Bearing life prediction method based on feature extraction and IBGC network

Abstract

The invention provides a bearing life prediction method based on feature extraction and an IBGC network. The method comprises the following steps: collecting a vibration signal group of a bearing by using a vibration sensor; processing any vibration signal group by adopting a completely adaptive noise set empirical mode decomposition method to obtain a reconstructed signal; extracting a time domain feature, a frequency domain feature and a time-frequency domain degradation feature from the reconstructed signal, and selecting A optimal degradation features from the time domain feature, the frequency domain feature and the time-frequency domain degradation feature; standardizing the data corresponding to the A preferred degradation features and setting RUL tags to obtain a degradation feature data set, and dividing a training set from the degradation feature data set; constructing an IBGC model, and training the IBGC model by adopting the training set; and inputting a test sample into the trained IBGC model, and outputting a prediction result by the trained IBGC model, thereby realizing accurate prediction of the residual service life of the bearing.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing remaining life prediction, and in particular, to a bearing life prediction method based on feature extraction and IBGC network. Background Art

[0002] Affected by factors such as the evolution of internal mechanisms and changes in the external environment, rotating mechanical equipment such as truck gearboxes often inevitably degrade, resulting in reduced performance and deterioration of their health status. In particular, if bearings, as key components, degrade, it may cause the entire equipment to malfunction, or even cause failure and cause immeasurable loss of life and property. As the core of the field of prognostic health management (PHM), the remaining useful life (RUL) prediction technology aims to accurately predict the remaining useful life of equipment through equipment degradation monitoring data. It is a key technology to reduce the risk of unplanned equipment failure, maintain degraded equipment in a timely manner, and reduce equipment maintenance time and costs.

[0003] In the process of feature extraction and RUL prediction of bearings, a single degradation feature may only contain information about a specific degree of degradation, that is, the degradation feature has different sensitivities to different degrees of degradation and cannot reflect the performance changes of rolling bearings throughout their life cycle. Therefore, it is necessary to pay attention to features with high degradation sensitivity at different monitoring moments. In addition, the traditional Bi-GRU network model does not have the ability to pay attention to time series and channels. It treats the input of each dimension and each time step equally, resulting in many irrelevant factors being considered in the RUL prediction process, which has caused certain interference to the prediction results.

[0004] In view of this, the present invention provides a bearing life prediction method based on feature extraction and IBGC network. An improved self-attention mechanism is added to the Bi-GRU network, and combined with the channel attention mechanism (CAM), an ISA-Bi-GRU-CAM (IBGC) network model is constructed, which helps to improve the prediction ability of the Bi-GRU network in the RUL prediction of rolling bearings. Summary of the invention

[0005] According to the above-mentioned deficiencies, a bearing life prediction method based on feature extraction and IBGC network is provided. The present invention mainly utilizes CEEMDAN for signal decomposition and combines feature criteria for signal reconstruction, which can effectively remove noise and impact in the original vibration signal. Feature extraction is performed based on time domain, frequency domain, and time-frequency domain, and feature optimization is performed based on comprehensive indicators to provide high-quality degradation features for the subsequent RUL. The constructed IBGC network model uses Bi-GRU to achieve deep mapping of degradation timing information and RUL. Combined with the improved self-attention mechanism, the input features of the time step closely related to RUL are enhanced, thereby improving the efficiency of feature extraction in the network. The channel attention mechanism is adopted to learn the importance of deep features, and different weights are assigned to deep features, so that the network model pays attention to important features, thereby enhancing the stability of the network model and improving the RUL prediction accuracy.

[0006] The technical means adopted by the present invention are as follows:

[0007] The present invention provides a bearing life prediction method based on feature extraction and IBGC network, comprising:

[0008] A vibration sensor is used to collect a vibration signal group of the bearing;

[0009] Using a fully adaptive noise ensemble empirical mode decomposition method to process any of the vibration signal groups to obtain a reconstructed signal;

[0010] Extracting time domain features, frequency domain features and time-frequency domain degradation features from the reconstructed signal, and selecting A preferred degradation features from the time domain features, the frequency domain features and the time-frequency domain degradation features by using a monotonicity index, a trend index and a robustness index, where A is an integer and A≥2;

[0011] Standardizing the data corresponding to the A preferred degradation features in the vibration signal group and setting RUL labels to obtain a degradation feature data set, and dividing a training set from the degradation feature data set;

[0012] Constructing an IBGC model, and training the IBGC model using the training set;

[0013] The test sample is input into the trained IBGC model, and the trained IBGC model outputs a prediction result.

[0014] Preferably, the method of using a fully adaptive noise set empirical mode decomposition method to process any of the vibration signal groups to obtain a reconstructed signal includes:

[0015] Adding Gaussian white noise to the vibration signal group to obtain the processed vibration signal group;

[0016] Iterating according to the processed vibration signal group, including: in the jth round of iteration, obtaining the jth intrinsic mode component and the corresponding jth residual component, when k rounds of iteration are performed and the k residual components obtained are monotonic and the number of extreme points is less than 3, the iteration is stopped, and k intrinsic mode components are obtained, 1≤j≤k;

[0017] Calculating the normalized kurtosis and normalized cross-correlation coefficient of the k eigenmode components respectively, arranging the k normalized kurtosis and normalized cross-correlation coefficients from large to small, and selecting the eigenmode components corresponding to the first five normalized kurtosis and normalized cross-correlation coefficients;

[0018] Signal reconstruction is performed according to the selected five intrinsic mode components to obtain the reconstructed signal.

[0019] Preferably, the adding of Gaussian white noise to the vibration signal group to obtain the processed vibration signal group is calculated in the following manner:

[0020] x(n) ′ =x(n)+ε 0 ω i (n)

[0021] Where x(n) ′ is the vibration signal group after processing, x(n) is the vibration signal group, ε 0 is the amplitude of the Gaussian white noise, ω i (n) is the Gaussian white noise added for the ith time.

[0022] Preferably, in the j-th iteration, the j-th intrinsic mode component and the corresponding j-th residual component are obtained, and are calculated in the following manner:

[0023] When j=1, the processed vibration signal group is subjected to EMD decomposition processing to obtain I components of the first round of iteration, and the average value of the I components of the first round of iteration is calculated as the first intrinsic mode component; the processed vibration signal group is subtracted from the first intrinsic mode component to obtain the corresponding first residual component;

[0024] When 2≤j≤k, the j-th eigenmode component and the corresponding j-th residual component are calculated according to the j-1-th residual component.

[0025] Preferably, when 2≤j≤k, the j-th intrinsic mode component and the corresponding j-th residual component are calculated according to the j-1-th residual component, and are calculated in the following manner:

[0026]

[0027] R j(n) = R j-1 (n)-IMF j (n)

[0028] Among them, E j [] is the j-th EMD decomposition operation, I is the I component of the j-th iteration, IMF j (n) is the jth eigenmode component, R j-1 (n) is the j-1th residual component, R j (n) is the jth residual component, ε j is the amplitude of the Gaussian white noise added for the jth time.

[0029] Preferably, the selecting A preferred degradation features from the time domain features, the frequency domain features and the time-frequency domain degradation features by using the monotonicity index, the trend index and the robustness index comprises:

[0030] The time domain feature, the frequency domain feature and the time-frequency domain degradation feature are B features in total, where B is an integer and B≥3, and A≤B;

[0031] Calculating the multi-angle comprehensive evaluation scores of the features respectively, including: calculating the monotonicity index, trend index and robustness index of the features; calculating the multi-angle comprehensive evaluation scores of the features according to the monotonicity index, trend index and robustness index of the features;

[0032] Arrange the B multi-angle comprehensive evaluation scores from large to small, and select the features corresponding to the first A multi-angle comprehensive evaluation scores as the preferred degradation features.

[0033] Preferably, the multi-angle comprehensive evaluation score of the feature is calculated in the following manner:

[0034] HI criteria(X) =0.3×Mon(X)+0.3×Tre(T,X)+0.3×Rob(X)

[0035] Among them, HI criteria(X) is the multi-angle comprehensive evaluation score of the Xth feature, Mon(X) is the monotonicity index of the Xth feature, Tre(T,X) is the trend index of the Xth feature, and Rob(X) is the robustness index of the Xth feature.

[0036] Preferably, the step of normalizing the data corresponding to the A preferred degradation features in the vibration signal group comprises:

[0037] Any of the above data is standardized and calculated as follows:

[0038]

[0039] Among them, X 1:k is the data after standardization, X i is the characteristic value of the data at time i, μ i:k is the mean of the characteristic values ​​of the data at all times, δ 1:k is the standard deviation of the characteristic values ​​of the data at all times.

[0040] Preferably, the constructing of the IBGC model comprises:

[0041] The query matrix, key matrix, and value matrix in the self-attention mechanism are replaced by three 1×1 convolutions to obtain an improved self-attention mechanism.

[0042] The improved self-attention mechanism, the stacked Bi-GRU layer and the channel attention layer are sequentially cascaded to obtain the IBGC model.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] 1. The bearing life prediction method based on feature extraction and IBGC network provided by the present invention can effectively remove noise and impact in the original vibration signal by using CEEMDAN for signal decomposition and combining feature criteria for signal reconstruction. Feature extraction is performed based on the time domain, frequency domain, and time-frequency domain, and feature optimization is performed based on comprehensive indicators to provide high-quality degradation features for the subsequent RUL. The constructed IBGC network model uses Bi-GRU to achieve deep mapping of degradation timing information and RUL. Combined with the improved self-attention mechanism, the input features of the time steps closely related to RUL are enhanced, thereby improving the efficiency of feature extraction in the network. The channel attention mechanism is used to learn the importance of deep features, and different weights are assigned to deep features, so that the network model pays attention to important features, thereby enhancing the stability of the network model and improving the RUL prediction accuracy.

[0045] 2. The bearing life prediction method based on feature extraction and IBGC network provided by the present invention is applied to the fault prediction and health management of rotating machinery such as bearings and gears, and realizes the online and offline prediction of the remaining service life of the equipment. It provides a theoretical basis and technical support for the intelligent operation and maintenance decision-making of rotating machinery. The technical application of this invention can improve the safety, reliability and availability of enterprise mechanical equipment and reduce the maintenance cost of equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0047] Figure 1 A schematic flow chart of a bearing life prediction method based on feature extraction and IBGC network provided by the present invention.

[0048] Figure 2 A flowchart of the improved self-attention mechanism.

[0049] Figure 3 A schematic diagram of a flow chart of the IBGC model provided by the present invention.

[0050] Figure 4 A schematic diagram of the monotonicity of a bearing.

[0051] Figure 5 A schematic diagram of the bearing trend.

[0052] Figure 6 A schematic diagram of the bearing robustness.

[0053] Figure 7 A schematic diagram of the weighted evaluation index of the bearing

[0054] Figure 8 A schematic diagram of the prediction results of the method of the present invention. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0056] It should be noted that the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or apparatus.

[0057] Combination Figure 1 , Figure 1A flow chart of the bearing life prediction method based on feature extraction and IBGC network provided by the present invention is provided to illustrate a specific embodiment of the bearing life prediction method based on feature extraction and IBGC network provided by the present invention, which is characterized by comprising:

[0058] A vibration sensor is used to collect a vibration signal group of the bearing;

[0059] A fully adaptive noise ensemble empirical mode decomposition method is used to process any vibration signal group to obtain a reconstructed signal;

[0060] Extract time domain features, frequency domain features and time-frequency domain degradation features from the reconstructed signal, and select A preferred degradation features from the time domain features, frequency domain features and time-frequency domain degradation features using monotonicity index, trend index and robustness index, where A is an integer and A≥2;

[0061] The data corresponding to the A preferred degradation features in the vibration signal group are standardized and RUL labels are set to obtain a degradation feature data set, and a training set is divided from the degradation feature data set;

[0062] Construct an IBGC model and use the training set to train the IBGC model;

[0063] The test sample is input into the trained IBGC model, and the trained IBGC model outputs the prediction result.

[0064] It should be noted that the vibration signal group is state data that can reflect bearing degradation information. The fully adaptive noise ensemble empirical mode decomposition method, namely CEEMDAN, uses CEEMDAN for decomposition and can filter and reduce noise on vibration signals with serious nonlinear, non-stationary characteristics and noise interference.

[0065] It can be understood that the bearing life prediction method based on feature extraction and IBGC network provided in this embodiment can effectively remove noise and impact in the original vibration signal by using CEEMDAN for signal decomposition and combining feature criteria for signal reconstruction. Feature extraction is performed based on the time domain, frequency domain, and time-frequency domain, and feature optimization is performed based on comprehensive indicators to provide high-quality degradation features for the subsequent RUL. The constructed IBGC network model uses Bi-GRU to achieve deep mapping of degradation timing information and RUL. Combined with the improved self-attention mechanism, the input features of the time step closely related to RUL are enhanced, thereby improving the efficiency of feature extraction in the network. The channel attention mechanism is used to learn the importance of deep features, and different weights are assigned to deep features, so that the network model focuses on important features, thereby enhancing the stability of the network model and improving the RUL prediction accuracy.

[0066] In some optional embodiments, a fully adaptive noise set empirical mode decomposition method is used to process any vibration signal group to obtain a reconstructed signal, including:

[0067] Adding Gaussian white noise to the vibration signal group to obtain a processed vibration signal group;

[0068] Iteration is performed according to the processed vibration signal group, including: in the jth round of iteration, the jth intrinsic mode component and the corresponding jth residual component are obtained, when the kth round of iteration is performed and the k residual components obtained are monotonic and the number of extreme value points is less than 3, the iteration is stopped, and k intrinsic mode components are obtained, 1≤j≤k;

[0069] Calculate the normalized kurtosis and normalized cross-correlation coefficient of k eigenmode components respectively, arrange the k normalized kurtosis and normalized cross-correlation coefficient from large to small, and select the eigenmode components corresponding to the first five normalized kurtosis and normalized cross-correlation coefficients;

[0070] The signal is reconstructed according to the selected five eigenmode components to obtain a reconstructed signal.

[0071] It is understandable that it cannot be guaranteed that every intrinsic mode component (IMF) is useful. Therefore, in order to comprehensively and accurately reflect the signal fault characteristic information in the time domain and frequency domain, in this embodiment, in a possible implementation method, the normalized kurtosis and the normalized cross-correlation coefficient of each IMF are calculated respectively in combination with the kurtosis and the cross-correlation coefficient. The IMFs corresponding to the first five weighted maximum values ​​should contain the most fault information. Therefore, these five IMFs are selected for signal reconstruction.

[0072] The determination process of the 5 IMFs is calculated as follows:

[0073] imp i =Normal(kur(IMF i ))+Normal(mur(IMF i ))

[0074] IMF target =max 1:5 (imp i )

[0075] Among them, kur is the kurtosis, mur is the correlation coefficient, Normal is the standardization operation, imp i is the importance of the i-th eigenmode component, IMF target is the selected eigenmode component group.

[0076] In some optional embodiments, Gaussian white noise is added to the vibration signal group to obtain a processed vibration signal group, which is calculated in the following manner:

[0077] x(n) ′ =x(n)+ε 0 ω i (n)

[0078] Where x(n) ′ is the processed vibration signal group, x(n) is the vibration signal group, ε 0 is the amplitude of Gaussian white noise, ω i (n) is the Gaussian white noise added for the i-th time.

[0079] In some optional embodiments, in the jth iteration, the jth eigenmode component and the corresponding jth residual component are obtained and calculated as follows:

[0080] When j=1, the processed vibration signal group is subjected to EMD decomposition processing to obtain I components of the first round of iteration, and the average value of the I components of the first round of iteration is calculated as the first intrinsic mode component; the processed vibration signal group is subtracted from the first intrinsic mode component to obtain the corresponding first residual component;

[0081] It can be understood that when j=1, it can be calculated as follows:

[0082]

[0083] R 1 (n) = x(n) ′ -IMF 1 (n)

[0084] Among them, IMF 1 (n) is the first eigenmode component, R 1 (n) is the first residual component, and I is the I component of the first iteration.

[0085] When 2≤j≤k, the jth intrinsic mode component and the corresponding jth residual component are calculated based on the j-1th residual component.

[0086] Specifically, when 2≤j≤k, the jth intrinsic mode component and the corresponding jth residual component are calculated according to the j-1th residual component, and are calculated in the following manner:

[0087]

[0088] R j (n) = R j-1 (n)-IMF j (n)

[0089] Among them, E j [] is the j-th EMD decomposition operation, I is the I component of the j-th iteration, IMFj (n) is the jth eigenmode component, R j-1 (n) is the j-1th residual component, R j (n) is the jth residual component, ε j is the amplitude of the Gaussian white noise added for the jth time.

[0090] In some optional embodiments, a monotonicity index, a trend index, and a robustness index are used to select A preferred degradation features from time domain features, frequency domain features, and time-frequency domain degradation features, including:

[0091] There are B features in total, namely, time domain features, frequency domain features and time-frequency domain degradation features, where B is an integer and B≥3, A≤B;

[0092] Calculating the multi-angle comprehensive evaluation scores of the features respectively, including: calculating the monotonicity index, trend index and robustness index of the features; calculating the multi-angle comprehensive evaluation scores of the features according to the monotonicity index, trend index and robustness index of the features;

[0093] Arrange the B multi-angle comprehensive evaluation scores from large to small, and select the features corresponding to the first A multi-angle comprehensive evaluation scores as the preferred degradation features.

[0094] It can be understood that this embodiment extracts multiple time domain degradation indicators from the reconstructed signal; extracts the spectrum effective value, spectrum mean, and four bearing fault characteristic frequencies as degradation characteristics from the frequency domain, and the four bearing fault characteristic frequencies are specifically the outer ring fault frequency, the inner ring fault frequency, the cage fault frequency, and the rolling element fault frequency. The specific characteristics and calculation formulas of the time domain and the frequency domain are shown in Table 1 below:

[0095] Table 1. Time domain and frequency domain feature calculation table

[0096]

[0097] Then select bior3.1 basis function and 4-layer decomposition to obtain 16 wavelet packet coefficients, calculate the energy of each node and normalize it, and select the normalized energy of the first 8 nodes as the time-frequency domain degradation feature. Therefore, there are 24 features in total, including time domain features, frequency domain features and time-frequency domain degradation features. Of course, it is not limited to this.

[0098] In some optional embodiments, the multi-angle comprehensive evaluation score of the feature is calculated as follows:

[0099] HI criteria(X) =0.3×Mon(X)+0.3×Tre(T,X)+0.3×Rob(X)

[0100] Among them, HI criteria(X)is the multi-angle comprehensive evaluation score of the Xth feature, Mon(X) is the monotonicity index of the Xth feature, Tre(T,X) is the trend index of the Xth feature, and Rob(X) is the robustness index of the Xth feature.

[0101] It can be understood that monotonicity measures the tendency of a feature sequence to increase or decrease monotonically and is calculated as follows:

[0102]

[0103] Among them, K is the length of the feature, d / dx is the derivative, and X is the Xth feature.

[0104] Trendiness measures the degree of correlation between feature sequences and time and is calculated as follows:

[0105]

[0106] Among them, T is time, cov() is covariance, σ is standard deviation, and rg is the rank of the abnormality.

[0107] The robustness index of a feature is an indicator that measures the stability and resistance of the feature to interference, noise or other changes. Better robustness helps the stability of RUL prediction and is calculated as follows:

[0108]

[0109] Among them, Y(k) and They are the original features and the features after adding noise.

[0110] A single evaluation index can only measure some characteristics of the degradation feature. Relying on a single index is not enough to select good degradation features for model prediction. In order to achieve a comprehensive evaluation of the degradation features from multiple angles, this embodiment uses a weighted combination of the above evaluation indicators as the standard for selecting the preferred features. After obtaining the comprehensive evaluation of each indicator, the top 14 features with the highest evaluation are selected as the preferred degradation features.

[0111] In some optional embodiments, the data corresponding to the A preferred degradation features in the vibration signal group are standardized, including:

[0112] To standardize any data, calculate as follows:

[0113]

[0114] Among them, X 1:k is the data after standardization, X i is the eigenvalue at time i in the data, μ i:k is the mean of the eigenvalues ​​of all the data at all times, δ1:k is the standard deviation of the eigenvalues ​​of the data at all times.

[0115] It is understandable that the data is standardized (Z-score standardization) to eliminate the impact of inconsistent orders of magnitude and dimensions between degenerate features on the network model.

[0116] During long-term use, rolling bearings are often in a stable operating state. At this time, the degradation indicators change slowly, and the linear change trend of RUL cannot be accurately predicted. If the actual RUL is directly used for model training, the accuracy of early predictions may be reduced. Therefore, in this application, a piecewise function is used to represent the RUL change of the rolling bearing, and the vibration amount of the rolling bearing in the healthy period is used as the standard value. When the vibration amount of the rolling bearing continuously exceeds the standard value, it is defined as the degradation starting point. Set the RUL label to "0" before the starting point, then increase the RUL label linearly, and set the RUL mark to "1" when it fails, that is, it is set to 0 when the bearing begins to fail, and it is set to 1 when it fails completely. Only with the features and corresponding labels can the next model be trained.

[0117] In some optional embodiments, referring to Figure 2 and Figure 3 , Figure 2 A flowchart of the improved self-attention mechanism. Figure 3 A schematic diagram of a flow chart of the IBGC model provided by the present invention, constructing the IBGC model, including:

[0118] The query matrix, key matrix, and value matrix in the self-attention mechanism are replaced by three 1×1 convolutions to obtain an improved self-attention mechanism.

[0119] The improved self-attention mechanism, stacked Bi-GRU layer and channel attention layer are cascaded sequentially to obtain the IBGC model.

[0120] It can be understood that the IBGC network model consists of stacked Bi-GRU layers, improved self-attention layers, channel attention layers, and fully connected layers. First, the degraded feature set is input into the improved self-attention layer, and different attention weights are assigned to different time steps of the input features to capture the contribution of time series at different time steps. Next, the Bi-GRU layers are stacked for feature learning to obtain deeper hidden features. The learned deep features are then input into the channel attention layer to learn the importance of different features and generate attention weights. The deep features and attention weights are then fused to generate deep features with importance. Finally, the deep features with importance are input into the fully connected layer network, and the deep features are mapped to RUL labels to obtain the final RUL prediction.

[0121] Among them, the Bi-GRU layer, namely the Bi-GRU model, is a recurrent neural network that consists of two independent GRU units, one that processes data forward in a time series, and the other that processes data in reverse in a time series. Through this bidirectional structure, the Bi-GRU model can simultaneously capture the forward and backward information of the sequence data, thereby better understanding and predicting patterns in the sequence. GRU is a gated recurrent neural network unit. In GRU, the forget gate and the input gate are combined into a unit called the update gate, which fuses the cell state and the hidden state. Therefore, its formula includes an update gate, a reset gate, and a new candidate state. The following is the calculation process of the Bi-GRU model:

[0122] Reset Gate R t It can be calculated as follows:

[0123] R t =sig(W r H t-1 +U r X t +b r )

[0124] Candidate hidden states It can be calculated as follows:

[0125]

[0126] Update Gate Z t It can be calculated as follows:

[0127] Z t =sig(W z H t-1 +U z X t +b r )

[0128] Update hidden state H t It can be calculated as follows:

[0129]

[0130] Among them, sig is the sigmoid activation function, tanh is the Tanh activation function, W r , U r , W, U are weight matrices, b r and b h is the bias term.

[0131] In order to obtain the contribution of each time step from the input features, this application adopts an improved self-attention mechanism (ISA). The self-attention mechanism randomly initializes three weight matrices w q , w k and wv To calculate the feature matrix, namely the query matrix (Q), key matrix (K) and value matrix (V). This calculation depends on the weight matrix, so that the self-attention mechanism cannot fully utilize the internal correlation of the input features. To solve this problem, in one possible implementation, the randomly initialized weight matrix in the self-attention mechanism is replaced by three 1×1 convolutions. Convolution increases the depth of ISA and introduces nonlinear features to enhance the expressiveness of ISA. At the same time, the residual connection structure is introduced to improve the stability of ISA. Input feature X t It is calculated by three 1×1 convolutions to construct three feature matrices Q, K, and V, calculated as follows:

[0132] Q=con 1×1 (X t )

[0133] K=con 1×1 (X t )

[0134] V=con 1×1 (X t )

[0135] Among them, con 1×1 is a 1×1 convolution, X t is the input feature.

[0136] Attention score (A score ) The time step of the feature is obtained through the Softmax function, which is calculated as follows:

[0137]

[0138] Among them, d k is the number of channels for Q and K.

[0139] Score the attention score Assign different time steps to V and output new features E, which are calculated as follows:

[0140] E=A score V

[0141] The output feature Y of ISA is calculated as follows:

[0142] Y=X t +W a E

[0143] Among them, W a is the weight parameter to be learned.

[0144] Given that the input degradation feature set is composed of multiple feature dimensions (channels), and the features of each channel contribute differently to RUL prediction, this application uses a channel attention mechanism to capture the contribution of channel features. In one possible implementation, a squeeze excitation mechanism (SE) is used as the channel attention mechanism. SE consists of two parts, namely squeeze (F sq (.)) and incentives (F ex (.)), calculated as follows:

[0145]

[0146] F ex (p) = sig(H w2 δ(H w1 p))

[0147] F scale (u c ,s c )=s c u c

[0148] Among them, u c =[u c (1),u c (2), …u c (i), …u c (W)], (c = 1, 2, ..., C), W is the length of the sliding window, c is the number of channels, and p is F sq (.), δ is the output of the ReLU function, sig is the Sigmoid function, and is the weight vector of the two hidden layers, r is the dimensionality reduction ratio of the dimensional transformation, and s is the attention weight, that is, F ex The output of (.), s c is the cth weight value in the attention weight s.

[0149] The SE mechanism can capture the key points of information, thereby improving the efficiency and accuracy of model processing. Therefore, the extracted degradation feature set is input into the SE module, and different weights are assigned to the degradation features to focus on the key features.

[0150] In the IBGC model training process, this application introduces L 2 Regularization term and dropout layer. At the same time, an early stopping mechanism is added, and the number of early stopping iterations is 30. After that, the learning rate is adjusted to the original 0.5 every 40 rounds. The hyperparameters of the IBGC model are shown in Table 2 below:

[0151] Table 2. Hyperparameters of IBGC model

[0152]

[0153] In some optional embodiments, referring to Figures 4 to 8 , Figure 4 A schematic diagram of the monotonicity of a bearing. Figure 5 A schematic diagram of the bearing trend. Figure 6 A schematic diagram of bearing robustness. Figure 7 A schematic diagram of the weighted evaluation index of the bearing. Figure 8 A schematic diagram of the prediction results of the method of the present invention.

[0154] This embodiment uses the full life data of the XJTU-SY bearing to further verify the feasibility of the proposed method. The experiment applies external force to the test bearing through a hydraulic loading system to accelerate the degradation of the test bearing. At the same time, vibration sensors are placed in the horizontal and vertical directions of the test bearing to collect the vibration signal of the bearing. When the maximum amplitude of the vibration signal exceeds 10 times the amplitude of the vibration signal in the normal operation stage, the experiment ends. The sampling frequency of the experimental setting is 25.6kHz, and 32768 points are recorded every 1min (ie 1.28s). The division of the training set and the test set, as well as the starting point are shown in Table 3:

[0155] Table 3 Bearing grouping results and starting points

[0156]

[0157] Among them, Bearing 1-2 (Bearing 1-1), Bearing 1-3 (Bearing 1-3) and Bearing 1-4 (Bearing1-3) are used as test bearings for characteristic parameter selection. Bearing 1-2 (Bearing 1-2) is used as the test bearing for RUL prediction.

[0158] Reference Figures 4 to 8 , it can be obtained that: the method of the present invention can be applied to the fault prediction and health management of rotating machinery such as bearings and gears, and realize the online and offline prediction of the remaining service life of the equipment. It provides a theoretical basis and technical support for the intelligent operation and maintenance decision-making of rotating machinery. The technical application of this invention can improve the safety, reliability and availability of enterprise mechanical equipment and reduce the maintenance cost of equipment.

[0159] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0160] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A bearing life prediction method based on feature extraction and IBGC network, characterized in that: include: A vibration sensor is used to collect a vibration signal group of the bearing; Using a fully adaptive noise ensemble empirical mode decomposition method to process any of the vibration signal groups to obtain a reconstructed signal; Extracting time domain features, frequency domain features and time-frequency domain degradation features from the reconstructed signal, and selecting A preferred degradation features from the time domain features, the frequency domain features and the time-frequency domain degradation features by using a monotonicity index, a trend index and a robustness index, where A is an integer and A≥2; Standardizing the data corresponding to the A preferred degradation features in the vibration signal group and setting RUL labels to obtain a degradation feature data set, and dividing a training set from the degradation feature data set; Constructing an IBGC model, and training the IBGC model using the training set; The test sample is input into the trained IBGC model, and the trained IBGC model outputs a prediction result.

2. The bearing life prediction method based on feature extraction and IBGC network according to claim 1 is characterized in that: The method of using a fully adaptive noise set empirical mode decomposition method to process any of the vibration signal groups to obtain a reconstructed signal includes: Adding Gaussian white noise to the vibration signal group to obtain the processed vibration signal group; Iterating according to the processed vibration signal group, including: in the jth round of iteration, obtaining the jth intrinsic mode component and the corresponding jth residual component, when k rounds of iteration are performed and the k residual components obtained are monotonic and the number of extreme points is less than 3, the iteration is stopped, and k intrinsic mode components are obtained, 1≤j≤k; Calculating the normalized kurtosis and normalized cross-correlation coefficient of the k eigenmode components respectively, arranging the k normalized kurtosis and normalized cross-correlation coefficients from large to small, and selecting the eigenmode components corresponding to the first five normalized kurtosis and normalized cross-correlation coefficients; Signal reconstruction is performed according to the selected five intrinsic mode components to obtain the reconstructed signal.

3. The bearing life prediction method based on feature extraction and IBGC network according to claim 2 is characterized in that: The adding of Gaussian white noise to the vibration signal group to obtain the processed vibration signal group is calculated in the following manner: x(n) ′ =x(n)+ε0ω i (n) Where x(n) ′ is the vibration signal group after processing, x(n) is the vibration signal group, ε0 is the amplitude of the Gaussian white noise, ω i (n) is the Gaussian white noise added for the ith time.

4. The bearing life prediction method based on feature extraction and IBGC network according to claim 2 is characterized in that: In the jth iteration, the jth eigenmode component and the corresponding jth residual component are obtained and calculated as follows: When j=1, the processed vibration signal group is subjected to EMD decomposition processing to obtain I components of the first round of iteration, and the average value of the I components of the first round of iteration is calculated as the first intrinsic mode component; the processed vibration signal group is subtracted from the first intrinsic mode component to obtain the corresponding first residual component; When 2≤j≤k, the j-th eigenmode component and the corresponding j-th residual component are calculated according to the j-1-th residual component.

5. The bearing life prediction method based on feature extraction and IBGC network according to claim 4 is characterized in that: When 2≤j≤k, the j-th eigenmode component and the corresponding j-th residual component are calculated according to the j-1-th residual component, and are calculated in the following manner: R j (n)=R j-1 (n)-IMF j (n) Among them, E j [] is the j-th EMD decomposition operation, I is the I component of the j-th iteration, IMF j (n) is the jth eigenmode component, R j-1 (n) is the j-1th residual component, R j (n) is the jth residual component, ε j is the amplitude of the Gaussian white noise added for the jth time.

6. The bearing life prediction method based on feature extraction and IBGC network according to claim 1 is characterized in that: The selecting A preferred degradation features from the time domain features, the frequency domain features and the time-frequency domain degradation features by using the monotonicity index, the trend index and the robustness index comprises: The time domain feature, the frequency domain feature and the time-frequency domain degradation feature are B features in total, where B is an integer and B≥3, and A≤B; Calculating the multi-angle comprehensive evaluation scores of the features respectively, including: calculating the monotonicity index, trend index and robustness index of the features; calculating the multi-angle comprehensive evaluation scores of the features according to the monotonicity index, trend index and robustness index of the features; Arrange the B multi-angle comprehensive evaluation scores from large to small, and select the features corresponding to the first A multi-angle comprehensive evaluation scores as the preferred degradation features.

7. The bearing life prediction method based on feature extraction and IBGC network according to claim 6 is characterized in that: The multi-angle comprehensive evaluation score of the features is calculated as follows: HI criteria(X) =0.3×Mon(X)+0.3×Tre(T,X)+0.3×Rob(X) Among them, HI criteria(X) is the multi-angle comprehensive evaluation score of the Xth feature, Mon(X) is the monotonicity index of the Xth feature, Tre(T,X) is the trend index of the Xth feature, and Rob(X) is the robustness index of the Xth feature.

8. The bearing life prediction method based on feature extraction and IBGC network according to claim 1 is characterized in that: The step of normalizing the data corresponding to the A preferred degradation features in the vibration signal group includes: Any of the above data is standardized and calculated as follows: Among them, X 1:k is the data after standardization, X i is the characteristic value of the data at time i, μ i:k is the mean of the characteristic values ​​of the data at all times, δ 1:k is the standard deviation of the characteristic values ​​of the data at all times.

9. The bearing life prediction method based on feature extraction and IBGC network according to claim 1 is characterized in that: The construction of the IBGC model includes: The query matrix, key matrix, and value matrix in the self-attention mechanism are replaced by three 1×1 convolutions to obtain an improved self-attention mechanism. The improved self-attention mechanism, the stacked Bi-GRU layer and the channel attention layer are sequentially cascaded to obtain the IBGC model.

Images