CN-116541687-B - Rolling bearing RUL prediction method

Abstract

The invention relates to the technical field of bearing detection and discloses a prediction method of a rolling bearing RUL, which comprises the following steps of S1, S2, feature compression fusion, primary extraction, feature size reduction, acquisition of compact features X f , S3, self-adaptive team feature response X f , acquisition of attention weighted features X ' f , S4, addition of Sinusoidal position codes to X' f to obtain features X PE with global position information, and S5, further learning of X PE by a stacked gating expansion causality convolution (GATED DILATED Causal Convolutional, GDCC) module, so that prediction of the bearing RUL is obtained. The accurate prediction of the bearing RUL is realized.

Inventors

• HE JING
• SUN WEI
• YU YUE
• ZHANG CHANGFAN

Assignees

• 湖南工业大学

Dates

Publication Date

20260505

Application Date

20230323

Claims (7)

1. 1. A method of predicting a rolling bearing RUL, comprising the steps of: S1, collecting vibration characteristics X of a bearing; s2, feature compression fusion is carried out on the primary extracted features, feature size is reduced, and compact features are obtained ; S3, for the said Performing self-adaptive correction on characteristic response to obtain attention weighted characteristic ; S4, pairing Adding Sinusoidal position codes to obtain features with global position information ; S5 stacked GDCC Module pairs Further learning is carried out, so that the prediction of the bearing RUL is obtained; the step S3 includes the steps of: S30, calculating a set of characteristic variable modulation weights under each time step , S31, calculating the set of the modulating weight of the characteristic map under the total time step of aggregation , S32, will be , And Fusion and calculate ; The said The calculation method of (1) is as follows: The said The calculation method of (1) is as follows: The said The calculation method of (1) is as follows: wherein N is the number of convolution kernels of the last convolution layer, and the convolution kernels , For explicitly modeling the correlation between feature variables at each time step; And Representing Relu and sigmoid activation functions, M being average pooling, BN representing batch normalization, convolution kernel , For explicit modeling of correlations between channels under a global distribution; the GDCC modules are stacked with an increasing expansion factor d, wherein the expansion factor d=1 for the first GDCC module, the expansion factor d=2 for the second GDCC module, and the expansion factor d=4 for the third GDCC module.
2. 2. The method for predicting a rolling bearing RUL according to claim 1, further comprising step S6, wherein the step S6 comprises updating the network weights by a gradient descent algorithm Until the iteration is terminated, the algorithm ends.
3. 3. The method according to claim 2, wherein said step S6 further comprises training the network by Adam optimizer and preserving the optimal weights during training 。
4. 4. The method according to claim 1, characterized in that said step S4 comprises the steps of: S40, calculating a position vector PE; s41, obtaining by utilizing Sinusoidal position coding module 。
5. 5. The method of claim 4, wherein the method of calculating the position vector PE is: wherein: An integer of the number of the two, Is that Is a function of the total time step of (1); Is the dimension of the position vector, and The dimension values are the same; an integer value in between; The said The calculation method of (1) is as follows: 。
6. 6. The method according to claim 1, wherein in step S5, sinusoidal position-coding module output is first inputted to GDCC Then output the predicted value of RUL 。
7. 7. The method of predicting a rolling bearing RUL according to claim 1, wherein step S5 comprises the steps of: S50: extracting deep features by sequentially passing through three GDCC modules with continuously increased expansion factors d ; S51 compression Feature map, obtaining final predicted features ; S52 processing using Relu functions Output of 。

Description

Rolling bearing RUL prediction method Technical Field The invention relates to the field of bearing detection, in particular to a prediction method of a rolling bearing RUL. Background In the modern industry, RUL (remaining life) prediction of mechanical equipment is becoming a key technology to ensure maximum continuous operation time and efficiency of mechanical equipment, reducing equipment maintenance costs. The existing mechanical RUL prediction method mainly comprises two main types, namely a model-based method and a data driving method. Model-based methods model the degradation process of a machine using first principles and failure mechanisms, and then use statistical estimation techniques, such as linear least squares, maximum likelihood estimation, and sequential Monte Carlo, to identify model parameters and predict RUL. However, in reality, the actual crack growth of the rolling bearing is extremely irregular, and it is difficult to establish an accurate failure model. Therefore, it is difficult to build an accurate mathematical statistics or physical degradation model in practical applications. In contrast, the data-driven approach does not require knowledge of the explicit failure mechanism of the machine, which can automatically infer causal relationships hidden in the data. In recent years, deep learning has become increasingly popular in data-driven RUL prediction. Deep learning has a more powerful representation learning ability than conventional machine learning techniques, and is capable of automatically learning multi-level representations from raw data. Therefore, with the help of the deep learning technology, a prediction model can be directly built based on the original sensor data, so that the complex process of artificial feature extraction is eliminated. Time series modeling is mostly implemented by a recurrent neural network (Recurrent Neural Network, RNN), where long short-term memory (LSTM) networks are most popular because it can solve the problems of gradient explosion and vanishing gradient of RNN before. However, LSTM has a limitation in that LSTM is a chain structure, and thus parallel operation cannot be achieved, resulting in slow running speed. In contrast, convolutional neural networks (Convolutional Neural Network, CNN) are able to process data in parallel, computing more efficiently. The hierarchical structure of CNNs provides a shorter path to capture remote dependencies between time steps than the chain structure of a cyclic network, and thus can also capture complex relationships well. Some results indicate that simple convolution structures are superior to canonical RNNs such as LSTM in both efficiency and accuracy. However, CNNs have two disadvantages, 1) the design of conventional CNNs cannot flexibly accommodate various time window sizes, and 2) CNNs have a corresponding deep structure to obtain sufficient receptive fields, and gradient vanishing easily occurs when the network is too deep. The prior art CN202210429223.4 discloses a rolling bearing RUL prediction method based on a model migration and wiener process, which is characterized by comprising the steps of acquiring a full-life rolling bearing time domain vibration signal under a working condition A as source domain data, acquiring the full-life rolling bearing time domain vibration signal under the working condition B as target domain data, inputting the source domain data and the target domain data into a health index model based on a single-layer non-negative constraint self-encoder network and a self-organizing feature mapping network, respectively acquiring health index labels of the source domain data and health index labels of the target domain data, preprocessing the source domain data and the target domain data, combining the preprocessed health index labels of the source domain data and the source domain data into a source domain pre-training model based on a depth non-negative constraint self-encoder network and a feedforward neural network, acquiring the source domain pre-training model parameters, wherein the source domain pre-training model parameters comprise weight parameters, migrating the source domain pre-training model parameters to the target domain network based on the depth non-negative constraint self-encoder network and the feedforward neural network as initial network parameters, respectively acquiring the health index labels of the source domain data and the target domain data, preprocessing the target domain data and the target domain data, performing the rolling bearing performance of the rolling bearing after the preprocessing and the target domain data are subjected to be subjected to a full-life degradation calculation according to the full-life rolling bearing performance degradation performance of the rolling bearing time domain pre-domain data, performing the rolling bearing performance degradation prediction model, and the rolling bearing performance of the rolling be