Online prediction method for residual service life of bearing under condition of no service life label

Abstract

The invention relates to an online prediction method for the residual service life of a bearing under the condition of no service life label, and the method can obtain the probability prediction result of the residual service life (RUL) of the bearing in the operation stage of the bearing. The method comprises the following steps: firstly, constructing a generative adversarial neural network considering cognitive uncertainty by utilizing health vibration signals at the initial stage of bearing operation, outputting multiple groups of bearing health reference vibration signals by utilizing a generator and adopting a Monte Carlo method, and calculating a Wasserstein distance between the vibration signal at the current moment and the reference vibration signals by utilizing a discriminator; the method comprises the following steps of: obtaining a bearing health index (Health Index, HI) of the bearing; secondly, identifying first prediction time (FPT) by utilizing a 3-sigma rule, carrying out degradation modeling by adopting a nonlinear Wiener process in a degradation stage, and carrying out parameter estimation under the condition of considering individual uncertainty, so as to realize RUL probability prediction in a bearing operation stage; finally, after a new HI monitoring result is obtained, on-line parameter estimation is carried out on the Wiener process, model parameters are updated, and the bearing RUL prediction precision is improved.

Description

(I) Technical field

[0001] The present invention belongs to the field of fault prognostics and health management (PHM), and specifically is an online prediction method for the remaining service life of a bearing without a life label. (II) Background technology

[0002] Complex industrial equipment that has been running for a long time often shows performance degradation over time. When the equipment state approaches the critical failure point, it may cause serious safety risks or economic losses. In order to improve the reliability and safety of the system, it is necessary to use the remaining useful life (RUL) prediction technology to estimate the failure time of the equipment in advance based on the degradation characterization quantity. By analyzing the current and historical operation data of the equipment, RUL prediction can infer when the health state reaches the critical value, providing a basis for formulating maintenance strategies and optimizing operation management, thereby reducing the risk of sudden failures and ensuring efficient and stable operation of the system. As a core component of mechanical equipment, the performance state of bearings directly affects the reliability and operation efficiency of the equipment. In the industrial field, the operation failure of bearings may cause equipment shutdown or even cause major safety accidents. Implementing online prediction of bearing RUL can not only improve the operation and maintenance efficiency of equipment, but also provide early warning of potential failures, avoid major economic losses and safety accidents, and has important engineering practice significance.

[0003] Existing bearing degradation modeling and life prediction technologies can generally be divided into two categories: labeled prediction and unlabeled prediction according to the data label situation. Labeled prediction relies on the complete historical data of the bearing, which includes clear failure time or life label. However, in actual industrial scenarios, customized small batch equipment often has unlabeled scenarios, and at this time, it is necessary to carry out RUL prediction under the condition of no life label. By building a prediction model based on historical performance data and operating status, it is possible to speculate on the future state of the bearing in an unlabeled scenario, thereby providing reliable support for the health management of the equipment and reducing safety hazards and economic losses. However, there are also many difficulties in carrying out online prediction of bearing RUL under the condition of no life label: 1) The health index (HI) of the bearing is difficult to construct through traditional deep learning methods, and the traditional construction method based on expert experience often cannot fully characterize its health status; 2) The RUL prediction method based on deep learning faces the dilemma of no training data in the absence of life labels, and it is difficult to apply directly.

[0004] In view of this, the present invention provides an online method for health state monitoring and remaining useful life prediction during the operation stage of a bearing. By utilizing the learning and representation ability of the generative adversarial network and the superior difference quantification ability of the Wasserstein distance, the health index (HI) of the bearing is constructed. Based on the generative adversarial network and the stochastic process, the epistemic uncertainty in the model is considered, enabling the probabilistic prediction of the remaining life of the bearing under the condition of no life labels, providing reliable life information support for the online operation and maintenance of the bearing. (III) Summary of the Invention

[0005] The present invention is an online prediction method for the remaining useful life of a bearing under the condition of no life labels, which conducts probabilistic prediction of the online remaining useful life (RUL) based on the generative adversarial neural network and the stochastic process. The process of the method of the present invention is as Figure 1 shown and specifically includes the following:

[0006] Step 1: Collect the vibration signals of the bearing, and use the window preprocessing technology to expand one frame of data collected by the sensor monitoring into multiple frames of data, and perform operations such as normalizing the data.

[0007] Step 2: Add a Dropout network structure to the Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP) for uncertainty quantification based on the Wasserstein distance. Design the parameter structures of the generator and the discriminator.

[0008] Step 3: Generate noise data from the normal distribution, conduct adversarial training of the generative network, and train the generator and the discriminator until the Wasserstein distance reaches the standard. Finally, the generator can augment a set of input healthy reference signals into multiple sets of enhanced healthy signals.

[0009] Step 4: Collect the current vibration monitoring signal x i of the bearing operation, and perform data preprocessing, including normalization and data slicing.

[0010] Step 5: Monitor the original vibration data of the bearing, and select the early normal operation stage as the initial detection stage; calculate the Wasserstein distance between the current moment of the bearing and the healthy vibration signal, which is defined as the health index HI, and obtain multiple sets of sample HI data.

[0011] Step 6: Use the 3-σ rule to identify the First Prediction time (FPT) of the bearing. Judge the HI i at the current t jWhether it is within the healthy range, it is defined that when five consecutive samples continuously exceed the 3σ range, it is determined that the bearing enters the degradation stage and FPT is identified.

[0012] Step 7: Considering the non-linear degradation characteristics of bearing degradation, use a univariate non-linear Wiener process with drift to model the bearing degradation and obtain the RUL distribution function of the bearing.

[0013] Step 8: Utilize multiple groups of sample HI data and adopt maximum likelihood estimation to estimate the unknown parameters of the Wiener process to obtain the complete RUL probability density function.

[0014] Step 9: Based on the obtained RUL distribution function, predict the expected value of RUL and its probability distribution.

[0015] Step 10: In the personalized online operation stage, according to the new bearing vibration index, input it into WGAN-GP to obtain the new augmented data result, and the Wiener process model parameters can be estimated and updated online, and the online updated RUL mean prediction and probability prediction results are output.

[0016] Among them, in Step 1, the method of normalization is:

[0017]

[0018] In the formula, N is the length of each frame of vibration signal.

[0019] In Step 2, the generator network G consists of 4 one-dimensional convolutional layers, 4 one-dimensional regularization layers, 1 Dropout layer and 1 fully connected layer. Its input is a 1×N random noise, and the output is a 1×N bearing vibration signal; the discriminator consists of 8 one-dimensional convolutional layers, 1 Dropout layer and 1 fully connected layer. Its input is a 1×N bearing vibration signal, and the output is a 1×N eigenvalue.

[0020] The mathematical expression of the Relu function in the generator Relu activation layer is:

[0021]

[0022] The mathematical expression of the Sigmoid function in the generator Sigmoid activation layer is:

[0023]

[0024] The mathematical expression of the LeakyRelu function in the discriminator LeakyRelu activation layer is:

[0025]

[0026] In step 3, noise data z is generated from a normal distribution and input into the generator G i to generate fake sample data which is mixed with real training sample data and then input into the discriminator; according to the output of the discriminator, its loss function is calculated and the Adam optimizer is used for parameter update; the generated fake samples are input into the discriminator again, the loss function of the generator is calculated according to its output, and the Adam optimizer is also used to optimize the generator parameters. The above process is iterated until the Wasserstein distance between the fake samples and the real samples output by the discriminator meets the preset threshold. Finally, for the trained generator G, when random noise is input, the purpose of generating any number of augmented vibration signals in the healthy state of the bearing is achieved.

[0027] The Wasserstein distance between the fake data samples and the real data samples can be calculated as follows:

[0028]

[0029] The threshold ρ can generally be taken as:

[0030] ρ ≤ 0.2

[0031] In step 6, the average value μ and the standard deviation σ of HI within the healthy range are calculated, and the bearing HI interval in the healthy state is established as [μ - 3σ, μ + 3σ]. The formula for judging that the bearing enters the degradation stage is expressed as:

[0032] |HI j+i , HI j+i+1 ,..., HI j+i+l - μ j | > 3σ (5)

[0033] where i = 1, 2, 3,..., I max , and at this time t i is the FPT.

[0034] In step 8, the nonlinear drift Wiener process can be expressed as:

[0035]

[0036] where μ(t; θ) represents a nonlinear function with the independent variable t, μ is the drift parameter, θ represents the parameter vector, with θ = (a, b), where the parameter a is a random parameter characterizing individual differences, the parameter b is a fixed parameter characterizing the common characteristics of the device, σ is the diffusion parameter, and B(·) is a standard Brownian motion process. A power function is selected to characterize the degradation process of the bearing HI, that is, μ(t; θ) = abt b-1 .

[0037] The converted first-passage-time PDF is:

[0038]

[0039] Wherein, ω is the failure threshold.

[0040] When the drift coefficient is a power function (μ(t;θ) = abt b-1 ), the first-passage-time probability density function is:

[0041]

[0042] Wherein, θ = (a, b).

[0043] When t k The corresponding device degradation state at the moment is x k = X(t k ), and the PDF of the remaining life l k is:

[0044]

[0045] Wherein, w k = w - x k .

[0046] In step 8, considering the randomness of the parameters, the parameter a in θ can be expressed as a random parameter subject to a normal distribution Then the first-passage-time PDF and the PDF of the remaining life at time t k are respectively:

[0047]

[0048] The unknown parameter is a vector The maximum likelihood estimation method is used for parameter estimation. Assume that N degradation curves are obtained, and the observed data on the nth degradation curve are measured at moments respectively, X represents the set of all degradation data. The profile likelihood function of the maximum likelihood estimation of σ B , b with respect to μ a and is:

[0049]

[0050] Wherein: N represents the sample size,

[0051] The maximum value is obtained to get σ BAfter obtaining the estimated values of a and b, the parameter μ to be estimated can be obtained. a and The estimation results are as follows:

[0052]

[0053] In step 9, after obtaining the estimated value of the vector and substituting it back into formulas (15) and (16), according to the probability density function, the probability distribution function prediction and mean value of RUL can be obtained.

[0054] In step 10, the bearing vibration index monitored during the online operation stage is input into WGAN-GP to obtain new augmented data, thereby online estimating the unknown parameter and updating the mean prediction and probability prediction results of RUL.

[0055] The present invention is an online prediction method for the remaining service life of a bearing under the condition of no life label, and its advantages and effects are as follows:

[0056] 1. The present invention provides a method for predicting the remaining service life under the condition of no life label data, avoiding the additional cost caused by collecting life labels.

[0057] 2. Compared with traditional physical simulation, the present invention uses WGAN to construct bearing health data. By training data to generate vibration signals of bearing health states, it avoids complex physical simulation modeling and more quickly and conveniently realizes the generation of high-fidelity signals, providing a sufficient amount of samples for subsequent random process model parameter estimation.

[0058] 3. The present invention adopts the method based on WGAN and takes uncertainty into account at the same time. It uses a loss function with practical physical significance to realize the online evaluation of the bearing health state, and the Drop-out network layer to realize the quantification of the uncertainty of the bearing health state. Generally, it realizes a neural network model that can identify the bearing health state only by training with health data, providing a reliable basis for subsequent random process parameter estimation.

[0059] 4. The present invention adopts the method based on Wiener random process to carry out bearing RUL prediction, and can use the Wiener process to realize online personalized RUL prediction. In this process, new health state prediction samples can be generated to continuously update the random process model parameters, continuously improve the life prediction accuracy, and provide reliable and high-precision life prediction services. (IV) BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is the flow chart of the method described in the present invention;

[0061] Figure 2 is the probability distribution of real sample data and generated samplesFigure 1 Consistency check

[0062] Figure 3 For the change trend of HI of the tested bearing

[0063] Figure 4 For the predicted results of the average RUL at all times of the bearing

[0064] Figure 5 For the predicted results of the RUL of the bearing at all times (t = 73) (V) Specific implementation manners

[0065] In order to explain the specific process of the present invention in more detail, the present invention will be described in more detail below in combination with the accompanying drawings and case verification.

[0066] The present invention uses the XJTU-SY rolling bearing accelerated life test data set for detailed description and verification. This data set adjusts the radial force through a hydraulic loading system and the rotation speed of the rotating shaft through a rotation speed controller to conduct bearing accelerated life tests under different load conditions combinations, and can collect and save the vibration signal data in real time during the entire test process. In the test, the data acquisition frequency is set to 25.6 kHz, the sampling time for each acquisition is 1.28 s, and the time interval between adjacent two acquisitions is 1 min. The length of each frame of vibration data collected in the test is 32768, and each frame of data is regarded as a sample number. The relative method is used in the test to define the failure threshold of the tested bearing, that is, when the maximum amplitude of the bearing in the horizontal or vertical direction exceeds ten times the maximum amplitude in the normal operation stage, the tested bearing is regarded as failed and the test is immediately stopped.

[0067] In the case, the tested bearing numbered Bearing1_3 under the condition of working condition 1 in the data set is selected, and there is no life label data, which meets the life prediction condition without life label in the present invention. The vibration signal is selected as the degradation feature.

[0068] S1: Data preprocessing: Slice the vibration data with a length of 32768 for each frame into 16 frames of vibration data with a length of 2048, and normalize all the data.

[0069] S2: Set the number of hidden layer nodes of the generator and discriminator of the WGAN according to the size of the vibration data, and set the model hyperparameters of the WGAN-GP. Conduct adversarial training of the generation network, train the generator and the discriminator until the Wasserstein distance reaches the standard, and generate 10 groups of enhanced vibration signals as the bearing reference signals for case construction.

[0070] Specifically, the detailed parameters are shown in Table 1, Table 2 and Table 3 respectively.

[0071] Table 1 Generator network parameter settings

[0072]

[0073]

[0074] Table 2 Discriminator Network D Parameter Settings

[0075]

[0076] Table 2 Discriminator Network D Parameter Settings (Continued)

[0077]

[0078] Table 3 WGAN-GP Model Parameters

[0079]

[0080] Specifically, according to the above hyperparameter settings, network training is started. The probability distribution graphs of the real sample data and the generated sample data are plotted. As Figure 2 shown, the probability distribution graphs of the two almost overlap, and the generated vibration signals have data distribution characteristics very similar to those of the original vibration signals, passing the consistency check.

[0081] Furthermore, the generator G has learned the data characteristics of the bearing healthy vibration signals. At the same time, due to the addition of the Dropout layer in the network structure design, the generator has the ability to express cognitive uncertainty. Ten sets of enhanced vibration signals are generated using it as the bearing reference signals for case construction.

[0082] S3: Based on the WGAN discriminator, calculate the Wasserstein distance between the bearing at the current moment and the healthy vibration signal to obtain multiple sets of Bearing1_3 sample HI data.

[0083] Specifically, the discriminator D already has the ability to distinguish whether the input signal is a healthy vibration signal, that is, the ability to quantify the distance between the input signal and the healthy vibration signal. Use the discriminator to calculate the Wasserstein distance between the monitored vibration signal of the bearing at the current moment and the 10 sets of enhanced healthy signals to realize the online extraction of bearing HI, and obtain the HI change trend graph of the Bearing1_3 test bearing as Figure 3 shown.

[0084] S4: Use the 3-σ rule to identify the FPT points of the bearing. The calculation result is that the Bearing1_3 test bearing starts to degenerate from the 79th time point until it fails at the 158th time point. Take the 79th moment as the first moment of bearing degradation and start preparing for the RUL prediction of the bearing.

[0085] Specifically, first, a small part of the degradation data in the bearing degradation stage is collected for Wiener process parameter estimation. The first ten points are taken for the first parameter estimation. After more degradation data is obtained through subsequent online monitoring, the parameters are re - estimated every 1 moment. (Since this process consists of a series of explicit mathematical formulas, re - estimating the parameters and calculating the RUL at the current moment can be completed in an extremely short time, thus meeting the requirements of online real - time prediction).

[0086] Furthermore, the Wiener process parameter estimation results and RUL prediction results at different moments are shown in Table 4.

[0087] Table 4 Wiener process parameter estimation results

[0088]

[0089] At this time, the error from the true RUL of the bearing, which is 69, is relatively large. Therefore, more degradation data needs to be collected for parameter update. According to the above - mentioned parameter update strategy, the RUL prediction results at all moments are obtained as Figure 4 shown. Although the RUL prediction results in the early stage of degradation have relatively large errors, as the amount of data increases and the number of parameter updates increases, the predicted RUL value gradually approaches the true value. When t = 73, the predicted RUL value is 1.10 times. It can be considered that the prediction result is acceptable and can give an early warning of the bearing failure 8 monitoring moments before the actual failure occurs. As Figure 5 shown, when t = 73, estimating the RUL at all moments based on the parameters estimated at this time, it can be seen that the RUL of the bearing can be estimated more accurately at this time, and the Wiener process parameter estimation results at this time are already relatively accurate.

Claims

1. A method for online prediction of the remaining service life of a bearing without a life label, characterized in that: The following steps are involved: Step 1: Collect the bearing vibration signal, use window preprocessing technology to expand one frame of data collected by the sensor monitoring into multiple frames of data, and perform normalization and other operations on the data. Step 2: Add the Dropout network structure to the Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP) based on the Wsserstein distance to quantify uncertainty. Design the generator and discriminator parameter structures. Step 3: Generate noise data from a normal distribution, perform adversarial training on the generative network, train the generator and the discriminator until the Wasserstein distance reaches the standard, and finally the generator can augment a set of input healthy reference signals into multiple sets of enhanced healthy signals. Step 4: Collect the current vibration monitoring signal x of the bearing i , and perform data preprocessing, including normalization and data slicing. Step 5: Monitor the original vibration data of the bearing and select the early normal operation stage as the initial stage of detection; The Wasserstein distance between the current moment of the bearing and the healthy vibration signal is calculated and defined as the health index HI, and multiple groups of sample HI data are obtained. Step 6: Use the 3-σ rule to identify the FPT of the bearing. Determine the current t i HI at the moment j Whether it is within the healthy interval, it is defined that when 5 consecutive samples exceed the 3σ interval, the bearing is judged to have entered the degradation stage and the FPT is identified. Step 7: Considering the nonlinear degradation characteristics of bearing degradation, a univariate nonlinear Wiener process with drift is used to model the bearing degradation and obtain the RUL distribution function of the bearing. Step 8: Using multiple groups of sample HI data, the maximum likelihood estimation is used to estimate the unknown parameters of the Wiener process and obtain the complete RUL probability density function. Step 9: Based on the obtained distribution function of RUL, predict the expected value of RUL and its probability distribution. Step 10: In the personalized online operation stage, according to the new bearing vibration index, the new augmented data results are input into WGAN-GP, the Wiener process model parameters can be estimated and updated online, and the online updated RUL mean prediction and probability prediction results are output. Among them, in step 1, the normalization method is: Where N is the length of each frame of vibration signal. In step 2, the generator network G consists of 4 one-dimensional convolutional layers, 4 one-dimensional regularization layers, 1 Dropout layer and 1 fully connected layer, its input is 1×N random noise, and its output is 1×N bearing vibration signal; the identification consists of 8 one-dimensional convolutional layers, 1 Dropout layer and 1 fully connected layer, its input is 1×N bearing vibration signal, and its output is 1×N eigenvalue. The mathematical expression of the Relu function in the generator Relu activation layer is: The mathematical expression of the Sigmoid function in the generator Sigmoid activation layer is: The mathematical expression of the LeakyRelu function in the discriminator LeakyRelu activation layer is: In step 3, noise data z is generated from a normal distribution and fed into the generator G i Generate fake sample data The generated fake samples are mixed with real training sample data and then input into the discriminator. Based on the output of the discriminator, its loss function is calculated and the parameters are updated using the Adam optimizer. The discriminator is input again, and the loss function of the generator is calculated based on its output. The Adam optimizer is also used to optimize the generator parameters. The above process is iterated until the Wasserstein distance between the false sample output by the discriminator and the real sample meets the preset threshold. Finally, the trained generator G is input with random noise to achieve the purpose of generating any number of augmented vibration signals of the bearing in a healthy state. The Wasserstein distance between the fake data sample and the real data sample can be calculated as: The threshold ρ can generally be taken as: ρ≤0.2 In step 6, the mean value μ and standard deviation σ of HI within the healthy range are calculated, and the HI interval of the bearing in a healthy state is established as [μ-3σ,μ+3σ]. The formula for judging whether the bearing has entered the degradation stage is expressed as: |HI j+i ,HI j+i+1 ,...,HI j+i+l -μ j |>3σ (5) In the formula, i=1,2,3,...,I max , at this time t i is FPT. In step 8, the nonlinear drift Wiener process can be expressed as: In the formula, μ(t; θ) represents a nonlinear function with independent variable t, μ is a drift parameter, θ represents a parameter vector, and θ = (a, b), where parameter a is a random parameter that describes individual differences, parameter b is a fixed parameter that describes the common characteristics of the equipment, σ is a diffusion parameter, and B(·) is a standard Brownian motion process. The power function is used to describe the degradation process of bearing HI, that is, μ(t; θ) = abt b-1 . The converted first arrival time PDF is: In the formula, ω is the failure threshold. The drift coefficient is a power function (μ(t;θ) = abt b-1 ) is: Where θ = (a, b). When t k The device degradation state corresponding to the time is x k =X(t k ), remaining life l k The PDF is: In the formula, w k =wx k . In step 8, considering the randomness of the parameters, the parameter a in θ can be expressed as a random parameter that obeys the normal distribution Then the first arrival time PDF and t k The remaining life PDFs at the time are: Unknown parameters are vectors The maximum likelihood estimation method is used for parameter estimation. Assume that N degradation curves are obtained, and the observed data on the nth degradation curve are Measured at the moment, X represents the set of all degraded data. B , b about μ a and The profile likelihood function of the maximum likelihood estimate is: Where: N represents the number of samples, Get the maximum value to get σ B After the estimated values ​​of and b, the estimated parameter μ can be obtained a and The estimated results are: In step 9, we get the vector After obtaining the estimated value, we bring it back to formula (15) and (16), and then we can get the probability distribution function prediction and mean of RUL based on the probability density function. In step 10, the bearing vibration index monitored during the online operation phase is input into WGAN-GP to obtain new augmented data, thereby estimating the unknown parameters online. Update the mean prediction and probability prediction results of RUL.

Images