An intelligent monitoring method for bearing lubrication status in rotating machinery

Abstract

The present invention relates to the technical field of signal processing and condition monitoring, and specifically to a method for intelligently monitoring the lubrication condition of bearings in rotating machinery. The method comprises the following steps: building an online monitoring test bench, designing an online monitoring system for a host computer, and forming a method for collecting vibration and electrostatic signals; collecting vibration signals and electrostatic signals of bearings under different lubrication conditions in real time, constructing a vibration-electrostatic-lubrication condition correlation model, and combining a signal processing method with a deep learning algorithm to form a bearing lubrication condition monitoring method. The present invention not only enhances the real-time and comprehensiveness of lubrication condition assessment, but also improves equipment operation and maintenance efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing and state monitoring, and in particular to an intelligent monitoring method for the lubrication state of a bearing in a rotating machine. Background Art

[0002] Bearings are one of the most critical components in rotating machinery, and their lubrication status directly determines their operating efficiency, wear rate, and service life. Currently, monitoring of bearing lubrication status primarily relies on two technologies: offline oil analysis and online vibration, temperature, and other signal detection. Offline oil analysis uses regular sampling to detect indicators such as lubricating oil viscosity and contaminant content. Although it can assess oil degradation trends, its real-time performance is poor and it is difficult to reflect the lubrication condition of bearings during dynamic operation. Online vibration and temperature detection can identify abnormal vibrations or temperature changes caused by poor lubrication through spectral analysis and time-domain feature extraction. However, due to significant environmental noise interference and insufficient sensitivity to early lubrication degradation (such as oil film rupture and boundary lubrication), monitoring accuracy is limited.

[0003] Traditional lubrication monitoring methods suffer from limitations such as a single data source (disconnecting oil analysis from mechanical dynamic signals), difficulty identifying early signs of lubrication degradation (e.g., micro-scaling without a noticeable vibration signal), and low data fusion between different detection technologies. These factors can easily lead to missed faults or misjudgments, thus compromising the accuracy of maintenance decisions. The industry urgently needs a multimodal monitoring method that integrates oil characteristics (e.g., viscosity loss, contaminant buildup) with bearing operating conditions (e.g., vibration, shock, and electrostatic discharge signals). By integrating multi-sensor information, the method dynamically correlates lubrication degradation with bearing anomalies, enabling accurate assessment and early warning of bearing lubrication status.

[0004] Therefore, developing a high-precision, multi-information fusion bearing lubrication status monitoring method and combining it with intelligent algorithms for real-time analysis can not only optimize lubrication management and reduce unplanned downtime, but also extend bearing life and improve the reliability and economy of equipment operation. Summary of the Invention

[0005] Based on the above objectives, the present invention provides a method for intelligently monitoring the lubrication status of bearings in a rotating machine, comprising the following steps:

[0006] S1: Build an online monitoring test bench, design a host computer online monitoring system, and develop a method for collecting vibration and electrostatic dual signals. Arrange temperature and vibration sensors, electrostatic sensors, servo motors, and oil pumps in the test bench to simulate different bearing lubrication conditions and synchronously collect the vibration and electrostatic signals of the bearings under different lubrication conditions.

[0007] S2: The vibration characteristics of the bearing under different lubrication conditions, clarifying the correlation between the vibration characteristics and the bearing lubrication condition. The electrostatic characteristics of the bearing under different lubrication conditions, clarifying the correlation between the charge and the bearing lubrication condition;

[0008] S3: De-noise the collected vibration and electrostatic signals, analyze the vibration signal characteristics, including time domain, frequency domain, and time-frequency domain characteristic parameters, and extract key vibration characteristic indicators related to the bearing lubrication status. At the same time, based on the amplitude, spectrum, and statistical characteristics of the electrostatic signal, analyze how the electrostatic signal changes with the lubrication status.

[0009] S4: Build a vibration-static-lubrication status correlation model, combining deep learning algorithms with data fusion technology to classify bearing lubrication status and achieve intelligent identification of lubrication status. Based on the lubrication status monitoring results, it outputs bearing lubrication status assessment information and, combined with equipment operating conditions, provides lubrication maintenance recommendations to improve equipment operational stability and maintenance efficiency.

[0010] Further, the S1 specifically includes:

[0011] S11: Use serial communication to establish communication between LabVIEW software and the inverter, sensor, oil pump, and flow meter to ensure the stability and real-time performance of the data acquisition system; design the host computer control logic to achieve dynamic control of motor start and stop, speed regulation, and oil pump output;

[0012] S12: By adjusting the motor speed and controlling the oil supply of the oil pump, the operating states of the bearing under three lubrication conditions, namely, good lubrication, insufficient lubrication, and overlubrication, are simulated, and the original signal data of the corresponding sensors are recorded.

[0013] Furthermore, the simulated bearing lubrication state specifically includes:

[0014] The corresponding relationship between the grease filling ratio and the lubrication status is as follows: if the filling volume accounts for less than 20-30% of the bearing free space, it is insufficient lubrication; if the filling volume accounts for 30-50% of the free space, it is good lubrication; if the filling volume exceeds 50% of the free space, it is overlubrication.

[0015] The inverter frequency is set to a constant range of 10Hz-50Hz to correspond to different motor speeds. A flow meter monitors and records the grease output from the oil pump, measuring both total and instantaneous flow. The ratio of grease to the free space within the bearing, combined with temperature data from the sensor and the set motor speed, simulates three bearing lubrication states: well-lubricated, under-lubricated, and over-lubricated.

[0016] Furthermore, the S2 specifically includes:

[0017] S21: Lubrication conditions directly affect bearing wear, thereby changing the characteristics of the vibration signal. Grease quality, motor speed, and oil film stability play a significant role in the vibration characteristics. Oil film rupture or insufficient lubrication can cause significant changes in the vibration signal. Different lubrication conditions correspond to different vibration signal intensities. Lubrication conditions directly affect bearing friction, thereby changing the characteristics of the electrostatic signal. The electrostatic intensity is determined by the charge generated by the friction of the rotating pair and is also affected by different grease types and motor speed.

[0018] Furthermore, the S21 specifically includes:

[0019] The lubrication status of a bearing directly affects its vibration and electrostatic signal characteristics. Under normal lubrication, the oil film is uniform and stable, the vibration signal amplitude is low, and the electrostatic signal is smooth, with no obvious abnormal components. Underlubrication leads to increased friction, increased vibration signal impact, and enhanced high-frequency components in the spectrum. Simultaneously, the electrostatic signal amplitude increases, fluctuates frequently, and may experience transient voltage spikes. Overlubrication can hinder rolling element motion due to excess grease, resulting in abnormal low-frequency components in the vibration signal, a reduced electrostatic signal amplitude, and generally stable fluctuations. By monitoring the changing characteristics of the vibration and electrostatic signals, the lubrication status of the bearing can be effectively identified and failures prevented.

[0020] Furthermore, the S3 specifically includes:

[0021] S31: Wavelet denoising is used for vibration signals. Wavelet decomposition is used to perform multi-scale analysis of the signal and threshold processing is performed on the high-frequency noise coefficient to retain the main characteristic components. An appropriate wavelet basis function (such as Daubechies wavelet dbN) is selected and discrete wavelet transform (DWT) is performed on the vibration signal to decompose it into approximation signals and detail signals in different frequency bands:

[0022] ;

[0023] in, is the low-frequency approximation signal, It is a high frequency detail signal.

[0024] Use soft threshold or hard threshold method to remove high-frequency noise. The typical soft threshold function is:

[0025] ;

[0026] in, is the original signal A quantity, is the threshold parameter;

[0027] Then perform wavelet reconstruction to restore the denoised signal; analyze the denoised vibration signal, extract time domain features (such as root mean square value RMS, peak factor, etc.), frequency domain features (such as power spectral density PSD, frequency center, etc.) and time-frequency domain features (such as wavelet energy distribution, etc.), and extract key vibration characteristic indicators related to the lubrication status.

[0028] For electrostatic signals, the Wiener Filtering method is used to optimize the signal with the minimum mean square error criterion to reduce noise interference. By real signal and noise Composition, namely:

[0029] ;

[0030] Calculate the autocorrelation function of the signal and power spectral density , the signal-to-noise ratio of the signal is obtained based on the noise estimation.

[0031] The transfer function of the Wiener filter is defined as:

[0032] ;

[0033] in, and are the power spectral densities of the signal and noise, respectively.

[0034] The electrostatic signal is processed by the Wiener filtering method to extract key features, such as amplitude characteristics (mean, total charge), spectral characteristics (main frequency distribution) and statistical characteristics (skewness, kurtosis, etc.), and then the variation pattern of the electrostatic signal with the lubrication state is analyzed.

[0035] S32: Based on the time domain, frequency domain, and time-frequency domain analysis methods of the vibration signal, key vibration characteristic indicators related to the bearing lubrication status are extracted, including root mean square value, kurtosis, peak-to-peak value, etc.; combined with experimental data, the correlation between the vibration characteristics and the bearing lubrication status is clarified, and a mapping relationship between the vibration characteristic parameters and the lubrication status is established. Based on the amplitude, spectral characteristics, and statistical characteristics of the electrostatic signal, key electrostatic signal indicators that characterize the lubrication status, such as kurtosis, power spectral density, and energy distribution, are extracted; combined with experimental data, the variation pattern of the electrostatic signal with the lubrication status is studied, and a corresponding relationship between the electrostatic signal characteristic parameters and the lubrication status is established.

[0036] Furthermore, the feature parameter extraction specifically includes:

[0037] For feature extraction of vibration signals after noise reduction, in terms of time domain feature extraction, the feature indicators include:

[0038] Root Mean Square (RMS): reflects the overall energy of the signal. The calculation formula is:

[0039] ;

[0040] Peak-to-Peak (PP): The difference between the maximum and minimum values of a signal:

[0041] ;

[0042] in, is the maximum value of the signal within a certain period of time. It is the minimum value of the signal in the same period of time;

[0043] Skewness ) and Kurtosis, ): Mainly used to detect impact signals, calculation formula:

[0044] ;

[0045] ;

[0046] in, It is signal sample data points, is the mean value of the signal samples, is the standard deviation of the signal samples, is the total number of signal samples;

[0047] In terms of frequency domain feature extraction, the energy distribution of the signal in the spectrum is analyzed through the Fast Fourier Transform (FFT). Common indicators include:

[0048] Dominant Frequency, ): Frequency component with the largest energy:

[0049] ;

[0050] in, is the power spectral density of the signal, i.e. the frequency The magnitude of the signal energy at

[0051] Center Frequency ):

[0052] ;

[0053] in, It is frequency points.

[0054] Bandwidth (B): reflects the spectrum expansion of the signal. The calculation formula is:

[0055] ;

[0056] In terms of time-frequency domain feature extraction, wavelet packet transform (WPT) or short-time Fourier transform (STFT) is used to jointly analyze the signal evolution process in the time and frequency dimensions.

[0057] The electrostatic signal originates from the charge generated by bearing friction. Its feature extraction can be carried out from three aspects: amplitude, spectrum and statistical characteristics. In terms of amplitude feature extraction, the characteristic indicators include: mean voltage, ), the calculation formula is:

[0058] ;

[0059] in, It is The voltage value of each data point;

[0060] Variance of electrostatic signal ), reflecting the volatility of the signal:

[0061] ;

[0062] In terms of spectrum feature extraction, Power Spectrum Analysis (PSA) is used to observe the frequency distribution of electrostatic signals. ) is calculated by Fourier transform to obtain the maximum energy frequency, and the calculation formula is:

[0063] ;

[0064] in, is the power spectral density of the electrostatic signal.

[0065] In terms of statistical feature extraction, skewness (Skewness, ) and Kurtosis, ), similar to vibration signals, is used to detect abnormal pulse phenomena in electrostatic signals.

[0066] Furthermore, the S4 specifically includes:

[0067] S41: In terms of signal feature fusion, a vibration feature fusion classification model was established using the random forest (RF) algorithm, and an evaluation model for the lubrication status of bearings monitored by vibration based on the random forest (RF) algorithm was constructed. An electrostatic feature fusion classification model was established using the radial basis function (RBF) neural network, and an evaluation model for the lubrication status of bearings monitored by electrostatics based on the radial basis function (RBF) neural network was constructed.

[0068] S42: In terms of decision fusion, the output results of the vibration and electrostatic lubrication state models are processed based on the weighted DS evidence theory, different weights are assigned to the two signals, and the final result is determined.

[0069] Furthermore, the S41 specifically includes:

[0070] Random Forest (RF) is an ensemble learning method that is an extension of the decision tree. It improves the accuracy and generalization ability of the model by constructing multiple decision trees and performing voting (classification) or averaging (regression).

[0071] The vibration and electrostatic signals after denoising and feature extraction are saved as vibration signal datasets and electrostatic signal datasets respectively. Both datasets consist of several feature parameters and three classification labels, and the datasets are divided into a 30% test set and a 70% training set. The data are normalized, and the mapmaxmin function is used to normalize the training set features to the [0,1] interval to improve the model convergence speed. The test set is normalized using the same normalization parameters. 50 decision trees are set, and the minimum number of leaf nodes is set to 1. The out-of-bag error calculation is enabled to evaluate the generalization ability. Function training classification random forest model; use the trained model to predict the classification results of the training set and test set, that is, the lubrication status of the bearing; calculate the classification accuracy of the training set and test set to measure the classification performance of the model; in order to predict the classification results more intuitively and clearly, calculate the importance of each feature to the classification task, and draw a bar chart for visualization, and draw the results under different numbers of decision trees. Error curve, plotting the true values of the training set and test set Compare the predicted values to the curve, observe the classification effect, and draw the confusion matrix of the training set and the test set to analyze the classification error;

[0072] The RBF neural network structure consists of an input layer, a hidden layer, and an output layer. The input layer is used to receive the feature vector of the electrostatic signal, and uses a radial basis function (such as a Gaussian function) to calculate the distance between the input data and the center point, and calculates the output value based on the distance. The output of each hidden layer node depends on the Euclidean distance from the input sample to the center of the node. The Gaussian basis function is:

[0073] ;

[0074] in, is the input vector, is the center vector, which is the position center of the radial basis function, is the width parameter;

[0075] The hidden layer nodes and output nodes are represented by weighted algebraic sum, using The function is used for classification; the output layer corresponds to the classification result of the lubrication state;

[0076] After the electrostatic data set is divided, the data is first normalized and scaled to the [0,1] interval. A radial basis function neural network is created to control the expansion speed of the radial basis function and ensure that the number of hidden layer neurons is consistent with the number of training samples. Conduct simulation tests and Perform forward propagation and make predictions on the training set and test set respectively; finally, calculate the classification accuracy of the training set and test set, visualize the structure of the RBF neural network, draw a comparison curve between the true value and the predicted value, generate a confusion matrix, and analyze the correct and incorrect classification situations.

[0077] Furthermore, the S42 specifically includes:

[0078] DS evidence theory is an uncertain reasoning method that can be used to fuse decisions from multiple information sources to improve classification accuracy. After training a random forest algorithm and a radial basis function neural network and determining the correlation between vibration and electrostatic signals and the bearing lubrication status, the DS evidence theory method is used to fuse these two signals at the decision level. The lubrication status classification probabilities output by the two models are first calculated, namely the basic probability allocation (BPA). These are then fused using the Dempster combination rule, which combines the results of the RF and RBFNN. Ultimately, the decision is made using the maximum confidence (Bel) or maximum BPA, resulting in more reliable lubrication status classification results. This allows for effective judgment of different bearing lubrication conditions.

[0079] Beneficial effects of the present invention:

[0080] The present invention provides a bearing lubrication condition monitoring method for rotating machinery. By integrating vibration and electrostatic dual-modal signal analysis techniques, combined with multidimensional feature modeling and decision-level information fusion, it achieves high-precision and robust monitoring of bearing lubrication conditions. The method offers the following significant advantages: First, it employs a complementary monitoring mechanism for vibration and electrostatic signals, overcoming the limitations of traditional single-signal analysis. The vibration signal constructs lubrication condition correlation features using a random forest model, effectively capturing the physical correlation between the dynamic response of the mechanical structure and lubrication failure. The electrostatic signal, based on a radial basis function neural network model, can sensitively characterize the charge migration characteristics induced by micro-damage on the friction pair surface. The two methods form a cross-validation from the two dimensions of mechanical dynamics and triboelectrics, significantly improving the ability to characterize early-stage minor lubrication anomalies. Second, the method innovatively introduces DS evidence theory for decision-level fusion, addressing the challenges of confidence quantification and conflict resolution for heterogeneous multi-source data. By constructing basic probability distribution functions for the outputs of the vibration and electrostatic models and dynamically optimizing weights using the Dempster synthesis rule, the method achieves probabilistic representation and collaborative reasoning of uncertain information, improving classification accuracy by over 15% compared to single models. The method demonstrates particularly strong discrimination capabilities in transitional lubrication states with fuzzy boundaries. Furthermore, the hybrid signal processing pipeline embedded in this method (including improved wavelet threshold noise reduction and joint time-frequency domain feature extraction) effectively suppresses strong noise interference in industrial sites, achieving a feature dimension compression ratio exceeding 60% while retaining over 98% of the effective information. The resulting lightweight model meets real-time monitoring requirements, with a computational latency of less than 50ms, representing a threefold improvement in efficiency compared to traditional BP neural networks.

[0081] This invention monitors the bearing lubrication status in real time and promptly adjusts the grease replacement cycle or adds grease, ensuring that the bearing is always in the optimal lubrication state. This reduces equipment friction, wear, and failures caused by insufficient lubrication, thereby improving the operating efficiency of rotating machinery. By promptly identifying potential problems such as poor lubrication, it can effectively delay bearing wear and reduce bearing failures caused by lubrication issues, thereby extending the service life of the equipment and reducing equipment replacement frequency and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0083] Figure 1 This is a logic block diagram of lubrication status monitoring according to an embodiment of the present invention;

[0084] Figure 2 Schematic diagram of a method flow in an embodiment of the present invention. DETAILED DESCRIPTION

[0085] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0086] like Figure 1-Figure 2 As shown, a method for intelligently monitoring the lubrication status of bearings in a rotating machine comprises the following steps:

[0087] S1: Build an online monitoring test bench, design a host computer online monitoring system, develop a vibration and electrostatic dual signal acquisition method, arrange temperature and vibration sensors, electrostatic sensors, servo motors, and oil pumps in the test bench, simulate different bearing lubrication conditions, and synchronously acquire the vibration and electrostatic signals of the bearings under different lubrication conditions;

[0088] S2: The vibration characteristics of the bearing under different lubrication conditions, clarifying the correlation between the vibration characteristics and the bearing lubrication condition; the electrostatic characteristics of the bearing under different lubrication conditions, clarifying the correlation between the charge amount and the bearing lubrication condition;

[0089] S3: De-noise the collected vibration and electrostatic signals, analyze the vibration signal characteristics, including time domain, frequency domain, and time-frequency domain characteristic parameters, and extract key vibration characteristic indicators related to the bearing lubrication status. At the same time, based on the amplitude, spectrum, and statistical characteristics of the electrostatic signal, analyze how the electrostatic signal changes with the lubrication status.

[0090] S4: Construct a vibration-static-lubrication status correlation model, combine deep learning algorithms with data fusion technology, classify the bearing lubrication status, and realize intelligent identification of lubrication status.

[0091] S1 specifically includes:

[0092] S11: Use serial communication to establish communication between LabVIEW software and the inverter, sensor, oil pump, and flow meter to ensure the stability and real-time performance of the data acquisition system; design the host computer control logic to achieve dynamic control of motor start and stop, speed regulation, and oil pump output;

[0093] S12: By adjusting the motor speed and controlling the oil supply of the oil pump, the operating states of the bearing under three lubrication conditions, namely, good lubrication, insufficient lubrication, and overlubrication, are simulated, and the original signal data of the corresponding sensors are recorded.

[0094] The corresponding relationship between the grease filling ratio and the lubrication status is as follows: if the filling volume accounts for less than 20-30% of the bearing free space, it is insufficient lubrication; if the filling volume accounts for 30-50% of the free space, it is good lubrication; if the filling volume exceeds 50% of the free space, it is overlubrication.

[0095] The inverter frequency is set to a constant range of 10Hz-50Hz to correspond to different motor speeds. A flow meter monitors and records the grease output from the oil pump, measuring both total and instantaneous flow. The ratio of grease to the free space within the bearing, combined with temperature data from the sensor and the set motor speed, simulates three bearing lubrication states: well-lubricated, under-lubricated, and over-lubricated.

[0096] S2 specifically includes:

[0097] S21: Lubrication conditions directly affect bearing wear, thereby changing the characteristics of the vibration signal. Grease quality, motor speed, and oil film stability play a significant role in the vibration characteristics. Oil film rupture or insufficient lubrication can cause significant changes in the vibration signal. Different lubrication conditions correspond to different vibration signal intensities. Lubrication conditions directly affect bearing friction, thereby changing the characteristics of the electrostatic signal. The electrostatic intensity is determined by the charge generated by the friction of the rotating pair and is also affected by different grease types and motor speed.

[0098] The lubrication status of a bearing directly affects its vibration and electrostatic signal characteristics. Under normal lubrication, the oil film is uniform and stable, the vibration signal amplitude is low, and the electrostatic signal is smooth, with no obvious abnormal components. Underlubrication leads to increased friction, increased vibration signal impact, and enhanced high-frequency components in the spectrum. Simultaneously, the electrostatic signal amplitude increases, fluctuates frequently, and may experience transient voltage spikes. Overlubrication can hinder rolling element motion due to excess grease, resulting in abnormal low-frequency components in the vibration signal, a reduced electrostatic signal amplitude, and generally stable fluctuations. By monitoring the changing characteristics of the vibration and electrostatic signals, the lubrication status of the bearing can be effectively identified and failures prevented.

[0099] S3 specifically includes:

[0100] S31: Wavelet denoising is used for vibration signals. Wavelet decomposition is used to perform multi-scale analysis of the signal and threshold processing is performed on the high-frequency noise coefficient to retain the main characteristic components. An appropriate wavelet basis function (such as Daubechies wavelet dbN) is selected and discrete wavelet transform (DWT) is performed on the vibration signal to decompose it into approximation signals and detail signals in different frequency bands:

[0101] ;

[0102] in, is the low-frequency approximation signal, It is a high frequency detail signal.

[0103] Use soft threshold or hard threshold method to remove high-frequency noise. The typical soft threshold function is:

[0104] ;

[0105] in, is the original signal A quantity, is the threshold parameter;

[0106] Then perform wavelet reconstruction to restore the denoised signal; analyze the denoised vibration signal, extract time domain features (such as root mean square value RMS, peak factor, etc.), frequency domain features (such as power spectral density PSD, frequency center, etc.) and time-frequency domain features (such as wavelet energy distribution, etc.), and extract key vibration characteristic indicators related to the lubrication status.

[0107] For electrostatic signals, the Wiener Filtering method is used to optimize the signal with the minimum mean square error criterion to reduce noise interference. By real signal and noise Composition, namely:

[0108] ;

[0109] Calculate the autocorrelation function of the signal and power spectral density , the signal-to-noise ratio of the signal is obtained based on the noise estimation.

[0110] The transfer function of the Wiener filter is defined as:

[0111] ;

[0112] in, and are the power spectral densities of the signal and noise, respectively.

[0113] The electrostatic signal is processed by the Wiener filtering method to extract key features, such as amplitude characteristics (mean, total charge), spectral characteristics (main frequency distribution) and statistical characteristics (skewness, kurtosis, etc.), and then the variation pattern of the electrostatic signal with the lubrication state is analyzed.

[0114] S32: Based on the time domain, frequency domain, and time-frequency domain analysis methods of the vibration signal, key vibration characteristic indicators related to the bearing lubrication status are extracted, including root mean square value, kurtosis, peak-to-peak value, etc.; combined with experimental data, the correlation between the vibration characteristics and the bearing lubrication status is clarified, and a mapping relationship between the vibration characteristic parameters and the lubrication status is established. Based on the amplitude, spectral characteristics, and statistical characteristics of the electrostatic signal, key electrostatic signal indicators that characterize the lubrication status, such as kurtosis, power spectral density, and energy distribution, are extracted; combined with experimental data, the variation pattern of the electrostatic signal with the lubrication status is studied, and a corresponding relationship between the electrostatic signal characteristic parameters and the lubrication status is established.

[0115] For feature extraction of vibration signals after noise reduction, in terms of time domain feature extraction, the feature indicators include:

[0116] Root Mean Square (RMS): reflects the overall energy of the signal. The calculation formula is:

[0117] ;

[0118] Peak-to-Peak (PP): The difference between the maximum and minimum values of a signal:

[0119] ;

[0120] in, is the maximum value of the signal within a certain period of time. It is the minimum value of the signal in the same period of time;

[0121] Skewness ) and Kurtosis, ): Mainly used to detect impact signals, calculation formula:

[0122] ;

[0123] ;

[0124] in, It is signal sample data points, is the mean value of the signal samples, is the standard deviation of the signal samples, is the total number of signal samples;

[0125] In terms of frequency domain feature extraction, the energy distribution of the signal in the spectrum is analyzed through the Fast Fourier Transform (FFT). Common indicators include:

[0126] Dominant Frequency, ): Frequency component with the largest energy:

[0127] ;

[0128] in, is the power spectral density of the signal, i.e. the frequency The magnitude of the signal energy at

[0129] Center Frequency ):

[0130] ;

[0131] in, It is frequency points.

[0132] Bandwidth (B): reflects the spectrum expansion of the signal. The calculation formula is:

[0133] ;

[0134] In terms of time-frequency domain feature extraction, wavelet packet transform (WPT) or short-time Fourier transform (STFT) is used to jointly analyze the signal evolution process in the time and frequency dimensions.

[0135] The electrostatic signal originates from the charge generated by bearing friction. Its feature extraction can be carried out from three aspects: amplitude, spectrum and statistical characteristics. In terms of amplitude feature extraction, the characteristic indicators include: mean voltage, ), the calculation formula is:

[0136] ;

[0137] Variance of electrostatic signal ), reflecting the volatility of the signal:

[0138] ;

[0139] In terms of spectrum feature extraction, Power Spectrum Analysis (PSA) is used to observe the frequency distribution of electrostatic signals. ) is calculated by Fourier transform to obtain the maximum energy frequency, and the calculation formula is:

[0140] ;

[0141] in, is the power spectral density of the electrostatic signal.

[0142] In terms of statistical feature extraction, skewness (Skewness, ) and Kurtosis, ), similar to vibration signals, is used to detect abnormal pulse phenomena in electrostatic signals.

[0143] S4 specifically includes:

[0144] S41: In terms of signal feature fusion, a vibration feature fusion classification model was established using the random forest (RF) algorithm, and an evaluation model for the lubrication status of bearings monitored by vibration based on the random forest (RF) algorithm was constructed. An electrostatic feature fusion classification model was established using the radial basis function (RBF) neural network, and an evaluation model for the lubrication status of bearings monitored by electrostatics based on the radial basis function (RBF) neural network was constructed.

[0145] S42: In terms of decision fusion, the output results of the vibration and electrostatic lubrication state models are processed based on the weighted DS evidence theory, different weights are assigned to the two signals, and the final result is determined.

[0146] Random Forest (RF) is an ensemble learning method that is an extension of the decision tree. It improves the accuracy and generalization ability of the model by constructing multiple decision trees and performing voting (classification) or averaging (regression).

[0147] The vibration and electrostatic signals after denoising and feature extraction are saved as vibration signal datasets and electrostatic signal datasets respectively. Both datasets consist of several feature parameters and three classification labels, and the datasets are divided into a 30% test set and a 70% training set. The data are normalized, and the mapmaxmin function is used to normalize the training set features to the [0,1] interval to improve the model convergence speed. The test set is normalized using the same normalization parameters. 50 decision trees are set, and the minimum number of leaf nodes is set to 1. The out-of-bag error calculation is enabled to evaluate the generalization ability. Function training classification random forest model; use the trained model to predict the classification results of the training set and test set, that is, the lubrication status of the bearing; calculate the classification accuracy of the training set and test set to measure the classification performance of the model; in order to predict the classification results more intuitively and clearly, calculate the importance of each feature to the classification task, and draw a bar chart for visualization, and draw the results under different numbers of decision trees. Error curve, plotting the true values of the training set and test set Compare the predicted values to the curve, observe the classification effect, and draw the confusion matrix of the training set and the test set to analyze the classification error;

[0148] The RBF neural network structure consists of an input layer, a hidden layer, and an output layer. The input layer is used to receive the feature vector of the electrostatic signal, and uses a radial basis function (such as a Gaussian function) to calculate the distance between the input data and the center point, and calculates the output value based on the distance. The output of each hidden layer node depends on the Euclidean distance from the input sample to the center of the node. The Gaussian basis function is:

[0149] ;

[0150] in, is the input vector, is the center vector, which is the position center of the radial basis function, is the width parameter;

[0151] The hidden layer nodes and output nodes are represented by weighted algebraic sum, using The function is used for classification; the output layer corresponds to the classification result of the lubrication state;

[0152] After the electrostatic data set is divided, the data is first normalized and scaled to the [0,1] interval. A radial basis function neural network is created to control the expansion speed of the radial basis function and ensure that the number of hidden layer neurons is consistent with the number of training samples. Conduct simulation tests and Perform forward propagation and make predictions on the training set and test set respectively; finally, calculate the classification accuracy of the training set and test set, visualize the structure of the RBF neural network, draw a comparison curve between the true value and the predicted value, generate a confusion matrix, and analyze the correct and incorrect classification situations.

[0153] DS evidence theory is an uncertain reasoning method that can be used to fuse decisions from multiple information sources to improve classification accuracy. After training a random forest algorithm and a radial basis function neural network and determining the correlation between vibration and electrostatic signals and the bearing lubrication status, the DS evidence theory method is used to fuse these two signals at the decision level. The lubrication status classification probabilities output by the two models are first calculated, namely the basic probability allocation (BPA). These are then fused using the Dempster combination rule, which combines the results of the RF and RBFNN. Ultimately, the decision is made using the maximum confidence (Bel) or maximum BPA, resulting in more reliable lubrication status classification results. This allows for effective judgment of different bearing lubrication conditions.

[0154] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention is limited to these examples. Within the scope of the present invention, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity.

[0155] The present invention is intended to cover all such substitutions, modifications and variations that fall within the broad scope of the claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for intelligently monitoring the lubrication status of bearings in rotating machinery, characterized in that: The following steps are involved: S1: Build an online monitoring test bench, design a host computer online monitoring system, develop a vibration and electrostatic dual signal acquisition method, arrange temperature and vibration sensors, electrostatic sensors, servo motors, and oil pumps in the test bench, simulate different bearing lubrication conditions, and synchronously acquire the vibration and electrostatic signals of the bearings under different lubrication conditions; S2: The vibration characteristics of the bearing under different lubrication conditions, clarifying the correlation between the vibration characteristics and the bearing lubrication condition; the electrostatic characteristics of the bearing under different lubrication conditions, clarifying the correlation between the charge amount and the bearing lubrication condition; S3: De-noise the collected vibration and electrostatic signals, analyze the vibration signal characteristics, including time domain, frequency domain, and time-frequency domain characteristic parameters, and extract key vibration characteristic indicators related to the bearing lubrication status. At the same time, based on the amplitude, spectrum, and statistical characteristics of the electrostatic signal, analyze how the electrostatic signal changes with the lubrication status. S4: Build a vibration-static-lubrication state correlation model, combine deep learning algorithms with data fusion technology, classify bearing lubrication states, and achieve intelligent identification of lubrication states; S41: In terms of signal feature fusion, a vibration feature fusion classification model was established using the random forest algorithm, and an evaluation model for the lubrication status of bearings monitored by vibration based on the random forest algorithm was constructed. An electrostatic feature fusion classification model was established using the radial basis function neural network, and an evaluation model for the lubrication status of bearings monitored by electrostatics based on the radial basis function neural network was constructed. S42: In terms of decision fusion, the output results of the vibration and electrostatic lubrication state models are processed based on the weighted DS evidence theory, different weights are assigned to the two signals, and the final result is determined.

2. The intelligent monitoring method for bearing lubrication status in a rotating machine according to claim 1, characterized in that: Said S1 specifically includes: S11: Use serial communication to establish communication between LabVIEW software and the inverter, sensor, oil pump, and flow meter to ensure the stability and real-time performance of the data acquisition system; design the host computer control logic to achieve dynamic control of motor start and stop, speed regulation, and oil pump output; S12: By adjusting the motor speed and controlling the oil supply of the oil pump, the operating states of the bearing under three lubrication conditions, namely, good lubrication, insufficient lubrication, and overlubrication, are simulated, and the original signal data of the corresponding sensors are recorded.

3. The method for intelligently monitoring the lubrication status of bearings in a rotating machine according to claim 2, characterized in that: The simulated bearing lubrication state specifically includes: The corresponding relationship between the grease filling ratio and the lubrication status is as follows: if the filling volume accounts for less than 20-30% of the bearing free space, it is insufficient lubrication; if the filling volume accounts for 30-50% of the free space, it is good lubrication; if the filling volume exceeds 50% of the free space, it is overlubrication. The inverter frequency is set to an equidistant interval of 10Hz-50Hz to correspond to different motor speeds; the flow meter observes and records the grease output of the oil pump from two aspects: total flow and instantaneous flow; the proportion of grease to the free space inside the bearing is combined with the temperature data fed back by the sensor and the set motor speed to jointly simulate the three lubrication states of the bearing, among which the three lubrication states are good lubrication, insufficient lubrication and overlubrication.

4. The method for intelligently monitoring the lubrication status of bearings in a rotating machine according to claim 3, characterized in that: The S3 specifically includes: S31: Using a wavelet noise reduction method for the vibration signal, performing multi-scale analysis on the signal through wavelet decomposition, and performing threshold processing on the high-frequency noise coefficient to retain the main characteristic components; Select appropriate wavelet basis functions, perform discrete wavelet transform on the vibration signal, and decompose it into approximation signals and detail signals in different frequency bands: ； in, is the low-frequency approximation signal, It is a high-frequency detail signal; Use soft threshold or hard threshold method to remove high-frequency noise. The typical soft threshold function is: ； in, is the original signal A quantity, is the threshold parameter; Then perform wavelet reconstruction to restore the denoised signal; analyze the denoised vibration signal, extract time domain features, frequency domain features and time-frequency domain features, and extract key vibration characteristic indicators related to the lubrication status; For electrostatic signals, the Wiener filtering method is used to optimize the signal with the minimum mean square error criterion to reduce noise interference; By real signal and noise Composition, namely: ； Calculate the autocorrelation function of the signal and power spectral density , obtain the signal-to-noise ratio of the signal based on noise estimation; The transfer function of the Wiener filter is defined as: ； in, and are the power spectral densities of the signal and noise, respectively; The electrostatic signal is processed by the Wiener filtering method to extract key features such as amplitude, spectrum and statistical features, and then the variation of the electrostatic signal with the lubrication state is analyzed; S32: Based on the time domain, frequency domain, and time-frequency domain analysis methods of the vibration signal, extract key vibration characteristic indicators related to the bearing lubrication status, including root mean square value, kurtosis, and peak-to-peak value; combine experimental data to clarify the correlation between the vibration characteristics and the bearing lubrication status, and construct a mapping relationship between the vibration characteristic parameters and the lubrication status; Based on the amplitude, spectral characteristics and statistical characteristics of the electrostatic signal, key electrostatic signal indicators characterizing the lubrication state are extracted, including kurtosis, power spectral density and energy distribution; combined with experimental data, the change law of the electrostatic signal with the lubrication state is studied, and the corresponding relationship between the characteristic parameters of the electrostatic signal and the lubrication state is established.

5. The intelligent monitoring method for bearing lubrication status in a rotating machine according to claim 4, characterized in that: The feature parameter extraction specifically includes: For feature extraction of vibration signals after noise reduction, in terms of time domain feature extraction, the feature indicators include: Root mean square value RMS: reflects the overall energy of the signal. The calculation formula is: ； Peak-to-peak value PP: the difference between the maximum and minimum values of the signal: ； in, is the maximum value of the signal within a certain period of time. It is the minimum value of the signal in the same period of time; Skewness , kurtosis : Mainly used to detect impact signals, calculation formula: ； ； in, It is signal sample data points, is the mean value of the signal samples, is the standard deviation of the signal samples, is the total number of signal samples; In terms of frequency domain feature extraction, the energy distribution of the signal in the spectrum is analyzed through Fourier transform. Common indicators include: Main frequency : Frequency component with maximum energy: ； in, is the power spectral density of the signal, i.e. the frequency The magnitude of the signal energy at The center frequency is : ； in, It is frequency points; Bandwidth B: reflects the spectrum expansion of the signal, calculated by the formula: ； In terms of time-frequency domain feature extraction, wavelet packet decomposition or short-time Fourier transform is used to jointly analyze the signal evolution process in the time and frequency dimensions; The electrostatic signal comes from the charge generated by bearing friction. Its feature extraction can be carried out from three aspects: amplitude, spectrum and statistical characteristics. In terms of amplitude feature extraction, the characteristic indicators include: electrostatic signal mean , the calculation formula is: ； in, It is The voltage value of each data point; Electrostatic signal variance , reflecting the volatility of the signal: ； In terms of spectrum feature extraction, power spectrum analysis PSA is used to observe the frequency distribution of electrostatic signals; The main frequency components The maximum energy frequency is calculated by Fourier transform, and the calculation formula is: ； in, is the power spectral density of the electrostatic signal; In terms of statistical feature extraction, skewness and kurtosis , similar to vibration signals, is used to detect abnormal pulse phenomena in electrostatic signals.

6. The method for intelligently monitoring the lubrication status of bearings in a rotating machine according to claim 5, characterized in that: The S41 specifically includes: Random forest is an ensemble learning method that is an extension of decision tree; Improve the accuracy and generalization ability of the model by building multiple decision trees and voting or averaging them; The vibration and electrostatic signals after denoising and feature extraction are saved as vibration signal datasets and electrostatic signal datasets respectively. Both datasets consist of several feature parameters and three classification labels, and the datasets are divided into a 30% test set and a 70% training set. The data are normalized, and the mapmaxmin function is used to normalize the training set features to the [0,1] interval to improve the model convergence speed. The test set is normalized using the same normalization parameters. 50 decision trees are set, and the minimum number of leaf nodes is set to 1. The out-of-bag error calculation is enabled to evaluate the generalization ability. Function training classification random forest model; The trained model is used to predict the classification results of the training set and the test set, that is, the lubrication status of the bearing; the classification accuracy of the training set and the test set is calculated to measure the classification performance of the model; in order to predict the classification results more intuitively and clearly, the importance of each feature to the classification task is calculated, and a bar chart is drawn for visualization, and the results under different numbers of decision trees are plotted. Error curve, plotting the true values of the training set and test set Compare the predicted values to the curve, observe the classification effect, and draw the confusion matrix of the training set and the test set to analyze the classification error; The RBF neural network structure consists of an input layer, a hidden layer, and an output layer. The input layer is used to receive the feature vector of the electrostatic signal, and the radial basis function is used to calculate the distance between the input data and the center point, and the output value is calculated based on the distance. The output of each hidden layer node depends on the Euclidean distance from the input sample to the center of the node. Among them, the radial basis function is a Gaussian function, and the Gaussian basis function is: ； in, is the input vector, is the center vector, which is the position center of the radial basis function, is the width parameter; The hidden layer nodes and output nodes are represented by weighted algebraic sum, using The function is used for classification; the output layer corresponds to the classification result of the lubrication state; After the electrostatic data set is divided, the data is first normalized and scaled to the [0,1] interval. A radial basis function neural network is created to control the expansion speed of the radial basis function and ensure that the number of hidden layer neurons is consistent with the number of training samples. Conduct simulation tests and Perform forward propagation and make predictions on the training set and test set respectively; Finally, the classification accuracy of the training set and test set is calculated, the structure of the RBF neural network is visualized, a comparison curve between the true value and the predicted value is drawn, a confusion matrix is generated, and the correct and incorrect classification situations are analyzed.

7. The method for intelligently monitoring the lubrication status of bearings in a rotating machine according to claim 6, characterized in that: The S42 specifically includes: DS evidence theory is an uncertain reasoning method used to fuse decisions from multiple information sources to improve classification accuracy. After training the random forest algorithm and radial basis function neural network and determining the correlation between vibration and electrostatic signals and bearing lubrication status, the DS evidence theory method is used to fuse the two signals at the decision level. First, the lubrication status classification probabilities output by the two models are calculated, that is, the basic probability distribution is calculated, and the Dempster combination rule is used for fusion. The results of RF and RBFNN can be integrated, and the maximum confidence or maximum BPA is finally used for decision making, thereby obtaining a more reliable lubrication status classification result and realizing effective judgment of different lubrication status of bearings.

Images