Method for predicting residual useful life of vibrating component based on deep multi-feature learning

Abstract

The present application relates to the technical field of failure prediction, in particular to a vibration component residual service life prediction method based on deep multi-feature learning, comprising the following steps: determining the first prediction time of the collected component original vibration signal to determine the start time of the component degradation period; using a sliding window to slide along the degradation period to obtain all vibration signal data within a period of time for time-frequency analysis to obtain time-frequency domain feature information of the degradation failure as input; further performing multi-scale feature extraction on the input time-frequency domain information and outputting the result; the beneficial effects are that the problem of unstable component residual life prediction results under multiple types of working conditions is solved; the ability to extract component degradation mechanism related features is strong, and the residual life prediction precision is high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault prediction, in particular to a vibration component residual service life prediction method based on deep multi-feature learning. BACKGROUND

[0002] With the continuous improvement of the intelligence and complexity of equipment, it faces great challenges to improve its reliability and stability in the service process. In the service process, the equipment usually produces vibration. Therefore, the research on the related predictive maintenance of the key components and systems of the equipment in the field of fault prediction and health management (PHM) based on vibration signals, including fault detection, fault diagnosis and life prediction, becomes increasingly important. As an important research branch of the field of fault prediction and health management, the residual life estimation (RUL) of the components has great engineering research significance. Therefore, it is very necessary to accurately predict the residual life of the components before the components fail.

[0003] In the prior art, data-driven methods can mine the direct relationship between historical degradation data and system failure mechanisms, thereby reducing the impact of human factors and accurately establishing the mapping relationship between the target system degradation state and the remaining useful life, which is gradually becoming the mainstream solution for RUL prediction. In recent years, some representative works have achieved good results around the application of deep learning in RUL of PHM. Literature [1] X. Li, W. Zhang, and Q. Ding, "Deep learning-based remaining useful life estimation of bearings using multi-scale feature extraction," Reliability Engineering & System Safety, vol. 182, pp. 208-218, 2019, doi: 10.1016 / j.ress.2018.11.011, [2] J. Zhu, N. Chen, and W. Peng, "Estimation of Bearing Remaining Useful Life Based on Multiscale Convolutional Neural Network," IEEE Transactions on Industrial Electronics, vol. 66, no. 4, pp. 3208-3216, 2019, doi: 10.1109 / tie.2018.2844856 and [3] B. Liu, Z. Gao, B. Lu, H. Dong, and Z. An, "SAL-CNN: Estimate the Remaining Useful Life of Bearings Using Time-frequency Information," arXiv preprint arXiv:2204.05045, 2022 all propose a RUL estimation method combining time-frequency analysis and convolutional neural network. After obtaining the time series degradation signal, time-frequency analysis technology is used to obtain the time-frequency representation of the component degradation signal, a multi-scale feature extractor composed of multiple convolutional layers is used to extract multi-scale features for each sequence, and then the feature maps of different feedforward convolutional layers are connected to further extract higher-level abstract features. Finally, the compressed features are sent to the fully connected layer for RUL estimation.

[0004] However, from the above literature research results, these methods combined with convolutional neural networks have difficulties in processing long sequences and filtering invalid features, are limited by the limited receptive field and weak global context modeling ability during convolution calculation, lack correlation ability in multi-scale feature extraction, and inevitably ignore important degradation information as more degradation information is ignored, resulting in low efficiency in capturing the correlation between long-term degradation features, and too much redundancy and invalid multi-scale degradation information hinder the model from learning accurate and effective high-level representations, thereby affecting the model generalization ability and prediction performance. In addition, in industrial practical applications, more and more operating conditions and failure modes make the degradation features more complex, and these traditional methods based on convolutional neural networks have great difficulty in multi-scale feature extraction related to remaining useful life prediction when facing such complex working conditions, and the RUL estimation result is not satisfactory. SUMMARY

[0005] The present application aims to provide a vibration component remaining useful life prediction method based on deep multi-feature learning to solve the problems raised in the above background.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: a vibration component remaining useful life prediction method based on deep multi-feature learning, the prediction method comprising the following steps:

[0007] The first prediction time determination is performed on the collected component original vibration signal to determine the start time of the degradation period of the component;

[0008] The time-frequency analysis is performed on all vibration signal data in a period of time using a sliding window along the degradation period to obtain time-frequency domain feature information of the degradation failure as input;

[0009] The input time-frequency domain information is further subjected to multi-scale feature extraction and the results are output.

[0010] Preferably, the first prediction time satisfies the formula:

[0011] {|k t-i -μ|>3σ} i＝0,1,2 ;

[0012] Where k t-i represents the kurtosis value of the t-i vibration sequence from monitoring, the average value μ and the kurtosis standard deviation σ are calculated according to the historical data of the early signal sampling; when the continuous three kurtosis values exceed the 3σ interval, it is determined that the component has entered the degradation period.

[0013] Preferably, the vibration signal is preprocessed using short-time Fourier transform to obtain a two-dimensional time-frequency graph for extracting degradation features.

[0014] Preferably, the Fourier transform is performed on the entire signal segment by window segmentation, and each small segment under a time window is approximated as a locally stationary signal, then the signal is converted into a time-frequency domain by sliding the window function along the time axis.

[0015] Preferably, for a non-stationary signal x(t), the expression after short-time Fourier transform is:

[0016]

[0017] Where h(τ-t) is the analysis window function.

[0018] Preferably, the time-frequency domain information input by the deep multi-layer feature extractor is further subjected to multi-scale feature extraction and result output, and the deep multi-layer feature extractor is composed of four layers of moving window transformer blocks, wherein each block has an array feature extraction processor composed of two Swin Transformer encoders.

[0019] Preferably, in a standard window self-attention mechanism, multiple attention heads of different heads are connected.

[0020] Preferably, a sliding window is introduced in the model, and the time-frequency graph of the sequence vibration signal in each sliding window is cut into different patch labels while sliding along the time axis, and then sequentially sent to multiple moving window transformer modules for processing.

[0021] Preferably, the attention calculation formula is:

[0022]

[0023] Wherein Query, key and value matrices, M 2 Represents the number of small pieces of each window, and d is the dimension of the query or key, B∈R L×L Represents the relative position bias, wherein L is the length of the input sequence.

[0024] Preferably, in the multi-layer moving window transformer structure, four processing blocks are set up, each block is composed of a plurality of moving window transformers and slice merging operations, which are used to learn the degradation mechanism of the input data, fuse the high-level features extracted at different frequencies and different time scales, and finally send the results to a multi-layer perceptron.

[0025] Compared with the prior art, the beneficial effects of the present application are:

[0026] This invention proposes a method for predicting the remaining service life of vibrating components based on deep multi-feature learning, which solves the problem of unstable prediction results under various working conditions. It improves upon existing methods for predicting the remaining service life of components based on multi-scale feature learning using convolutional neural networks. Existing methods are limited by the narrow window range of convolutional computation, resulting in a limited receptive field and difficulties in processing input sequence information for global modeling and filtering redundant and invalid feature information. Therefore, we introduce a transformer model, which effectively addresses the common problems of traditional methods. Through moving windows and window self-attention mechanisms, it can not only effectively establish long-distance relationships between features at different scales but also improve the ability to filter redundant and invalid information, effectively enhancing the generalization ability of multi-scale feature extraction methods under complex working conditions.

[0027] This novel deep RUL prediction method boasts strong capabilities in extracting features relevant to component degradation mechanisms and high accuracy in remaining lifetime prediction. It is entirely based on a window self-attention mechanism, fully considering the impact of excessive invalid information learning on prediction performance. A multi-scale feature extractor starts from the time and frequency scales of historical time-frequency domain sequences, effectively extracting feature representations of different granularities using a self-attention mechanism. Finally, a multilayer perceptron is employed to fuse the extracted multi-scale degradation features and output the remaining lifetime prediction result. This method can learn the features most relevant to the degradation mechanism of vibration components under complex operating conditions and provide accurate remaining lifetime prediction results. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the feature extraction processor based on Swing Transformer of the present invention;

[0029] Figure 2 This is a schematic diagram of the prediction method of the present invention;

[0030] Figure 3 The complete lifecycle vibration signal diagrams for components 1-4 are shown in the PHM 2012 dataset.

[0031] Figure 4 This is a schematic diagram illustrating the predicted curve effect of the present invention on some test components. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the present invention clear and complete, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only some, not all, embodiments of the present invention, and are merely illustrative of the embodiments of the present invention. They are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Example 1

[0034] Please see Figures 1 to 2 This invention provides a technical solution: a method for predicting the remaining service life of vibrating components based on deep multi-feature learning, the prediction method comprising the following steps:

[0035] The initial prediction time of the acquired original vibration signal of the component is determined to determine the start time of the component's decay period;

[0036] A sliding window is used to slide along the decay period to collect all vibration signal data within a certain time period for time-frequency analysis, so as to obtain the time-frequency domain characteristic information of the degradation fault as input;

[0037] The input time-frequency domain information will be further processed by multi-scale feature extraction and result output.

[0038] Example 2

[0039] A method for predicting the remaining service life of vibrating components based on deep multi-feature learning, the method comprising the following steps:

[0040] 1. First Prediction Time Determination (FPT) Algorithm

[0041] Since the research focuses on the raw vibration signals of components, it typically involves collecting a complete signal sample from the start of operation until the equipment stops due to a failure. This sample is usually divided into a healthy phase and a degradation phase. During the early stages of operation, the equipment is in a healthy state, and the data from this period does not show significant differences, so RUL estimation is unnecessary. However, using data from this phase can actually lead to interference noise that affects the learning of degradation characteristics. Once a certain threshold is reached, i.e., the equipment fails, the degradation index will increase exponentially as the failure worsens.

[0042] Therefore, to ensure the accuracy of RUL prediction, we must perform fault detection to determine the stage at which degradation begins, i.e., the first prediction time. The first prediction time must satisfy formula (1).

[0043] {|k t-i -μ|>3σ} i＝0,1,2 (1)

[0044] Where k t-i This indicates that the kurtosis value of the ti-th vibration sequence from the monitoring is calculated, and its corresponding average value μ and kurtosis standard deviation σ are calculated based on historical data from earlier signal sampling. When three consecutive kurtosis values ​​exceed the 3σ interval, the component is considered to have entered the degradation period.

[0045] 2. Short-time Fourier Transform Principle

[0046] Time-frequency analysis is an effective method for studying non-stationary signals. In this paper, short-time Fourier transform (SFT) is used to preprocess the vibration signal to obtain a two-dimensional time-frequency plot for extracting degradation features. SFT is a joint time-frequency analysis method that performs Fourier transform on the entire signal segment through windowing; each small segment within a time window can be approximated as a locally stationary signal. The window function slides along the time axis, and the signal can then be converted to the time-frequency domain. For a non-stationary signal x(t), its representation after SFT is as follows:

[0047]

[0048] Where h(τ-t) is the analysis window function. In particular, the Hamming window is chosen as the window function.

[0049] 3. Deep Hierarchical Extractor Algorithm and Self-Attention Mechanism Principle

[0050] The depth layer extractor consists of several stacked transformer encoder structures, forming a four-layer moving window transformer module. Its key component is the design of the window self-attention mechanism. Figure 2 The second part details the design of the multi-layered moving window transformer structure, which consists of four layers of moving window transformer blocks. Each block contains an array feature extraction processor consisting of two SwinTransformer encoders as a group, such as... Figure 1 As shown.

[0051] In a standard window self-attention mechanism, the attention of multiple different heads is connected. We introduce a sliding window into the model, which slides along the time axis while the time-frequency plot of the sequence vibration signal within each sliding window is segmented into different patch markers and sequentially fed into multiple moving window transformer modules for processing. The self-attention calculation can be performed in the following form:

[0052]

[0053] in These are the query, key, and value matrix, M. 2 B represents the number of small pieces in each window, while d is the dimension of the query or key, where B∈RL×L This represents the relative positional deviation, where L is the length of the input sequence.

[0054] Example 3

[0055] A method for predicting the remaining service life of vibrating components based on deep multi-feature learning, the method comprising the following steps:

[0056] The system consists of two parts: data preprocessing and a deep multilayer feature extractor. In the data preprocessing stage, the initial prediction time of the acquired raw vibration signal of the component is first determined to identify the start time of the component's degradation period. Then, a sliding window is used to slide along the degradation period to extract all vibration signal data within a certain time period for time-frequency analysis to obtain the time-frequency domain feature information of the degradation fault as input.

[0057] In the deep multilayer feature extractor section, the input time-frequency domain information is further processed for multi-scale feature extraction and output. The multilayer moving window transformer structure comprises four processing blocks, each consisting of several moving window transformers and slice merging operations. These blocks are used to learn the degradation mechanism of the input data, fuse high-level features extracted from different frequencies and time scales, and finally feed the results into the multilayer perceptron.

[0058] Example 4

[0059] Experimental data were obtained from two popular vibration component (rolling bearing) degradation datasets: the PHM 2012 Data Challenge and the XJTU-SY bearing degradation dataset from Xi'an Jiaotong University. Twenty groups of bearings under four different operating conditions were selected, representing data from healthy operation to failure. Each bearing group faced different fault types and intensities during operation. The raw horizontal vibration signals from these datasets were used to verify the prediction accuracy of the remaining service life method for rolling bearings under complex operating conditions. Figure 3 This shows the raw horizontal vibration signal of a rolling bearing throughout its entire life cycle, including the healthy and deteriorating periods.

[0060] Once the FPT (Front-Term Physical Time) of each test component is determined, the operating-to-failure data for each bearing can be categorized into healthy and degraded states, allowing for the extraction of time-frequency domain information between the FPT and end-of-life. Furthermore, label normalization has been widely used in studies related to RUL (Remaining Usage Limit) prediction. Therefore, we define the normalized remaining service life label during degradation as follows:

[0061]

[0062] in This represents the RUL label value at the nth time after the start of the decay period, where N is the total number of samples in the decay period. Thus, RUL is converted into a ratio between 0 and 1, which is used to represent the actual and predicted RUL. Corresponding labels can be established for all test samples in the decay state, thereby obtaining a supervised RUL prediction training set and validation set.

[0063] The network training process involves updating parameters using the backpropagation algorithm. Since the RUL prediction problem is a regression problem, the mean square error is used as the loss function to train the proposed model. As shown in Equation 5, the more accurate the network's prediction, the lower the loss error.

[0064]

[0065] in This represents the RUL label value at the nth time after the start of the recession, where N is the total number of samples during the recession. It is the labeled RUL value output by the prediction model.

[0066] For training, all test bearings were used as prediction components in turn. The remaining bearing data in the other datasets were labeled and used as the training and validation sets for each experiment to obtain cross-validation results. Training samples were randomly divided into mini-batches of a fixed size of 48. In each epoch, such a mini-batch of 48 samples was fed into the network. We used an improved AdamW optimizer, which adaptively adjusts the learning rate and iteratively optimizes the network parameters, and set the epoch to 300.

[0067] Table 1 and Figure 4 The presentation showcases the partial RUL prediction performance of our patented technology on test data. It can be observed that the overall prediction results of the proposed method can track the actual RUL label value well, and can accurately estimate the remaining life trend of the test bearing. In most cases, the prediction deviation of a single test bearing is controlled within 10% (i.e., 0.010).

[0068] Table 1 shows partial performance results of the patented technology on the test dataset.

[0069] .

[0070] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the remaining service life of vibrating components based on deep multi-feature learning, characterized in that, The prediction method includes the following steps: The initial prediction time of the acquired original vibration signal of the component is determined to determine the start time of the component's decay period; A sliding window is used to slide along the decay period to collect all vibration signal data within a certain time period for time-frequency analysis, so as to obtain the time-frequency domain characteristic information of the degradation fault as input; The input time-frequency domain information will be further processed by multi-scale feature extraction and the results will be output. The time-frequency domain information input by the deep multi-layer feature extractor is further used for multi-scale feature extraction and output. The deep multi-layer feature extractor consists of four moving window transformer blocks, each of which contains an array feature extraction processor consisting of two SwingTransformer encoders. In a standard window self-attention mechanism, the attention of multiple different heads is connected. A sliding window is introduced into the model. While sliding along the time axis, the time-frequency map of the sequence vibration signal within each sliding window is cut into different patch markers and sent sequentially to multiple moving window transformer modules for processing.

2. The method for predicting the remaining service life of vibrating components based on deep multi-feature learning according to claim 1, characterized in that: The first prediction time satisfies the formula: ； in This represents the calculation of the kurtosis value of the monitored vibration sequence at time ti, calculated based on the corresponding average value from historical data of earlier signal sampling. and kurtosis standard deviation ; When three consecutive kurtosis values ​​exceed 3 If the value is within a certain range, then the component is considered to have entered the degradation period.

3. The method for predicting the remaining service life of vibrating components based on deep multi-feature learning according to claim 1, characterized in that: The vibration signal was preprocessed using short-time Fourier transform to obtain a two-dimensional time-frequency diagram for extracting degradation features.

4. The method for predicting the remaining service life of vibrating components based on deep multi-feature learning according to claim 1, characterized in that: The entire signal segment is subjected to Fourier transform by windowing. The small segment under each time window is approximated as a locally stationary signal. The window function slides along the time axis, and then the signal is converted into the time-frequency domain.

5. The method for predicting the remaining service life of vibrating components based on deep multi-feature learning according to claim 1, characterized in that: For non-stationary signals Its representation after short-time Fourier transform is as follows: ； in For analysis window functions.

6. The method for predicting the remaining service life of vibrating components based on deep multi-feature learning according to claim 1, characterized in that: Attention calculation formula: in , , These are query, key, and value matrices, respectively. This represents the number of small pieces in each window, while d is the dimension of the query or key. This represents the relative positional deviation, where L is the length of the input sequence.

7. The method for predicting the remaining service life of vibrating components based on deep multi-feature learning according to claim 1, characterized in that: In the multilayer moving window transformer structure, four processing blocks are set up. Each block consists of several moving window transformers and slice merging operations. It is used to learn the degradation mechanism of input data, fuse high-level features extracted from different frequencies and time scales, and finally send them to the multilayer perceptron to output the results.

Images