A multi-model driven life prediction method for extra-large bearings at different degradation stages

Abstract

The present invention discloses a multi-model driven life prediction method for super-large bearings under different degradation stages, comprising the following steps: (1) performing signal decomposition on the full-life vibration acceleration data of the super-large bearing, selecting appropriate components for fusion based on the correlation of each component, and obtaining a performance degradation index; (2) calculating the gradient of the performance degradation index, and dividing the degradation stages based on the threshold of the gradient and the degradation index value; (3) selecting a bidirectional gated recurrent unit network suitable for time series prediction in deep learning to construct a first prediction model for the super-large bearing, and a load-life model as the second prediction model, and combining the two to form a third prediction model; (4) formulating an appropriate model use and update strategy based on the first three steps to perform life prediction for different degradation times. The present invention combines sensor monitoring signal data with the mechanism of the super-large bearing itself, and the model has strong robustness and generalization ability, and has certain application value.

Description

Technical Field

[0001] The present invention relates to a multi-model driven extra-large bearing life prediction method under different degradation stages, which is a high-precision remaining service life prediction method based on a mechanism model and a data-driven model. Background Art

[0002] As key transmission components of large-scale mechanical equipment, oversized bearings are widely used in various large rotating equipment. With the vigorous development of fields such as wind power generation and energy extraction, the demand for this type of bearings continues to increase. Since the reliability of oversized bearings determines the stable operation of mechanical equipment to a certain extent, their fault prognostics and health management (PHM) becomes crucial. However, the working environment of oversized bearings is harsh and the working conditions are complex, which makes their PHM somewhat difficult. When a fault occurs, accurate RUL prediction can not only provide a basis for the replacement of parts, but also reduce safety risks and economic losses.

[0003] Scholars both domestically and internationally have conducted extensive research on this topic, with two broad approaches proposed: mechanism-based and data-based. The former studies the damage and degradation mechanisms of equipment, analyzes the causes of failure, and deduces a mathematical model linking lifespan and failure modes. However, this approach fails to account for operational realities and uncertainties. The latter, leveraging condition monitoring technology and modeling using collected sensor data, is suitable for complex mechanical systems or emerging components with unknown damage mechanisms. These approaches can be categorized as statistical models or artificial intelligence algorithms. However, their disadvantages include poor interpretability and the tendency to create a "black box" effect.

[0004] Through the continuous efforts of researchers, the life status assessment of extra-large bearings has made good progress. However, according to the existing literature, the following points still need to be further explored:

[0005] 1) The working conditions of extra-large bearings are complex, and the damage mechanism is difficult to analyze. The use of a single mechanism model for life prediction has the problem of insufficient accuracy.

[0006] 2) There is a lack of historical data on the full lifespan of oversized bearings. Consequently, the data used for training and the data used as input for lifespan prediction have significantly different distributions. Typical data-driven models have poor generalization capabilities, which results in poor performance of older models on new tasks.

[0007] 3) The complete prediction process includes four parts: data acquisition, indicator construction, stage division, and remaining useful life prediction. Many scholars have only proposed new methods or made improvements to one or several of these parts, and these methods do not fully utilize and connect all the previous and subsequent parts. Summary of the Invention

[0008] The purpose of the present invention is to propose a multi-model driven life prediction method for extra-large bearings in different degradation stages. Specifically, the vibration acceleration signal of the extra-large bearing is combined to construct degradation indicators and divide the degradation stages, and multiple models are used to perform update or prediction tasks at different stages. The proposed method utilizes the data of the extra-large bearing in the early stage of degradation for optimization. The model has strong generalization ability and can be used to solve the problem of insufficient historical data on extra-large bearings. The fused model has the advantages of both the mechanism model and the data-driven model, and the prediction accuracy is improved and has a certain degree of interpretability.

[0009] Models trained on different types of bearing data sets cannot meet the prediction requirements of extra-large bearings that require PHM, and the experimental costs of the same type of bearings are too high, resulting in a lack of historical data. The present invention is dedicated to the life assessment method of extra-large bearings under small sample or no sample conditions, and combines the bearing mechanism and monitored data to construct and update the model. The vibration acceleration data is subjected to signal decomposition, and appropriate components are selected according to the correlation to fuse and obtain the performance degradation index; the gradient of the performance degradation index is calculated, and the degradation stage is divided according to the threshold of the gradient and the degradation index value; a bidirectional gated recurrent unit network suitable for time series prediction in deep learning is selected to construct the first prediction model for extra-large bearings, and the load-life model is the second prediction model, and the two are combined to form the third prediction model; a suitable model use and update strategy is formulated, and a mature model is used to predict the life of the degradation intensification stage, which has certain application value.

[0010] The technical solutions of the present invention are as follows:

[0011] A multi-model driven life prediction method for large bearings at different degradation stages includes the following steps:

[0012] Step (1) Obtaining the performance degradation index of the extra-large bearing: Decompose the collected extra-large bearing full life cycle signal using the improved fully adaptive noise ensemble empirical mode decomposition (ICEEMDAN), calculate the correlation coefficient between each component and the original signal, select the root mean square RMS of the four components with the highest correlation coefficient, and fuse them using kernel principal component analysis (KPCA) to obtain the degradation index;

[0013] Step (2), dividing the degradation stage: the gradient of the degradation index in step (1) is calculated every 50 points, and two sets of thresholds are set for the index and gradient values respectively. When both the index and the gradient trigger the set thresholds, two trigger points are recorded; the trigger points divide the degradation of the oversized bearing into three stages;

[0014] Step (3), establishing multiple life prediction models: establishing a data model based on the bidirectional gated cyclic unit, establishing a mechanism model based on the load-life model, and fusing the results of the two to establish a data-mechanism model;

[0015] Step (4), update the life prediction model: use the full life degradation index of the old extra-large bearing to pre-train the data model, then update the model with the prediction results of the two-stage mechanism model by modifying the loss function, and then use the results of the one and two-stage data models to correct the mechanism model; in the third stage, the model is continuously updated, and the data-mechanism model finally obtained can evaluate the remaining service life of the full degradation cycle.

[0016] The steps for obtaining the performance degradation index of the extra-large bearing in step (1) are as follows:

[0017] The signal is decomposed into nine intrinsic mode functions (IMFs) using the improved fully adaptive noise ensemble empirical mode decomposition (ICEEMDAN), denoted as IMF1 to IMF9.

[0018] Calculate the Spearman correlation coefficient between the IMF and the original signal:

[0019]

[0020] Select the four components with the highest ρ values and calculate the RMS for each:

[0021] Kernel principal component analysis (KPCA) was used to reduce the standardized four-dimensional root mean square to one dimension, and the decay index was obtained, which was recorded as DI.

[0022] Among them, ρ i is the value of the Spearman correlation coefficient, i∈(1,2,3,…,9); x is the original vibration signal, and N is the length of the signal data; eliminating the intrinsic mode function IMF components with low correlation can reduce the impact of noise on the signal, and normalizing the root mean square RMS of each component can reduce the impact of the intrinsic mode function IMF with large distribution differences on the fusion result; the improved fully adaptive noise ensemble empirical mode decomposition (ICEENDAN) algorithm is used to better solve the modal aliasing phenomenon existing in traditional empirical mode decomposition, while the kernel principal component analysis (KPCA) algorithm has stronger expressive ability when processing nonlinear data.

[0023] The steps of dividing the degradation stages in step (2) are as follows:

[0024] The gradient of DI is calculated every 10 points, and a total of Gradient values, is the rounding function, the value is expressed as N g ; The formula for obtaining the gradient G(s) is: s is each point in the gradient, s∈(1,2,3,…,N g ),;;

[0025] Two sets of thresholds {Td1, Td2} and {Tg1, Tg2} are set for the degradation index and gradient, respectively. Td and Tg are the thresholds for the degradation index and gradient, respectively. When both the degradation index and gradient trigger thresholds Td1 and Tg1 for the first time, the trigger point t1 is recorded. Similarly, the trigger point t2 can be obtained for thresholds Td2 and Tg2.

[0026] The trigger point divides the degradation of oversized bearings into three stages: healthy stage, initial degradation stage and aggravated degradation stage. t1 is the dividing point between the first two stages, while t2 is the dividing point between the last two stages.

[0027] The proposed degradation stage division method imposes two thresholds on the numerical values of the degradation index DI and the gradient G to constrain them. This improvement can simultaneously constrain both the degree of degradation and the rate of degradation, making the method more robust and rational. Accurate stage division ensures the accuracy of subsequent life prediction.

[0028] The steps for establishing the life prediction model in step (3) are as follows:

[0029] Data-driven model construction: The original vibration signal and DI of the extra-large bearing are both time series data, so the BiGRU network suitable for time series prediction is used as the basis for building the model; compared with traditional recurrent neural networks, the bidirectional structure of BiGRU can capture both past and future information at the same time, thereby more comprehensively exploring the relationship between time series data; the model is trained using the existing historical data of rolling bearings, and the mean square error is used as the loss function to obtain the initial model F and the prediction result y pre The relationship with input DI is: pre =F(feature). Where feature is a four-dimensional feature constructed from the DI and the variance, kurtosis, and waveform indicators of the original signal;

[0030] Mechanism model construction: The mechanism model is established based on the load-life model; the input of the model is the operating parameters of the extra-large bearing and the operating time t, and the calculation formula is: where y phy is the model output, a1, a2, and a3 are life adjustment coefficients, which can be obtained by referring to the ISO 281 manual; C a and P a Represents the rated dynamic load and equivalent thrust load, R is the speed of the oversized bearing, the unit is r / min;

[0031] Model fusion: Calculate the results of the data-driven model and the mechanism model separately, and find the average of the two. The calculation formula is:

[0032] The scheme for updating the life prediction model in step (4) is as follows:

[0033] The lifespan of the first two degradation stages is predicted using the data-driven model and the mechanism model, and y pre and y phy ; Then modify the loss function LF of the data-driven model to update the model; the modified loss function is:

[0034] In the formula, N 1-2 is the data length of the first two degradation stages, y train 、 They represent the model's output for the training set, the training set label, and the model's output for the feature respectively;

[0035] For the updated data-driven model, the feature of the initial degradation stage is used as input to obtain the prediction result y pre , using the least squares method to fit the power function and the initial degradation stage pre Finally, the optimal values of parameters a, b, and c are obtained, and the mechanism model is updated. The new model is:

[0036] After the model is updated, the two new models are brought in for model fusion; the fused model is used to predict the remaining service life in the initial degradation stage, y mul (t) is the prediction result, which can be used to obtain the estimated remaining life at the current moment; after entering the stage of intensified degradation, the new model and signal data are used to continuously repeat steps (1)-(3) to update the model, and the RUL prediction is performed from time t1 to the current time point.

[0037] The beneficial effects of the present invention are:

[0038] 1. The method for constructing degradation indicators for oversized bearings disclosed in this invention allows for intuitive visualization of the bearing's degradation process and trends. This method simplifies signal noise removal and indicator construction into a single step, resulting in smoother degradation indicators and more pronounced trends.

[0039] 2. The disclosed method for classifying degradation stages for oversized bearings uses a threshold-based triggering algorithm based on the DI and DI gradient to create three degradation stages. This method is relatively simple and can be used to classify degradation based on both the degree of degradation and the rate of degradation. Furthermore, two sets of thresholds are used for triggering, reducing the probability of misjudgment.

[0040] 3. The proposed method for predicting the RUL of extra-large bearings combines a mechanistic model with a data-driven model, and provides a method for updating and growing the model at different degradation stages. The resulting model combines the advantages of both mechanistic and data-driven models, fully utilizing data from degradation engineering, and provides insights for predicting the life of rotating machinery under small sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is the prediction flow chart of the present invention.

[0042] Figure 2 This is a data model structure diagram based on BiGRU in the present invention.

[0043] Figure 3 It is the updating process of the model in the present invention.

[0044] Figure 4 It is the original vibration signal diagram in the present invention.

[0045] Figure 5 It is a schematic diagram of the degradation index and degradation stage division in the present invention.

[0046] Figure 6 It is a schematic diagram of the result of life prediction using the method adopted in the present invention.

[0047] Figure 7 It is a schematic diagram of the life prediction results of other existing comparative methods used in the present invention. DETAILED DESCRIPTION

[0048] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0049] Glossary:

[0050] 1. Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN);

[0051] 2. Root Mean Square (RMS);

[0052] 3. Kernel Principal Component Analysis (KPCA);

[0053] 4. Bidirectional Gate Recurrent Unit (BiGRU);

[0054] 5. The data-mechanism model can evaluate the remaining useful life (RUL) of the entire degradation cycle.

[0055] like Figures 1 to 7 As shown, this embodiment describes a method for predicting the life of an extra-large bearing based on a mechanism model and a data model, comprising the following steps:

[0056] Step (1) Obtaining the performance degradation index of extra-large bearings: Using the improved complete adaptive noise ensemble empirical mode decomposition (ICEEMDAN), the collected extra-large bearing full life cycle signal is decomposed to obtain the correlation coefficient between each component and the original signal. The root mean square (RMS) of the four components with the highest correlation coefficient is selected and fused using kernel principal component analysis (KPCA) to obtain the degradation index.

[0057] Step (2) Degradation Stages: The gradient of the degradation index in step (1) is calculated every 50 points. Two thresholds are set for the index and gradient values, respectively. When both the index and gradient trigger the set thresholds, two trigger points are recorded. The trigger points divide the degradation of the oversized bearing into three stages.

[0058] Step (3) establishes multiple life prediction models: a data model is established based on a bidirectional gate recurrent unit (BiGRU), a mechanism model is established based on a load-life model, and the results of the two are fused to establish a data-mechanism model.

[0059] Step (4) Update the life prediction model: Pre-train the data model using the full-life degradation indicators of old oversized bearings. Then, by modifying the loss function, update the model with the prediction results of the two-stage mechanism model. Then, use the results of the first and second stage data models to correct the mechanism model. In the third stage, the model is continuously updated. The resulting data-mechanism model can be used to estimate the remaining useful life (RUL) of the full degradation cycle.

[0060] The steps for obtaining the performance degradation index of the extra-large bearing in step (1) are as follows:

[0061] 1) Use ICEEMDAN to decompose the signal into nine intrinsic mode functions (IMFs), denoted as IMF1 to IMF9;

[0062] 2) Calculate the Spearman correlation coefficient between the IMF and the original signal:

[0063]

[0064] 3) Select the four components with the highest ρ values and calculate the RMS:

[0065] 4) KPCA is used to reduce the standardized four-dimensional RMS matrix to one dimension to obtain the decay index, which is recorded as DI.

[0066] Among them, ρ i is the value of the Spearman correlation coefficient, i∈(1,2,3,…,9). x is the original vibration signal, and N is the length of the signal data. Eliminating the IMF components with low correlation can reduce the impact of noise on the signal, and normalizing the RMS of each component can reduce the impact of IMF with large distribution differences on the fusion result. Figure 4 It can be seen that due to the harsh working environment, the effective information in the monitored vibration acceleration signal of the original vibration signal is often covered by a large amount of noise, which affects the accuracy of the signal indicator construction and stage division. Signal denoising depends on efficient decomposition and fusion, so the ICEENDAN algorithm is used for signal decomposition. This method solves the modal aliasing phenomenon existing in traditional empirical mode decomposition. First, the original signal x is decomposed by EMD to obtain 9 IMF components and a residual term R. Then, each IMF component is used as the target component to perform an adaptive noise set operation to obtain 9 sets containing several noise components, and finally the components in the noise set are added to the target component. Among the obtained IMF components, IMF1, IMF2, IMF3, and IMF5 have the largest correlation. The KPCA algorithm is used for signal fusion, which has stronger expression ability when processing nonlinear data. By Figure 5 It can be seen that the degradation index obtained is as follows: points 300-1100 and points 1300-2100 are relatively stable, points 1100-1300 and after point 2100 have two rapid rises, and the overall monotonicity of the curve is obvious, which is in line with the degradation law of extra-large bearings.

[0067] The steps of dividing the degradation stages in step (2) are as follows:

[0068] 1) Obtain the gradient of DI every 10 points, and obtain Gradient values, is the rounding function, the value is expressed as N gThe formula for obtaining the gradient G(s) is:

[0069] s is each point in the gradient, s∈(1,2,3,…,N g ),;

[0070] 2) Set two sets of thresholds {Td1, Td2} and {Tg1, Tg2} for the degradation index and gradient, where Td and Tg are the thresholds for the degradation index and gradient, respectively. When both the degradation index and gradient trigger thresholds Td1 and Tg1 for the first time, record the trigger point t1. Similarly, for thresholds Td2 and Tg2, the trigger point t2 can be obtained.

[0071] 3) Trigger point The degradation of oversized bearings is divided into three stages: healthy stage, initial degradation stage and aggravated degradation stage. t1 is the dividing point between the first two stages, while t2 is the dividing point between the last two stages.

[0072] The proposed degradation stage division method imposes two thresholds on the values of the degradation index DI and the gradient G. This improvement can simultaneously constrain both the degree of degradation and the rate of degradation, making the method more robust and reasonable. Accurate stage division ensures the accuracy of subsequent life prediction. Figure 5 As shown, the two thresholds are set as {0.5, 0.7} and {2×10 -3 , 5×10 -4}, trigger points t1 and t2 are 1190 and 2220, respectively. Constraining both the degradation index DI and the gradient G can prevent false triggering of the gradient due to a surge in data within a short period, as well as premature triggering due to DI value fluctuations during a stable phase. As can be seen, the three distinct phases have different data distributions, and each phase is of similar length.

[0073] The steps for establishing the life prediction model in step (3) are as follows:

[0074] 1) Data-driven model construction:

[0075] The original vibration signal and DI of the extra-large bearing are both time series data, so the BiGRU network suitable for time series prediction is used as the basis for building the model. Compared with the traditional recurrent neural network, the bidirectional structure of BiGRU can capture both past and future information at the same time, thereby more comprehensively exploring the relationship between time series data. The model structure is as follows: Figure 2 As shown. The model is trained using the existing rolling bearing historical data, and the loss function uses the mean square error to obtain the initial model F and the prediction result y pre The relationship with input DI is: pre =F(feature). Feature is a four-dimensional feature constructed from the DI and the variance, kurtosis, and waveform indicators of the original signal.

[0076] Figure 2 In the GRU network module on the right, there is the following calculation formula: t =σ(W z ·[h t-1 ,x t ]), r t =σ(W r ·[h t-1 ,x t ]), x t is the current input, and the hidden layer outputs h t , σ is the sigmoid function. W represents the weight of each part, Used to summarize input and historical hidden layers.

[0077] 2) Mechanism model construction:

[0078] The mechanism model is established based on the load-life model. The input of the model is the operating parameters of the oversized bearing and the operating time t. The calculation formula is: where y phy is the model output, a1, a2, and a3 are life adjustment factors, which can be obtained by referring to the ISO 281 manual. a and P a Represents the rated dynamic load and equivalent thrust load, R is the speed of the oversized bearing, the unit is r / min.

[0079] 3) Model fusion: Calculate the results of the data-driven model and the mechanism model separately, and find the average of the two. The calculation formula is:

[0080] The scheme for updating the life prediction model in step (4) is as follows:

[0081] The lifespan of the first two degradation stages is predicted using the data-driven model and the mechanism model, and y pre and y phy Then modify the loss function LF of the data-driven model to update the model. The modified loss function is:

[0082]

[0083] In the formula, N 1-2 is the data length of the first two degradation stages, y train 、 They represent the model's output for the training set, the training set label, and the model's output for the feature.

[0084] For the updated data-driven model F', the feature of the initial degradation stage is used as input to obtain the prediction result y pre , using the least squares method to fit the power function and the initial degradation stage pre Finally, the optimal values of parameters a, b, and c are obtained, and the mechanism model is updated. The new model is:

[0085] After the model is updated, the two new models are brought in for model fusion. The fused model is used to predict the remaining service life in the initial degradation stage. mul (t) is the prediction result, which can be used to obtain the estimated remaining life at the current moment. After entering the degradation intensification stage, the new model and signal data are used to continuously repeat steps (1)-(3) to update the model, and the RUL prediction is performed from time t1 to the current time point. The update process and update strategy of the life prediction model are as follows: Figure 3 As shown in the figure, the degradation index DI is divided into multiple stages, and different models and tasks are assigned to each stage. Each step, such as degradation index establishment, degradation stage division, and model establishment, is fully connected and applied, realizing the RUL prediction of extra-large bearings based on multi-model data-mechanism under small samples.

[0086] The prediction results are as follows Figure 6 As shown in the figure, the predicted RUL results fluctuate around the actual value, indicating high prediction accuracy. After the bearing enters the stage of severe degradation, the prediction accuracy increases significantly. This is because the incorporation of the mechanism model reduces the fluctuation of the curve, and the sampling points almost all fall on the actual RUL line.

[0087] In order to prove the effectiveness of the proposed method, the initial BiGRU data model without update, the long short-term memory network combined with transfer learning (Transfer Learning Long Short-term Memory, referred to as TL-LSTM), and the BiGRU combined with transfer learning (TL-BiGRU) were introduced for comparative analysis. The prediction results are as follows Figure 7 The results further demonstrate the superiority and reliability of the proposed method. As can be seen from the figure, while the initial data model, which has not been updated and adjusted, has high accuracy on training samples, it performs very poorly on datasets with significantly different data distributions. While the LSTM and BiGRU network models fine-tuned using model transfer have significantly improved prediction accuracy, the prediction curves show significant fluctuations, and the accuracy is low during periods of severe degradation.

[0088] In order to more intuitively compare the prediction performance of each model, the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are used for comparison and measurement. The calculation formulas are as follows:

[0089]

[0090]

[0091] Where n is the total number of prediction points, y real (t) is the actual RUL, and y(t) is the predicted RUL. MAE reflects the magnitude of the prediction error, RMSE characterizes the dispersion of the prediction, and MAPE reflects the degree of deviation between the prediction and the actual RUL. The prediction errors of each model are shown in Table 1.

[0092] Table 1 Prediction error of each model

[0093]

[0094] As can be seen from the table, compared with the BiGRU initial model, TL-LSTM and TL-BiGRU, the three prediction errors of the model proposed in this paper are significantly reduced, verifying the advantages of this model in the RUL prediction task of extra-large bearings. It effectively solves the problems of poor prediction results caused by the lack of full life history data of extra-large bearings and the large difference in distribution between training samples and actual data, and provides new ideas for the intelligent operation and maintenance of bearings.

[0095] The above description is merely a preferred embodiment of the present invention and does not limit the technical scope of the present invention. Therefore, any minor modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A multi-model driven life prediction method for large bearings at different degradation stages, characterized by The following steps are involved: Step (1) Obtaining the performance degradation index of the extra-large bearing: Decomposing the collected extra-large bearing full life cycle signal using the improved fully adaptive noise ensemble empirical mode decomposition (ICEEMDAN). The specific implementation of the signal decomposition by the ICEEMDAN algorithm is as follows: First, perform EMD decomposition on the original signal x to obtain 9 IMF components and a residual term R. Then, perform adaptive noise set operation on each IMF component as the target component to obtain 9 sets containing several noise components. Finally, the components in the noise set are added to the target component. Calculate the correlation coefficient between each component and the original signal, select the root mean square RMS of the four components with the highest correlation coefficient, and use kernel principal component analysis KPCA to fuse them to obtain the decay index; Step (2), dividing the degradation stage: the gradient of the degradation index in step (1) is calculated every 10 points, and two sets of thresholds are set for the index and gradient values respectively. When both the index and the gradient trigger the set thresholds, two trigger points are recorded; the trigger points divide the degradation of the oversized bearing into three stages; Step (3), establish multiple life prediction models: establish a data model based on the bidirectional gated recurrent unit BiGRU, establish a mechanism model based on the load-life model, and fuse the results of the two to establish a data-mechanism model; Step (4), update the life prediction model: use the full life degradation index of old extra-large bearings to pre-train the data model, then update the model with the prediction results of the two-stage mechanism model by modifying the loss function, and then use the results of the first and second stage data models to correct the mechanism model; in the third stage, the model is continuously updated, and the data-mechanism model finally obtained can evaluate the remaining service life RUL of the full degradation cycle.

2. The multi-model driven extra-large bearing life prediction method at different degradation stages according to claim 1 is characterized in that: The steps for obtaining the performance degradation index of the extra-large bearing in step (1) are as follows: 1) Use ICEEMDAN to decompose the signal into 9 intrinsic mode functions (IMFs), denoted as IMF1 to IMF9; 2) Calculate the Spearman correlation coefficient between the IMF and the original signal: ； 3) Select ρ Calculate the RMS of the four components with the highest values: ; 4) Use KPCA to reduce the standardized four-dimensional root mean square to one dimension and obtain the decay index, recorded as DI, where ρ i is the value of the Spearman correlation coefficient, i∈(1,2,3,…,9), x is the original vibration signal, and N is the length of the signal data. Eliminating the IMF components with low correlation can reduce the impact of noise on the signal. Normalizing the RMS of each component can reduce the impact of IMFs with large distribution differences on the fusion result. The ICEENDAN algorithm is used to better solve the modal aliasing phenomenon existing in traditional empirical mode decomposition, while the KPCA algorithm has stronger expression ability when processing nonlinear data.

3. The multi-model driven extra-large bearing life prediction method at different degradation stages according to claim 2 is characterized in that: The steps for dividing the degradation stage in step (2) are as follows: 1) Calculate the gradient of DI every 10 points, and obtain a total of ⌈N / 10⌉ gradient values. ⌈·⌉ is the upward rounding function, and the value is expressed as N g ; The formula for obtaining the gradient G(s) is: ; s is each point in the gradient, s∈(1,2,3,…,N g ); 2) Set two sets of thresholds {Td1, Td2} and {Tg1, Tg2} for the degradation index and gradient, where Td and Tg are the thresholds for the degradation index and gradient, respectively. When both the degradation index and gradient trigger thresholds Td1 and Tg1 for the first time, record the trigger point t1. Similarly, for thresholds Td2 and Tg2, the trigger point t2 can be obtained. 3) Trigger point: The degradation of oversized bearings is divided into three stages: healthy stage, initial degradation stage and aggravated degradation stage, where t1 is the dividing point between the first two stages, and t2 is the dividing point between the last two stages; The proposed degradation stage division method imposes two threshold constraints on the values of the degradation index DI and the gradient G. This improvement can simultaneously constrain both the degree of degradation and the rate of degradation. Therefore, this method has better robustness and rationality. Accurate stage division ensures the accuracy of subsequent life prediction.

4. The multi-model driven extra-large bearing life prediction method at different degradation stages according to claim 2 is characterized in that: The steps for establishing the life prediction model in step (3) are as follows: 1) Data-driven model construction: The original vibration signal and DI of the extra-large bearing are both time series data. Therefore, the BiGRU network suitable for time series prediction is used as the basis for building the model. Compared with the traditional recurrent neural network, the bidirectional structure of BiGRU can capture both past and future information at the same time, thereby more comprehensively mining the relationship between time series data. The model is trained using the existing historical data of rolling bearings, and the loss function uses the mean square error to obtain the initial model F and the prediction results. y pre The relationship with input DI is: , where feature is a four-dimensional feature constructed by DI and the variance, kurtosis, and waveform indicators of the original signal; 2) Mechanism model construction: The mechanism model is established based on the load-life model. The input of the model is the operating parameters of the extra-large bearing and the operating time t. The calculation formula is: ,in y phy is the model output, a1, a2, and a3 are life adjustment coefficients, which can be obtained by referring to the ISO 281 manual, and C a and P a Represents the rated dynamic load and equivalent thrust load, R is the speed of the oversized bearing, the unit is r / min; 3) Model fusion: Calculate the results of the data-driven model and the mechanism model separately, and find the average of the two. The calculation formula is: .

5. The multi-model driven extra-large bearing life prediction method at different degradation stages according to claim 4 is characterized in that: The scheme for updating the life prediction model in step (4) is as follows: 1) Use data-driven model and mechanism model to predict the life of the first two degradation stages and obtain y pre and y phy Then modify the loss function LF of the data-driven model to update the model. The modified loss function is: ;In the formula, N 1-2 is the data length of the first two degradation stages, 、 、 They represent the model's output for the training set, the training set label, and the model's output for the feature respectively; 2) For the updated data-driven model, the feature of the initial degradation stage is used as input to obtain the prediction result y pre , using the least squares method to fit the power function and the initial degeneration stage y pre ， Finally, the optimal values of parameters a, b, and c are obtained, and the mechanism model is updated. The new model is: ; 3) After the model is updated, the two new models are brought in for model fusion, and the fused model is used to predict the remaining service life in the initial degradation stage. That is the prediction result, which can be used to obtain the estimated remaining life at the current moment. After entering the stage of intensified degradation, the new model and signal data are used to continuously repeat steps (1)-(3) to update the model and perform RUL prediction from time t1 to the current time point.

Images