Friction coefficient volatility and mean value index layered modeling prediction method

Abstract

The invention discloses a friction coefficient volatility and mean value index layered modeling prediction method, which belongs to the technical field of mechanical engineering and intelligent monitoring, and comprises the following steps: S1, multi-source information acquisition and preprocessing; s2, multi-source information feature extraction and correlation analysis are carried out; s3, predicting a friction coefficient volatility index based on the step S2; and S4, performing friction coefficient mean value index prediction based on the step S3. According to the friction coefficient volatility and mean value index layered modeling prediction method, by introducing a layered modeling mechanism and a mixed training strategy, modeling with higher physical rationality and higher prediction precision for various indexes of the friction coefficient is achieved, and the method has good physical interpretability, popularization applicability and engineering practical value.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mechanical engineering and intelligent monitoring technology, and particularly relates to a friction coefficient fluctuation and mean value index hierarchical modeling and prediction method. BACKGROUND

[0002] In the prior art, rotating parts such as rotor-bearing and meshing gears, and sliding members such as sliding rails and piston-cylinder sleeves are widely used in the mechanical, transportation and energy industries. During operation, friction and wear inevitably occur, resulting in significant energy loss and economic loss. Therefore, monitoring the state of the friction interface is of great significance to ensure the stability of the mechanical system and prolong the service life.

[0003] Under actual working conditions, direct measurement of friction is challenging due to the difficulty in directly obtaining friction at the contact interface and the limitations of sensor placement. Acoustic vibration-derived signals are closely related to the friction state during sliding friction, and therefore can be used for indirect monitoring of the friction state. However, the complex nonlinear relationship between acoustic vibration features and friction coefficient is not well understood.

[0004] In recent years, multi-source signal fusion methods can achieve acoustic signal and vibration signal fusion modeling, improving the accuracy and robustness of friction state monitoring. However, existing methods focus on predicting the mean value of the friction coefficient, without distinguishing between the mean value and the fluctuation component, making it difficult to fully reflect the dynamic evolution law and potential fault features. SUMMARY

[0005] The purpose of the present application is to provide a friction coefficient fluctuation and mean value index hierarchical modeling and prediction method, which introduces a hierarchical modeling mechanism and a hybrid training strategy to achieve more physically reasonable modeling of various indicators of the friction coefficient with higher prediction accuracy, and has good physical interpretability, applicability, and engineering practical value.

[0006] To achieve the above purpose, the present application provides a friction coefficient fluctuation and mean value index hierarchical modeling and prediction method, comprising the following steps:

[0007] Step S1, multi-source information acquisition and preprocessing;

[0008] Step S2, multi-source information feature extraction and correlation analysis;

[0009] Step S3, friction coefficient fluctuation index prediction based on step S2;

[0010] Step S4, friction coefficient mean value index prediction based on step S3.

[0011] Preferably, step S1 specifically comprises the following steps:

[0012] Step S101, based on a standard friction and wear tester, a multi-source friction information synchronous acquisition system is built for acquiring multi-source information, the multi-source information including friction force, friction coefficient, and acoustic vibration information;

[0013] The acoustic vibration information includes X-direction vibration, Y-direction vibration, and sound pressure signal; the X-direction vibration is vibration along the friction direction, and the Y-direction vibration is vibration perpendicular to the friction direction in the horizontal plane;

[0014] Step S102, acoustic vibration signals under an empty load working condition are acquired before the test starts, as background noise baseline data;

[0015] Step S103, a test sample is installed, a test load, a speed, and a duration are set, and then the friction test is started;

[0016] Step S104, multi-source data under dry friction process of a set material and working condition are acquired;

[0017] Step S105, the excitation threshold detected in the X-direction vibration signal is used as a time anchor point, and time stamp alignment of the multi-source information is realized;

[0018] Step S106, a least mean square (LMS) adaptive filter is used to filter the acquired signals, background noise is suppressed, and filtered multi-source friction information data aligned with the time stamp are obtained.

[0019] Preferably, step S2 specifically includes the following steps:

[0020] Step S201, a moving average method is used to extract time domain features, frequency domain features, and time-frequency domain features of the multi-source information;

[0021] Step S202, a Pearson correlation coefficient between the acoustic vibration information and the friction coefficient is calculated, and a correlation matrix between the acoustic vibration information and the friction coefficient is obtained;

[0022] Step S203, each correlation coefficient in the correlation matrix is sorted in descending order according to the absolute value, and a friction coefficient feature having an absolute correlation coefficient greater than 0.8 with any acoustic vibration signal feature is screened out, a plurality of high-correlation friction features are obtained, and analysis shows that the plurality of high-correlation features are all fluctuation features of the friction coefficient;

[0023] Step S204, in the plurality of high-correlation friction features obtained, the highest 5 acoustic vibration features corresponding to each feature are further retained, and a corresponding relationship table between the acoustic vibration signal features and the friction coefficient fluctuation features is constructed.

[0024] Preferably, step S3 specifically includes the following steps:

[0025] Step S301, regression training of the friction coefficient fluctuation is carried out by using a basic regression model, and the basic regression model is integrated and fused by using different fusion algorithms;

[0026] Step S302, the fitting effect of the regression model is evaluated by using evaluation indexes;

[0027] Step S303, the model with the highest evaluation index is selected as the prediction model of the friction coefficient fluctuation index, and is saved.

[0028] Preferably, step S4 specifically comprises the following steps:

[0029] Step S401, according to the distribution characteristics and statistical laws of the friction coefficient fluctuation index, a physical heuristic feature reflecting the relationship between the friction coefficient fluctuation index and the mean value index is constructed;

[0030] Step S402, the extracted fluctuation index and the physical heuristic feature are taken as input features, and the mean value index of the friction coefficient is taken as an output variable, and a data set for regression modeling is constructed;

[0031] Step S403, a hybrid training strategy is adopted, and a random forest regression model is used to regress and predict the mean value index of the friction coefficient;

[0032] Step S404, model fitting is completed based on the training result, and the trained random forest model is saved as the prediction model of the friction coefficient mean value index.

[0033] Preferably, the physical heuristic feature is specifically:

[0034] The derived estimated mean value:

[0035]

[0036] Wherein, peak is the peak value in the fluctuation index; min is the minimum value in the fluctuation index;

[0037] The derived estimated root mean square value:

[0038]

[0039] Wherein, std is the standard deviation in the fluctuation index;

[0040] The derived frequency energy value:

[0041] freq energy =psd mean *hjorth_activity;

[0042] Wherein, psd meanThe energy spectrum density mean is a fluctuation index, and the hjorth_activity is a signal variance.

[0043] Preferably, in step S201, the time domain features, the frequency domain features and the time-frequency domain features of the multi-source information are extracted by using a moving window method, the width of the moving window is set as a time length corresponding to one rotation of the friction, and the step length is one quarter of the width.

[0044] Preferably, in step 301, different fluctuation indexes of the friction coefficient are used as target variables, and corresponding vibration and sound features are selected as input variables according to a corresponding relationship table between the vibration and sound signal features and the fluctuation features of the friction coefficient.

[0045] Preferably, the basic regression model includes a plurality of representative regression algorithms such as a linear model (such as Linear regression, Ridge regression, Lasso regression), a support vector regression (SVR) based on a kernel function, a non-parametric model K nearest neighbor regression (KNN), a tree-based ensemble model (random forest RF, gradient boosting GBDT, extreme gradient boosting XGBoost), a Bayesian regression (Bayesian Ridge) and an artificial neural network (ANN).

[0046] Preferably, the integrated fusion strategy includes:

[0047] A weighted average method, which assigns weights according to the performance of each model and performs weighted fusion on the prediction results;

[0048] A soft voting regression method, which performs simple average on the output results of each model to alleviate the overfitting problem;

[0049] Stacked fusion, which constructs a meta-regressor to re-learn the prediction results of the basic model to capture higher-order nonlinear relationships.

[0050] Therefore, the present application adopts the above-mentioned layered modeling and prediction method of the fluctuation and mean value indexes of the friction coefficient, and the technical effects are as follows:

[0051] The layered modeling method is used to model and predict the fluctuation and mean value of the friction coefficient, so that the statistical modeling based on the vibration and sound information-fluctuation of the friction coefficient, fluctuation-mean value is more physically reasonable and interpretable;

[0052] Based on the basic regression algorithm and the fusion algorithm, an algorithm model more suitable for the prediction of the fluctuation of the friction coefficient is selected;

[0053] A physically inspired fluctuation feature is proposed, and a hybrid training method is proposed to meet the statistical modeling needs of the fluctuation-mean value of the friction coefficient and improve the training speed.

[0054] The technical solutions of the present application are further described below with reference to the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is a flow chart of an embodiment of the present application, a friction coefficient fluctuation and mean value index layered modeling prediction method;

[0056] Figure 2 is a framework diagram of an embodiment of the present application, a friction coefficient fluctuation and mean value index layered modeling prediction method;

[0057] Figure 3 is a correspondence table between acoustic vibration signal features and friction coefficient fluctuation features of an embodiment of the present application, a friction coefficient fluctuation and mean value index layered modeling prediction method. DETAILED DESCRIPTION

[0058] The technical solutions of the present application are further described below with reference to the drawings and examples.

[0059] Unless otherwise defined, the technical terms or scientific terms used in the present application shall have the usual meanings understood by those with ordinary skills in the art to which the present application belongs.

[0060] Example 1

[0061] As shown in Figure 1 , the present application provides a friction coefficient fluctuation and mean value index layered modeling prediction method, Figure 2 is a framework diagram of the prediction method, including the following steps:

[0062] Step S1, multi-source information acquisition and preprocessing; based on a standard friction and wear tester (hereinafter referred to as a tester), a multi-source friction information synchronous acquisition system is built by combining various sensing devices, such as force sensors, acoustic pressure sensors, and acceleration vibration sensors. Taking a ball-disk rotary friction test as an example, multi-source information under specified working conditions is synchronously acquired, including friction force, friction coefficient, acoustic vibration information, including X-direction vibration, Y-direction vibration, and acoustic pressure signal; wherein the X-direction vibration is the vibration along the friction direction, and the Y-direction vibration is the vibration perpendicular to the friction direction in the horizontal plane.

[0063] According to the X-direction vibration excitation threshold, the timestamps of the multi-source information are aligned. The vibration signal and acoustic pressure signal under the no-friction state are taken as background noise, and the features thereof are statistically modeled to determine the energy-based suppression threshold. A least mean square adaptive filter is used to suppress the background noise to obtain the filtered timestamp-aligned multi-source friction information data.

[0064] Step S2, multi-source information feature extraction and correlation analysis; using a moving window to extract the time domain features of multi-source information (including 16 time domain features such as mean, root mean square, standard deviation, peak value, maximum value, peak-to-peak value, energy waveform factor, and sound pressure level), frequency domain features (including 8 frequency domain features such as maximum frequency, maximum frequency amplitude, frequency mean, and energy spectrum peak value), and time-frequency domain features (obtained according to the short-time Fourier transform method). The width of the moving window is set to the time length corresponding to one rotation of the friction, and the step is one quarter of the width, ensuring that the extracted features can represent the signal characteristics of each friction process.

[0065] The Pearson correlation coefficient between the acoustic vibration information and the friction coefficient is calculated to obtain a correlation matrix between the acoustic vibration information and the friction coefficient; each correlation coefficient in the correlation matrix is sorted in descending order according to the absolute value, and friction coefficient features with an absolute correlation coefficient greater than 0.8 with any acoustic vibration signal feature are selected, a total of 10 high-correlation friction features. Analysis shows that these high-correlation features are all volatility features of the friction coefficient;

[0066] Among the above-mentioned 10 high-correlation friction volatility features, the 5 acoustic vibration features with the highest correlation corresponding to each feature are further retained, and a corresponding relationship table between the acoustic vibration signal features and the friction coefficient volatility features is constructed, as shown in Figure 3 , to provide a high-quality feature combination for subsequent modeling and prediction.

[0067] Step S3, friction coefficient volatility index prediction based on step S2;

[0068] A variety of basic regression models are used for regression training of the friction coefficient volatility, with different indicators of the friction coefficient volatility as the target values and the acoustic vibration features selected according to the relationship table as the input values. The basic regression models include a variety of representative regression algorithms such as linear models (such as Linear regression, Ridge regression, Lasso regression), support vector regression (SVR) based on kernel functions, non-parametric model K nearest neighbor regression (KNN), tree-based ensemble models (random forest RF, gradient boosting GBDT, extreme gradient boosting XGBoost), Bayesian regression (Bayesian Ridge), and artificial neural networks (ANN);

[0069] To improve the prediction accuracy and model generalization ability, the prediction results output by the above-mentioned multiple basic regression models are integrated and fused. The fusion strategies include:

[0070] Weighted average method, according to the performance of each model to allocate weights, weighted fusion of prediction results;

[0071] Soft voting regression method, simple average of the output results of each model to alleviate the overfitting problem;

[0072] Stacked fusion, the meta-regressor relearns the base model prediction results to capture higher-order nonlinear relationships.

[0073] Specific evaluation indicators are used to evaluate the fitting effect of the regression model, such as mean square error (MSE), root mean square error (RMSE), goodness of fit (R2), etc. The data set is divided into training set and test set using the leave-out method. The test set does not participate in the training process, and is only used to verify the model obtained by training. The model with the highest evaluation index is selected as the prediction model of the friction coefficient volatility index, and is saved.

[0074] Step S4, based on step S3, the mean value index of the friction coefficient is predicted; based on the friction coefficient volatility feature, combined with the statistical properties and the distribution characteristics of the friction coefficient, three physical heuristic derived features from volatility to mean value are constructed to establish the mapping relationship between the volatility index and the central tendency index, including the mean value mean and the root mean square RMS; the volatility index and the physical heuristic feature are input, and the mean value index is output to construct the data set;

[0075] The physical heuristic feature is specifically:

[0076] The derived estimated mean value is:

[0077]

[0078] Wherein, peak is the peak value in the volatility index; min is the minimum value in the volatility index;

[0079] The derived estimated root mean square value is:

[0080]

[0081] Wherein, std is the standard deviation in the volatility index;

[0082] The derived frequency energy value is:

[0083] freq energy =psd mean *hjorth_activity;

[0084] Wherein, psd mean is the energy spectrum density mean in the volatility index; hjorth_activity is the signal variance in the volatility index.

[0085] To improve the robustness of the model in the actual deployment scenario, a hybrid training strategy is introduced: the true value of part of the intermediate features in the training set is randomly replaced by the corresponding predicted value;

[0086] Based on the constructed physical heuristic features, a random forest regression model is trained, the evaluation index is consistent with S3, the model with the highest evaluation index is selected as the prediction model of the friction coefficient volatility index, and is saved;

[0087] Finally, the saved model with the highest accuracy is combined to realize accurate prediction from the sound and vibration features to the friction coefficient volatility index and the friction coefficient central tendency index with physical interpretability.

[0088] Therefore, the application adopts the above-mentioned layered modeling and prediction method of the friction coefficient volatility and mean value index, realizes accurate modeling of the multi-dimensional dynamic behavior of the friction coefficient by introducing a layered modeling mechanism and a mixed training strategy, and has good physical interpretability, generalization and engineering practical value.

[0089] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: the technical solutions of the present application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for modeling and predicting the coefficient of friction fluctuation and mean value index layering, characterized in that, Comprising the following steps: Step S1, multi-source information collection and preprocessing; Step S2, multi-source information feature extraction and correlation analysis; Step S3, based on step S2, friction coefficient fluctuation index prediction; Step S4, based on step S3, friction coefficient mean index prediction.

2. The method according to claim 1, characterized in that, Step S1 specifically includes the following steps: Step S101, based on a standard friction and wear tester, a multi-source friction information synchronous acquisition system is built for collecting multi-source information, including friction force, friction coefficient, and acoustic vibration information; The acoustic vibration information includes X-direction vibration, Y-direction vibration, and sound pressure signal; wherein the X-direction vibration is the vibration along the friction direction, and the Y-direction vibration is the vibration perpendicular to the friction direction in the horizontal plane; Step S102, collect the acoustic vibration signal under the no-load working condition before the test starts, as the background noise baseline data; Step S103, install the test sample, set the test load, speed, and duration, and then start the friction test; Step S104, collect the multi-source data under the dry friction process of the set material and working condition; Step S105, use the excitation threshold detected in the X-direction vibration signal as the time anchor point to realize the timestamp alignment of the multi-source information; Step S106, use the least mean square adaptive filter to filter the collected signal to suppress the background noise and obtain the filtered multi-source friction information data aligned with the timestamp.

3. The method according to claim 2, wherein, Step S2 specifically includes the following steps: Step S201, use the moving average method to extract the time domain features, frequency domain features, and time-frequency domain features of the multi-source information; Step S202, calculate the Pearson correlation coefficient between the acoustic vibration information and the friction coefficient to obtain the correlation matrix between the acoustic vibration information and the friction coefficient; Step S203, sort the correlation coefficients in the correlation matrix in descending order according to the absolute value, select the friction coefficient features with absolute correlation coefficients greater than 0.8 with any acoustic vibration signal feature, obtain multiple high-correlation friction features, and analyze to obtain multiple high-correlation features, which are all fluctuation features of the friction coefficient; Step S204, in the obtained multiple high-correlation friction features, further retain the 5 acoustic vibration features with the highest correlation corresponding to each feature, and construct the corresponding relationship table between the acoustic vibration signal features and the friction coefficient fluctuation features.

4. The method according to claim 3, wherein, Step S3 specifically includes the following steps: Step S301, use a basic regression model to regress and train the friction coefficient fluctuation, and use a fusion algorithm to integrate and fuse the basic regression model; Step S302, use evaluation indexes to evaluate the fitting effect of the regression model; Step S303, select the model with the highest evaluation index as the prediction model of the friction coefficient fluctuation index, and save it.

5. The method of claim 4, wherein the method further comprises: Step S4 specifically includes the following steps: Step S401, according to the distribution characteristics and statistical laws of the friction coefficient fluctuation index, construct a physical heuristic feature reflecting the relationship between the friction coefficient fluctuation index and the mean index; Step S402, use the extracted fluctuation index and physical heuristic feature as input features, and use the mean index of the friction coefficient as the output variable to construct a data set for regression modeling; Step S403, a mixed training strategy is adopted, and a random forest regression model is used to regress and predict the mean value of the friction coefficient; Step S404, based on the training result, the model fitting is completed, and the trained random forest model is saved as the prediction model of the mean value of the friction coefficient.

6. The method according to claim 5, wherein, The physical heuristic feature is specifically: The derived estimated mean value: Wherein, peak is the peak value in the volatility index; min is the minimum value in the volatility index; The derived estimated root mean square value: Wherein, std is the standard deviation in the volatility index The derived frequency energy value: freq energy = psd mean *hjorth_activity; wherein psd mean is the mean energy spectral density in the volatility indicator; and hjorth_activity is the signal variance in the volatility indicator.

7. The method of claim 3, wherein the method further comprises: In step S201, the time domain feature, the frequency domain feature and the time-frequency domain feature of the multi-source information are extracted by using the moving window method, the width of the moving window is set to the time length corresponding to one rotation of the friction, and the step length is one quarter of the width.

8. The method of claim 4, wherein the method further comprises: In step S301, different volatility indexes of the friction coefficient are taken as target variables, and corresponding acoustic vibration features are selected as input variables according to the corresponding relationship table between the acoustic vibration signal features and the friction coefficient volatility features.

9. The method of claim 8, wherein the method further comprises: The basic regression model includes multiple representative regression algorithms such as linear model, support vector regression based on kernel function, non-parametric model K nearest neighbor regression, tree-based ensemble model, Bayesian regression and artificial neural network.

10. The method of claim 8, wherein the method further comprises: The integrated fusion strategy includes: Weighted average method, according to the performance of each model, the weight is allocated, and the prediction result is weighted and fused; Soft voting regression method, the output results of each model are simply averaged to alleviate the overfitting problem; Stacked fusion, the meta-regressor is constructed to relearn the prediction results of the basic model to capture higher-order nonlinear relationships.

Images