Single-source domain generalization based cross-domain bearing life prediction method and system

Abstract

The application relates to the field of bearing life prediction, in particular to a cross-domain bearing life prediction method and system based on single-source field generalization. In order to solve the defect that the bearing life prediction precision is low in the prior art, the predictor used in the application is composed of a data preprocessing module and a life prediction module, the data preprocessing module is used for constructing a data pair sequence composed of a time domain characteristic peak-peak value time sequence of vibration data samples and normalized values of the time sequence as input of the life prediction module, so as to predict the residual life of the bearing at the vibration data sample collection time. The application is favorable for avoiding the data distribution difference of different bearings, and ensures the accuracy of the predictor obtained by learning the source bearing for predicting the target bearing. In this way, while increasing the diversity of the data distribution through the source bearing data, the application also avoids generating counterfactual data on the basis of increasing the data diversity, and improves the generalization ability of the prediction model.

Description

Technical Field

[0001] This invention relates to the field of bearing life prediction, and in particular to a cross-domain bearing life prediction method and system based on single-source domain generalization. Background Technology

[0002] Bearings, as a crucial component of rotating machinery, are vital to the operation of industrial equipment. Bearing failure directly impacts the safety of the entire equipment and results in economic losses. Therefore, predicting the remaining life of bearings plays a significant role in equipment prognosis and health management.

[0003] Existing deep learning models have achieved great success in predicting the remaining life of bearings, but they assume that training and testing data follow the same data distribution. However, in real-world scenarios involving the prediction of the remaining life of different bearings, vibration data collected on unknown bearings cannot be guaranteed to be independent and identically distributed with the training vibration data due to variations in operating conditions or materials. Further collecting full-lifecycle vibration data for all types of bearings is impractical. If existing learning models are trained on a single bearing and then directly applied to unknown bearings, the model's performance will significantly degrade.

[0004] In existing technologies, single-source domain generalization methods have broad application prospects for predicting the remaining life of unknown bearings. These methods can model from a single source domain with high generalization ability, increasing the diversity of data distribution and leading to better results in other unknown target domains. However, currently, single-source domain generalization methods are not applied to bearing remaining life prediction, nor can they be directly applied to it. There are two main reasons for this: First, bearing vibration datasets consist of time-series data covering the entire lifespan, with a large amount of small fluctuations in the early stages. These early data have little or no impact on describing the bearing degradation trend, and may even have a negative effect. Therefore, for more accurate lifespan prediction, the vibration dataset for the entire lifespan should be more rationally divided to reduce interference from early data. Second, existing single-source domain generalization methods increase the diversity of data distribution through data augmentation techniques. However, this image-oriented augmentation technique often introduces the uncertainty of generating "counterfactual data," which is completely inconsistent with the actual vibration data distribution of the bearing, thus degrading the model's performance. Summary of the Invention

[0005] To address the shortcomings of low bearing life prediction accuracy in the existing technologies, this invention proposes a cross-domain bearing life prediction method and system based on single-source domain generalization, which can achieve high-precision cross-domain prediction.

[0006] The present invention adopts the following technical solution:

[0007] A cross-domain bearing life prediction method based on single-source domain generalization is proposed. First, a predictor is obtained through machine learning. The predictor includes a data preprocessing module and a life prediction module. The data preprocessing module processes the vibration data of the bearing, and the life prediction module uses the processed data as input data to predict the remaining life of the bearing. Then, the vibration data of the target bearing is obtained and input into the predictor to obtain the predicted life value output by the predictor.

[0008] The data preprocessing module processes the bearing vibration data in the following steps:

[0009] SA1. Obtain the bearing vibration data D = {d} 1 ,d 2 ,…,d n ,…,d N}, d n This represents the nth vibration data sample of the bearing; the vibration data sample is composed of multiple data points collected per unit time according to the set acquisition frequency, and D is composed of vibration data samples collected in continuous time; n is the ordinal number, and N is the number of vibration data samples in the vibration data D, 1≦n≦N;

[0010] SA2, Obtain the input time series X = {x 1 ,x 2 ,…,x i ,…,x I}, i is the ordinal number, 1 ≤ i ≤ I, m is the step size, I = N - m + 1; x i This represents the i-th time series sample; x represents i The j-th data point, i.e., vibration data sample d i+j-1 The peak-to-peak value of the time-domain characteristic; 1≦j≦m;

[0011] SA3. Normalize the input time series X, let x i The value after normalization is

[0012] SA4, the data preprocessing module outputs the following data:

[0013] Preferably, the vibration data D of the bearing is obtained by filtering from the original data H of the bearing. The filtering of vibration data D includes the following steps:

[0014] SB1. The original data H of the bearing is denoted as H = {h} 1 ,h 2 ,…,h r ,…,hR}, h r Let represent the r-th vibration data sample collected during bearing monitoring, 1 ≦ r ≦ R, where R represents the total number of vibration data samples collected; extract the time-domain characteristic peak-to-peak value set V = {v} from the original data H. 1 ,v 2 ,…,v r ,…,v R The isotonic regression algorithm is used on the time-domain characteristic peak-to-peak set V to obtain degraded data with a monotonic trend. Indicates degraded data The r-th value in;

[0015] SB2, Construct a sliding window of length z for degenerate data Perform sliding values ​​to obtain the degradation gradient of the data in each window, and take the value in the k-th window. The corresponding degradation gradient is Δ k , 1≦k≦R-z+1;

[0016] SB3, Constructing a Degenerate Continuous Domain and Representing the degenerate continuous domain The starting and ending values, The initial value is The initial value is

[0017] SB4, Calculation Degenerate gradient Δ for all window values s Δ s+1 , ..., Δ d-z+1 The mean is denoted as Av. (The last part, "judging," is incomplete and requires further context.) Does the following condition 1-2 meet? If yes, proceed to step SB5; if not, proceed to step SB6.

[0018] Condition 1: There exists a continuous fluctuation domain. and Representing the continuous wave domain The starting and ending values, The degradation gradient Δ corresponding to all window values. b Δ b+1 , ..., Δ c-z+1 All are greater than 0, and The number of values ​​that the window can take, c-z+1-b+1, is greater than or equal to the set value C0, where C0>0;

[0019] Condition 2: The number of values ​​in the window corresponding to the degradation gradient greater than the mean Av is Q, where Q is greater than or equal to a set value Q0, 0 <Q0≤C0;

[0020] SB5, Update the degenerate continuum Then return to SB4;

[0021] SB6, with As the final degradation point, obtain the original data {h d ,h d+1 ,…,h R} as vibration data D={d 1 ,d 2 ,…,d n ,…,d N}, that is, d n =h d+n-1 .

[0022] Preferably, step SB4 specifically includes the following sub-steps:

[0023] SB41, Calculation The mean Av of the degradation gradients for all window values;

[0024] SB42, Order The initial value of t is s;

[0025] SB43, Calculation Values ​​retrieved from the upper window The degradation gradient Δ t ;

[0026] SB44, Determine Δ t Is it greater than 0? If yes, update t = t + 1 and return to step SB43; otherwise, execute step SB45.

[0027] SB45. Determine if tb is greater than or equal to C0; if not, update t = t + 1 and return to step SB42; if yes, execute the following step SB46.

[0028] SB46, Data The corresponding degradation gradient Δ b Δ b+1 , ..., Δ t-1 Let Q be the number of data points greater than the mean Av. Determine if Q is greater than or equal to a set value Q0; if not, update t = t + 1 and return to step SB42; if yes, let the data... As a continuous wave domain Right now

[0029] Preferably, the lifetime prediction module includes a preliminary prediction network, an inverse normalization unit, and an output unit; the preliminary prediction network is used to obtain x i The corresponding first predicted value y i and The corresponding second predicted value The inverse normalization unit is used for the second predicted value Perform inverse normalization to obtain inverse normalized predicted values. The output of the output unit is the output of the predictor, and the output result of the output unit is the remaining lifetime sequence. For bearing vibration data sample d i+m-1 The remaining lifetime prediction value over the acquisition time;

[0030]

[0031] λ is a learnable affine parameter.

[0032] Preferred:

[0033]

[0034]

[0035]

[0036] Ex i x represents i The average value, Varx i x represents i The standard deviation; γ and β are learnable affine parameters; ε is a set constant, ε is in the interval

[10] -6 10 -5 [Values ​​can be taken from the above].

[0037] Preferred:

[0038]

[0039] Preferably, the acquisition of the predictor includes the following steps:

[0040] S1. Construct and initialize a predictor consisting of a data preprocessing module and a lifetime prediction module;

[0041] S2. Obtain vibration data D from the source bearing. S , Construct multiple learning samples I = N - m + 1, the vibration data sample consists of multiple data points collected per unit time according to the set sampling frequency, D S It consists of vibration data samples collected over a continuous period of time; D represents S The nth vibration data sample in the data; For the source bearing in vibration data samples The remaining lifetime in terms of collection time, The data is manually labeled as true values; the predictor is then subjected to machine learning on the training samples to iterate the parameters until the predictor's loss reaches the set loss threshold, at which point the predictor is fixed.

[0042] The loss function of the predictor is:

[0043]

[0044] The source bearing for the predictor output in the vibration data sample The remaining lifetime prediction value over the acquisition time.

[0045] Preferably, the preliminary prediction network is constructed based on a single-layer gated recurrent unit (GRU).

[0046] Preferably, the present invention also proposes a cross-domain bearing life prediction system based on single-source domain generalization, used to implement the aforementioned cross-domain bearing life prediction method based on single-source domain generalization. The cross-domain bearing life prediction system based on single-source domain generalization includes a memory storing a computer program, which, when executed, implements the aforementioned cross-domain bearing life prediction method based on single-source domain generalization.

[0047] Preferably, it also includes a processor connected to a memory, the processor being used to execute the computer program to implement the above-described cross-domain bearing life prediction method based on single-source domain generalization.

[0048] The advantages of this invention are:

[0049] (1) The predictor used in this invention consists of a data preprocessing module and a life prediction module. The data preprocessing module is used to construct a data pair sequence consisting of the time-domain characteristic peak-to-peak time series of vibration data samples and the normalized value of the time series, which serves as the input to the life prediction module to predict the remaining life of the bearing at the time of vibration data sample acquisition. This invention helps to avoid the differences in data distribution among different bearings, ensuring the accuracy of the predictor obtained by learning from the source bearing in predicting the target bearing. In this way, while increasing the diversity of data distribution through source bearing data, it also avoids generating counterfactual data based on increasing data diversity, and improves the generalization ability of the prediction model.

[0050] (2) Before the bearing life prediction, the present invention designs an adaptive threshold stage division method based on the isotonic regression algorithm, which simply and effectively identifies the degradation points of different bearing vibration data, and further divides the degradation stage through the degradation points, so as to select the original data of the degradation stage as vibration data for the remaining life prediction, thereby reducing the interference of early random fluctuations on the remaining life prediction of different bearings.

[0051] (3) By screening the data degradation stage, this invention overcomes the defects of large differences in the data distribution between training bearings and unknown test bearings and large amplitude changes in the rapid degradation stage, and greatly improves the accuracy of cross-domain prediction.

[0052] (4) In this invention, the data is normalized in the data preprocessing stage and the prediction results are inversely normalized in the lifetime prediction stage. In this way, the bearing data that is normalized by design instance and reversibly normalized in parallel is combined into a unified lifetime prediction learning framework by a predictor based on gated loop unit. This avoids generating counterfactual data on the basis of increasing data diversity, and at the same time improves the generalization ability of the unknown bearing model. Attached Figure Description

[0053] Figure 1 Here is a flowchart of a vibration data screening method;

[0054] Figure 2 for Upper continuous wave domain Iterative flowchart;

[0055] Figure 3 Here is a flowchart of the data preprocessing process;

[0056] Figure 4 This is a graph of degraded data processed using the isotonic regression algorithm;

[0057] Figure 5 Iterative graph of adaptive threshold algorithm;

[0058] M1: The average value Av from the first algorithm iteration;

[0059] M2: The average value Av from the second algorithm iteration;

[0060] M3: The average value Av from the 3rd algorithm iteration;

[0061] Mn: The average value Av of the nth algorithm iteration;

[0062] D1: First Algorithm Iteration Right now

[0063] D2: Second Algorithm Iteration

[0064] D3: 3rd Algorithm Iteration

[0065] Dn: the nth algorithm iteration

[0066] Figure 6 Predicted remaining service life of different tested bearings. Detailed Implementation

[0067] Vibration data screening method

[0068] Reference Figure 1 This embodiment proposes a vibration data filtering method for filtering vibration data D from the raw bearing data H to predict the remaining life of the bearing.

[0069] H = {h} 1 ,h 2 ,…,h r ,…,h R}, h r H represents the r-th vibration data sample collected in bearing monitoring, where 1 ≦ r ≦ R, and R represents the total number of vibration data samples collected. A vibration data sample is composed of multiple data points collected per unit time according to a set collection frequency. H is composed of vibration data samples collected over a continuous time period.

[0070] Vibration data is denoted as D = {d} 1 ,d 2 ,…,d n ,…,d N}, where N represents the number of vibration data samples in vibration data D, and d n Let n represent the nth vibration data sample in vibration data D, where 1 ≦ n ≦ N.

[0071] In this embodiment, the method for filtering vibration data D from the original data H includes the following steps SB1-SB5.

[0072] SB1. Extract the time-domain characteristic peak-to-peak value set V = {v} from the original data H. 1 ,v 2 ,…,v r ,…,v R}, using the isotonic regression algorithm on the time-domain characteristic peak-to-peak set V, we obtain, as follows Figure 4 The degradation data shown exhibits a monotonic trend. Indicates degraded data The r-th value in the array.

[0073] SB2, Construct a sliding window of length z for degenerate data Perform sliding value extraction to obtain the degradation gradients of the data in each window;

[0074] In this step, the total number of window value extractions obtained is R - z + 1, and the value extracted from the k-th window is The corresponding degradation gradient is Δ k , where 1 ≤ k ≤ R - z + 1.

[0075] SB3. Construct the degradation continuous domain and respectively represent the starting value and the ending value of the degradation continuous domain , The initial value of The initial value of

[0076] SB4. Calculate The mean value of the degradation gradients of all the window value extractions on is denoted as Av; determine whether

[0077] satisfies the following conditions 1 - 2; if it does, execute step SB5; if not, execute step SB6; There exists a continuous fluctuation domain and respectively represent the starting value and the ending value of the continuous fluctuation domain , For all the window value extractions on b 、Δ b+1 、……、Δ<于 c-z+1 are all greater than 0, and The number of window value extractions c - z + 1 - b + 1 on

[0078] Condition 2: The number of window value extractions on b 、Δ b+1 、……、Δ c-z+1 whose corresponding degradation gradients are greater than the mean value Av is Q, and Q is greater than or equal to the set value Q0, 0 < Q0 ≤ C0; Q is equal to the number of data in Δ

[0079] SB5. Update the degradation continuous domain Then return to SB4. <000038​​​​​​​​​​,d 2 ,…,d n ,…,d N}, that is, d n =h d+n-1 .

[0081] In SB3, The correspondence between all window values ​​and the degradation gradient is as follows:

[0082]

[0083]

[0084] ...

[0085]

[0086] Av=(Δ s +Δ s+1 +……+Δ d-z+1 ) / (d-z+1-s+1)

[0087] Reference Figure 2 Step SB4 specifically includes the following sub-steps:

[0088] SB41, Calculation The mean Av of the degradation gradients for all window values;

[0089] SB42, Order The initial value of t is s;

[0090] SB43, Calculation Values ​​retrieved from the upper window The degradation gradient Δ t ;

[0091] SB44, Determine Δ t Is it greater than 0? If yes, update t = t + 1 and return to step SB43; otherwise, execute step SB45.

[0092] SB45. Determine if tb is greater than or equal to C0; if not, update t = t + 1 and return to step SB42; if yes, execute the following step SB46.

[0093] SB46, Data The corresponding degradation gradient Δ b Δ b+1 , ..., Δ t-1 Let Q be the number of data points greater than the mean Av. Determine if Q is greater than or equal to a set value Q0; if not, update t = t + 1 and return to step SB42; if yes, let the data... As a continuous wave domain Right now

[0094] The iterative process is as follows: Figure 5 As shown.

[0095] Predictor

[0096] The predictor provided in this embodiment is obtained through machine learning. The predictor includes a data preprocessing module and a lifetime prediction module. The lifetime prediction module includes a preliminary prediction network, an inverse normalization unit, and an output unit.

[0097] Reference Figure 3 The data preprocessing module is used to process the vibration data of the bearing. Its input is the vibration data of the bearing to be predicted, D = {d}. 1 ,d 2 ,…,d n ,…,d N}, its output data x i All are transition parameters; x represents i The j-th data point, i.e., vibration data sample d i+j-1 The corresponding time-domain characteristic peak value, i.e., vibration data sample d i+j-1 The difference between the maximum amplitude and the minimum amplitude, where m is the step size; For x i The value after normalization; I = N-m+1, 1≦i≦I, 1≦j≦m.

[0098]

[0099]

[0100]

[0101] Ex i x represents i The average value, Varx i x represents i The standard deviation; γ and β are learnable affine parameters; ε represents a set constant; ε takes a very small value to prevent errors in square root calculations, specifically a value of 10. -6 .

[0102] The initial prediction network is used to obtain x i The corresponding first predicted value y i and The corresponding second predicted value

[0103] The inverse normalization unit is used for the second predicted value Perform inverse normalization to obtain inverse normalized predicted values.

[0104]

[0105] The output of the output unit is the output of the predictor, and the output result of the output unit is the remaining lifetime sequence. Corresponding data For vibration data sample d i+m-1 The corresponding time-domain characteristic peak value, therefore For the bearing vibration data sample d in vibration data D i+m-1 The remaining lifetime in terms of collection time.

[0106]

[0107] Where λ represents the learnable affine parameter;

[0108] In this embodiment, the predictor updates its parameters by learning the vibration data of the source bearing in the initial state.

[0109] The learning process of a predictor includes the following steps:

[0110] SC1. Obtain vibration data of the source bearing D S , Construct multiple learning samples I = N - m + 1, where N is the number of vibration data samples in the vibration data, and each vibration data sample consists of multiple data points collected per unit time according to a set sampling frequency. D S It consists of vibration data samples collected over a continuous period of time; D represents S The nth vibration data sample in the data; For the source bearing in vibration data D S Vibration data samples The remaining lifetime in terms of collection time, The data is manually labeled as true values; multiple learning samples are split into training samples and validation samples; the predictor is initialized.

[0111] SC2. Select n1 training samples, and let the predictor learn from the selected training samples and update the predictor parameters.

[0112] SC3. Select one validation sample to calculate the updated predictor loss. The loss function is:

[0113]

[0114] in, The source bearing for the predictor output in the vibration data sample The remaining lifetime prediction value over the acquisition time; and The predicted and actual values ​​of remaining lifespan per unit of time;

[0115] SC4. Determine if RMSELoss is less than or equal to the set loss threshold; if not, return to SC2; if yes, the predictor training is complete.

[0116] A cross-domain bearing life prediction method based on single-source domain generalization

[0117] In this embodiment, vibration data of the source bearing is first acquired to train the predictor; then, vibration data of the target bearing D = {d} is acquired. 1 ,d 2 ,…,d n ,…,d N The vibration data of the target bearing is input into the predictor to obtain the remaining life sequence output by the predictor. For vibration data sample d in vibration data D i+m-1 The remaining lifetime prediction value over the acquisition time.

[0118] It is worth noting that in this embodiment, the input data of the predictor can be vibration data filtered from the original data, or the original data can be used directly; using the filtered vibration data is beneficial to further improve the accuracy of the predictor.

[0119] Example

[0120] In this embodiment, the IEEE PHMChallenge2012 bearing dataset provided by the PRONOSTIA test platform is used to verify the performance of the predictor provided by this invention.

[0121] The PHMChallenge2012 dataset: This test bench consists of three main parts: a rotating section, a degradation generation section (used to apply radial force to the bearing under test), and a measurement section. The rotating section consists of an asynchronous motor with a gearbox and two shafts with a power of 250W, which transmit rotational motion through the gearbox. The measurement section includes a load assembly and multiple sensors. The load assembly provides a load of 4000N, which can rapidly degrade bearing performance. The characteristics of bearing degradation are determined based on two data types from the sensors: vibration and temperature. The accelerometer samples at a frequency of 25.6kHz, recording the measured vibration signal every 10 seconds, with each acquisition lasting 0.1 seconds, and 2560 data points acquired per acquisition. The experiment stops when the amplitude of the vibration signal reaches 20g, assuming complete bearing failure; g represents gravitational acceleration, g = 9.8 m / s². 2 .

[0122] In this embodiment, bearing degradation experiments under three working conditions were conducted on this test bench, as shown in Table 1.

[0123] Table 1 Basic Information of PHM2012 Bearing

[0124]

[0125] In this embodiment, the first bearing under the first working condition is first used to train the predictor. Then, the trained predictor is used to predict the remaining service life of the bearings 1_2, 1_3, 1_5, and 1_6 under the first working condition; the bearings 2_1, 2_2, 2_3, 2_4, 2_5, and 2_6 under the second working condition; and the bearings 1 and 2, 3_1 and 3_2 under the third working condition.

[0126] First, adaptive threshold stage division was performed on the 12 bearings in Table 1 to obtain the vibration data from the rapid degradation stage (i.e., the period between the final degradation point and the end point of the original data). For example... Figure 5 The diagram illustrates the iterative process of adaptive threshold stage division for bearings 1_3 and 2_2, as well as the location of the final degradation point. Our algorithm can dynamically lock the location of the degradation point in each iteration.

[0127] In this embodiment, the results of the 12 cross-bearing fault diagnosis tasks are as follows: Figure 6As shown, multiple experiments were conducted on the remaining 12 test sets using this predictor to verify its effectiveness. The figure shows that the lowest fitted R-squared value was 38.77%, and the highest reached 89.8%. Except for bearings 1_6, 2_1, and 2_5, the model achieved good results for the other bearings, indicating that it has good predictive and generalization capabilities for most unknown test bearings.

[0128] In addition, it can be seen that the RMSE range of the 12 tested bearings is 0 to 6.5. The minimum RMSE of bearing 2_6 can reach 0.43, while the maximum RMSE of bearing 2_3 is only 6.5, which is relatively stable.

[0129] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cross-domain bearing life prediction method based on single-source domain generalization, characterized in that, First, a predictor is obtained, which is acquired through machine learning. The predictor includes a data preprocessing module and a life prediction module. The data preprocessing module processes the vibration data of the bearing, and the life prediction module takes the processed data as input data and predicts the remaining life of the bearing based on the input data. Then, the vibration data of the target bearing is obtained and input into the predictor to obtain the life prediction value output by the predictor. The data preprocessing module processes the bearing vibration data in the following steps: SA1, Obtain bearing vibration data , Indicates the bearing's first A vibration data sample; a vibration data sample consists of multiple data points collected per unit time according to a set sampling frequency. It consists of vibration data samples collected over continuous time; n is the ordinal number, and N is the vibration data. The number of vibration data samples, 1≦n≦N; SA2, Obtain the input time series , ={ , ,…, ,… }, where i is an ordinal number, 1 ≤ i ≤ I, Let I be the step size, and I = N - m + 1; This represents the i-th time series sample; express The j-th data point, i.e., the vibration data sample The peak-to-peak value of the time-domain characteristic; 1≦j≦m; SA3, for the input time series Perform normalization processing, so that The value after normalization is ; SA4, the data preprocessing module outputs the following data: , ; The vibration data D of the bearing is obtained by filtering from the original data H of the bearing. The filtering of vibration data D includes the following steps: SB1, denoted as the original bearing data H. , This represents the r-th vibration data sample collected during bearing monitoring, where 1 ≤ r ≤ R, and R represents the total number of vibration data samples collected; the time-domain characteristic peak-to-peak value set is extracted from the original data H. ,right Using the isotonic regression algorithm, degenerate data with a monotonic trend is obtained. ; Indicates degraded data The r-th value in; SB2, Construct a sliding window of length z for degenerate data Perform sliding values ​​to obtain the degradation gradient of the data in each window, and take the value in the k-th window. The corresponding degradation gradient is , 1≦k≦R-z+1; SB3, Constructing a Degenerate Continuous Domain , and Representing the degenerate continuous domain The starting and ending values, The initial value is , The initial value is ; SB4, Calculation Degenerate gradient for all window values The mean is denoted as Av. (The last part, "judging," is incomplete and requires further context.) Does the following condition 1-2 meet? If yes, proceed to step SB5; if not, proceed to step SB6. Condition 1: There exists a continuous fluctuation domain. , Representing the continuous wave domain The starting and ending values, Degenerate gradients corresponding to all window values , ... All are greater than 0, and The number of values ​​that the window can take, c-z+1-b+1, is greater than or equal to the set value C0, where C0>0; Condition 2: The number of values ​​in the window corresponding to the degradation gradient greater than the mean Av is Q, where Q is greater than or equal to a set value Q0, 0 <Q0≤C0; SB5, Update the degenerate continuum , = , = Then return to SB4; SB6, with Final degradation point, obtain raw data As vibration data ,Right now = .

2. The cross-domain bearing life prediction method based on single-source domain generalization as described in claim 1, characterized in that, Step SB4 specifically includes the following sub-steps: SB41, Calculation The mean Av of the degradation gradients for all window values; SB42, Order = The initial value of t is s; SB43, Calculation Values ​​retrieved from the upper window degradation gradient ; SB44, judgment Is it greater than 0? If yes, update t=t+1 and return to step SB43. No, proceed to step SB45; SB45. Determine if tb is greater than or equal to C0; if not, update t=t+1 and return to step SB42. If yes, then proceed with the following step SB46; SB46, Data Corresponding degradation gradient , ... The number of data points greater than the mean Av is denoted as Q. Determine whether Q is greater than or equal to the set value Q0; if not, update t=t+1 and return to step SB42. Yes, then let the data As a continuous wave domain ,Right now = .

3. The cross-domain bearing life prediction method based on single-source domain generalization as described in claim 1, characterized in that, The lifetime prediction module includes a preliminary prediction network, an inverse normalization unit, and an output unit; the preliminary prediction network is used to obtain... The corresponding first predicted value and The corresponding second predicted value The inverse normalization unit is used to process the second predicted value. Perform inverse normalization to obtain inverse normalized predicted values. The output of the output unit is the output of the predictor, and the output result of the output unit is the remaining lifetime sequence. , For bearing vibration data samples The remaining lifetime prediction value over the acquisition time; These are learnable affine parameters.

4. The cross-domain bearing life prediction method based on single-source domain generalization as described in claim 3, characterized in that: express The average value, express The standard deviation of ; γ and β are both learnable affine parameters; For the set constant, In the interval [10] -6 10 -5 [Values ​​can be taken from the above].

5. The cross-domain bearing life prediction method based on single-source domain generalization as described in claim 4, characterized in that: 。 6. The cross-domain bearing life prediction method based on single-source domain generalization as described in claim 3, characterized in that, Obtaining the predictor involves the following steps: S1. Construct and initialize a predictor consisting of a data preprocessing module and a lifetime prediction module; S2. Obtain vibration data D from the source bearing. S , Construct multiple learning samples { | }, I=N-m+1, the vibration data sample consists of multiple data points collected per unit time according to the set sampling frequency. It consists of vibration data samples collected over a continuous period of time; express The nth vibration data sample in the data; For the source bearing in vibration data samples The remaining lifetime in terms of collection time, The data is labeled with human annotations; the predictor performs machine learning on the training samples to iterate the parameters until the predictor's loss reaches the set loss threshold, at which point the predictor is fixed. The loss function of the predictor is: The source bearing for the predictor output in the vibration data sample The remaining lifetime prediction value over the acquisition time.

7. The cross-domain bearing life prediction method based on single-source domain generalization as described in claim 3, characterized in that, The preliminary prediction network is constructed based on a single-layer gated recurrent unit (GRU).

8. A cross-domain bearing life prediction system based on single-source domain generalization, characterized in that, It includes a memory storing a computer program, which, when executed, implements the cross-domain bearing life prediction method based on single-source domain generalization as described in any one of claims 1-7.

9. The cross-domain bearing life prediction system based on single-source domain generalization as described in claim 8, characterized in that, It also includes a processor connected to a memory, the processor being used to execute a computer program to implement the cross-domain bearing life prediction method based on single-source domain generalization as described in any one of claims 1-7.

Images