A method for evaluating degradation trend of a variable operating condition bearing
By constructing a bearing degradation trend assessment model under varying operating conditions, and utilizing random convolution kernel transformation and a linear classifier, the problems of poor generalization and high computational complexity of existing bearing degradation trend assessment models are solved, thus achieving efficient online assessment.
Patent Information
- Application Number
- CN202211493210.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-25
- Publication Date
- 2026-01-02
- Estimated Expiration
- 2042-11-25
AI Technical Summary
Existing bearing degradation trend assessment models suffer from poor generalization, high computational complexity, and insufficient computing power at the equipment edge, making it difficult to achieve online industrial bearing degradation trend assessment.
A bearing degradation trend assessment model under varying operating conditions is constructed. It employs a feature extraction module, a feature fusion module, a random convolution kernel transformation module, and a label classifier. By fusing modal features with time-frequency features, the model reduces dimensionality and maps health status. The model initializes the bearing data using the probability distribution of random convolution kernels and trains the model using random convolution kernel transformation and a linear classifier.
The model's generalization ability and evaluation accuracy under different operating conditions have been improved, while computational complexity and resource requirements have been reduced, enabling lightweight online bearing degradation trend assessment.
Smart Images
Figure CN115878986B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the technical field of equipment state monitoring, and more particularly to a degradation trend evaluation method of a variable working condition bearing. BACKGROUND
[0002] As the most commonly used core bearing mechanical components in rotating machinery, bearings become high-frequency failure parts due to the discreteness of service life. When running in normal working environment, bearings are inevitably affected by load, speed, temperature, lubrication and other objective conditions, and different positions, degrees and reasons of failure may occur, such as wear, gluing and fracture of inner ring, outer ring, roller and retainer. As the most basic and widely used mechanical parts in rotating structures, whether the bearing works normally is the core problem of rotating machinery state monitoring. Therefore, the prediction of rolling bearing degradation trend and the evaluation of health state have a crucial influence on the entire mechanical industry.
[0003] The health condition characterization parameters of the evaluation method of bearing degradation trend are all from the vibration signals in the working process of the bearing. The feature extraction method of the vibration signal is divided into two categories according to the feature dimension: only one-dimensional feature vector is used and multi-dimensional feature vector is used. The one-dimensional feature vector itself has a series of problems such as poor interference and less information, and it is difficult to accurately and effectively evaluate the health condition and degradation trend of the bearing throughout its life cycle. The method using multi-dimensional feature vector has stronger anti-interference ability and more comprehensive feature expression, and the prediction accuracy of the multi-dimensional vector input model is higher than that of the one-dimensional feature vector input model, but the calculation complexity is higher than that of the one-dimensional feature input. The existing bearing degradation trend machine learning model has poor generalization, the deep learning model has large sample demand, long model training and reasoning time, and the edge computing capability of the equipment is insufficient. The current situation of lightweight and high-accuracy online trend evaluation model is urgent, which leads to the difficulty of realizing the industrial online bearing degradation trend evaluation by the existing technical solutions. SUMMARY
[0004] In view of the above defects or improvement needs of the prior art, the present application provides a degradation trend evaluation method of a variable working condition bearing, which aims to improve the generalization of the bearing degradation trend machine learning model and improve the evaluation accuracy.
[0005] To achieve the above-mentioned purpose, according to one aspect of the present application, a degradation trend evaluation method of a variable working condition bearing is provided, comprising:
[0006] S1. Construct a bearing degradation trend evaluation model; the bearing degradation trend evaluation model comprises a feature extraction module, a feature fusion module, a random convolution kernel transformation module, a dimension reduction module and a label classifier; the feature extraction module is used for extracting modal features and time-frequency features of input data; the feature fusion module is used for fusing the time-frequency features and the modal features through canonical correlation analysis; the random convolution kernel transformation module is used for performing convolution operation on the fused feature parameters through a large number of multi-scale convolution kernels, and extracting multi-scale features related to bearing failure and irrelevant to bearing failure; the dimension reduction module is used for dimension reduction projection of the multi-scale features, and constructing a health state feature; and the label classifier is used for mapping the dimension-reduced health state feature to a corresponding predicted bearing health state;
[0007] S2. The existing multi-condition labeled bearing vibration signal data is used as source domain data, and the vibration signal data of the bearing without a new condition label is used as target domain data to form a training sample set.
[0008] S3. The bearing degradation trend evaluation model is iteratively trained by using the training sample set; in the training process, the joint probability distribution of the random convolution kernel is initialized to be consistent with the distribution of the bearing training set data; the length of the random convolution kernel is set to be the product of a value greater than the number of sampling points in one circle of the rolling bearing and a normal distribution
[0009] S4. The vibration data of the bearing to be measured under the current condition is input into the trained model to obtain the bearing degradation trend evaluation result.
[0010] Further, the fusion process of the modal features and the time-frequency features is as follows:
[0011] The time-frequency features are fused through canonical correlation analysis.
[0012] After the structural modal features are multiplied by the corresponding coefficients, the time-frequency features fused through canonical correlation analysis are subjected to secondary fusion through concat.
[0013] Further, the modal features are obtained by a random subspace identification method SSICOV based on covariance; the modal features include modal frequency, modal mode and modal damping.
[0014] Further, the modal features are further preprocessed as follows: the first order natural frequency is taken as an effective value from the multi-order natural frequency parameters; and interpolation processing is performed on the missing values.
[0015] Further, the time domain features are root mean square value and kurtosis index.
[0016] Further, the frequency domain feature is a root mean square frequency.
[0017] The application also provides a bearing degradation trend evaluation device for variable working conditions, comprising:
[0018] The evaluation model construction module is configured to construct a bearing degradation trend evaluation model, which comprises a feature extraction module, a feature fusion module, a random convolution kernel transformation module, a dimension reduction module, and a label classifier.
[0019] The training sample construction module is configured to use the existing multi-condition labeled bearing vibration signal data as source domain data and the unlabeled bearing vibration signal data of a new condition as target domain data to construct a training sample set.
[0020] The iterative training module is configured to use the training sample set to iteratively train the bearing degradation trend evaluation model.
[0021] The online evaluation module is configured to input the vibration data of the bearing to be measured under the current condition into the trained model to obtain a bearing degradation trend evaluation result.
[0022] Overall, the above technical solutions of the present application can achieve the following beneficial effects compared with the prior art.
[0023] (1) Unlike the conventional random initialization of the random convolution kernel, the probability distribution of the random convolution kernel is initialized as the probability distribution of the bearing data, which is beneficial to the decoupling of the noisy coupled bearing signal into bearing fault related features and bearing fault independent multi-scale feature representation by the random convolution kernel, thereby improving the model generalization for the unseen bearing condition (i.e., the unknown domain) and accurately judging the degradation trend under different conditions; at the same time, since the random convolution kernel transformation is equivalent to only a single layer of network, the parameter quantity is 1-2 orders of magnitude lower than that of the feature extraction network, which has a significant speed advantage.
[0024] (2) The application adopts modal parameters and time domain and frequency domain parameters to jointly construct the health state characteristics of the bearing vibration signal. Due to the degradation of the mechanical system, the structural dynamics characteristics of the system will change, and the natural frequency and damping ratio of the system will change with the degradation of the system. Therefore, the modal characteristics represented by the natural frequency and the damping ratio can be used as the characteristics with prior information to represent the degradation trend of the bearing. By incorporating the prior characteristics, the solution space of the model can be effectively reduced, so that the model pays more attention to the generalizable characteristics related to the degradation of the bearing, thereby improving the generalization and accuracy of the model. BRIEF DESCRIPTION OF DRAWINGS
[0025] Figure 1 A flow chart of a construction method of a variable working condition bearing degradation trend evaluation model provided for embodiment 1 of the application is shown in the figure.
[0026] Figure 2 A variable working condition bearing degradation trend evaluation model structure schematic diagram provided for embodiment 1 of the application is shown in the figure.
[0027] Figure 3 A schematic diagram of a time series data prediction logic of a random convolution kernel transformation model provided for embodiment 1 of the application is shown in the figure. DETAILED DESCRIPTION
[0028] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other.
[0029] Embodiment 1
[0030] A variable working condition bearing degradation trend evaluation method uses the data of the bearing data set of Xi'an Jiaotong University as input, then performs feature extraction, uses a plurality of other linearly independent characteristic quantities (including time domain, frequency domain and modal characteristic quantities) to mechanically learn and deduce the selected health index parameter, then uses random convolution kernel transformation to train data, thereby obtaining an accurate label classifier, and further establishes a variable working condition bearing degradation trend evaluation model.
[0031] Specifically, as shown in Figure 1 , the following steps are included:
[0032] S1. Building a bearing degradation trend evaluation model;
[0033] Referring to Figure 2 , the bearing degradation trend evaluation model includes a feature extraction module, a feature fusion module, a random convolution kernel transformation module, a dimension reduction module and a label classifier;
[0034] The feature extraction module is configured to extract modalities, time-frequency features of the input data; the feature fusion module is configured to fuse the time-frequency features and the modality features through canonical correlation analysis; the random convolution kernel transformation module is configured to process the fused feature parameters to obtain multi-scale features; the dimension reduction module is configured to project the multi-scale features to reduce dimensions and construct health state features; and the label classifier is configured to map the health state features after dimension reduction to corresponding predicted bearing health states.
[0035] The establishment of the health state index will be described from several aspects below.
[0036] Modal information acquisition: modal parameters usually include modal frequencies, modal dampings and modal shapes.
[0037] Due to the degradation of the system, the covariance of the vibration signal will shift, and the Toeplitz matrix constructed by the covariance-based stochastic subspace identification method SSI-COV reflects the covariance change, therefore, the modal frequencies, modal shapes and modal dampings representing the current equipment structure state are obtained by the SSICOV function. After a complete set of files (the full life cycle vibration signal of a bearing in the data set) are all processed according to the same SSICOV modal parameter identification method, the modal frequency, modal damping and modal shape health state index parameter change trend matrix of the full life cycle can be obtained.
[0038] It should be noted that the feature extraction module also pre-processes the modal parameters, for example, the natural frequency parameters obtained by the SSICOV function are mostly more than one, and there may be no corresponding natural frequency at a certain moment. Therefore, only the first order natural frequency is taken as the basic parameter for model experiment, for example, the first data of each item is selected as the sample for the modal frequency, and in the case of missing value, interpolation processing is considered for the missing value, and the modal damping and modal shape health state index are also converted into column vectors by a similar method as part of the input of the feature fusion module.
[0039] Time domain information acquisition: In the time domain characteristics of bearing vibration data, the dimensional parameters and the dimensionless parameters have different focuses: the dimensional characteristics are more related to the objective working environment such as load and rotating speed, while the dimensionless characteristics tend to directly show the impact information of the bearing parts during operation. If the bearing vibration data is classified according to the characteristics, the early degradation damage of the bearing is roughly divided into two types: surface damage and wear. The dimensionless parameter can effectively identify the former, but it is not sensitive to the latter; the dimensional parameter is the opposite, which is very sensitive to the latter but almost has no response to the former. In addition, the time domain characteristic parameters of bearing vibration data also have some differences in the focus of describing the bearing health status. For example, the root mean square can effectively reflect the vibration of the bearing parts, describe the amplitude and energy change trend of the vibration signal in the time domain, and therefore is more suitable for characterizing the current health status of the bearing. The kurtosis index and skewness index can reflect the difference between the current bearing vibration signal and the normal distribution (if the bearing is in normal operating condition, the amplitude of the vibration signal should basically follow the normal distribution), which can sensitively find the surface damage of the bearing. Generally, any time domain characteristic data cannot fully reflect the degradation trend of the bearing throughout its life cycle, so one dimensional parameter and one dimensionless parameter are considered as references. In this embodiment, the root mean square value and the kurtosis index are taken.
[0040] The bearing vibration acceleration data is input, and the root mean square value and the kurtosis index are calculated by formula (number).
[0041]
[0042] wherein
[0043]
[0044] x(n) and N represent N data at each collection point, and the vibration displacement corresponding to each data is x(n).
[0045] After a complete set of files (the full life cycle vibration signal of a bearing in the data set) is processed in a similar manner, the root mean square value and the kurtosis value parameters of this experiment can be obtained.
[0046] Frequency domain information acquisition: The frequency domain characteristics of the vibration data reflect the frequency components and their energy sizes contained in the signal. The use of frequency domain characteristic data to describe the health status of the bearing is mainly through the changes of the frequency components, energy sizes and main frequency positions within a certain time, and different frequency domain characteristics mostly focus on the changes of one of them. The root mean square frequency is selected as the characteristic data used in this experiment.
[0047] The acquisition mode is basically the same as that of the time series data. After Fourier transform of the data, the required root mean square frequency parameters are obtained through formula processing.
[0048]
[0049] wherein s(k) is the frequency spectrum signal obtained by Fourier transform of the time domain signal x(n); f k is the frequency amplitude size of the corresponding point; N fft is the length of the frequency spectrum signal.
[0050] Then, feature fusion is needed, and a bearing health condition representation parameter (health state index) is constructed. As the mechanical system degrades, the system structure dynamics characteristics change, and the natural frequency and damping ratio of the system change with the degradation of the system. Therefore, the modal characteristics represented by the natural frequency and damping ratio can be used as features representing the degradation trend of the bearing with prior information. By incorporating prior features, the solution space of the model can be effectively reduced, so that the model pays more attention to the generalizable features related to bearing degradation, thereby improving the generalization and accuracy of the model.
[0051] Directly performing splicing feature fusion can cause modal characteristics to be submerged in numerous time-frequency characteristics, and the change of the structural modal parameters with the degradation of the system cannot guide the change. Therefore, the present application proposes a two-stage feature fusion method, which first fuses time-frequency characteristics through first-stage feature correlation analysis, and then multiplies the structural modal characteristics by a coefficient and the time-frequency characteristics after typical correlation analysis to perform concat secondary fusion. The modal characteristic coefficient is determined according to the bearing modal contribution size of the application scenario. The two-stage feature fusion can effectively improve the prominence of the fusion of the dynamic characteristics and the time-frequency domain characteristics.
[0052] S2. The vibration signal data of the existing multi-condition labeled bearing is used as the source domain data, and the vibration signal data of the bearing without a new condition label is used as the target domain data to form a training sample set.
[0053] S3. Referring to Figure 3 , the training data is input into the model for model training, and the modeling effect is preliminarily verified under the same distribution test data, such as taking different bearing fault data under the same condition for modeling effect test. After analysis, the model is used to predict the vibration signal data of the bearing without a new condition label, and finally the coincidence degree of the predicted value and the actual value is compared to determine whether the model meets the expectation.
[0054] The specific implementation of the method is as follows:
[0055] First, the training dimensions in the random convolution kernel transformation module are defined, and the classifier is fitted.
[0056] Unlike the conventional random convolution kernel random initialization, the application initializes the joint probability distribution of the random convolution kernel to be consistent with the bearing training set data distribution. By initializing the probability distribution of the random convolution kernel to the probability distribution of the bearing data, the random convolution kernel can be beneficially decoupled from the noisy coupled bearing signal to the bearing fault related feature and the bearing fault irrelevant feature multi-scale feature representation, thereby improving the model generalization for the unseen bearing working condition (i.e. unknown domain), and realizing accurate judgment of the degradation trend under different working conditions.
[0057] In order to make the model focus on the periodic feature law of the rolling bearing fault, the length of the random convolution kernel is set to the product of a value C exceeding the number of sampling points of a rolling bearing circle and a normal distribution. The weight, bias, dilation and stride of the random convolution kernel are initialized by the normal distribution. When the random convolution kernel transformation module fills in to generate each kernel, it will randomly decide whether to use padding when applying the kernel with equal probability.
[0058] The features after fusion and dimensionality reduction will be used to train a linear classifier. The linear classifier is divided into ridge regression classifier and logistic regression classifier. When the number of training samples is less than the number of features, the ridge regression classifier is used, and vice versa.
[0059] In addition, when the random convolution kernel transformation model and the classifier are combined, a single-layer convolutional neural network with random kernel weights is actually formed, and the parameter amount of the model is 1-2 orders of magnitude lower than that of the feature extraction network model in the conventional technology, which has a significant speed advantage for online application scenarios and requires less computing resources.
[0060] The random convolution kernel transformation process is:
[0061]
[0062] Where, X i is the i-th position of the time series; ω is the convolution kernel; l kernel is the length of the convolution kernel; d is the dilation coefficient.
[0063] After random convolution kernel transformation, two feature values will be obtained: maximum value (equivalent to global maximum pooling) and proportion of positive values.
[0064] Finally, the health state feature is input into the label classifier for training. Since the degradation stage of bearing degradation trend evaluation is an integer value, the decimalization of the model prediction output needs to be converted to an integer during the training process. Considering that the bearing degradation process is often divided into five stages, i.e., normal stage, initial failure stage, development stage, rapid development stage and end stage, the model prediction result is first normalized to 0-1, and then multiplied by the dimensionless constant 5 to obtain the health state corresponding to the true bearing degradation trend.
[0065] S4. After the model training is completed, the to-be-tested data is input into the trained model to evaluate the bearing degradation trend under the current working condition.
[0066] Example 2
[0067] A method for evaluating the degradation trend of a variable working condition bearing, comprising:
[0068] The bearing vibration signal data to be diagnosed is input into the variable working condition bearing degradation trend evaluation model constructed by the method for constructing a variable working condition bearing degradation trend evaluation model provided in Embodiment 1 of the present application. After modal and time domain feature extraction based on the feature extraction module in the variable working condition bearing degradation trend evaluation model, feature fusion is performed to construct a health state index, which is input into the label classifier in the variable working condition bearing degradation trend evaluation model, so as to obtain the fault category label of the bearing vibration signal data to be diagnosed.
[0069] 1) Prepare the data set:
[0070] The Xi'an Jiaotong University bearing data set is taken. Each data set has multiple bearing degradation stages, such as initial failure stage, development stage, rapid development stage and end stage.
[0071] Load and divide the multi-period data into training set and test set; standardize the data in the divided training set, and select part of the working condition combination as the source domain, such as the data of 35Hz 12kN Bearing1_1 and 37.5Hz 11kN Bearing2_1 as the source domain, and the data of 35Hz 12kN Bearing1_2 and 37.5Hz 11kN Bearing2_2 as the target domain.
[0072] 2) Train the variable working condition bearing degradation trend evaluation model:
[0073] The method described in Embodiment 1 of the application is used to train the variable working condition bearing fault diagnosis model. As in a typical convolutional neural network, random convolution kernel transformation uses a convolution kernel to convert a time series. Essentially, any aspect of the kernel is random, such as length, weight, bias, dilation and padding, etc. Through experimental testing, for each kernel, these values are set as follows:
[0074] Length is randomly and equally-probably chosen from {7, 9, 11} to ensure that the convolution kernel is much shorter than the input time series in most cases; weights are sampled from a standard normal distribution and mean-pooled after being set; bias is sampled from a uniform distribution over (-1, 1) and only correlates with the positive ratio in the feature map; dilation is sampled on an exponential scale with base 2 to ensure that the effective length of the kernel (including dilation) equals the length of the input time series, allowing kernels that are otherwise similar but have different dilations to match the same or similar patterns at different frequencies and scales; padding is (randomly and equally-probably) decided when generating each kernel whether to use padding or not, and if padding is used, a certain amount of zero padding is added at the beginning and end of each time series when applying the kernel, so that the "middle" elements of the kernel are centered around each point in the time series; the stride is always taken as 1.
[0075] 3) Degradation trend evaluation of bearings under variable working conditions:
[0076] The vibration signal data of the bearing data in the test set that has not been seen without labels is input into the trained degradation trend evaluation model of the bearing under variable working conditions, and after feature extraction, the extracted features are directly input into the label classifier to evaluate the degradation trend of the bearing under variable working conditions.
[0077] In summary, the bearing data set data of Xi'an Jiaotong University is used as input, and appropriate parameters are selected to characterize and verify the bearing health status. Based on the time series training model based on random convolution kernel transformation, a bearing degradation trend evaluation model is established, and the characterization parameters of the bearing health status are predicted. At the same time, the data given by the bearing data set of Xi'an Jiaotong University is used as the experimental and verification object, and is substituted into the established model to predict the degradation stage of the bearing. Finally, the applicability and effectiveness of the model are analyzed. Compared with the traditional bearing health status and degradation trend evaluation which only selects one-dimensional time domain feature vector, this paper proposes a two-stage fusion method of time-frequency domain features and modal feature parameters to prevent modal features from being submerged in numerous time-frequency domain features. Random convolution kernel transformation model is used, which uses a large number of convolution kernels. Benefiting from almost no need to consider learning weights, the cost of convolution calculation is relatively low, so a relatively small amount of calculation cost can be used to use a large number of kernels.
[0078] The related technical solutions are the same as in Embodiment 1, which will not be repeated here.
[0079] Embodiment 3
[0080] The application discloses a system for evaluating degradation trend of a variable working condition bearing, comprising a memory and a processor, the memory stores a computer program, and the processor executes the computer program to execute the method for evaluating degradation trend of the variable working condition bearing provided in Embodiment 1 of the application.
[0081] Embodiment 4
[0082] A machine readable storage medium stores machine executable instructions, when the machine executable instructions are invoked and executed by a processor, the machine executable instructions cause the processor to execute the method for constructing a variable working condition bearing fault diagnosis model provided in Embodiment 1 of the application and / or the variable working condition bearing fault diagnosis method provided in Embodiment 2 of the application.
[0083] It is easy for those skilled in the art to understand that the above description is only the preferred embodiment of the application, and is not intended to limit the application, and any modification, equivalent replacement and improvement made within the spirit and principle of the application should be included in the protection scope of the application.
Claims
1. A method of assessing the degradation trend of a variable operating condition bearing, characterized in that, Comprise: S1. Construct a bearing degradation trend evaluation model; the bearing degradation trend evaluation model comprises a feature extraction module, a feature fusion module, a random convolution kernel transformation module, a dimension reduction module and a label classifier; the feature extraction module is used for extracting modal features and time-frequency features of input data; The feature fusion module is used for fusing the time-frequency features and the modal features through canonical correlation analysis; the random convolution kernel transformation module is used for performing convolution operation on the fused feature parameters through a large number of multi-scale convolution kernels, and extracting multi-scale features related to bearing failure and unrelated to bearing failure; the dimension reduction module is used for dimension reduction projection of the multi-scale features, and constructing health state features; the label classifier is used for mapping the health state features after dimension reduction to corresponding predicted bearing health states; S2. The vibration signal data of the existing multi-condition labeled bearing is used as source domain data, and the vibration signal data of the bearing without new condition and label is used as target domain data to form a training sample set; S3. The training sample set is used for iterative training of the bearing degradation trend evaluation model; In the training process, the joint probability distribution of the random convolution kernel is initialized to be consistent with the data distribution of the bearing training set; the length of the random convolution kernel is set to be the product of a value greater than the number of sampling points in one revolution of the rolling bearing and a normal distribution S4. The vibration data of the bearing to be measured under the current working condition is input into the trained model to obtain the bearing degradation trend evaluation result.
2. A method of assessing the degradation trend of a variable operating condition bearing according to claim 1, characterized in that, The fusion process of the modal features and the time-frequency features is: Fusing the time-frequency features through canonical correlation analysis; After multiplying the structural modal features by the corresponding coefficients, the time-frequency features fused through canonical correlation analysis are subjected to secondary fusion through concat.
3. A method of assessing the degradation trend of a variable operating condition bearing according to claim 2, characterized in that, The modal features are obtained through the random subspace identification method SSICOV based on covariance; the modal features include modal frequency, modal shape and modal damping.
4. A method of assessing the degradation trend of a variable operating condition bearing according to claim 3, characterized in that, The modal features are further preprocessed as follows: the first order natural frequency is taken as the effective value from the multi-order natural frequency parameters; and interpolation processing is performed on the missing values.
5. A method of assessing the degradation trend of a variable operating condition bearing according to any one of claims 1 to 4, characterized in that, The time domain features are root mean square value and kurtosis index.
6. A method of assessing the degradation trend of a variable operating condition bearing according to claim 5, characterized in that, The frequency domain feature is the root mean square frequency.
7. A device for assessing the degradation trend of a variable operating condition bearing, characterized by Comprise: An evaluation model construction module is configured to construct a bearing degradation trend evaluation model; the bearing degradation trend evaluation model comprises a feature extraction module, a feature fusion module, a random convolution kernel transformation module, a dimension reduction module and a label classifier; the feature extraction module is used for extracting modal features and time-frequency features of input data; The feature fusion module is used for fusing the time-frequency features and the modal features through canonical correlation analysis; the random convolution kernel transformation module is used for performing convolution operation on the fused feature parameters through a large number of multi-scale convolution kernels, and extracting multi-scale features related to bearing failure and unrelated to bearing failure; the dimension reduction module is used for dimension reduction projection of the multi-scale features, and constructing health state features; the label classifier is used for mapping the health state features after dimension reduction to corresponding predicted bearing health states; A training sample construction module is configured to use the vibration signal data of the existing multi-condition labeled bearing as source domain data, and the vibration signal data of the bearing without new condition and label as target domain data to form a training sample set; An iteration training module is configured to iteratively train the bearing degradation trend assessment model by using the training sample set; In the training process, the joint probability distribution of the random convolution kernel is initialized to be consistent with the data distribution of the bearing training set, and the length of the random convolution kernel is set to be the product of a value greater than the number of sampling points in one revolution of the rolling bearing and a normal distribution An online assessment module is configured to input the vibration data of the bearing to be measured under the current working condition into the trained model to obtain the bearing degradation trend assessment result.
8. A computer-readable storage medium, characterized in that, The computer readable storage medium comprises a stored computer program, wherein the computer program controls the device where the storage medium is located to perform the variable working condition bearing degradation trend assessment method according to any one of claims 1 to 6 when the computer program is run by a processor.