Turntable bearing life prediction method based on data distribution adaptive transfer learning

Abstract

The invention relates to the field of turntable bearing service life prediction, in particular to a turntable bearing service life prediction method based on data distribution adaptive transfer learning, which comprises the following steps: S1, carrying out adaptive noise fully integrated robust local mean decomposition on a turntable bearing vibration signal for noise reduction, time domain, frequency domain and time-frequency domain features are extracted from the data of the source domain and the target domain to reconstruct a turntable bearing data set; according to the method, data distribution adaptive transfer learning is used, the distribution difference between the data can be balanced and reduced, the service life prediction precision can be improved by using the data after the distance is shortened, and meanwhile, the modeling problem caused by data shortage is solved.

Description

Technical Field

[0001] The present invention relates to the field of turntable bearing life prediction, and in particular to a turntable bearing life prediction method based on data distribution adaptive transfer learning. Background Art

[0002] Slewing bearings are large bearings capable of simultaneously supporting combined loads, including axial loads, radial loads, and overturning moments. They are primarily used in large equipment such as wind turbines, port machinery, construction machinery, and military equipment. These bearings operate in relatively harsh environments and are prone to frequent failures. Therefore, predicting their remaining service life can help prevent failures early, mitigating safety risks.

[0003] Because turntable bearings bear diverse loads and rotate at low speeds, weak fault signals can be easily masked by noise. Furthermore, research on turntable bearing failure mechanisms is relatively scarce, making previous methods for predicting the life of ordinary bearings inapplicable. With the rapid development of deep learning technology, bearing life prediction techniques based on deep learning have been widely used. For example, convolutional neural networks have a strong ability to discover implicit information in raw signals without having to know the component's damage mechanism.

[0004] Existing turntable bearing life prediction methods mostly rely on deep learning models, such as convolutional neural networks (CNNs). These methods typically require extensive hyperparameter tuning and task-specific network architecture design. To adapt to diverse task requirements, large amounts of data are required, and prediction performance can be significantly reduced when data distribution varies. More critically, existing methods often struggle to effectively transfer knowledge from the source domain when target domain data is scarce or inconsistently distributed. This results in insufficient model generalization and an inability to maintain good prediction performance in data-scarce environments.

[0005] Data distribution adaptive transfer learning aims to automatically adjust the data distribution differences between the source domain and the target domain, avoiding the need for a large number of manual adjustments. This method can effectively transfer the knowledge of the source domain to the target domain by adaptively processing the differences in data distribution, regardless of whether the data is scarce or not, significantly improving the prediction accuracy and generalization ability of the model, and does not need to rely on complex hyperparameter tuning, with higher flexibility and applicability. Therefore, the present invention proposes a method based on data distribution adaptive transfer learning, which can balance and reduce the distribution differences between data, and perform ant colony optimization on the balance factor to further reduce the distribution differences to achieve accurate knowledge transfer. Finally, the long short-term memory network (LSTM) is used to realize the life prediction of the turntable bearing and improve the accuracy of the prediction. Summary of the Invention

[0006] The main technical problem solved by the present invention is to provide a turntable bearing life prediction method based on data distribution adaptive transfer learning, which can not only overcome the data distribution differences between the source domain and the target domain, but also effectively deal with the modeling problems caused by the scarcity of target domain data. Through transfer learning, the effective transfer of source domain knowledge is realized, thereby improving the prediction accuracy and generalization ability under data scarcity conditions.

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

[0008] A turntable bearing life prediction method based on data distribution adaptive transfer learning includes the following steps:

[0009] S1: The slewing bearing vibration signal is subjected to denoising by the fully integrated robust local mean decomposition (CERLMDAN) with adaptive noise, and the time domain, frequency domain, and time-frequency domain features are extracted from the source and target domain data to reconstruct the slewing bearing dataset;

[0010] S2: Calculate the marginal distribution difference and conditional distribution difference between the source domain and the target domain, and use the balance factor to balance the marginal distribution difference and conditional distribution difference between the two domains;

[0011] S3: Through each round of iteration, the distribution difference between the two domains is changed and the balance factor is recalculated. The balance factor is optimized using the ant colony algorithm to further balance the distribution difference between the two domains, bringing the source domain data and the target domain data closer together.

[0012] S4: Use the balance factor obtained in S3 to reduce the distribution difference between the source domain and target domain data, and use the LSTM prediction model to predict the life of the turntable bearing using the source domain and target domain data with reduced distribution difference;

[0013] Furthermore, step S1 includes the following steps:

[0014] S1.1: The slewing bearing vibration signal is subjected to adaptive noise reduction using fully integrated robust local mean decomposition. The signal is decomposed into multiple product functions (PFs) using CERLMDAN. Based on whether the number of extreme points of the PF components is less than 3 and the energy ratio is less than 0.001, four PF components that meet the conditions are selected to reconstruct the signal to achieve filtering and noise reduction.

[0015] S1.2: Normalize the denoised signal to reduce computational complexity;

[0016] S1.3: Use the denoised signal to perform feature extraction in the time domain, frequency domain, and time-frequency domain, extracting 14 feature parameters: time domain: kurtosis, peak factor, impulse factor, crest factor, shape factor, margin factor, mean, variance, skewness, peak vibration, and root mean square vibration; frequency domain: spectral flatness and energy ratio; time-frequency domain: entropy and fractal dimension;

[0017] S1.4: Combine the 14 feature parameters extracted from the vibration data of the two turntable bearings and combine their respective features into source domain and target domain datasets;

[0018] Furthermore, step S2 includes the following steps:

[0019] S2.1: Calculate the marginal distribution difference and conditional distribution difference between the source domain and the target domain to calculate the balance factor to balance the marginal distribution difference and conditional distribution difference between the two domains;

[0020] S2.2: According to M0=ee T C calculates the marginal distribution difference between the source domain and the target domain to bring the overall distribution of the source domain and the target domain closer. A-distance is used to calculate the marginal distribution distance. A-distance is defined as the error obtained by establishing a binary classifier to classify two different fields. Calculate the conditional distribution difference to ensure that samples of the same category are distributed closer in the two domains, and calculate the conditional distribution distance by traversing each category;

[0021] Furthermore, in S2.2, a weight matrix e is constructed to assign positive weights (1 / n s ), target domain samples are given negative weights (-1 / n t ), then calculate the product of e and its transpose and multiply it by the number of categories C to obtain the marginal distribution adaptation matrix M0, where n s and n t are the number of samples in the source domain and the target domain respectively; for each category c, construct a weight vector e c , set the sample position of category c in the source domain to 1 / N s , the sample position of category c in the target domain is set to -α / N t , where α is the ratio coefficient of the number of samples in the source domain and the target domain: N s and N t are the number of samples in the source domain and target domain in each category, and then c Construct the conditional distribution adaptation matrix N with the product accumulation of its transpose;

[0022] Furthermore, step S3 includes the following steps:

[0023] S3.1: Evaluate the distance between the source and target domains using the A-distance calculation method. Use the SVM model for prediction, then calculate the mean absolute error (mean absolute error). Then, calculate the A-distance value (adist_m) based on error and adist = 2·(1-2·error). For each category, use A_distance to measure the distribution difference between samples of that category in the source and target domains. Then, take the average of the A-distance values for all categories as the overall conditional distribution difference metric (adist_c). Then, calculate the balance factor (μ) using the formula (μ = adist_c / (adist_c + adist_m).

[0024] S3.2: The ant colony algorithm is then used to optimize the balance factor calculated in S3.1. Ants select the balance factor μ based on the pheromone, calculate fitness, and update the pheromone, giving μ with high fitness more opportunities to be selected. Through continuous iteration, each time the ants update their selection strategy based on the current pheromone, the balance factor μ is ultimately optimized to minimize the difference between the source and target domains.

[0025] S3.3: Based on the marginal distribution difference and conditional distribution difference obtained in S2 and the balance factor μ obtained in S3.1, a total migration matrix is calculated according to the formula M = (1-μ)·M0 + μ·N, where M0 is the marginal distribution difference and N is the conditional distribution difference, to balance the distribution difference between the two domains.

[0026] S3.4: Finally, by solving the optimization problem min tr(A T KMK T A), K is the kernel matrix calculated by radial basis, and the optimal feature transformation matrix A is obtained, thereby obtaining a new feature representation that minimizes the distribution difference between the source domain and the target domain in the new feature space;

[0027] Furthermore, in step S3.2, an iterative loop is added to optimize the balance factor in each round and select the optimal balance factor to maximize the similarity between the source domain and the target domain;

[0028] Furthermore, in step S3.4, the source domain and target domain data after the distance is shortened are obtained and then predicted by LightGBM. The predicted category labels are used in the next epoch, and then the contents in steps S2 and S3 are repeated until the distance between the two domains is minimized.

[0029] Furthermore, in step S4, the obtained new source domain and target domain data are predicted using the LSTM model to achieve life prediction of the turntable bearing.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The present invention uses data distribution adaptive transfer learning, and through the optimization of the balance factor, makes the proportional weights of the source domain and the target domain more accurate, thereby realizing the migration of effective data; replaces the KNN classifier with the LightGBM classifier, thereby improving the classification accuracy; migrates the knowledge of the source domain data to the target domain, and reduces the data distribution difference by balancing the marginal distribution difference and conditional distribution difference between the source domain and the target domain, which can effectively improve the accuracy and generalization ability of the turntable bearing life prediction, improve the safety and economy in engineering production, and at the same time improve the stability and adaptability of the model, and has a strong cross-domain knowledge transfer capability, which is suitable for a variety of different prediction tasks and greatly reduces the time and cost of model development. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Schematic diagram of the method flow of the present invention;

[0033] Figure 2A This is a schematic diagram of a turntable bearing test bench according to the present invention;

[0034] Figure 2B This is the second schematic diagram of the turntable bearing test bench of the present invention;

[0035] In the figure: 1- hydraulic cylinder 1; 2- upper flange; 3- main test bearing; 4- companion test bearing; 5- lower flange; 6- hydraulic cylinder 2; 7- mounting back plate; 8- hydraulic cylinder 3; 9- driving pinion; 10- hydraulic motor; 11- base;

[0036] Figure 3 This is a physical picture of the turntable bearing test bench of the present invention;

[0037] Figure 4 This is a first set of turntable bearing vibration acceleration signal diagrams of the present invention;

[0038] Figure 5 This is a second set of turntable bearing vibration acceleration signal diagrams of the present invention;

[0039] Figure 6 This is a diagram of the vibration acceleration signal of the first group of turntable bearings after noise reduction according to the present invention;

[0040] Figure 7 This is a vibration acceleration signal diagram of the second group of turntable bearings after noise reduction according to the present invention;

[0041] Figure 8 is a first set of characteristic value curves of slewing bearing vibration signals according to the present invention;

[0042] Figure 9 is a second set of characteristic value curves of slewing bearing vibration signals according to the present invention;

[0043] Figure 10 This is a graph showing the lifespan prediction results of the present invention;

[0044] Figure 11 This is a comparison chart of the prediction results of the life prediction method of the present invention and those of TCA and JDA;

[0045] Figure 12 This is a comparison chart of the prediction results of the life prediction method of the present invention, BDA, and direct prediction methods; Figure 13 This is a comparison chart of evaluation indicators of the life prediction results of the present invention. DETAILED DESCRIPTION

[0046] The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more precise definition of the protection scope of the present invention.

[0047] Example 1

[0048] This embodiment discloses a turntable bearing life prediction method based on data distribution adaptive transfer learning, such as Figure 1 , including the following steps:

[0049] S1: The slewing bearing vibration signal is subjected to denoising by the fully integrated robust local mean decomposition (CERLMDAN) with adaptive noise, and the time domain, frequency domain, and time-frequency domain features are extracted from the source and target domain data to reconstruct the slewing bearing dataset;

[0050] S2: Calculate the marginal distribution difference and conditional distribution difference between the source domain and the target domain, and use the balance factor to balance the marginal distribution difference and conditional distribution difference between the two domains;

[0051] S3: Through each round of iteration, the distribution difference between the two domains is changed and the balance factor is recalculated. The balance factor is optimized using the ant colony algorithm to further balance the distribution difference between the two domains, bringing the source domain data and the target domain data closer together.

[0052] S4: Use the balance factor obtained in S3 to reduce the distribution difference between the source domain and target domain data, and use the LSTM prediction model to predict the life of the turntable bearing using the source domain and target domain data with reduced distribution difference;

[0053] In step S1, in order to obtain the vibration signal of the turntable bearing, the turntable bearing accelerated life test is carried out, which mainly includes mechanical, hydraulic and measurement and control parts. The test bench is driven by a hydraulic motor, and the turntable bearing engaged with it is driven to rotate through the small gear. The schematic diagram of the test bench is shown in Figure 2, and the actual picture is shown in Figure 2. Figure 3 shown.

[0054] In step S1, the experimental object is the QNA-730-22 single-row ball slewing bearing. In the first group of tests, the axial force of 96kN and the overturning force of 240kNm are selected as the ultimate load according to the specifications and load capacity of the slewing bearing. The experiment lasts for 11 days. The performance of each component of the slewing bearing has degraded until it is stuck. At this time, the slewing bearing is in a state of wear and failure. Its vibration acceleration signal is as follows: Figure 4 The second set of fatigue life experiments selected axial force of 330KN and overturning moment of 138KNm as the limit load according to the specifications and load capacity of the slewing bearing. The experiment lasted for 4 days. All parts of the slewing bearing showed performance degradation until the machine was shut down. At this time, the slewing bearing was in a state of wear and failure. Its vibration acceleration signal is shown as follows: Figure 5 As shown. The turntable bearing vibration signal is subjected to adaptive noise fully integrated robust local mean decomposition denoising, and then multiple RLMD decompositions are performed by adding different white noise sequences to decompose it into multiple product functions PF. In each decomposition process, the modal components are extracted by identifying local extreme points and constructing envelopes. According to whether the number of extreme points of the PF component is less than 3 and whether the energy ratio is less than 0.001, the multiple decomposition results are integrated and averaged to enhance the noise resistance performance. Subsequently, the obtained modes are scored based on the correlation index and energy ratio, and the top 4 modes with higher scores are selected for signal reconstruction to finally obtain the denoised signal. The reconstructed signal of the first group of turntable bearings after denoising is shown in FIG. Figure 6 As shown, the reconstructed signal of the second group of turntable bearings after noise reduction is as follows Figure 7 shown.

[0055] The noise-reduced signal is used to extract features in the time domain, frequency domain, and time-frequency domain, and 14 feature parameters are extracted: time domain: kurtosis, peak factor, impulse factor, crest factor, form factor, margin factor, mean, variance, skewness, peak vibration, and root mean square vibration; frequency domain: spectral flatness and energy ratio; time-frequency domain: entropy and fractal dimension. Some characteristic value curves extracted from the first group of turntable bearing vibration signals are shown in the figure below. Figure 8 As shown in the figure, some characteristic value curves extracted from the vibration signal of the second group of turntable bearings are shown in Figure 9 shown.

[0056] Using vibration data from two types of turntable bearings, the same data processing method was used for each. The first set of bearing vibration data used a 1-second sample, with a sample selected every 4 seconds. The second set of bearing vibration data also used a 1-second sample, but with a sample selected every 10 seconds. The first set of turntable bearing data served as the source domain, and the second set of turntable bearing data served as the target domain. Two labels were then used: one for classification (classification was achieved using multi-feature fusion, first calculating multiple health indicators including comprehensive energy, kurtosis skewness combination, peak pulse ratio, and energy entropy ratio. Then, the key change life prediction points of the data were detected using three methods: comprehensive health indicator method, energy mutation detection method, and statistical feature mutation detection method. Based on these change points, the bearing status was automatically divided into three stages as classification labels), and the other for (RUL) labeling (using the number of samples as the label).

[0057] During the implementation, the experiment used an RTX 3090 GPU, Python 3.9, and the deep learning framework PyTorch version 1.8.0. The experiment was divided into two main parts: data distribution adaptation and model prediction. An ant colony algorithm was used to optimize the balancing factor to minimize the distribution difference between the two domains.

[0058] In step 2, first construct the marginal distribution difference matrix M0=ee T C is used to measure the overall distribution difference, where e is a weight vector matrix. By calculating the product of e and its transpose and multiplying it by the number of categories C, the marginal distribution adaptation matrix M0 is obtained; the weight vector e is constructed, and the source domain samples are given positive weights (1 / n s ), target domain samples are given negative weights (-1 / n t ), where n s and n t are the number of samples in the source domain and the target domain respectively; then, by constructing the conditional distribution difference matrix To measure the conditional distribution difference, where e c is the weighted vector matrix of category c; for each category c, construct a weight vector e and set the sample position of category c in the source domain to 1 / N s , the sample position of category c in the target domain is set to -α / N t , where α is the ratio coefficient of the number of samples in the source domain and the target domain: Then the conditional distribution adaptation matrix N is constructed by multiplying and accumulating e and its transpose, where N s and N t are the number of samples in the source domain and target domain in each category respectively;

[0059] In step 3, the overall distribution difference (adist_m) and intra-class distribution difference (adist_c) between the source and target domains are first evaluated by A-distance, where A-distance is defined as the error obtained by establishing a binary classifier to classify two different domains. Specifically, for each category i, samples belonging to the category in the source and target domains are extracted (X si and X tj ), calculate the A-distance between them, and then take the average of all class PAD (that is, A_distance) values as the intra-class distribution difference. Finally, use the formula μ = adist_c / (adist_c + adist_m) to calculate the balance factor.

[0060] The ant colony algorithm optimizes the value of μ by simulating the foraging behavior of ants. Each ant selects a value of μ (in the range [0.0, 1.0]) based on the pheromone concentration along the path, and then uses an objective function to evaluate the fitness of this μ value. The objective function calculates the difference between the weighted distribution weighted_dist = (1-μ)·source_dist + μ·target_dist and the ideal distribution (the average of the source and target domain distributions). The smaller the difference, the higher the fitness; source_dist and target_dist are the distributions of the source and target domain data, respectively. In each iteration, the ant leaves pheromones on the path that are proportional to its fitness, while the existing pheromones gradually evaporate. After multiple iterations, the algorithm converges to a more optimal value of μ.

[0061] Then, based on the calculated marginal distribution difference M0, conditional distribution difference N, and balance factor μ, the two distribution difference matrices are combined using the formula M = (1-μ)·M0+μ·N, and Frobenius norm normalization is performed.

[0062] Finally, by solving the optimization problem min tr(A T KMK T A), K is the kernel matrix calculated by radial basis, and the optimal feature transformation matrix A is obtained, thereby obtaining a new feature representation that minimizes the distribution difference between the source domain and the target domain in the new feature space;

[0063] Add an iterative loop, optimize the balance factor of each round, select the optimal balance factor, and maximize the similarity between the source domain and the target domain; obtain the source domain and target domain data after the distance is shortened and then use LightGBM to predict, and use the predicted category label for the next epoch, and then repeat the contents in steps S2 and S3 until the distance between the two domains is minimized, and finally obtain the source domain and target domain data with balanced distribution differences.

[0064] The obtained source domain and target domain data are used to predict the life of the turntable bearing using the LSTM prediction model, and the turntable bearing life prediction curve of the target domain is obtained, as shown in Figure 10 As shown in the figure, the LSTM model is directly trained using the source domain data to predict the life of the turntable bearing in the target domain (that is, without using transfer learning - DirtRUL). At the same time, five methods including traditional transfer component analysis (TCA), joint distribution adaptation (JDA) and balanced distribution adaptation (BDA) before optimization are selected for comparison. Figure 11 and 12 As shown, the mean absolute error (MAE) and mean square error (MSE) of several prediction methods are calculated for comparison, as shown in Figure 13 As shown in FIG, the results show that the method proposed in the present invention can better reduce the data distribution difference, resulting in better migration effect and higher prediction accuracy.

[0065] The embodiments disclosed in the present invention are only preferred implementation methods for explaining the technical solutions, and are not intended to limit the scope of patent protection. Those skilled in the art should understand that, without departing from the core principles and spirit of the present invention, technical adjustments made for specific application scenarios, equivalent replacements of embodiments, or structural deformations and process optimizations based on the contents of the specification and drawings all fall within the scope of protection of the present invention. Whether directly implemented or indirectly applied to other fields through technical associations, all implementation methods that conform to the technical concept of the present invention and fall within the scope defined in the claims are protected by patent law.

Claims

1. A turntable bearing life prediction method based on data distribution adaptive transfer learning, characterized in that: The steps include: S1: The slewing bearing vibration signal is subjected to denoising by the fully integrated robust local mean decomposition (CERLMDAN) with adaptive noise, and the time domain, frequency domain, and time-frequency domain features are extracted from the source and target domain data to reconstruct the slewing bearing dataset; S2: Calculate the marginal distribution difference and conditional distribution difference between the source domain and the target domain, and use the balance factor to balance the marginal distribution difference and conditional distribution difference between the two domains; S3: Through each round of iteration, the distribution difference between the two domains is changed and the balance factor is recalculated. The balance factor is optimized using the ant colony algorithm to further balance the distribution difference between the two domains, bringing the source domain data and the target domain data closer together. S4: Use the balance factor obtained in S3 to reduce the distribution difference between the source domain and target domain data, and use the LSTM prediction model to predict the life of the turntable bearing using the source domain and target domain data with reduced distribution difference; Wherein, the step S1 specifically includes: S1.1: The slewing bearing vibration signal is subjected to adaptive noise reduction using fully integrated robust local mean decomposition. The signal is decomposed into multiple product functions (PFs) using CERLMDAN. Based on whether the number of extreme points of the PF components is less than 3 and the energy ratio is less than 0.001, four PF components that meet the conditions are selected for signal reconstruction to achieve filtering and noise reduction. S1.2: Normalize the denoised signal to reduce computational complexity; S1.3: Use the denoised signal to perform feature extraction in the time domain, frequency domain, and time-frequency domain, extracting 14 feature parameters: time domain: kurtosis, peak factor, impulse factor, crest factor, margin factor, mean, variance, skewness, peak vibration, and root mean square vibration; frequency domain: spectral flatness and energy ratio; time-frequency domain: entropy and fractal dimension; S1.4: Combine the 14 features extracted from the source and target domain data to form the source and target domain datasets. The step S2 specifically includes: S2.1: Calculate the marginal distribution difference and conditional distribution difference between the source domain and the target domain to calculate the balance factor to balance the marginal distribution difference and conditional distribution difference between the two domains; S2.2: According to M0=ee T C calculates the marginal distribution difference between the source domain and the target domain, where e is the category weight matrix and C is the number of categories. It is used to bring the overall source domain and target domain distribution closer together. A-distance is used to calculate the marginal distribution distance. A-distance is defined as the error obtained by establishing a binary classifier to classify two different fields. Calculate the conditional distribution difference, where e c is a weight matrix for each category; ensure that samples of the same category are distributed closer in the two domains, and calculate the conditional distribution distance by traversing each category; 2. The method for predicting the life of a turntable bearing based on data distribution adaptive transfer learning according to claim 1, characterized in that: In step S1.4, a dataset consisting of a source domain and a target domain is used for training and prediction, and the target domain is used as the final validation set.

3. The method for predicting the life of a turntable bearing based on data distribution adaptive transfer learning according to claim 1, characterized in that: In step S2.2, it is necessary to calculate the marginal distribution difference between the source domain and the target domain, according to M0=ee T C, calculate the marginal distribution difference; 4. The method for predicting the life of a turntable bearing based on data distribution adaptive transfer learning according to claim 1, characterized in that: In step S2.2, it is necessary to calculate the conditional distribution difference between the source domain and the target domain, by traversing each category to build Ensure that samples of the same category are distributed more closely in the two domains.

5. The method for predicting the life of a turntable bearing based on data distribution adaptive transfer learning according to claim 1, characterized in that: The step S3 specifically includes the following steps: S3.1: Evaluate the distance between the source and target domains using the A-distance calculation method. Use the SVM model for prediction, then calculate the mean absolute error (mean absolute error). Then, calculate the A-distance value (adist_m) based on error and adist = 2·(1-2·error). For each category, use A_distance to measure the distribution difference of samples of that category in the source domain and the target domain, and then take the average of all categories of A-distance as the overall conditional distribution difference measure (adist_c); then calculate the balance factor μ according to the formula μ = adist_c / (adist_c + adist_m); S3.2: The ant colony algorithm is then used to optimize the balance factor calculated in S3.

1. Ants select the balance factor μ based on the pheromone, calculate fitness, and update the pheromone, giving μ with high fitness more opportunities to be selected. Through continuous iteration, each time the ants update their selection strategy based on the current pheromone, the balance factor μ is ultimately optimized to minimize the difference between the source and target domains. S3.3: Based on the marginal distribution difference and conditional distribution difference obtained in S2 and the balance factor μ obtained in S3.2, calculate a total migration matrix according to the formula M = (1-μ)·M0 + μ·N, where M0 is the marginal distribution difference and N is the conditional distribution difference, to balance the distribution difference between the two domains. S3.4: Finally, by solving the optimization problem min tr(A T KMK T A), K is the kernel matrix calculated by radial basis, and the optimal feature transformation matrix A is obtained, thereby obtaining a new feature representation that minimizes the distribution difference between the source domain and the target domain in the new feature space; 6. The method for predicting the life of a turntable bearing based on data distribution adaptive transfer learning according to claim 1, characterized in that: In step S4, the balance factor obtained in step S3 is used to reduce the distribution difference between the source domain and the target domain data, and the source domain and target domain data with reduced distribution difference are used to predict the turntable bearing life using the LSTM prediction model.

Images