A method and system for predicting failure of a rolling bearing

Abstract

The application relates to the field of fault detection, and provides a bearing fault detection algorithm and system. An original vibration signal is decomposed through a local mean decomposition algorithm, and an effective PF component is screened out by using a Pearson correlation coefficient, and then signal reconstruction is carried out, irrelevant noise signal interference is removed, feature extraction of a bearing degradation process is realized based on a product form of a harmonic noise ratio and a root mean square value, the sensitivity of the harmonic noise ratio to periodic impact is fully utilized, and the problem that the root mean square value has low sensitivity to initial bearing degradation is avoided; a state space model of different bearing degradation processes is constructed based on a Paris model and a Foreman model, compared with a single model prediction model, the prediction precision of RUL is more effectively improved; regularization of a particle filtering resampling process is realized based on an Euclidean distance, the diversity of particles is effectively improved, the particle depletion phenomenon of basic particle filtering is avoided, and the filtering estimation precision of a nonlinear state is improved.

Description

Technical Field

[0001] This invention relates to the field of bearing failure prediction, and in particular to a method and system for predicting rolling bearing failures. Background Technology

[0002] my country attaches great importance to nuclear safety, clearly stating that "risk prevention" should be the core focus, and continuously improving the safe utilization level of nuclear facilities. my country's marine nuclear power plants generally employ corrective or preventative maintenance, resulting in a huge annual maintenance volume, and a large number of spare parts severely occupy the space required for other resources. Statistics show that maintenance costs account for 73% of the total lifecycle cost, with improper maintenance accounting for more than one-third of this cost, indicating a significant waste of human and material resources. According to statistics from the U.S. Electric Power Institute, implementing predictive maintenance in fields such as aerospace can reduce operating costs by more than 20%. Therefore, it is necessary to accelerate the application of predictive maintenance to reduce equipment failure rates, avoid unexpected downtime, and improve equipment lifespan; simultaneously, it can reduce maintenance costs and improve economic efficiency. Currently, the bottleneck of predictive maintenance is how to accurately predict the remaining service life of equipment, thereby allocating relevant resources based on its lifespan distribution. Fault prediction is closely related to multiple front-end factors, such as component aging mechanisms, sensing and measurement, characteristic parameter analysis, and prediction algorithms, with complex coupling relationships between them. Therefore, it is essential to comprehensively analyze and rationally optimize the fault prediction model.

[0003] To address these issues, scholars both domestically and internationally have conducted extensive research on fault prediction technologies. These research methods can be broadly categorized into three types. The first type utilizes multivariate statistical analysis, typically presenting RUL (Recovery Duration and Lifetime) prediction results as conditional failure probabilities based on observational data. A common characteristic of this approach is that it fits the observed data using probability theory and mathematical statistics without relying on any physical mechanisms to form an RUL prediction model. However, this method requires assumptions about lifetime distribution, which often deviates significantly from reality. The second type employs machine learning algorithms, essentially a pattern regression analysis. With the rapid development of artificial intelligence and big data technologies, it has become possible to learn component degradation patterns from historical data using machine learning and deep learning without building complex physical models. However, compared to other methods, its "black box" nature makes its computational results unreliable. Furthermore, its complete reliance on data for modeling, coupled with the difficulty of obtaining degradation data in practice, severely limits its application in RUL prediction. The third type combines physical mechanisms to establish mathematical models to describe the fault process and ultimately predict the operating state based on these models. Although it has some drawbacks, such as high model complexity and difficulty in analyzing the failure mechanism of some complex devices, once the physical mechanism model is established, its analysis results have the highest accuracy among all methods.

[0004] Currently, physical process models are widely used for material performance degradation and crack propagation. The Paris-Erdogan (PE) model is one of the most widely used physical models in the field of mechanical structural materials, mainly used to describe the crack propagation process. Later, Melgar et al. applied this model to the field of fault prediction for aircraft structural materials. In China, Wang Jin et al. transformed the PE model into an empirical model for fault prediction; Lu Guanghui used the inner ring of a bearing as the research object to predict the fatigue wear of this part. Physical models and formulas themselves do not have predictive and recursive capabilities, so this invention combines a Bayesian framework, starting from statistics and probability models, to study fault prediction technology based on regularized particle filtering and physical mechanisms. Due to its superior performance in solving nonlinear and non-Gaussian system behaviors, particle filtering algorithms have been widely applied to various fault prediction scenarios. An et al. established a mathematical model to track the degradation state of a machine, while unknown parameters were optimized using the PF method. To further reduce the variance of RUL estimation and load calculation, Haque et al. proposed an auxiliary particle filter to predict the remaining lifetime of insulated gate bipolar transistors. Michael Pecht et al. used a particle filter algorithm to predict faults in circuit systems.

[0005] Based on the above challenges and opportunities, this invention takes rolling bearings as the research object. The reason is that the operating load of most rotating machinery in nuclear power plants is transmitted to the base through rolling bearings. Therefore, in order to avoid sudden bearing failure and equipment failure, which would affect the safe and reliable operation of nuclear power plants, it is necessary to effectively predict their remaining service life. Finally, the laboratory has an experimental platform for studying rolling bearings, which can be adapted and improved on the basis of the previous work to meet the needs of fault prediction. It is worth emphasizing that the research work in this invention has both basic scientific research value and practical application prospects. From the scientific perspective, this invention relates to the multidisciplinary field of fault mechanism analysis and artificial intelligence technology, and is based on the study of the formation mechanism of typical rolling bearing fault modes and the fault prediction model of regularized particle filtering. From the perspective of technical application, the rolling bearing fault prediction technology based on local mean decomposition and regularized particle filtering has the following characteristics and advantages: (1) A method combining LMD and Pearson correlation coefficient is proposed to extract degradation features based on the characteristics of rolling bearing fault signals. The extracted mean square harmonic noise ratio (MSHNR) index can well reflect the degradation trend of the bearing. Compared with traditional indicators such as effective value, it has higher sensitivity to each degradation stage of the bearing. (2) Based on Euclidean distance, the regularization of the particle filter resampling process is realized, and the influence of the particle update time scale on the filter estimation is considered, which effectively improves the diversity of particles and improves the estimation accuracy of nonlinear states. (3) The rolling bearing life prediction method based on different degradation stages is more consistent with the bearing degradation process. Compared with the single model prediction method, it can improve the prediction accuracy of RUL. It is hoped that through the research of this invention, a breakthrough in fault prediction technology can be achieved. The research results can not only provide new ideas for improving the safety and economy of nuclear power systems and other complex systems, but also promote the establishment and development of related theoretical systems. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for predicting rolling bearing failures, thereby improving the accuracy of bearing failure prediction.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A method for predicting rolling bearing failures, the prediction method comprising:

[0009] Step 1: Acquire the raw vibration signal;

[0010] Step 2: Preprocess the original vibration signal to obtain the preprocessed vibration signal. ;

[0011] Step 3: Based on the preprocessed vibration signal All local extreme points of the vibration signal were calculated. ;

[0012] Step 4: Based on the local extreme points Find the average value of all adjacent local extrema. and envelope estimate ;

[0013] Step 5: Average the values ​​of all adjacent local extreme points. and envelope estimate Connecting them, we get the local mean function. and envelope estimation function ;

[0014] Step 6: From the pre-processed vibration signal Lieutenant General Local Mean Function Separate it to obtain the function Then, the envelope estimation function is used. Demodulation is performed to obtain the function. ;

[0015] Step 7: The function Repeat steps 2-6 as a new function until... It is a pure frequency modulation function;

[0016] Step 8: Multiply the envelope functions obtained in the iterative processes of Steps 2-7 to obtain the envelope signal, i.e.: ;

[0017] Step 9: Convert the pure frequency modulation signal Envelope function and envelope signal Multiplying them together yields the first component of the product function: ;

[0018] Step 10: From the original vibration signal Separate from it to obtain a new signal ,Will Repeat steps 2-8 using the original data, looping. Next, until Until it becomes a monotonic function or the number of extreme points is less than 3;

[0019] Step 11: Original vibration signal After the iterative process of steps 2-10 above, it is decomposed into a series of PF components and a residual component, namely: ;

[0020] Step 12: Calculate the original vibration signal for each PF component. Based on the correlation degree, select the PF component with a correlation degree greater than or equal to the set threshold that can fully interpret the vibration characteristics of the signal, and then reconstruct the signal based on the selected PF component;

[0021] Step 13: Based on the reconstructed signal, calculate the root mean square (RMS) and harmonic noise ratio (HNR) of the reconstructed signal, and extract features of the reconstructed signal based on the product of RMS and HNR, MSHNR, to generate the MSHNR degradation curve. Use the instantaneous rate of change at each point of the degradation curve to divide the rolling bearing into the stable degradation stage and the rapid degradation stage.

[0022] Step 14: In the stable degradation stage of the bearing, construct the state space model of the stable degradation stage based on the Paris model; in the rapid degradation stage, construct the state space model of the rapid degradation stage based on the Foreman model.

[0023] Step 15: Design the basic parameters of the particle filter algorithm; the basic parameters include: number of particles N, noise variance Q, initial particle variance P, and step size t. n Resampling threshold N tv And the spatial dimension n;

[0024] Step 16: Initialize the particle set in the regularized particle filter, starting from the prior distribution. The initial particle set is obtained by sampling N particles. The initial weights of each particle are ;

[0025] Step 17: Input the test dataset into the regularized particle filter, from the importance density function. The particle set at time k is obtained by mid-sampling. ,Right now , The particle representing time k, represent k The observation values ​​at time k are used to calculate the weights of each particle at time k. After obtaining the weights of each particle, the weights are normalized to obtain... ;

[0026] Step 18: Calculate the effective number of particles based on the normalized particle weights. If the number of effective particles is greater than the threshold value N tv Execute step 17; if it is less than the threshold value, execute step 19.

[0027] Step 19: Obtain the system observation sequence and the sample estimator observation matrix. For the actual observations of the system, take a sequence of length L. For the observed values ​​calculated from each sample value in the sample set, take each sample in the time series within the interval [j=k-L+1, k] and calculate the corresponding observed values ​​according to the observation equation.

[0028] Calculate the Euclidean distance between the system observation sequence and the estimated observation value for each particle. ;

[0029] The Euclidean distance Normalizing to the (0,1) interval yields ;

[0030] based on Reassign sample weights to ;

[0031] from Resampling is performed to obtain a new set of particles after regularization. ;in, The samples obtained in step 16 k Momentary particle set, This involves adjusting the normalization weights of each particle in step 16 through step 18.

[0032] Step 20: From Calculate the filtered value for the current state estimate;

[0033] Step 21: Based on the filtered value of the current state estimate, the rolling bearing fault prediction result at the current moment is obtained by recursively using the state space model of the bearing stable degradation stage and the state space model of the rapid degradation stage;

[0034] Step 22: Determine if the maximum number of prediction steps has been reached at the current time. t f If yes, the prediction ends and the algorithm exits; otherwise... K = K +1, return to step 15 to continue the prediction.

[0035] Optional, the average of adjacent local extrema. m i The expression is as follows:

[0036]

[0037] in, n i Represents a local extreme point. n i+1 Indicates and n i Adjacent local extreme points.

[0038] Optionally, the envelope estimate The expression is as follows:

[0039] .

[0040] Optional, function sum function The expression is as follows:

[0041]

[0042] .

[0043] Optionally, the set threshold is 0.7.

[0044] Optionally, the state-space model for the bearing's stable degradation stage can be expressed as follows:

[0045] ,in, The length of the crack. The number of load cycles, This represents the crack propagation rate. For material parameters, This represents the stress intensity factor amplitude.

[0046] Optionally, the expression for the state-space model of the rapid degradation stage is as follows:

[0047] ,in, and These are stress ratio and fracture toughness, respectively. The length of the crack. For the number of load cycles, This represents the crack propagation rate. For material parameters, This represents the stress intensity factor amplitude.

[0048] Based on the above-described method of this invention, this invention further provides a rolling bearing fault prediction system, the system comprising:

[0049] The original vibration signal acquisition module is used to acquire the original vibration signal;

[0050] The preprocessing module is used to preprocess the original vibration signal to obtain the preprocessed vibration signal. ;

[0051] The local extremum point calculation module is used to calculate the local extremum point based on the preprocessed vibration signal. All local extreme points of the vibration signal were calculated. ;

[0052] The local extreme point average and envelope estimation calculation module is used to calculate the local extreme point average and envelope estimation values ​​based on the local extreme points. Find the average value of all adjacent local extrema. and envelope estimate ;

[0053] The module for determining the local mean function and envelope estimation function is used to average the values ​​of all adjacent local extrema. and envelope estimate Connecting them, we get the local mean function. and envelope estimation function ;

[0054] The separation and demodulation module is used to extract the preprocessed vibration signal. Lieutenant General Local Mean Function Separate it to obtain the function Then, the envelope estimation function is used. Demodulation is performed to obtain the function. ;

[0055] The loop module is used to loop the function. The computation of the preprocessing module—separation and demodulation module—is repeated as a new function until... It is a pure frequency modulation function;

[0056] The envelope signal calculation module is used to multiply the envelope functions obtained from the iterative process of the preprocessing module and the loop module to obtain the envelope signal, that is: ;

[0057] The product function component calculation module is used to calculate the pure frequency modulated signal. Envelope function and envelope signal Multiplying them together yields the first component of the product function: ;

[0058] Separation module, used to separate From the original vibration signal Separate from it to obtain a new signal ,Will As a step in the original data repetition preprocessing module - envelope signal calculation module, the loop... Next, until Until it becomes a monotonic function or the number of extreme points is less than 3;

[0059] Decomposition module for raw vibration signals After the iterative process of the preprocessing module-separation module described above, it is decomposed into a series of PF components and a residual component, namely: ;

[0060] The correlation calculation module is used to calculate the correlation between each PF component and the original vibration signal. Based on the correlation degree, select the PF component with a correlation degree greater than or equal to the set threshold that can fully interpret the vibration characteristics of the signal, and then reconstruct the signal based on the selected PF component;

[0061] The segmentation module is used to calculate the root mean square (RMS) and harmonic noise ratio (HNR) of the reconstructed signal based on the reconstructed signal, and to extract features of the reconstructed signal based on the product of RMS and HNR, MSHNR, to generate an MSHNR degradation curve. The instantaneous rate of change at each point of the degradation curve is used to divide the rolling bearing into a stable degradation stage and a rapid degradation stage.

[0062] The model building module is used to build a state-space model for the bearing stable degradation stage based on the Paris model, and to build a state-space model for the rapid degradation stage based on the Foreman model.

[0063] The parameter setting module is used to set the basic parameters of the particle filter algorithm; the basic parameters include: number of particles N, noise variance Q, initial particle variance P, and step size t. n Resampling threshold N tv And the spatial dimension n;

[0064] The initialization module is used to initialize the particle set in the regularized particle filter, starting from the prior distribution. The initial particle set is obtained by sampling N particles. The initial weights of each particle are ;

[0065] The normalization module takes the test dataset as input and feeds it into the regularized particle filter, from the importance density function. The particle set at time k is obtained by mid-sampling. ,Right now , The particle representing time k, represent k The observation values ​​at time k are used to calculate the weights of each particle at time k. After obtaining the weights of each particle, the weights are normalized to obtain... ;

[0066] The first judgment module is used to calculate the effective number of particles based on the normalized particle weights. If the number of effective particles is greater than the threshold value N tv Execute the normalization module; if the value is less than the threshold, execute the next module.

[0067] The new particle set determination module is used to obtain the system observation sequence and the sample estimate observation matrix. For the actual observations of the system, a sequence of length L is taken. For the observed values ​​calculated from each sample value in the sample set, take each sample in the time series within the interval [j=k-L+1, k] and calculate the corresponding observed values ​​according to the observation equation.

[0068] Calculate the Euclidean distance between the system observation sequence and the estimated observation value for each particle. ;

[0069] The Euclidean distance Normalizing to the (0,1) interval yields ;

[0070] based on Reassign sample weights to ;

[0071] from Resampling is performed to obtain a new set of particles after regularization. ;in, Let k be the set of particles. These are the normalized weights of each particle;

[0072] The filter value calculation module is used by... Calculate the filtered value for the current state estimate;

[0073] The fault prediction module is used to recursively obtain the rolling bearing fault prediction result at the current moment based on the filtered value of the current state estimate using the state space model of the bearing stable degradation stage and the state space model of the rapid degradation stage.

[0074] The second judgment module is used to determine whether the maximum number of prediction steps has been reached at the current moment. t f If yes, the prediction ends and the algorithm exits; otherwise... K = K +1, return to the parameter setting module to continue prediction.

[0075] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0076] The method described in this invention decomposes the original vibration signal using a local mean decomposition algorithm and filters out the effective PF components using the Pearson correlation coefficient, thereby reconstructing the signal, removing interference from irrelevant noise signals, and improving the signal-to-noise ratio (SNR). It extracts features of the bearing degradation process based on the product of harmonic noise ratio (HNR) and root mean square (RMS) value, fully utilizing the sensitivity of HNR to periodic impacts to multiply the RMS value, highlighting the degradation fluctuation trend, and avoiding the problem of low sensitivity of RMS value to early bearing degradation. It constructs state-space models of different bearing degradation processes based on the Paris and Foreman models, which better reflect the bearing degradation process compared to single-model prediction models, effectively improving the prediction accuracy of RUL. It regularizes the particle filter resampling process based on Euclidean distance, considering the impact of particle update time scale on filter estimation, effectively improving particle diversity, avoiding particle depletion in basic particle filtering, and improving the filter estimation accuracy for nonlinear states. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 This is a flowchart of the rolling bearing failure prediction method according to an embodiment of the present invention. Detailed Implementation

[0079] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] The purpose of this invention is to provide a method and system for predicting rolling bearing failures, thereby improving the accuracy of bearing failure prediction.

[0081] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0082] Figure 1 This is a flowchart of the rolling bearing fault prediction method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0083] Step 1: The raw data obtained from the acceleration sensor and process parameter sensors such as differential pressure, temperature and flow rate on the rolling bearing of the circulating water pump are stored in the calculation through the data acquisition board. At the same time, the raw parameters are preprocessed and data feature engineering is performed to remove noise interference and features that are irrelevant to fault prediction. The data are also normalized and standardized to avoid the influence of dimensions on subsequent fault prediction.

[0084] Step 2: Based on the preprocessed vibration signal The local mean points of the vibration signal were calculated. At the same time, based on local extreme points Find the average value of all adjacent local extrema. and envelope estimate ,Right now , .

[0085] Step 3: Based on cubic Hermite interpolation, average all adjacent local extrema. and envelope estimate Connecting them, we get the local mean function. and envelope estimation function .

[0086] Step 4: From the original vibration signal Lieutenant General Local Mean Function Separate it to obtain the function Then, the envelope estimation function is used. Demodulation was performed to obtain .

[0087]

[0088]

[0089] Step 5: Repeat steps 2-4 as a new function until... Since it is a pure frequency modulation function, meaning its envelope estimation function is a constant function 1, in practical applications, a variable can be set. When satisfied The iteration terminates when the time is reached.

[0090] Step 6: Multiply the envelope functions obtained in the iteration process of Steps 2-5 to obtain the envelope signal, that is:

[0091]

[0092] Then the pure frequency modulation signal Envelope function and envelope signal Multiplying them together yields the first component of the product function:

[0093]

[0094] Step 7: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] From the original vibration signal Separate from it to obtain a new signal ,Will Repeat steps 2-6 using the original data, looping. Next, until Until it becomes a monotonic function or the number of extreme points is less than 3.

[0095]

[0096] Step 8: Original vibration signal After the iterative process of steps 2-7 above, it is decomposed into a series of PF components and a residual component, namely:

[0097]

[0098] Step 9: Calculate the correlation between each PF component and the original vibration signal based on the Pearson correlation coefficient (PCC). Based on the degree of correlation, a power factor (PF) with PCC ≥ 0.7 that can fully interpret the signal vibration characteristics is selected, and then the signal is reconstructed based on the selected PF.

[0099] Step 10: Based on the reconstructed signal, calculate the root mean square (RMS) and harmonic to noise ratio (HNR). Then, based on the product of RMS and HNR, perform feature extraction of the reconstructed signal using MSHNR. Finally, based on the MSHNR degradation curve, utilize... Set an alarm threshold to determine if a bearing has transitioned from normal operation to a degradation phase (wherein, The mean and standard deviation of the MSHNR index (representing the bearing in its normal operating period) are used to divide the rolling bearing into stable degradation stage and rapid degradation stage by using the instantaneous rate of change at each point of the degradation curve.

[0100] Step 11: In the stable degradation stage of the bearing, a state-space model of its degradation process is constructed based on the Paris model. In the rapid degradation stage, a state-space model of its degradation process is constructed based on the Foreman model. Then, based on the selection of multi-stage state-space models, rolling bearing fault prediction is performed based on regularized particle filtering. The model construction for the stable degradation stage and the rapid degradation stage of the bearing is as follows:

[0101] (1) Stable degradation stage of rolling bearings—Paris model

[0102] The Paris model is: In the formula, The length of the crack. For the number of load cycles, This represents the crack propagation rate. For material parameters, This represents the stress intensity factor amplitude. When the load cycle interval of the rolling bearing... When the value is sufficiently small, approximating the above equation using the difference formula yields the following model for the bearing's stable degradation stage: , This represents the stress range at the bearing crack.

[0103] (2) Rapid degradation stage of rolling bearings—Foreman model

[0104] The Foreman model is: In the formula, and These are stress ratio and fracture toughness, respectively. When the load cycle interval of the rolling bearing... When the crack rate is sufficiently small, the differential term of the Forman model in the above equation can be approximated by a difference formula. In this case, the fatigue crack propagation rapid propagation model based on the Forman model is: .

[0105] Step 12: Set the basic parameters of the particle filter algorithm. The time-invariant parameters that need to be set for the regularized particle filter in this invention are shown in the table below.

[0106] Table 1 Basic parameters of the particle filter algorithm

[0107]

[0108] Step 13: Initialize the particle set in the regularized particle filter, starting from the prior distribution. The initial particle set is obtained by sampling N particles. Initial weights for each particle .

[0109] Step 14: Input the test dataset into the regularized particle filter, from the importance density function. (in the formula) The particle representing time k, The particle set at time k is obtained by sampling from the observations at time k. ,Right now Simultaneously calculate the weight of each particle at time k: In the formula Let be the weight of the i-th particle at time k. Let be the likelihood probability function of the system, which characterizes the state of the system as determined by . Transferred to The degree of similarity between the observed values ​​and the system's observation equation is determined by the system's observation equation (the system's observation equation is...). , representing the degree to which the observed value reflects the state of the system, in the formula It is the observation function. The decision is made based on the observed noise. The state transition function of the system can be derived from the system's state equation (the system's state equation is...). , represents the transition of the system state from the previous state to the next state, in the formula It is a state transition function. The weights are determined by process noise. After obtaining the weights of each particle, the weights are normalized. .

[0110] Step 15: Calculate the effective number of particles If it is greater than the threshold value N tv Proceed to step 14. If the value is less than the threshold, proceed to step 16.

[0111] Step 16: Regularized resampling based on Euclidean distance: Prepare the system observation sequence and the sample estimate observation matrix. For the actual observations of the system, take a length of... L sequence For the observed values ​​calculated from each sample value in the sample set, the time series is taken as [ j = k - L + 1, k For each sample within the interval, the corresponding observed value is calculated according to the observation equation:

[0112]

[0113] Euclidean distance is calculated based on the system observation sequence and the estimated observation value for each particle. Then Normalizing to the (0,1) interval yields .based on The reassigned sample weights are Then from Resampling is performed to obtain a new set of particles after regularization. ;in, Let k be the set of particles. The normalized weights of each particle.

[0114] Step 17: From Calculate the filtered value of the current state estimate, and then recursively obtain the rolling bearing fault prediction result at the current moment from the current state space model selected in step 11 and output it.

[0115] Step 18: Determine if the maximum prediction step t has been reached at the current time. f If so, the prediction ends and the algorithm exits; otherwise, K = K + 1, and the algorithm returns to step 12 to continue the prediction.

[0116] Step 19: The actual circulating water pump rolling bearing fault prediction results are obtained by using the local mean decomposition and regularized particle filtering obtained in steps 1-18. The relevant results can be used as a reference for maintenance and decision-making personnel to take relevant measures in a timely manner, which can improve the economy while ensuring safety.

[0117] Based on the above-described method of this invention, this invention further provides a rolling bearing fault prediction system, the system comprising:

[0118] The original vibration signal acquisition module is used to acquire the original vibration signal;

[0119] The preprocessing module is used to preprocess the original vibration signal to obtain the preprocessed vibration signal. ;

[0120] The local extremum point calculation module is used to calculate the local extremum point based on the preprocessed vibration signal. All local extreme points of the vibration signal were calculated. ;

[0121] The local extreme point average and envelope estimation calculation module is used to calculate the local extreme point average and envelope estimation values ​​based on the local extreme points. Find the average value of all adjacent local extrema. and envelope estimate ;

[0122] The module for determining the local mean function and envelope estimation function is used to average the values ​​of all adjacent local extrema. and envelope estimate Connecting them, we get the local mean function. and envelope estimation function ;

[0123] The separation and demodulation module is used to extract the preprocessed vibration signal. Lieutenant General Local Mean Function Separate it to obtain the function Then, the envelope estimation function is used. Demodulation is performed to obtain the function. ;

[0124] The loop module is used to loop the function. The computation of the preprocessing module—separation and demodulation module—is repeated as a new function until... It is a pure frequency modulation function;

[0125] The envelope signal calculation module is used to multiply the envelope functions obtained from the iterative process of the preprocessing module and the loop module to obtain the envelope signal, that is: ;

[0126] The product function component calculation module is used to calculate the pure frequency modulated signal. Envelope function and envelope signal Multiplying them together yields the first component of the product function: ;

[0127] Separation module, used to separate From the original vibration signal Separate from it to obtain a new signal ,Will As a step in the original data repetition preprocessing module - envelope signal calculation module, the loop... Next, until Until it becomes a monotonic function or the number of extreme points is less than 3;

[0128] Decomposition module for raw vibration signals After the iterative process of the preprocessing module-separation module described above, it is decomposed into a series of PF components and a residual component, namely: ;

[0129] The correlation calculation module is used to calculate the correlation between each PF component and the original vibration signal. Based on the correlation degree, select the PF component with a correlation degree greater than or equal to the set threshold that can fully interpret the vibration characteristics of the signal, and then reconstruct the signal based on the selected PF component;

[0130] The segmentation module is used to calculate the root mean square (RMS) and harmonic noise ratio (HNR) of the reconstructed signal based on the reconstructed signal, and to extract features of the reconstructed signal based on the product of RMS and HNR, MSHNR, to generate an MSHNR degradation curve. The instantaneous rate of change at each point of the degradation curve is used to divide the rolling bearing into a stable degradation stage and a rapid degradation stage.

[0131] The model building module is used to build a state-space model for the bearing stable degradation stage based on the Paris model, and to build a state-space model for the rapid degradation stage based on the Foreman model.

[0132] The parameter setting module is used to set the basic parameters of the particle filter algorithm; the basic parameters include: number of particles N, noise variance Q, initial particle variance P, and step size t. n Resampling threshold N tv And the spatial dimension n;

[0133] The initialization module is used to initialize the particle set in the regularized particle filter, starting from the prior distribution. The initial particle set is obtained by sampling N particles. The initial weights of each particle are ;

[0134] The normalization module takes the test dataset as input and feeds it into the regularized particle filter, from the importance density function. The particle set at time k is obtained by mid-sampling. ,Right now , The particle representing time k, represent k The observation values ​​at time k are used to calculate the weights of each particle at time k. After obtaining the weights of each particle, the weights are normalized to obtain... ;

[0135] The first judgment module is used to calculate the effective number of particles based on the normalized particle weights. If the number of effective particles is greater than the threshold value N tv Execute the normalization module; if the value is less than the threshold, execute the next module.

[0136] The new particle set determination module is used to obtain the system observation sequence and the sample estimate observation matrix. For the actual observations of the system, a sequence of length L is taken. For the observed values ​​calculated from each sample value in the sample set, take each sample in the time series within the interval [j=k-L+1, k] and calculate the corresponding observed values ​​according to the observation equation.

[0137] Calculate the Euclidean distance between the system observation sequence and the estimated observation value for each particle. ;

[0138] The Euclidean distance Normalizing to the (0,1) interval yields ;

[0139] based on Reassign sample weights to ;

[0140] from Resampling is performed to obtain a new set of particles after regularization. ;in, Let k be the set of particles. These are the normalized weights of each particle;

[0141] The filter value calculation module is used by... Calculate the filtered value for the current state estimate;

[0142] The fault prediction module is used to recursively obtain the rolling bearing fault prediction result at the current moment based on the filtered value of the current state estimate using the state space model of the bearing stable degradation stage and the state space model of the rapid degradation stage.

[0143] The second judgment module is used to determine whether the maximum number of prediction steps has been reached at the current moment. t f If yes, the prediction ends and the algorithm exits; otherwise... K = K +1, return to the parameter setting module to continue prediction.

[0144] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0145] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A rolling bearing failure prediction method, characterized in that, The prediction method comprises: Step 1: obtaining an original vibration signal; Step 2: preprocessing the original vibration signal to obtain a preprocessed vibration signal ; Step 3: Based on the pre-processed vibration signal , all local extreme points of the vibration signal are calculated ; Step 4: Based on the local extrema points , the average of all adjacent local extrema points is calculated and the envelope estimate ; Step 5: Average all adjacent local extrema and envelope estimates are connected to obtain a local mean function and envelope estimate function ; Step 6: Local mean function is separated from the pre-processed vibration signal Step 7: The function is obtained by demodulating the function using the envelope estimation function ;​ Step 7: The function Repeat the calculation of Step 2 - Step 6 as a new function until is a pure frequency modulation function; Step 8: The envelope signal is obtained by multiplying the envelope functions obtained from the iteration process of Step 2-Step 7, i.e. ; Step 9: multiplying the pure frequency modulated signal envelope function and envelope signal together, resulting in a first product function component: ; Step 10: repeat the steps of Step 2-Step 8 with the new signal as the original data Step 11: repeat the steps of Step 2-Step 8 with the new signal as the original data Step 12: repeat the steps of Step 2-Step 8 with the new signal as the original data Step 13: repeat the steps of Step 2-Step 8 with the new signal as the original data Step 14: repeat the steps of Step 2-Step 8 with the new signal as the original data Step 15: repeat the steps of Step 2-Step 8 with the new signal as the original data Step 16: repeat the steps of Step 2-Step 8 with the new signal as the original data Step 11 : Original vibration signal Through the iteration process of the above steps 2-10, the original vibration signal is decomposed into a series of PF components and a residual component, i.e.: ; Step 12: Calculate the relationship between each PF component and the original vibration signal. Based on the correlation degree, select the PF component with a correlation degree greater than or equal to the set threshold that can fully interpret the vibration characteristics of the signal, and then reconstruct the signal based on the selected PF component; Step 13: on the basis of the reconstructed signal, root mean square RMS and harmonic noise ratio HNR of the reconstructed signal are calculated, feature extraction of the reconstructed signal is realized based on a product MSHNR of the RMS and the HNR, an MSHNR degradation curve is generated, and instantaneous change rates at each point of the degradation curve are used to realize division of a stable degradation stage and a rapid degradation stage of the rolling bearing; Step 14: in the stable degradation stage of the bearing, a state space model of the stable degradation stage of the bearing is constructed based on a Paris model, and in the rapid degradation stage, a state space model of the rapid degradation stage is constructed based on a Foreman model; Step 15: setting the basic parameters of the particle filter algorithm; the basic parameters include: the number of particles N, the noise variance Q, the initial particle variance P, the step size t n , the resampling threshold N tv , and the spatial dimension n; Step 16: Initialize the particle set in the regularized particle filter from the prior distribution by sampling N particles from the prior distribution to obtain an initial particle set with an initial weight of Step 17: input the test data set as input into the regularized particle filter, sample from the importance density function to obtain the particle set at time k , representing the particle at time k, representing k the observation value at time k, while calculating the weight of each particle at time k , after obtaining the weight of each particle, weight normalization is performed to obtain ;​ Step 18: Calculate the number of valid particles based on the normalized particle weights If the number of valid particles is greater than a threshold value N tv Step 17 is performed if the number of valid particles is greater than a threshold value N, and step 19 is performed if the number of valid particles is less than a threshold value. Step 19: Obtain the system observation sequence and the sample estimated quantity observation matrix, take the sequence with length L for the actual observation value of the system ; take the samples in the interval [j=k-L+1, k] for the observation value calculated by the sample value in the sample set, and calculate the corresponding observation value according to the observation equation; Euclidean distance of the sequence of observations of the computing system and the estimate of each particle observation ; normalizing the Euclidean distance to the interval (0, 1) gives ; based on reassigning sample weights, for ; Resampling from the regularized new particle set ; where, is the particle set at time k, is the normalized weight of each particle; Step 20: from computing a filtered value of the state estimate at the current time instant; Step 21: based on the filtered value of the state estimation at the current moment, the state space model of the stable degradation stage of the bearing and the state space model of the rapid degradation stage are recursively obtained to obtain a rolling bearing fault prediction result at the current moment; Step 22: Determine if the maximum number of prediction steps has been reached at the current time. t f If yes, the prediction ends and the algorithm exits; otherwise... K = K +1, return to step 15 to continue the prediction.

2. The rolling bearing failure prediction method according to claim 1, characterized in that, the average of the adjacent local extrema m i The expression is as follows: wherein n i denotes a local extremum point, n i+1 denotes a local extremum point, n i denotes a local extremum point.

3. The rolling bearing failure prediction method according to claim 2, characterized in that, The envelope estimate The expression for the envelope estimate is as follows: 。 4. The rolling bearing failure prediction method according to claim 1, characterized in that, function and function The expression of the function 。 5. The rolling bearing failure prediction method according to claim 1, characterized in that, The set threshold is 0.

7.

6. The rolling bearing failure prediction method according to claim 1, characterized in that, The expression of the state space model of the stable degradation stage of the bearing is as follows: wherein, is the crack length, is the number of load cycles, is the crack growth rate, is the material parameter, is the stress intensity factor amplitude.

7. The rolling bearing failure prediction method according to claim 1, characterized in that, The expression of the state space model of the rapid degradation stage is as follows: wherein, and are the stress ratio and the fracture toughness, respectively, is the crack length, is the number of load cycles, is the crack propagation rate, is a material parameter, is the stress intensity factor amplitude.

8. A rolling bearing fault prediction system characterized by, The system comprises: An original vibration signal acquisition module is configured to acquire an original vibration signal; a preprocessing module configured to preprocess the original vibration signal to obtain a preprocessed vibration signal ; A local extreme point calculation module is configured to calculate all local extreme points of the preprocessed vibration signal ;​ a local extremum point average and envelope estimate calculation module for calculating an average value and an envelope estimate value from all adjacent local extremum points ;​​ a local mean function and an envelope estimation function determination module, configured to determine a mean value of all adjacent local extreme points and an envelope estimation value connected to obtain a local mean function and an envelope estimation function ; a separation and demodulation module for separating and demodulating the pre-processed vibration signal a local mean function is separated to obtain a function an envelope estimation function is demodulated to obtain a function ; The loop module is used to loop the function. The computation of the preprocessing module—separation and demodulation module—is repeated as a new function until... It is a pure frequency modulation function; An envelope signal calculation module is configured to multiply the envelope functions obtained by the pre-processing module and the iteration process of the loop module to obtain an envelope signal, i.e. ; a product function component calculation module for calculating a pure frequency modulation signal an envelope function and an envelope signal multiplying, to obtain a first product function component: ; Separating module for separating from the original vibration signal a new signal , and taking the new signal as the original data to repeat the steps of the preprocessing module-envelope signal calculation module, and looping times until it is a monotonic function or the number of extreme points is less than 3. decomposition module for the original vibration signal Through the above iterative process of the preprocessing module-separation module, the original vibration signal is decomposed into a series of PF components and a residual component, i.e. ; A correlation degree calculation module is configured to calculate the correlation degree of each PF component with the original vibration signal , select the PF component with a correlation degree greater than or equal to a set threshold value which can completely interpret the vibration characteristics of the signal, and then reconstruct the signal based on the selected PF component. A division module is configured to, on the basis of the reconstructed signal, calculate root mean square RMS and harmonic noise ratio HNR of the reconstructed signal, realize feature extraction of the reconstructed signal based on a product MSHNR of the RMS and the HNR, generate an MSHNR degradation curve, and realize division of a stable degradation stage and a rapid degradation stage of the rolling bearing by using instantaneous change rates at each point of the degradation curve; A model construction module is configured to, in the stable degradation stage of the bearing, construct a state space model of the stable degradation stage of the bearing based on a Paris model, and in the rapid degradation stage, construct a state space model of the rapid degradation stage based on a Foreman model; The parameter setting module is used to set the basic parameters of the particle filter algorithm; the basic parameters include: number of particles N, noise variance Q, initial particle variance P, and step size t. n Resampling threshold N tv And the spatial dimension n; An initialization module is configured to initialize a particle set in the regularized particle filter, sample N particles from a prior distribution to obtain an initial particle set The initialization weight of each particle is ;​ A normalization module is configured to input a test data set into the regularized particle filter to sample a particle set at time k from an importance density function k ;​​​​​​​ The first judging module is used for calculating the effective particle number based on the normalized particle weight If the effective particle number is greater than a threshold value N tv , executing a normalization module, if less than the threshold value, executing the next module; A new particle set determination module is configured to obtain a system observation value sequence and a sample estimated value observation matrix, wherein for an actual observation value of the system, a sequence with a length of L is taken ; and for an observation value calculated by a sample value in the sample set, samples in a time sequence in an interval of [j=k-L+1, k] are taken, and a corresponding observation value is calculated according to an observation equation. Euclidean distance of the sequence of system-observation values and each particle's estimate of the observation value ; normalizing the Euclidean distance to the interval (0, 1) gives ; based on reassigning sample weights, for ; Resampling from the regularized new particle set ; where, is the particle set at time k, is the normalized weight of each particle; The filter value calculation module is configured to calculate a filter value of a current time state estimation by calculating a filter value of a current time state estimation; A fault prediction module is configured to, based on a filtered value of state estimation at the current moment, recursively obtain a rolling bearing fault prediction result at the current moment by using the state space model of the stable degradation stage of the bearing and the state space model of the rapid degradation stage. A second judging module is used for judging whether the current time reaches the maximum prediction step number t f If yes, the prediction ends, the algorithm is exited, otherwise K = K +1, the parameter setting module is returned to continue the prediction.

Images