Rolling bearing life prediction method based on improved LSTMNN

Through the improved spider wasp optimization algorithm and core principal component analysis, the LSTMNN combined with the feature and time dual-head attention mechanism, the problem of low feature extraction difficulty and accuracy in bearing life prediction is solved, and more efficient feature reduction is achieved Peacekeeping accurate life expectancy prediction.

CN118332279BActive Publication Date: 2024-11-26DALIAN JIAOTONG UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410430596.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-10
Publication Date
2024-11-26
Estimated Expiration
2044-04-10

AI Technical Summary

Technical Problem

In the prediction of bearing life, the prior art has problems such as feature extraction difficulty, feature redundancy, low computing efficiency, insufficient nonlinear relationship capture and insufficient time information dependence acquisition, resulting in low prediction accuracy.

Method used

The improved Spider Wasp Optimization Algorithm (ISWO) optimizes the variational modal decomposition (VMD) parameters, combines the improved core principal component analysis (KPCA) for feature dimensionality reduction, and introduces long and short-term memory of the feature and time dual-ben-head attention mechanism Neural Network (LSTMNN) performs rolling bearing life prediction.

Benefits of technology

It reduces the difficulty of extracting bearing fault feature data sets, improves the prediction accuracy of the remaining service life of rolling bearings, effectively captures feature and time information, and improves the model's learning ability and prediction accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118332279B_ABST
    Figure CN118332279B_ABST
Patent Text Reader

Abstract

The present invention discloses a rolling bearing life prediction based on an improved LSTMNN, which belongs to the technical field of bearing life prediction, and includes the following steps: S1, feature extraction, using an improved spider wasp optimization algorithm ISWO to optimize the variational mode decomposition VMD parameters, and using ISWO-VMD to extract features to construct a feature data set; S2, feature dimension reduction, using an improved kernel principal component analysis KPCA to reduce the dimension of the features in the feature data set; S3, network prediction, the key features obtained after the dimension reduction are input into an improved long short-term memory neural network LSTMNN to predict the remaining service life of the rolling bearing; wherein the improved long short-term memory neural network LSTMNN adopts a feature and time dual multi-head attention mechanism. The present invention adopts the above-mentioned rolling bearing life prediction based on the improved LSTMNN, reduces the difficulty of extracting the bearing fault feature data set, and improves the prediction accuracy of the remaining service life of the rolling bearing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of bearing life prediction, and in particular to a rolling bearing life prediction method based on an improved LSTMNN. Background Art

[0002] For trains, safety is the top priority. As one of the important components of trains, the working condition of bearings is directly related to the safe operation of trains. Therefore, studying the prediction method of the remaining service life of bearings, thereby improving the accuracy of bearing life prediction and avoiding accidents, is of great significance to the safe operation of trains. At present, data-driven methods have become one of the important means of bearing life prediction. Among them, methods based on vibration signals have occupied a considerable position in the field of bearing life prediction. In order to supplement the time domain and frequency domain characteristics of vibration signals, it is necessary to extract the time and frequency domain features.

[0003] The variational mode decomposition method (VMD) with strong adaptability and better modal decomposition ability has been used to extract the time-frequency domain characteristics of bearing vibration signals. However, the manual selection of parameters may lead to over-decomposition or under-decomposition of VMD. Therefore, some scholars have used swarm intelligence algorithms to optimize VMD, but their optimization algorithms are prone to local optimality and slow convergence to varying degrees, which in turn causes the parameters of VMD to fall into local optimality. Compared with swarm intelligence optimization algorithms such as WOA and SSA, SWO has a faster convergence speed, so it can better obtain the decomposition parameters of VMD. Therefore, SWO is used to optimize VMD.

[0004] In addition, too many dimensions of feature extraction will inevitably cause redundancy of feature information, which may eventually result in low accuracy of bearing remaining life prediction, while the traditional kernel principal component analysis (KPCA) has large computational complexity when processing large samples. To solve this problem, Ladhiri et al. proposed a method based on a new simplified rank KPCA in 2018, which improved the dimensionality reduction efficiency of features. In 2020, Sun Hao et al. proposed a KPCA method based on adaptive updates to enhance the computational efficiency of data integration. Although the above studies have improved computational efficiency, they have ignored the nonlinear relationship of the data. In order to better capture this nonlinear relationship, it is very important to select an appropriate kernel function, so a nonlinear component is added to the kernel function.

[0005] With the development of deep learning and neural networks, Long Short-Term Memory Neural Network (LSTMNN) has been widely used in prediction. In response to the problem that it is difficult to determine the hyperparameters of neural networks, many scholars have proposed solutions. However, these methods do not fully explore important feature information, and further research should be conducted on the correlation of feature information.

[0006] In order to solve the problem that the machine's posture is difficult to control during the crossing process, Wang Lai et al. used bidirectional long short-term memory (Bi-LSTM) to predict the posture in 2023. In order to overcome the problem of insufficient accuracy of small sample training, Zhou Ciming et al. proposed an LSTMNN with transfer learning and ensemble learning for life prediction in 2023. However, the above method does not fully obtain the dependence on time information, and the time information needs to be further captured. In summary, the present invention proposes an LSTMNN with a dual multi-head attention mechanism of features and time to predict the remaining service life of rolling bearings. Summary of the invention

[0007] The purpose of the present invention is to provide a rolling bearing life prediction method based on an improved LSTMNN, which reduces the difficulty of extracting bearing fault feature data sets and improves the prediction accuracy of the remaining service life of the rolling bearing.

[0008] To achieve the above object, the present invention provides a rolling bearing life prediction method based on an improved LSTMNN, comprising the following steps:

[0009] S1, feature extraction, using the improved spider wasp optimization algorithm ISWO to optimize the variational mode decomposition VMD parameters, and using ISWO-VMD for feature extraction to construct a feature data set;

[0010] S2. Feature dimensionality reduction: Improved kernel principal component analysis KPCA is used to reduce the dimensionality of features in the feature data set. Improved kernel principal component analysis KPCA improves the kernel function of KPCA. After adding nonlinear components to the kernel function, we have:

[0011]

[0012] Where K(·) represents kernel function mapping; a+b=1, a, b represent component coefficients; x i , x j Represents the feature vector of the input sample; hyperparameter γ∈[0,1]; random number r∈[0,1];

[0013] S3, network prediction, the key features obtained after dimensionality reduction are input into the improved long short-term memory neural network LSTMNN to predict the remaining service life of the rolling bearing;

[0014] Among them, the improved long short-term memory neural network LSTMNN adopts feature and time dual multi-head attention mechanism.

[0015] Preferably, in step S1, feature extraction includes extracting vibration signal features in the time domain, frequency domain, and time-frequency domain;

[0016] The extracted time domain features include mean, root mean square, variance, standard deviation, kurtosis, skewness, maximum and minimum values;

[0017] Among the extracted frequency domain features, for a non-periodic signal x(t), Fourier transform is used to extract multiple frequency domain features from the transformed signal, including mean frequency, root mean square frequency, variance frequency, standard deviation frequency, kurtosis frequency, skewness frequency, maximum frequency and minimum frequency.

[0018] Preferably, an improved SWO is used to extract time-frequency domain features, and the improved method is as follows:

[0019] S101. Initialize the population using the cubic mapping method. The process is expressed as:

[0020] SW pop (n+1)=4SW pop (n) 3 -3SW pop (n) (1)

[0021] Among them, SW pop (n) represents the population at the previous moment; SW pop (n+1) represents the population at the next moment;

[0022] S102, using the population guidance mechanism to solve the search space, and enhance the exploration ability of the population individuals to the surrounding space. The process is expressed as:

[0023]

[0024] in, represents the solution of the population guidance mechanism; SW i t =(SW1 t , SW2 t ,…,SW n t ) represents an individual in n-dimensional space at time t; l i (f) u i (f) represents the dynamic upper and lower bounds of the entire population at f iterations, l i(f) = min(SW i t ),u i (f) = max(SW i t ); Random number rand()∈[0.1], representing the expansion and contraction control amount of the dynamic boundary;

[0025] S103, in order to further improve the convergence speed and accuracy of the algorithm, the inertia weight of the sine-cosine transform is introduced in the SWO position update, as shown in formula (3):

[0026]

[0027] Among them, the random number z∈[0,1]; ω represents the inertia weight coefficient; f is the current iteration number; f max Indicates the maximum number of iterations;

[0028] The introduced sine-cosine transform inertia weight is dynamically adjusted with the number of iterations, which helps SWO to jump out of the local optimum, thereby improving the convergence efficiency of the algorithm. The position update process after the introduction of ω is expressed as:

[0029]

[0030] in, Represents the position parameter at time t+1; are two random position parameters; μ is the search coefficient;

[0031] S104 and SWO searches are nonlinear optimization strategies. A nonlinear convergence factor p is introduced into the search coefficient, and its expression is:

[0032]

[0033] The search coefficient μ after the introduction of p is expressed as:

[0034] μ=p·|rn|·r1 (6)

[0035] Among them, the random number r1∈[0,1]; the random number |rn| obeys the normal distribution.

[0036] Preferably, the specific steps of feature extraction and feature dimensionality reduction are:

[0037] S201, extracting time domain, frequency domain, and time-frequency domain feature parameters, wherein the time-frequency domain parameters are extracted using the ISWO-VMD method;

[0038] S202, feature screening, screening the contribution rate of the extracted time domain and frequency domain features and the time and frequency domain output of ISWO-VMD, so as to screen out the principal component variables;

[0039] S203, feature dimensionality reduction, using the principal component variables selected by the contribution rate to perform KPCA feature fusion to achieve the purpose of feature dimensionality reduction, and judging the similarity of the reduced dimensionality features by cosine similarity, which is expressed as:

[0040]

[0041] Among them, x and y represent two eigenvectors; |x|, |y| represent the modulus of the corresponding eigenvectors.

[0042] Preferably, the extraction method in step S201 is specifically:

[0043] S2011, parameter initialization, initializing ISWO population parameters, location parameters, number of iterations, and optimal fitness value;

[0044] S2012, parameter update, using the population guidance mechanism and sine-cosine transform inertia weights to update the position parameters of the spider colony;

[0045] S2013, iterative optimization. If the maximum number of iterations is not reached, jump to step S2012, update the state of the spider colony and perform global optimization. If the maximum number of iterations is reached, stop the optimization and output the K value and α value of the optimized VMD. The K value represents the number of mode decompositions; the α value represents the penalty factor.

[0046] Preferably, in step S3, the feature multi-head attention mechanism FMA is used to extract input features of LSTMNN. At time t, for an input x with M features t =[x 1,t , x 2,t , ..., x M,t ], and its attention weight calculation formula is:

[0047] e t =σ(W c x t +b c ) (9)

[0048] Among them, e t =[e 1,t , e 2,t , ..., e M,t ] represents the combination of attention weight coefficients corresponding to each input feature at the current time t; W c represents the trainable weight matrix; b c is the bias vector; σ() represents the sigmoid activation function; the obtained feature attention weight coefficient e t After normalization, we have:

[0049]

[0050] Among them, α M,t represents the feature of the Mth attention weight; exp represents the exponential function expression; e M,t represents the attention weight coefficient of the Mth feature at time t; e i,t Represents the attention weight coefficient of the i-th feature at time t. After the above calculation, the feature attention weight α is obtained t =[α 1,t , α 2,t , ..., α M,t ], and transform equation (10) with x t Multiply them together to get the weighted calculation and then calculate the weighted vector

[0051]

[0052] The expression for weighted input long is:

[0053] Multihead feature =Concat(head1, head2,..., head h ) (12)

[0054] in, Concat is to concatenate multiple weighted vectors;

[0055] The expression of the feature multi-head attention mechanism is:

[0056]

[0057] in,

[0058] Preferably, in step S3, the temporal multi-head attention mechanism TMA assigns attention weights to the time information carried by each historical moment of the input sequence to distinguish the impact of the time information on the predicted output at the current moment;

[0059] The input of the temporal multi-head attention is the LSTMNN hidden layer state h T =[h 1,T ,h 2,T , ..., h C,T ] time series, where T is the time of model iteration and C is the time window length of the input sequence; the time attention weight vector corresponding to each historical moment in the current moment is:

[0060] l t =ReLU(W d h T +b d ) (14)

[0061] Among them, W d represents a trainable weight matrix; b d Represents the bias vector; ReLU() represents the activation function;

[0062] Normalize the attention weight coefficient each time to get the time attention weight β t =[β 1,t , β 2,t , β τ,t …, β k,t ], where β τ,t is the attention weight value of τ times, weighted by the hidden layer state of each corresponding historical time, and the comprehensive time series information state h′ is obtained T ; Its process expression is:

[0063]

[0064] Among them, l τ,t represents the time attention weight coefficient of the τth feature at time t; l i,t represents the time attention weight coefficient of the i-th feature at time t; β i , t represents the normalized attention weight coefficient of the ith feature at time t; h i,t represents the hidden layer input of the i-th feature at time t; N represents the number of time points;

[0065] The expression of the temporal multi-head attention mechanism is:

[0066] Multihead time =Concat(head1,head2,…,head h ) (17)

[0067] in

[0068] Preferably, the steps of network prediction are:

[0069] S301, weight initialization, initializing the parameters of the input layer attention weight and the hidden layer attention weight;

[0070] S302, calculating feature attention weights, using formula (9) to calculate the feature attention weights corresponding to each feature;

[0071] S303, constructing weighted features, using formula (11) to obtain weighted features, and using formula (13) to perform multiple feature concatenation;

[0072] S304, calculate the time attention weight, input the obtained features into LSTMNN, and use formula (14) to calculate the time attention weight corresponding to each historical moment;

[0073] S305, constructing a time-weighted feature, using formula (16) to obtain a weighted feature, and using formula (17) to concatenate multiple time-weighted features;

[0074] S306, update the hidden layer output state. If the maximum number of iterations has not been reached, jump to step S305. If the maximum number of iterations has been reached, stop updating the time-weighted features and output the predicted value of the remaining life of the rolling bearing.

[0075] Therefore, the present invention adopts the above-mentioned rolling bearing life prediction method based on improved LSTMNN, optimizes VMD parameters using improved spider wasp optimization algorithm, extracts time-frequency domain features to construct fault feature data set, reduces the difficulty of extracting bearing fault feature data set, and retains the sensitivity and stability of bearing fault characterization.

[0076] The improved KPCA is used to perform kernel principal component analysis to reduce the feature similarity after dimensionality reduction, thereby effectively achieving the dimensionality reduction of feature data.

[0077] The proposed improved neural network model effectively improves the model's learning of feature information and time information by introducing a dual multi-head attention mechanism for features and time. In ablation experiments and generalization experiments, the model has good accuracy in predicting the remaining service life of rolling bearings.

[0078] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a flow chart of a rolling bearing life prediction method in a rolling bearing life prediction embodiment based on improved LSTMNN of the present invention;

[0080] Figure 2 It is an IDSA-LSTMNN structure diagram of an embodiment of a rolling bearing life prediction method based on an improved LSTMNN of the present invention;

[0081] Figure 3 It is the iterative result of the fitness function of the embodiment of the rolling bearing life prediction method based on the improved LSTMNN of the present invention;

[0082] Figure 4 It is a decomposition result diagram of ISWO-VMD of an embodiment of a rolling bearing life prediction method based on an improved LSTMNN of the present invention;

[0083] Figure 5 It is a WOA-VMD, SWO-VMD and ISWO-VMD reconstruction result diagram of the rolling bearing life prediction method embodiment based on the improved LSTMNN of the present invention;

[0084] Figure 6 It is the characteristic data of the time domain, frequency domain, and time-frequency domain of the embodiment of the rolling bearing life prediction method based on the improved LSTMNN of the present invention;

[0085] Figure 7 It is the characteristic contribution rate result of the embodiment of the rolling bearing life prediction method based on the improved LSTMNN of the present invention;

[0086] Figure 8 It is the feature dimension reduction result of the embodiment of the rolling bearing life prediction method based on the improved LSTMNN of the present invention; Figure 8 (a) is the feature dimension reduction result of classic KPCA; Figure 8 (b) is the feature dimension reduction result of KPCA based on the improved function;

[0087] Fig. 9 It is the prediction result of multiple networks under bearing2_3 of the embodiment of the rolling bearing life prediction method based on improved LSTMNN of the present invention;

[0088] Fig.10 It is the prediction result of multiple networks in the embodiment of the rolling bearing life prediction method based on the improved LSTMNN of the present invention; Fig.10 (a) in the figure is the prediction result of multiple networks under bearing1_3; Fig.10 (b) in the figure is the prediction result of multiple networks under bearing3_3. DETAILED DESCRIPTION

[0089] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0090] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0091] Embodiment 1

[0092] The present invention provides a method for predicting the remaining useful life (RUL) of a rolling bearing based on an improved long short-term memory neural network (IDSA-LSTMNN), and its flow chart is as follows: Figure 1 As shown, the following steps are included:

[0093] S1. Feature extraction. The improved spider wasp optimization algorithm (ISWO) is used to optimize the variational mode decomposition (VMD) parameters. The feature extraction is performed using ISWO-VMD to construct a feature data set. Feature extraction is very important for the prediction of the remaining service life of bearings. The vibration signal features in the time domain, frequency domain, and time-frequency domain are extracted to further screen out the characteristic variables reflecting effective degradation.

[0094] Feature extraction includes extracting vibration signal features in the time domain, frequency domain, and time-frequency domain;

[0095] The extracted time domain features include mean, root mean square, variance, standard deviation, kurtosis, skewness, maximum value and minimum value, corresponding to feature 1 to feature 8 respectively.

[0096] Among the extracted frequency domain features, for a non-periodic signal x(t), Fourier transform is used to extract multiple frequency domain features from the transformed signal, including mean frequency, root mean square frequency, variance frequency, standard deviation frequency, kurtosis frequency, skewness frequency, maximum frequency and minimum frequency, corresponding to features 9 to 16 respectively.

[0097] The improved SWO is used to extract time-frequency domain features. The improvement method is as follows:

[0098] S101. Initialize the population using the cubic mapping method. The process is expressed as:

[0099] SW pop (n+1)=4SW pop (n) 3 -3SW pop (n) (1)

[0100] Among them, SW pop (n) represents the population at the previous moment; SW pop (n+1) represents the population at the next moment;

[0101] S102, using the population guidance mechanism to solve the search space, and enhance the exploration ability of the population individuals to the surrounding space. The process is expressed as:

[0102]

[0103] in, represents the solution of the population guidance mechanism; SW i t =(SW1 t , SW2 t ,…,SW n t ) represents an individual in n-dimensional space at time t; l i (f) u i (f) represents the dynamic upper and lower bounds of the entire population at f iterations, l i (f) = min(SW i t ),u i (f) = max(SW i t); Random number rand()∈[0.1], representing the expansion and contraction control amount of the dynamic boundary;

[0104] S103, in order to further improve the convergence speed and accuracy of the algorithm, the inertia weight of the sine-cosine transform is introduced in the SWO position update, as shown in formula (3):

[0105]

[0106] Among them, the random number z∈[0,1]; ω represents the inertia weight coefficient; f is the current iteration number; f max Indicates the maximum number of iterations;

[0107] The introduced sine-cosine transform inertia weight is dynamically adjusted with the number of iterations, which helps SWO to jump out of the local optimum, thereby improving the convergence efficiency of the algorithm. The position update process after the introduction of ω is expressed as:

[0108]

[0109] in, Represents the position parameter at time t+1; are two random position parameters; μ is the search coefficient;

[0110] S104 and SWO are nonlinear optimization strategies during search, and cannot adjust the global search capability and local development capability well. In order to prevent the search from falling into the local optimum, a nonlinear convergence factor p is introduced into the search coefficient, and its expression is:

[0111]

[0112] The search coefficient μ after the introduction of p is expressed as:

[0113] μ=p·|rn|·r1 (6)

[0114] Among them, the random number r1∈[0,1]; the random number |rn| obeys the normal distribution;

[0115] The time domain features extracted using this method include: IMF1 energy entropy and IMF2 energy entropy, corresponding to feature 17 and feature 18 respectively.

[0116] S2. Feature dimensionality reduction: Improved kernel principal component analysis (KPCA) is used to reduce the dimensionality of features in the feature data set. KPCA maps the input data to a high-dimensional feature space through a kernel function and performs nonlinear processing in the feature space. However, there is a strong nonlinear relationship between the feature data of the bearing, and traditional kernel functions may find it difficult to capture this nonlinear relationship. Improved KPCA improves the kernel function of KPCA. After adding nonlinear components to the kernel function, we have:

[0117]

[0118] Where K(·) represents kernel function mapping; a+b=1, a, b represent component coefficients; x i , x j Represents the feature vector of the input sample; hyperparameter γ∈[0,1]; random number r∈[0,1].

[0119] To sum up, the specific steps of feature extraction and feature dimensionality reduction are:

[0120] S201, extracting time domain, frequency domain, and time-frequency domain characteristic parameters, wherein the time-frequency domain parameters are extracted using the ISWO-VMD method; the specific extraction method is:

[0121] S2011, parameter initialization, initializing ISWO population parameters, location parameters, number of iterations, and optimal fitness value;

[0122] S2012, parameter update, using the population guidance mechanism and sine-cosine transform inertia weights to update the position parameters of the spider colony;

[0123] S2013, iterative optimization. If the maximum number of iterations is not reached, jump to step S2012, update the state of the spider colony and perform global optimization. If the maximum number of iterations is reached, stop the optimization and output the K value and α value of the optimized VMD. The K value represents the number of mode decompositions; the α value represents the penalty factor.

[0124] S202, feature screening, screening the contribution rate of the extracted time domain and frequency domain features and the time and frequency domain output of ISWO-VMD, so as to screen out the principal component variables;

[0125] S203, feature dimensionality reduction, using the principal component variables selected by the contribution rate to perform KPCA feature fusion to achieve the purpose of feature dimensionality reduction, and judging the similarity of the reduced dimensionality features by cosine similarity, which is expressed as:

[0126]

[0127] Among them, x and y represent two eigenvectors; |x|, |y| represent the modulus of the corresponding eigenvectors.

[0128] S3. The key features after dimensionality reduction are input into the improved long short-term memory neural network LSTMNN to predict the remaining service life of rolling bearings. The traditional LSTMNN has relatively weak ability to capture the correlation of feature information and the dependence of time information, which may lead to low network accuracy. In view of the above shortcomings of LSMTNN, IDSA-LSTMNN is proposed based on the dual multi-head attention mechanism of Feature Multi-head Attention Mechanism (FMA) and Time Multi-head Attention Mechanism (TMA), so that the network can better capture the characteristics and time information of the data, and finally improve the network prediction accuracy.

[0129] In step S3, the feature multi-head attention mechanism FMA is used to extract input features of LSTMNN. Taking time t as an example, for an input x with M features t =[x 1,t , x 2,t , ..., x M,t ], and its attention weight calculation formula is:

[0130] e t =σ(W c x t +b c ) (9)

[0131] Among them, e t =[e 1,t , e 2,t , ..., e M,t ] represents the combination of attention weight coefficients corresponding to each input feature at the current time t; W c represents the trainable weight matrix; b c is the bias vector; σ() represents the sigmoid activation function; the obtained feature attention weight coefficient e t After normalization, we have:

[0132]

[0133] Among them, α M,t represents the feature of the Mth attention weight; exp represents the exponential function expression; e M,t represents the attention weight coefficient of the Mth feature at time t; e i,t Represents the attention weight coefficient of the i-th feature at time t. After the above calculation, the feature attention weight α is obtained t =[α 1,t , α 2,t , ..., α M,t ], and transform equation (10) with xt Multiply them together to get the weighted calculation and then calculate the weighted vector

[0134]

[0135] Different from the previous attention mechanism (Dual Attention Mechanism Based LSTM Neural Network, DSA-LSTMNN), the feature multi-head attention mechanism not only focuses on the original feature information, but also on the weighted feature information, dynamically assigns attention weights to the input features of the model, and mines the correlation between the input features and the output, thereby further improving the LSTMNN prediction accuracy. The expression of weighted input multi-head is:

[0136] Multihead feature =Concat(head1, head2,..., head h ) (12)

[0137] in, Concat is to concatenate multiple weighted vectors;

[0138] The expression of the feature multi-head attention mechanism is:

[0139]

[0140] in,

[0141] Similarly, the temporal multi-head attention mechanism (TMA) assigns attention weights to the time information carried by each historical moment of the input sequence to distinguish the impact of time information on the predicted output at the current moment; at the same time, it independently extracts the time series information between data at different historical moments, enhancing the information expression of key moments.

[0142] The input of the temporal multi-head attention is the LSTMNN hidden layer state h T =[h 1,T ,h 2,T , ..., h C,T ] time series, where T is the time of model iteration and C is the time window length of the input sequence; the time attention weight vector corresponding to each historical moment in the current moment is:

[0143] l t =ReLU(W d h T +b d ) (14)

[0144] Among them, W drepresents a trainable weight matrix; b d Represents the bias vector; ReLU() represents the activation function;

[0145] Normalize the attention weight coefficient each time to get the time attention weight β t =[β 1,t , β 2,t , β τ,t ..., β k,t |, where β τ,t is the attention weight value of τ times, weighted by the hidden layer state of each corresponding historical time, and the comprehensive time series information state h′ is obtained T ; Its process expression is:

[0146]

[0147] Among them, l τ,t represents the time attention weight coefficient of the Tth feature at time t; l i,t represents the time attention weight coefficient of the i-th feature at time t; β i,t represents the normalized attention weight coefficient of the ith feature at time t; h i,t represents the hidden layer input of the i-th feature at time t; N represents the number of time points;

[0148] The expression of the temporal multi-head attention mechanism is:

[0149] Multihead time =Concat(head1, head2,..., head h ) (17)

[0150] in

[0151] Figure 2 The figure below shows the structure of the improved long short-term memory neural network (IDSA-LSTMNN). Combined with the IDSA-LSTMNN structure, the steps of network prediction are as follows:

[0152] S301, weight initialization, initializing the parameters of the input layer attention weight and the hidden layer attention weight;

[0153] S302, calculating feature attention weights, using formula (9) to calculate the feature attention weights corresponding to each feature;

[0154] S303, constructing weighted features, using formula (11) to obtain weighted features, and using formula (13) to perform multiple feature concatenation;

[0155] S304, calculate the time attention weight, input the obtained features into LSTMNN, and use formula (14) to calculate the time attention weight corresponding to each historical moment;

[0156] S305, constructing a time-weighted feature, using formula (16) to obtain a weighted feature, and using formula (17) to concatenate multiple time-weighted features;

[0157] S306, update the hidden layer output state. If the maximum number of iterations has not been reached, jump to step S305. If the maximum number of iterations has been reached, stop updating the time-weighted features and output the predicted value of the remaining life of the rolling bearing.

[0158] In order to evaluate the performance of the proposed IDSA-LSTMNN model in rolling bearing life prediction, the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (RCoefficient of Determination, R 2 ) is used as the evaluation index. The specific calculation formula is as follows:

[0159]

[0160] Among them, y i is the true value, is the predicted value, is the average value.

[0161] In the analysis of experimental results, the open source data of rolling bearings in IEEEPHM2012 was used to select bearing 1 (bearing1_3) under working condition 1 (4000N and 1800r / min), bearing 2 (bearing2_3) under working condition 2 (4200N and 1650r / min), and bearing 3 (bearing3_3) under working condition 3 (5000N and 1500r / min) for the experiment. During the experiment, only the horizontal vibration signal of the bearing was considered. The vibration signal was placed vertically on the outer ring of the bearing by the vibration sensor until the outer ring of the bearing failed. The sampling frequency of the experiment was 25.6kHz, and the sampling was performed every 10 seconds, with 2560 points sampled at a time, and a total of 3737600 sampling points.

[0162] Time-frequency domain feature extraction and analysis

[0163] In the process of VMD decomposition, there are two parameters that are extremely important for the decomposition of the signal. One is the number of modal decompositions K, and the other is the penalty factor α. To determine the K and α values, the sample entropy is used as the fitness function. The ranges of K and α are given as: K∈[3,10], α∈[100,2500], and ISWO is used to optimize the parameters. The iterative results of the fitness function are as follows: Figure 3 shown.

[0164] from Figure 3 It can be seen that ISWO-VMD has a lower fitness function score than WOA-VMD and SWO-VMD, and uses fewer iterations to find the optimal parameter value. The final optimized parameter K value and α value are K = 4 and α = 2490 respectively. The optimized K value and α value are input into VMD for time-frequency domain feature extraction. The decomposition results of ISWO-VMD are shown in Figure 4 The reconstructed signals ISWO-VMD, SWO-VMD and WOA-VMD are shown in Figure 5 shown.

[0165] analyze Figure 5 It can be seen that in the range of 0-1350s, the reconstructed signal amplitudes obtained by the three algorithms WOA-VMD, SWO-VMD, and ISWO-VMD all fit the original signal well. However, in the range of 1350s-1450s, the reconstructed signal amplitudes obtained by SWO-VMD and WOA-SWO showed a certain deviation; compared with the above two algorithms, the reconstructed signal amplitude obtained by the ISWO-VMD algorithm has a higher fit with the original signal. The above results show that after the population initialization strategy is improved and the position parameters are updated, the proposed ISWO can better optimize the K value and α value of VMD, thereby better extracting the time-frequency domain characteristics of the signal. Table 1 further shows the effectiveness of ISWO-VMD.

[0166] Table 1 Evaluation indicators of reconstructed signals

[0167] Model <![CDATA[f rmse ]]> <![CDATA[f mae ]]> <![CDATA[R 2 ]]> WOA-VMD <![CDATA[4.88×10 -5 ]]> <![CDATA[9.27×10 -5 ]]> 0.8433 SWO-VMD <![CDATA[2.50×10 -5 ]]> <![CDATA[6.53×10 -5 ]]> 0.9534 ISWO-VMD <![CDATA[1.26×10 -5 ]]> <![CDATA[3.29×10 -5 ]]> 0.9867

[0168] After calculating Table 1, it can be seen that compared with SWO-VMD and WOA-VMD, the signal reconstruction f of ISWO-VMD rmse They decreased by 49.68% and 74.26% respectively. mae They decreased by 49.61%, 64.51% and R 2 They increased by 3.49% and 17% respectively. Figure 5 and Table 1 comprehensively illustrate the effectiveness of the proposed ISWO-VMD in the field of time-frequency domain feature extraction.

[0169] Results and analysis of feature dimensionality reduction

[0170] By using the time-frequency domain features extracted by ISWO-VMD and combining them with the time-domain and frequency-domain features extracted by the time-domain and frequency-domain extraction method in step S1, 18 sets of feature data that can characterize bearing degradation are obtained. The feature data extraction results are shown in Figure 2. Figure 6 shown.

[0171] analyze Figure 6 It can be seen that the extracted feature data all reflect the degradation process of rolling bearings. The contribution rates of the above features are calculated using KPCA based on the improved kernel function. The calculation results are as follows: Figure 7 shown.

[0172] from Figure 7 It can be seen that the sum of the contribution rates of feature 1, feature 2, and feature 3 exceeds 90%. Therefore, the above three features are used as principal component variables, KPCA feature dimensionality reduction is performed, and the cosine similarity of formula (8) is used to calculate the similarity. The result of KPCA feature dimensionality reduction is as follows: Figure 8 shown.

[0173] J th = 0.5 as the similarity threshold for weak and strong distinction (J < 0.5 for strong distinction), we get Figure 8 The result of feature dimensionality reduction is shown in the figure. The cosine similarity calculation shows that the similarities between features 1, 2, features 1, 3, and features 2, 3 obtained after dimensionality reduction by traditional KPCA are 0.278, 0.095, and 0.982, respectively, that is, features 2 and 3 are weakly distinguished, indicating that traditional KPCA fails to effectively decouple features 2 and 3; using KPCA based on the improved kernel function, the similarities between features 1, 2, features 1, 3, and features 2, 3 obtained after dimensionality reduction are 0.341, 0.072, and 0.193, respectively, that is, every two features are strongly distinguished, indicating that the improved kernel function improves the decoupling ability of KPCA for features.

[0174] Results analysis of RUL prediction

[0175] Taking the data of bearing 2 (bearing2_3) under working condition 2 as an example, the prediction results of the remaining service life of the bearings are compared by using IDSA-LSTMNN and multi-scale convolutional neural network (MCNN), DSA-LSTMNN and LSTMNN. Fig. 9 shown.

[0176] analyze Fig. 9It can be seen that IDSA-LSTMNN fits the real value better. The reason is that the proposed feature and time dual multi-head attention mechanism not only focuses on the correlation of feature information weights, but also pays attention to the dependency of time information. Compared with the three algorithms of DSA-LSTMNN, MCNN, and LSTMNN, the prediction results of the IDSA-LSTMNN algorithm are more consistent with the actual value. The regression evaluation indicators are shown in Table 2.

[0177] Table 2 Regression evaluation indicators

[0178] index <![CDATA[f rmse ]]> <![CDATA[f mae ]]> <![CDATA[R 2 ]]> DSA-LSTMNN 2.4648 0.7053 0.9441 LSTMNN 2.8670 0.9381 0.9224 MCNN 2.0783 0.6078 0.9606 IDSA-LSTMNN 0.8543 0.2707 0.9890

[0179] After calculating Table 2, it can be seen that compared with LSTMNN, DSA-LSTMNN and MCNN, the prediction result f of IDSA-LSTMNN is rmse They were reduced by 65.34%, 70.2%, and 58.89%, respectively. mae They decreased by 61.62%, 71.14%, 55.46% and R 2 They increased by 4.76%, 7.22% and 2.96% respectively. In summary, Fig. 9 Table 2 and Table 3 comprehensively illustrate the effectiveness of the proposed IDSA-LSTMNN in the field of bearing remaining life prediction. In order to further demonstrate the effectiveness of the proposed method, the following four groups of models are obtained by controlling different variables. The specific settings are shown in Table 3.

[0180] Table 3 Combination settings of different variables

[0181] Model Feature multi-head attention mechanism Temporal Multi-Head Attention Mechanism IDSA-LSTMNN yes yes FMA-LSTMNN yes no TMA-LSTMNN no yes

[0182] The IDSA-LSTMNN model is the LSTMNN of the proposed dual multi-head attention mechanism. The IDSA-LSTMNN and FMA-LSTMNN models are used to verify the advantages of the temporal multi-head attention mechanism; the IDSA-LSTMNN and TMA-LSTMNN models are used to verify the advantages of the feature multi-head attention mechanism. Taking the bearing 2 (bearing2_3) data under working condition 2 as an example, the evaluation indicators of the ablation experiment are shown in Table 4.

[0183] Table 4 Regression evaluation indicators of ablation experiments

[0184] index <![CDATA[f rmse ]]> <![CDATA[f mae ]]> <![CDATA[R 2 ]]> IDSA-LSTMNN 0.9243 0.2865 0.9805 FMA-LSTMNN 2.3583 0.5624 0.9523 TMA-LSTMNN 2.0638 0.4363 0.9642

[0185] After calculating Table 4, it can be seen that compared with FMA-LSTMNN and TMA-LSTMNN, IDSA-LSTMNN has a higher f rmse They were reduced by 60.81%, 55.21%, and f mae They decreased by 49.06%, 34.33%, and R 2They increased by 2.96% and 1.69% respectively.

[0186] In summary, the proposed IDSA-LSTMNN has a higher accuracy in predicting the remaining service life of rolling bearings in terms of mining time information dependency and feature information correlation. In order to further demonstrate the generalization of the proposed method, the data of bearing 1 (bearing1_3) under working condition 1 and bearing 3 (bearing3_3) under working condition 3 were selected for generalization experiments. The prediction results of the remaining service life of the bearing are shown in Figure 2. Fig.10 The regression evaluation indicators are shown in Table 5.

[0187] Table 5 Evaluation index results of multiple networks

[0188]

[0189]

[0190] After calculating the bearing1_3 index in Table 5, it can be found that compared with LSTMNN, MCNN and DSA-LSTMNN, the prediction result f of IDSA-LSTMNN is rmse They were reduced by 70.57%, 61.21%, and 67.52%, respectively. mae They decreased by 77.58%, 52.55%, 63.78% and R 2 The results are 8.36%, 1.96% and 4.22% higher respectively. After calculating the bearing3_3 index in Table 5, it can be found that compared with LSTMNN, MCNN and DSA-LSTMNN, the prediction result f of IDSA-LSTMNN is rmse They were reduced by 69.78%, 63.3%, and 78.84%, respectively. mae They decreased by 69.67%, 39.84%, 55.3% and R 2 They increased by 6.94%, 1.73% and 2.95% respectively. Fig.10 and Table 5 comprehensively illustrate the effectiveness of the proposed IDSA-LSTMNN in the generalization experiment field.

[0191] Therefore, the present invention adopts the above-mentioned rolling bearing life prediction method based on improved LSTMNN, which reduces the difficulty of extracting bearing fault feature data sets and improves the prediction accuracy of the remaining service life of the rolling bearing.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. The rolling bearing life prediction method based on improved LSTMNN is characterized by: The following steps are involved: S1, feature extraction, using the improved spider wasp optimization algorithm ISWO to optimize the variational mode decomposition VMD parameters, and using ISWO-VMD for feature extraction to construct a feature data set; S2. Feature dimensionality reduction: Improved kernel principal component analysis KPCA is used to reduce the dimensionality of features in the feature data set. Improved kernel principal component analysis KPCA improves the kernel function of KPCA. After adding nonlinear components to the kernel function, we have: Where K(·) represents the kernel function mapping; a+b=1, a, b represent the component coefficients; x i , x j Represents the feature vector of the input sample; hyperparameter γ∈[0,1]; random number r∈[0,1]; S3, network prediction, the key features obtained after dimensionality reduction are input into the improved long short-term memory neural network LSTMNN to predict the remaining service life of the rolling bearing; Among them, the improved long short-term memory neural network LSTMNN adopts feature and time dual multi-head attention mechanism.

2. The rolling bearing life prediction method based on improved LSTMNN according to claim 1 is characterized in that: In step S1, feature extraction includes extracting vibration signal features in the time domain, frequency domain, and time-frequency domain; The extracted time domain features include mean, root mean square, variance, standard deviation, kurtosis, skewness, maximum and minimum values; Among the extracted frequency domain features, for a non-periodic signal x(t), Fourier transform is used to extract multiple frequency domain features from the transformed signal, including mean frequency, root mean square frequency, variance frequency, standard deviation frequency, kurtosis frequency, skewness frequency, maximum frequency and minimum frequency.

3. The rolling bearing life prediction method based on improved LSTMNN according to claim 2 is characterized in that: The improved SWO is used to extract time-frequency domain features. The improvement method is as follows: S101. Initialize the population using the cubic mapping method. The process is expressed as: SW pop (n+1)=4SW pop (n) 3 -3SW pop (n) (1) Among them, SW pop (n) represents the population at the previous moment; SW pop (n+1) represents the population at the next moment; S102, using the population guidance mechanism to solve the search space, and enhance the exploration ability of the population individuals to the surrounding space. The process is expressed as: in, represents the solution of the population guidance mechanism; SW i t =(SW1 t ,SW2 t ,…,SW n t ) represents an individual in n-dimensional space at time t; l i (f) u i (f) represents the dynamic upper and lower bounds of the entire population at f iterations, l i (f) = min(SW i t ),u i (f) = max(SW i t ); Random number rand()∈[0.1], representing the expansion and contraction control amount of the dynamic boundary; S103, in order to further improve the convergence speed and accuracy of the algorithm, the inertia weight of the sine-cosine transform is introduced in the SWO position update, as shown in formula (3): Among them, the random number z∈[0,1]; ω represents the inertia weight coefficient; f is the current iteration number; f max Indicates the maximum number of iterations; The introduced sine-cosine transform inertia weight is dynamically adjusted with the number of iterations, which helps SWO to jump out of the local optimum, thereby improving the convergence efficiency of the algorithm. The position update process after the introduction of ω is expressed as: in, Represents the position parameter at time t+1; are two random position parameters; μ is the search coefficient; S104 and SWO searches are nonlinear optimization strategies. A nonlinear convergence factor p is introduced into the search coefficient, and its expression is: The search coefficient μ after the introduction of p is expressed as: μ=p·|rn|·r1 (6) Among them, the random number r1∈[0,1]; the random number |rn| obeys the normal distribution; The time domain features extracted by this method include: IMF1 energy entropy and IMF2 energy entropy.

4. The rolling bearing life prediction method based on improved LSTMNN according to claim 3 is characterized in that: The specific steps of feature extraction and feature dimensionality reduction are: S201, extracting time domain, frequency domain, and time-frequency domain feature parameters, wherein the time-frequency domain parameters are extracted using the ISWO-VMD method; S202, feature screening, screening the contribution rate of the extracted time domain and frequency domain features and the time and frequency domain output of ISWO-VMD, so as to screen out the principal component variables; S203, feature dimensionality reduction, using the principal component variables selected by the contribution rate to perform KPCA feature fusion to achieve the purpose of feature dimensionality reduction, and judging the similarity of the reduced dimensionality features by cosine similarity, which is expressed as: Among them, x and y represent two eigenvectors; |x|, |y| represent the modulus of the corresponding eigenvectors.

5. The rolling bearing life prediction method based on improved LSTMNN according to claim 4 is characterized in that: The extraction method in step S201 is specifically: S2011, parameter initialization, initializing ISWO population parameters, location parameters, number of iterations, and optimal fitness value; S2012, parameter update, using the population guidance mechanism and sine-cosine transform inertia weights to update the position parameters of the spider colony; S2013, iterative optimization. If the maximum number of iterations is not reached, jump to step S2012, update the state of the spider colony and perform global optimization. If the maximum number of iterations is reached, stop the optimization and output the K value and α value of the optimized VMD. The K value represents the number of mode decompositions; the α value represents the penalty factor.

6. The rolling bearing life prediction method based on improved LSTMNN according to claim 1 is characterized in that: In step S3, the feature multi-head attention mechanism FMA is used to extract input features of LSTMNN. At time t, for an input x with M features t =[x 1,t ,x 2,t ,...,x M,t ], and its attention weight calculation formula is: e t =σ(W c x t +b c ) (9) Among them, e t =[e 1,t ,e 2,t ,...,e M,t ] represents the combination of attention weight coefficients corresponding to each input feature at the current time t; W c represents the trainable weight matrix; b c is the bias vector; σ() represents the sigmoid activation function; the obtained feature attention weight coefficient e t After normalization, we have: Among them, α M,t represents the feature of the Mth attention weight; exp represents the exponential function expression; e M,t represents the attention weight coefficient of the Mth feature at time t; e i,t Represents the attention weight coefficient of the i-th feature at time t. After the above calculation, the feature attention weight α is obtained t =[α 1,t ,α 2,t ,...,α M,t ], and transform equation (10) with x t Multiply them together to get the weighted calculation and then calculate the weighted vector The expression for weighted input long is: Multihead feature =Concat(head1,head2,…,head h ) (12) in, Concat is to concatenate multiple weighted vectors; The expression of the feature multi-head attention mechanism is: in, 7. The rolling bearing life prediction method based on improved LSTMNN according to claim 6 is characterized in that: In step S3, the temporal multi-head attention mechanism TMA assigns attention weights to the time information carried by each historical moment of the input sequence to distinguish the impact of time information on the predicted output at the current moment; The input of the temporal multi-head attention is the LSTMNN hidden layer state h T =[h 1,T ,h 2,T ,...,h C,T ] time series, where T is the time of model iteration and C is the time window length of the input sequence; the time attention weight vector corresponding to each historical moment in the current moment is: l t =ReLU(W d h T +b d ) (14) Among them, W d represents a trainable weight matrix; b d represents the bias vector; ReLU( ) represents the activation function; Normalize the attention weight coefficient each time to get the time attention weight β t =[β 1,t ,β 2,t ,β τ,t ...,β k,t ], where β τ,t is the attention weight value of τ times, weighted by the hidden layer state of each corresponding historical time, and the comprehensive time series information state h' T ; Its process expression is: Among them, l τ,t represents the time attention weight coefficient of the τth feature at time t; l i,t represents the time attention weight coefficient of the i-th feature at time t; β i,t represents the normalized attention weight coefficient of the ith feature at time t; h i,t represents the hidden layer input of the i-th feature at time t; N represents the number of time points; The expression of the temporal multi-head attention mechanism is: Multihead time =Concat(head1,head2,…,head h ) (17) in 8. The rolling bearing life prediction method based on improved LSTMNN according to claim 7 is characterized in that: The steps of network prediction are: S301, weight initialization, initializing the parameters of the input layer attention weight and the hidden layer attention weight; S302, calculating feature attention weights, using formula (9) to calculate the feature attention weights corresponding to each feature; S303, constructing weighted features, using formula (11) to obtain weighted features, and using formula (13) to perform multiple feature concatenation; S304, calculate the time attention weight, input the obtained features into LSTMNN, and use formula (14) to calculate the time attention weight corresponding to each historical moment; S305, constructing a time-weighted feature, using formula (16) to obtain a weighted feature, and using formula (17) to concatenate multiple time-weighted features; S306, update the hidden layer output state. If the maximum number of iterations has not been reached, jump to step S305. If the maximum number of iterations has been reached, stop updating the time-weighted features and output the predicted value of the remaining life of the rolling bearing.