the old working condition generation adversarial network to K. 1 , set the number of discriminator iterations of the old working condition generation adversarial network to Set the number of iterations of the generator of the old working condition generative adversarial network to Set the number of iterations k of the old working condition generative adversarial network 1 The initial value k 1 =1, the number of discriminator iterations of the old working condition generative adversarial network Initial value of The number of generator iterations of the old working condition generative adversarial network Initial value of

[0027] Step 2.3.2, D SR1 and D SS1 Input them into the discriminator D1 of the old working condition generation adversarial network for forward propagation to obtain the adversarial loss and the weight W of the discriminator D1 D1 and bias b D1 Make updates;

[0028] Step 2.3.3, judgment Is it true? If so, get the discriminator D′1 after stage training and execute step 2.3.4; otherwise, let And return to step 2.3.2;

[0029] Step 2.3.4, D SS1 Input the generator G1 of the old working condition generative adversarial network to perform forward propagation to obtain the old working condition supervision fake data set D SR1 and D SF1 They are respectively input into the discriminator D1 for forward propagation to obtain the adversarial loss, and the weight W of the generator G1 is respectively G1 and bias b G1 Make updates; D SF1 The i-th sample and its label in m SF1 D SF1 Number of samples;

[0030] Step 2.3.5, judgment Is it true? If so, get the generator G′1 after stage training and execute step 2.3.6; otherwise, let And return to step 2.3.4;

[0031] Step 2.3.6, determine k 1 =K 1 Is it true? If so, get the generator G1″ and the discriminator D1″ and execute step 2.4; otherwise, let k 1 =k 1 +1, return to step 2.3.3.

[0032] Furthermore, in step 2.3.2, the update is specifically as follows:

[0033]

[0034]

[0035] Among them, θ D Represents the model parameters of the discriminator, L D represents the discriminator loss, λ D represents the discriminator learning rate, m is the sample size, and D is the discriminator.

[0036] Furthermore, in step 2.3.4, the adversarial training loss of the old working condition generative adversarial network is:

[0037]

[0038] Where minmax V(D,G) includes max V(D,G) and min V(D,G). Max V(D,G) is to optimize the discriminator D to distinguish the generated sample G(Z), and min V(D,G) is to optimize the generator G so that the discriminator D cannot distinguish the generated sample G(Z). The expectation of the real data distribution, log D(X) is the result of the discriminator’s discrimination of the real data, is the expectation of the noise data distribution, and log D(G(Z)) is the discriminator's judgment result on the generated samples.

[0039] Furthermore, in step 2.3.4, the update is specifically as follows:

[0040]

[0041]

[0042] Among them, θ G Represents the generator model parameters, L G represents the generator loss, λ G Represents the generator learning rate.

[0043] Furthermore, the step 2.4 is specifically as follows:

[0044] Step 2.4.1. Construct a new working condition GAN, set the generator parameters of the new working condition GAN to be consistent with the generator parameters of the old working condition GAN, G2 = G1″, and set the discriminator parameters of the new working condition GAN to be consistent with the discriminator parameters of the old working condition GAN, D2 = D1″;

[0045] Step 2.4.2: Set the number of iterations of the new working condition generative adversarial network to K 2 , set the number of discriminator iterations of the new working condition generation adversarial network to Set the number of iterations of the generator of the new working condition generative adversarial network to Set the number of iterations k of the new working condition generative adversarial network 2 The initial value k 2 =1, the number of discriminator iterations of the new working condition generative adversarial network Initial value The number of iterations of the generator of the new working condition generative adversarial network Initial value

[0046] Step 2.4.3, D SC2 and D SS2 Input them into the discriminator D2 of the new working condition generation adversarial network for forward propagation to obtain the adversarial loss and the weight W of the discriminator D2 D2 and bias b D2 Make updates;

[0047] Step 2.4.4, judgment Is it true? If so, get the discriminator D′2 after stage training and execute step 2.4.5; otherwise, let And return to step 2.4.3;

[0048] Step 2.4.5, D SS2 Input the generator G2 of the new working condition generative adversarial network, propagate the previous term, and obtain the fake data set of the old working condition supervision D SR1 and D SF2 They are input into the discriminator D2 for forward propagation to obtain the adversarial loss and the weight W of the generator G2. G2 and bias b G2 Make updates; D SF2 The i-th sample and its label in m SF2 D SF2 Number of samples;

[0049] Step 2.4.6, judgment Is it true? If so, get the generator G′2 after stage training and execute step 2.4.7; otherwise, let And return to step 2.4.5;

[0050] Step 2.4.7, determine k 2 =K 2 Is it true? If so, get the generator G2″ and the discriminator D2″ and execute step 3; otherwise, let k 2 =k 2 +1, return to step 2.4.4.

[0051] Furthermore, in step 2.4.5, the adversarial training loss of the new working condition generative adversarial network is:

[0052]

[0053] Furthermore, in step 4, the distance is calculated as Euclidean distance:

[0054]

[0055] Where n is D SG2 Number of samples.

[0056] Furthermore, in step 2.3.1, the K 1 =1000, described

[0057] Furthermore, in step 1, the nonlinear dynamic model consists of four parts: an inner ring, a roller, an outer ring and a base; the simulation uses a lumped mass method to simplify the model.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] (1) The variable working condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling of the present invention first uses the old working condition dynamics simulation data and the old working condition real data to train a generative adversarial network based on the adversarial learning idea to correct the simulation data; secondly, uses the old working condition dynamics simulation data and the new working condition simulation data to train the working condition conversion model, and inputs the old working condition real data into the working condition conversion model to obtain the new working condition conversion data; then, uses the new working condition simulation data and the new working condition conversion data to fine-tune the generative adversarial network; finally, inputs the new working condition simulation data into the fine-tuned generative adversarial network to obtain the new working condition generated data, and trains the fault diagnosis model to diagnose the new working condition real data to obtain the diagnosis result; the present invention is suitable for fault diagnosis under new working condition without data, and through adversarial learning and correction of simulation data, the difference between simulation data and real data can be narrowed.

[0060] (2) The variable working condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling of the present invention proposes a fault sequence conversion model based on a quadratic polynomial function when establishing the working condition conversion relationship. It can effectively convert the real data of the old working condition into the converted data of the new working condition, which is used to fine-tune the parameters of the generative adversarial network and generate generated data similar to the real data of the new working condition.

[0061] (3) The zero-sample mechanical fault diagnosis method for variable working conditions driven by mechanism data coupling of the present invention has a strong diagnostic performance for fault diagnosis without data under new working conditions, which is different from the old working condition supervision real data set D SR1 , New working condition supervision simulation data set D SS2 , New working condition supervision conversion data set D SC2 , the total data set D generated without fine-tuning for new working condition supervision SG1 Compared with other diagnostic methods with training sets, the present invention performs well in fault diagnosis tasks with no data under new working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of the variable operating condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling of the present invention;

[0063] Figure 2 Schematic diagram of the simplified model of the lumped mass method in step 1.2 of an embodiment of the present invention;

[0064] Figure 3 This is a vibration time domain signal waveform diagram with label 0 under the old working condition in step 1.2 of the embodiment of the present invention;

[0065] Figure 4 This is a vibration time domain signal waveform diagram labeled 1 under the old working condition in step 1.2 of the embodiment of the present invention;

[0066] Figure 5This is a vibration time domain signal waveform diagram labeled 2 under the old working condition in step 1.2 of the embodiment of the present invention;

[0067] Figure 6 This is a vibration time domain signal waveform diagram with label 0 under the new working condition in step 1.2 of the embodiment of the present invention;

[0068] Figure 7 This is a vibration time domain signal waveform diagram labeled 1 under the new working condition in step 1.2 of the embodiment of the present invention;

[0069] Figure 8 This is a vibration time domain signal waveform diagram labeled 2 under the new working condition in step 1.2 of the embodiment of the present invention;

[0070] Figure 9 This is a schematic diagram of the principle of constructing a fault sequence conversion model in step 2.1 of an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The present invention will be further described below with reference to the accompanying drawings and exemplary embodiments.

[0072] The experimental conditions of this embodiment are as follows:

[0073] The central processing unit is Intel(R)Core(TM)i5-7500 3.40GHZ, the memory is 16G, the operating system is WINDOWS 10, the simulation software is MATLAB R2020b, and the experimental software is python3.7.

[0074] Reference Figure 1 The variable working condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling includes the following steps:

[0075] Step 1: Obtain the old working condition supervision simulation dataset D SS1 and new working condition supervision simulation data set D SS2 ;

[0076] Step 1.1: Establish a nonlinear dynamic model of the rolling bearing. The nonlinear dynamic model consists of four parts: inner ring, roller, outer ring, and base.

[0077] Step 1.2, reference Figure 2 , the parameters of the old working condition and the new working condition are input into the nonlinear dynamic model of the rolling bearing and the lumped mass method is used to simplify the model for simulation. The parameters include labels, and one or more of bearing type, speed, load, health status, and defect size; the corresponding vibration time domain signal obtained is converted into vibration frequency domain signal as a sample, and the samples are respectively organized into P old working condition supervision simulation data sets according to the labels. and P new working condition supervision simulation data sets As shown in Table 1; D SS1 The i-th sample and its label in m SS1 D SS1 Number of samples, m SS1 =50; D SS2 The i-th sample and its label in m SS2 D SS2 Number of samples, m SS2 =50;P=3;

[0078] Table 1

[0079]

[0080] Figures 3 to 8 The waveform diagram of the vibration time domain signal of the rolling bearing under two working conditions and three labels, where the vertical axis represents the amplitude of the vibration time domain signal and the horizontal axis represents time; Figure 3 The waveform of the vibration time domain signal with label 0 under the old working condition is shown in the figure. Figure 4 It represents the vibration time domain signal waveform with label 1 under the old working condition. Figure 5 The waveform diagram of the vibration time domain signal with label 2 under the old working condition is shown below. Figure 6 It represents the vibration time domain signal waveform with label 0 under the new working condition. Figure 7 It represents the vibration time domain signal waveform with label 1 under the new working condition. Figure 8 The waveform diagram of the vibration time domain signal with label 2 under the new working condition is shown;

[0081] Step 2: D SS1 、D SS2 Form a set of simulation data sets, there are three sets of simulation data sets, namely the simulation data set with label 0, the simulation data set with label 1, and the simulation data set with label 2; perform steps 2.1 to 2.4 on the three sets of simulation data sets respectively;

[0082] Step 2.1, build a fault sequence conversion model;

[0083] Step 2.1.1, reference Figure 9 , set the frequency domain threshold T to 95% quantile, filter the old working condition supervision simulation data set D obtained in step 1 SS1 , New working condition supervision simulation data set D SS2 , respectively get the filtered old working condition supervision simulation data set Filtering the new working condition supervision simulation data set for The i-th sample and its label in , for The i-th sample and its label in ;

[0084] Step 2.1.2: Use the quadratic polynomial function as the fitting function to fit and Obtain the fault sequence conversion model Ψ(·);

[0085] Step 2.2: Obtain new working condition supervision conversion data D SC2 :

[0086] Get I and D under the old working condition SS1 The real vibration time domain signal with the same parameters is converted into a vibration frequency domain signal as a sample to form the real data set of old working condition supervision D SR1 Input the fault sequence conversion model Ψ(·) obtained in step 2.1.2 to obtain the new working condition supervision conversion data I = 50; D SR1 The i-th sample and its label in m SR1 D SR1 Number of samples, D SC2 The i-th sample and its label in m SC2 D SC2 Number of samples;

[0087] Step 2.3: Build and train the old working condition generative adversarial network:

[0088] Step 2.3.1: Construct the old working condition generation adversarial network and set the number of iterations of the old working condition generation adversarial network to K. 1 =1000, set the number of discriminator iterations of the old working condition generative adversarial network to Set the number of iterations k of the old working condition generative adversarial network 1 The initial value k 1 =1, the number of discriminator iterations of the old working condition generative adversarial network Initial value of The number of generator iterations of the old working condition generative adversarial network Initial value of

[0089] Step 2.3.2: The old working condition supervision real data set D SR1 and old working condition supervision simulation data set D SS1 Input them into the discriminator D1 of the old working condition generation adversarial network for forward propagation to obtain the adversarial loss and the weight W of the discriminator D1 D1 and bias b D1 Make updates;

[0090] The update process is as follows:

[0091]

[0092]

[0093] Among them, θ D Represents the model parameters of the discriminator, L D represents the discriminator loss, λ D represents the discriminator learning rate, m is the sample size, and D is the discriminator;

[0094] Step 2.3.3, judgment Is it true? If so, get the discriminator D′1 after stage training and execute step 2.3.4; otherwise, let And return to step 2.3.2;

[0095] Step 2.3.4: The old working condition supervision simulation data set D SS1 Input the generator G1 of the old working condition generative adversarial network to perform forward propagation to obtain the old working condition supervision fake data set The old working condition supervision real data set D RR1 and the old working condition supervision fake data set D SF1 They are respectively input into the discriminator D1 for forward propagation to obtain the adversarial loss, and the weight W of the generator G1 is respectively G1 and bias b G1 Make updates; D SF1 The i-th sample and its label in m SF1 D SF1 Number of samples;

[0096] The adversarial training loss of the old working condition generative adversarial network is:

[0097]

[0098] Where minmax V(D,G) includes max V(D,G) and min V(D,G). Max V(D,G) is to optimize the discriminator D to distinguish the generated sample G(Z), and min V(D,G) is to optimize the generator G so that the discriminator D cannot distinguish the generated sample G(Z). The expectation of the real data distribution, log D(X) is the result of the discriminator’s discrimination of the real data, is the expectation of the noise data distribution, log D(G(Z)) is the discriminant result of the generated sample;

[0099] The updates are specifically:

[0100]

[0101]

[0102] Among them, θ G Represents the generator model parameters, L G represents the generator loss, λ G Represents the generator learning rate;

[0103] Step 2.3.5, judgment Is it true? If so, get the generator G′1 after stage training and execute step 2.3.6; otherwise, let And return to step 2.3.4;

[0104] Step 2.3.6, determine k 1 =K 1 Is it true? If so, get the fully trained generator G1″ and discriminator D1″ and execute step 2.4; otherwise, let k 1 =k 1 +1, return to step 2.3.3;

[0105] Step 2.4: Build and train a new working condition generative adversarial network;

[0106] Step 2.4.1. Construct a new working condition GAN, set the generator parameters of the new working condition GAN to be consistent with the generator parameters of the old working condition GAN, G2 = G1″, and set the discriminator parameters of the new working condition GAN to be consistent with the discriminator parameters of the old working condition GAN, D2 = D1″;

[0107] Step 2.4.2: Set the number of iterations of the new working condition generative adversarial network to K 2 , set the number of discriminator iterations of the new working condition generation adversarial network to Set the number of iterations k of the new working condition generative adversarial network 2 The initial value k 2 =1, the number of discriminator iterations of the new working condition generative adversarial network Initial value The number of iterations of the generator of the new working condition generative adversarial network Initial value

[0108] Step 2.4.3: Convert the new working condition supervision data set D SC2 and new working condition supervision simulation data set D SS2 Input them into the discriminator D2 of the new working condition generation adversarial network for forward propagation to obtain the adversarial loss and the weight W of the discriminator D2 D2 and bias b D2 Make updates;

[0109] Step 2.4.4, judgment Is it true? If so, get the discriminator D′2 after stage training and execute step 2.4.5; otherwise, let And return to step 2.4.3;

[0110] Step 2.4.5: The old working condition supervision simulation data set D SS2 Input the generator G2 of the new working condition generative adversarial network, propagate the previous term, and obtain the fake data set of the old working condition supervision The old working condition supervision real data set D SR1 and the old working condition supervision fake data set D SF2 They are input into the discriminator D2 for forward propagation to obtain the adversarial loss and the weight W of the generator G2. G2 and bias b G2 Make updates; D SF2 The i-th sample and its label in m SF2 D SF2 Number of samples;

[0111] The adversarial training loss of the new working condition generation adversarial network is:

[0112]

[0113] Step 2.4.6, judgment Is it true? If so, get the generator G′2 after stage training and execute step 2.4.7; otherwise, let And return to step 2.4.5;

[0114] Step 2.4.7, determine k 2 =K 2 Is it true? If so, get the fully trained generator G2″ and discriminator D2″ and execute step 3; otherwise, let k 2 =k 2 +1, return to step 2.4.4;

[0115] Step 3: Obtain new working condition supervision data set D SG2 ;

[0116] The three new working condition supervision simulation data sets D in step 1.2 SS2 Input the corresponding generator G1″ obtained in step 2.3.5 to obtain the corresponding new working condition supervision and no fine-tuning generated dataset Generate dataset D′ of three new working conditions without fine-tuning SG1 The new working condition supervision is merged to generate the total data set D without fine-tuning SG1 ;

[0117] The three new working condition supervision simulation data sets D in step 1.2 SS2 Input the corresponding generator G2″ obtained in step 2.4.7 to obtain the corresponding new working condition supervision fine-tuning generation dataset The three new working conditions are fine-tuned to generate the dataset D′ SG2 The new working condition supervision fine-tuning is combined to generate the total data set D SG2 ; D′ SG1 The i-th sample and its label in m SG1 D′ SG1 Number of samples, D′ SG2 The i-th sample and its label in m SG2 D′ SG2 Number of samples;

[0118] Step 4: Obtain fault diagnosis results;

[0119] Collect the vibration time domain signal of the rolling bearing under the new working condition and convert it into the vibration frequency domain signal as a sample to form an unsupervised real data set for the new working condition Calculate D SR2 Each sample in D SG2 The distance between each sample in And sort; D SR2 The label of each sample in D SG2 The I samples with the smallest distance from it have the highest proportion of labels, and we can get D SR2 The prediction result of the label to which each sample belongs; D SR2 The jth sample in m SR2 D SR2 Number of samples, D SG2 The i-th sample in .

[0120] The distance is calculated as Euclidean distance:

[0121]

[0122] Where n is D SG2 Number of samples.

[0123] Analysis of experimental results:

[0124] To prove the effectiveness of the embodiment of the present invention, the three old working condition supervision real data sets D SR1 , the three new working condition supervision simulation data sets D obtained in step 1 SS2 , the three new working condition supervision transformation data sets D obtained in step 2.2 SC2The three new working conditions obtained in step 3 are supervised without fine-tuning to generate the total data set D SG1 , the new working condition supervision fine-tuning obtained in step 3 generates the total data set D SG2 As the training set, the nearest neighbor algorithm is used for training to obtain the corresponding fault diagnosis model. Using each fault diagnosis model, databases 1, 2, 3, 4, and 5 are used as target domains for diagnosis, and the obtained diagnostic accuracy is shown in Table 2:

[0125] Table 2

[0126]

[0127] The diagnostic accuracy shown is obtained using the classification accuracy Acc, which is expressed as:

[0128]

[0129] Where, is the label predicted for the j-th target domain test sample, y j represents the actual label of the j-th target domain test sample.

[0130] As can be seen from Table 2, the total data set D generated by fine-tuning the new working condition supervision of the embodiment of the present invention is SG2 The classification accuracy of the training set is basically around 99%, which is significantly higher than the other four.

[0131] In summary, the present invention can generate generated data similar to real data, thereby improving the accuracy of rolling bearing fault diagnosis.

[0132] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. For ordinary professional and technical personnel in this field, the specific technical solutions recorded in the above embodiments can be modified, or some of the technical features therein can be replaced by equivalents. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions protected by the present invention.

Claims

1. A zero-sample mechanical fault diagnosis method for variable working conditions driven by mechanism data coupling, characterized by: The steps include: Step 1: Obtain the old working condition supervision simulation dataset D SS1 and new working condition supervision simulation data set D SS2 ; Establish a nonlinear dynamic model of the rolling bearing; input the parameters of the old working condition and the new working condition into the nonlinear dynamic model of the rolling bearing for simulation, wherein the parameters include labels, and one or more of bearing type, speed, load, health status, and defect size; convert the corresponding vibration time domain signal obtained into vibration frequency domain signal as samples, and organize the samples into P old working condition supervision simulation data sets according to the labels and P new working condition supervision simulation data sets D SS1 The i-th sample and its label in m SS1 D SS1 Number of samples, D SS2 The i-th sample and its label in m SS2 D SS2 Sample number; P ≥ 2; Step 2: D SS1 、D SS2 Form a set of simulation data sets, and perform steps 2.1 to 2.4 on each of the P sets of simulation data sets in sequence; Step 2.1, build a fault sequence conversion model; Set the frequency domain threshold T to filter the D obtained in step 1 SS1 、D SS2 , respectively get the filtered old working condition supervision simulation data set Filtering the new working condition supervision simulation data set Fitting using a quadratic polynomial function and Obtain the fault sequence conversion model Ψ(·); for The i-th sample and its label in , for The i-th sample and its label in ; Step 2.2: Obtain new working condition supervision conversion data D SC2 : Get I and D under the old working condition SS1 The real vibration frequency domain signals with the same parameters are used as samples to form the real data set of old working condition supervision D SR1 Input the fault sequence conversion model Ψ(·) obtained in step 2.1 to obtain the new working condition supervision conversion data I=m SS1 ; D SR1 The i-th sample and its label in m SR1 D SR1 Number of samples, D SC2 The i-th sample and its label in m SC2 D SC2 Number of samples; Step 2.3: Build and train the old working condition generative adversarial network; Construct the old working condition generation adversarial network and use D SR1 and D SS1 , update the parameters of the discriminator D1 and generator G1 of the old working condition generative adversarial network in turn, and finally obtain the generator G1″ and discriminator D1″; Step 2.4: Build and train a new working condition generative adversarial network; According to the generator G1″ and the discriminator D1″, the parameters of the generator G2 and the discriminator D2 of the new working condition generation adversarial network are set to construct the new working condition generation adversarial network; using d SC2 and D SS2 , update the parameters of the discriminator D2 and generator G2 of the new working condition generative adversarial network in turn, and finally obtain the generator G2″ and discriminator D2″; Step 3: Obtain new working condition supervision data set D SG2 ; Replace the three Ds in step 1 SS2 Input the corresponding generator G1″ obtained in step 2.3 to obtain the corresponding new working condition supervision and no fine-tuning generated dataset The three D' SG1 The new working condition supervision is merged to generate the total data set D without fine-tuning SG1 ; Replace the three D in step 1 SS2 Input the corresponding generator G2″ obtained in step 2.4 to obtain the corresponding new working condition supervision fine-tuning generation dataset The three D' SG2 The new working condition supervision fine-tuning is combined to generate the total data set D SG2 ; D′ SG1 The i-th sample and its label in m SG1 D′ SG1 Number of samples, D′ SG2 The i-th sample and its label in m SG2 D′ SG2 Number of samples; Step 4: Obtain fault diagnosis results; Collect the vibration time domain signal of the rolling bearing under the new working condition and convert it into the vibration frequency domain signal as a sample to form an unsupervised real data set for the new working condition Calculate D SR2 Each sample in D SG2 The distance between each sample in And sort; D SR2 The label of each sample in D SG2 The I samples with the smallest distance from it have the highest proportion of labels, and we can get D SR2 The prediction result of the label to which each sample belongs; D SR2 The jth sample in m SR2 D SR2 Number of samples, D SG2 The i-th sample in .

2. The variable working condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling according to claim 1 is characterized in that: The step 2.3 is specifically as follows: Step 2.3.1: Construct the old working condition generation adversarial network and set the number of iterations of the old working condition generation adversarial network to K. 1 , set the number of discriminator iterations of the old working condition generation adversarial network to Set the number of iterations of the generator of the old working condition generative adversarial network to Set the number of iterations k of the old working condition generative adversarial network 1 The initial value k 1 =1, the number of discriminator iterations of the old working condition generative adversarial network Initial value of The number of generator iterations of the old working condition generative adversarial network Initial value of Step 2.3.2, D SR1 and D SS1 Input them into the discriminator D1 of the old working condition generation adversarial network for forward propagation to obtain the adversarial loss and the weight W of the discriminator D1 D1 and bias b D1 Make updates; Step 2.3.3, judgment Is it true? If so, get the discriminator D′1 after stage training and execute step 2.3.4; otherwise, let And return to step 2.3.2; Step 2.3.4, D SS1 Input the generator G1 of the old working condition generative adversarial network to perform forward propagation to obtain the old working condition supervision fake data set D SR1 and D SF1 They are respectively input into the discriminator D1 for forward propagation to obtain the adversarial loss, and the weight W of the generator G1 is respectively G1 and bias b G1 Make updates; D SF1 The i-th sample and its label in m SF1 D SF1 Number of samples; Step 2.3.5, judgment Is it true? If so, get the generator G′1 after stage training and execute step 2.3.6; otherwise, let And return to step 2.3.4; Step 2.3.6, determine k 1 =K 1 Is it true? If so, get the generator G1″ and the discriminator D1″ and execute step 2.4; otherwise, let k 1 =k 1 +1, return to step 2.3.

3.

3. The variable operating condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling according to claim 2 is characterized by: In step 2.3.2, the update is specifically as follows: Among them, θ D Represents the model parameters of the discriminator, L D represents the discriminator loss, λ D represents the discriminator learning rate, m is the sample size, and D is the discriminator.

4. The variable operating condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling according to claim 3 is characterized in that: In step 2.3.4, the adversarial training loss of the old working condition generative adversarial network is: Where minmax V(D,G) includes max V(D,G) and min V(D,G). Max V(D,G) is to optimize the discriminator D to distinguish the generated sample G(Z), and minV(D,G) is to optimize the generator G so that the discriminator D cannot distinguish the generated sample G(Z). The expectation of the real data distribution, log D(X) is the result of the discriminator’s discrimination of the real data, is the expectation of the noise data distribution, and log D(G(Z)) is the discriminator's judgment result on the generated samples.

5. The variable operating condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling according to claim 4 is characterized by: In step 2.3.4, the update is specifically as follows: Among them, θ G Represents the generator model parameters, L G represents the generator loss, λ G Represents the generator learning rate.

6. The variable operating condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling according to any one of claims 2 to 5, characterized in that: The step 2.4 is specifically as follows: Step 2.4.

1. Construct a new working condition GAN, set the generator parameters of the new working condition GAN to be consistent with the generator parameters of the old working condition GAN, G2 = G1″, and set the discriminator parameters of the new working condition GAN to be consistent with the discriminator parameters of the old working condition GAN, D2 = D1″; Step 2.4.2: Set the number of iterations of the new working condition generative adversarial network to K 2 , set the number of discriminator iterations of the new working condition generation adversarial network to Set the number of iterations of the generator of the new working condition generative adversarial network to Set the number of iterations k of the new working condition generative adversarial network 2 The initial value k 2 =1, the number of discriminator iterations of the new working condition generative adversarial network Initial value The number of iterations of the generator of the new working condition generative adversarial network Initial value Step 2.4.3, D SC2 and D SS2 Input them into the discriminator D2 of the new working condition generation adversarial network for forward propagation to obtain the adversarial loss and the weight W of the discriminator D2 D2 and bias b D2 Make updates; Step 2.4.4, judgment Is it true? If so, get the discriminator D′2 after stage training and execute step 2.4.5; otherwise, let And return to step 2.4.3; Step 2.4.5, D SS2 Input the generator G2 of the new working condition generative adversarial network, propagate the previous term, and obtain the fake data set of the old working condition supervision D SR1 and D SF2 They are input into the discriminator D2 for forward propagation to obtain the adversarial loss and the weight W of the generator G2. G2 and bias b G2 Make updates; D SF2 The i-th sample and its label in m SF2 D SF2 Number of samples; Step 2.4.6, judgment Is it true? If so, get the generator G′2 after stage training and execute step 2.4.7; otherwise, let And return to step 2.4.5; Step 2.4.7, determine k 2 =K 2 Is it true? If so, get the generator G2″ and the discriminator D2″ and execute step 3; otherwise, let k 2 =k 2 +1, return to step 2.4.

4.

7. The variable operating condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling according to claim 6 is characterized by: In step 2.4.5, the adversarial training loss of the new working condition generative adversarial network is:

8. The variable operating condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling according to claim 7 is characterized by: In step 4, the distance is calculated as Euclidean distance: Where n is D SG2 Number of samples.

9. The variable operating condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling according to any one of claims 2 to 5, characterized in that: In step 2.3.1, the K 1 =1000, described 10. The variable operating condition zero-sample mechanical fault diagnosis method driven by mechanism data coupling according to claim 8, characterized in that: In step 1, the nonlinear dynamic model consists of four parts: an inner ring, a roller, an outer ring, and a base; the simulation uses a lumped mass method to simplify the model.

Images