Slipper bearing life prediction method based on adaptive enhanced VMD-BILSTM model

Abstract

The invention is suitable for the technical field of large-scale double-drive mill slipper bearing life prediction, and particularly relates to a large-scale double-drive mill slipper bearing life prediction method based on an adaptive enhanced VMD-BILSTM model, and the method comprises the steps: obtaining an original vibration signal, carrying out the noise reduction processing of the original vibration signal, and obtaining a preprocessing vibration signal; constructing an adaptive enhanced VMD-BILSTM model, extracting degradation features from two dimensions of a time domain and a frequency domain, compressing feature dimensions, and establishing a degradation feature matrix as the input of the adaptive enhanced VMD-BILSTM model; and adaptively determining optimal parameters of the adaptive enhanced VMD-BILSTM model by using a Bayesian network to obtain a residual service life prediction model of the sliding shoe bearing of the large-scale double-drive mill, and predicting the service life of the bearing by using the residual service life prediction model of the sliding shoe bearing of the large-scale double-drive mill. According to the method, WOA-VMD filtering processing is carried out on the original vibration signals, the influence of noise interference in the signals is reduced, the feasibility of noise elimination is proved, and the problem that [K,] parameters are difficult to select is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of large-scale double-drive mill sliding shoe bearing life prediction, and particularly relates to a large-scale double-drive mill sliding shoe bearing life prediction method based on an adaptive enhanced VMD-BILSTM model. BACKGROUND

[0002] The large-scale double-drive mill is a core grinding equipment in the fields of cement, metallurgy and mining. The sliding shoe bearing of the large-scale double-drive mill bears huge load by relying on external high-pressure oil film. Once the sliding shoe bearing fails accidentally, a disastrous consequence will be caused. Severe dry friction causes the bearing alloy to melt and ablate, damages the bearing body and the expensive hollow shaft, and even causes the support structure to break; the whole line is forced to stop urgently, which brings high maintenance cost, long construction period and great loss of production capacity.

[0003] The residual useful life (RUL) prediction of the sliding shoe bearing is crucial. The RUL refers to the time during which the equipment normally operates according to the predetermined function before failure. Accurate prediction of the RUL can help to make maintenance plans in advance and avoid production interruption and safety accidents caused by sudden failure.

[0004] Traditional signal processing methods such as Fourier transform and wavelet analysis have limited effect on processing such strong coupling non-stationary signals, and it is difficult to separate effective degradation features. The empirical mode decomposition method has the problems of mode aliasing, end effect and false components, and the effect of decomposing the complex mixed signals of the double-drive system is unstable. The existing deep learning methods usually directly use original or simply processed signals, and cannot solve the problem of feature ambiguity caused by the coupling of double-drive signals; the signal decomposition parameters and prediction models are mostly fixed, and cannot adapt to the significant changes in signal characteristics during the entire degradation process from health to failure of the bearing, which leads to a decrease in prediction accuracy and insufficient robustness.

[0005] The adaptive enhanced VMD-BILSTM model can effectively solve the above problems. The model can accurately separate noise and fault features by dynamically optimizing the modal decomposition parameters. The adaptive enhancement mechanism can adjust the feature extraction and prediction weights in real time to adapt to the changes in signal characteristics under various complex working conditions. The method provides a high-precision and strong-robustness solution for the RUL prediction of the sliding shoe bearing. SUMMARY

[0006] The purpose of the present application is to provide a sliding shoe bearing life prediction method based on an adaptive enhanced VMD-BILSTM model, which aims to solve the problem that traditional methods cannot effectively capture weak fault features, and at the same time, the disturbance of noise will also reduce the accuracy of evaluating the degradation of the sliding shoe bearing.

[0007] The present application is realized in the following way: the sliding shoe bearing life prediction method based on the adaptive enhanced VMD-BILSTM model comprises the following steps: The original vibration signal is acquired, and noise reduction processing is performed on the original vibration signal to obtain the preprocessed vibration signal; An adaptive augmented VMD-BILSTM model is constructed, and degradation features are extracted from both the time and frequency domains and the feature dimensions are compressed. The degradation feature matrix is ​​then used as the input to the adaptive augmented VMD-BILSTM model. The optimal parameters of the adaptively enhanced VMD-BILSTM model are determined by using a Bayesian network to obtain a prediction model for the remaining service life of the sliding bearing of a large dual-drive mill. The bearing life is then predicted using this prediction model.

[0008] Preferably, in the step of acquiring the original vibration signal, performing noise reduction processing on the original vibration signal to obtain the preprocessed vibration signal, the variational mode filtering method is used for noise reduction processing, and the mathematical expression is: in: It is obtained through filtering. K One intrinsic mode function (IMF). The first one extracted by VMD k The central angular frequency of each IMF It is the Dirac delta function. It is the impulse response of the Hilbert transform. t The sampling time of the vibration signal is the fundamental variable for constructing time-domain or time-frequency-domain features; The optimal solution is calculated by introducing an augmented Lagrange function, which is expressed as follows:

[0009] In the formula, It is a penalty parameter. For Lagrange multipliers, It is the input raw signal to be decomposed.

[0010] Preferably, the steps for calculating the optimal solution include: (1) Initialization n =0、 , and ; (2) n = n +1, execute the entire loop; (3) Execute the first inner loop, by renew ; (4) Order k = k +1, repeat step (3), whenk = K , break the first iteration loop; (5) execute the second inner loop, update ; ; (6) let k = k +1, repeat step (5), when k = K , end the second loop; (7) update ; ; (8) execute steps (2-7), when the iteration stopping condition is met, output K IMFs.

[0011] wherein: n represents the number of iterations of the outer loop, controls the number of iterations of the entire decomposition process; k represents the index number of the modal component, used to traverse and update all modal components; respectively represent the values of the n th modal component, center frequency and Lagrange multiplier in the k th iteration; K represents the total number of preset modal components, i.e. the number of IMFs; represents the step size parameter of the Lagrange multiplier update; represents the iteration stopping threshold.

[0012] Preferably, in the step of noise reduction processing by the variational mode filtering method, the whale optimization algorithm is used to calculate the global optimal solution: (a) initialization of whale individuals, population, position and iteration number, the i th individual position is as follows: wherein: is a random number, the search space size is , and are the minimum and maximum values of the boundary parameters, respectively; (b) according to the coefficient vector A , select the path of searching prey, use the predation mechanism probability p to determine the final predation mechanism, when and , the best agent performs contraction and encirclement, represented as: (4-6) wherein, , and p The range of values ​​is random numbers, t and t max These are the current iteration and the maximum number of iterations. and These are vector coefficients. It decreases linearly to 0 during the iteration process. and This represents the vectors representing the current solution position and the optimal solution position; when and Select a random search agent for iterative updates, the expression is: In the formula: To randomly select individual position vectors; The iteration uses a spiral contraction method, and the expression is:

[0013] In the formula, Indicating the distance between whale pods and their prey, l yes Random numbers between b It is the helical coefficient; (c) Determine whether the termination condition is met, i.e. The maximum number of iterations is reached, and the convergence factor is... Reduce the value to 0. If the condition is not met, return to step (b). If the condition is met, output the best search agent.

[0014] Preferably, the fitness value is calculated each time the whale updates its position. The fitness value is updated by comparing it with the fitness value of the new individual whale. The convergence factor is determined, and an iterative formula is selected for iterative updates until the termination condition is met. The optimal parameters are then output. In the step of calculating the global optimum using the whale optimization algorithm, the envelope entropy is used as the fitness function, and the number of iterations is set to 50 for zero-mean signals. Its envelope entropy The calculation formula is: In the formula, , yes The signal after Hilbert envelope demodulation yes The normalized expression.

[0015] Preferably, in the step of adaptively determining the optimal parameters of the adaptive enhanced VMD-BILSTM model by the Bayesian network, in the optimization process, the search space of the hyperparameters is optimized by limited iteration experiments on the objective function to obtain the optimal hyperparameter combination. In the formula, is the final optimal hyperparameter combination, is the acquisition function. The probability distribution representation of the optimization process is: In the formula, P ( E ) is a Gaussian distribution, and the Gaussian regression process P(D|E) is determined by the kernel matrix , and the expression is: .

[0016] Preferably, the adaptive enhanced VMD-BILSTM model adopts an LSTM unit structure.

[0017] Preferably, in the step of denoising processing by the variational mode decomposition filtering method, it is defined that the signal is composed of IMFs with limited bandwidth and amplitude modulation frequency modulation, and the amplitude modulation frequency modulation signal is defined as:

[0018]

[0019] In the formula, , is the instantaneous amplitude and frequency of , which is simplified to a harmonic signal with an amplitude of and a frequency of in a time interval range .

[0020] Preferably, in the step of calculating the global optimal solution by the whale optimization algorithm, there are N only whales in space, and the position of each whale in space is recorded as: .

[0021] The large-scale double-drive mill shoe bearing life prediction method based on the adaptive enhanced VMD-BILSTM model has the following beneficial effects: (1) The original vibration signal is subjected to WOA-VMD filtering processing, the influence of noise interference in the signal is reduced, the feasibility of denoising is proved, and the problems of [K,​ the problem of difficult parameter selection; (2) The BO-BILSTM prediction model has improved training speed by autonomously selecting time domain and frequency domain features and using a comprehensive degradation function to reduce redundant information. The BO optimizes the BILSTM network to adaptively determine model parameters, reducing the cost of manual debugging. Meanwhile, the effectiveness of the enhanced VMD-BILSTM combination is verified by the XJTU-SY data set. Compared with the BILSTM, CNN, and VMD-GRU regression models, the root mean square error is the smallest, and the determination coefficient is the largest, which shows that the adaptively enhanced VMD-BILSTM model can more accurately provide real-time maintenance strategies for equipment. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The RNN network loop structure diagram provided for the embodiment of the application; Figure 2 The LSTM unit structure diagram provided for the embodiment of the application; Figure 3 The improved LSTM unit structure diagram provided for the embodiment of the application; Figure 4 The prediction model building process schematic diagram provided for the embodiment of the application. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0024] In actual working conditions, due to the influence of the environment and the equipment, the collected vibration signals of the large double-drive mill sliding shoe bearing often have a large amount of noise, resulting in low signal-to-noise ratio and a large amount of invalid characteristic information. At the same time, the key hyperparameters of the prediction network need good prior knowledge to achieve better regression results. In view of the above problems, a prediction method of adaptively enhanced VMD fusion bidirectional long short-term memory network (VMD-BILSTM) is proposed.

[0025] Firstly, the model uses whale optimization algorithm (WOA) to optimize the VMD parameters for filtering and denoising of the vibration acceleration signal.

[0026] Secondly, the minimum envelope entropy is used to select the best modal component to extract the degradation features in the time domain and frequency domain and compress the feature dimension, and a degradation feature matrix is established as the input of the RUL model.

[0027] Then, the Bayesian network (BO) is used to adaptively determine the optimal parameters of the BILSTM model, and a large double-drive mill sliding shoe bearing remaining useful life prediction model is established.

[0028] Finally, the remaining life prediction verification is performed using the XJTU-SY bearing dataset.

[0029] Experiments show that compared with VMD-GRU, BILSM and CNN prediction models, the prediction accuracy and stability of the self-adaptive enhanced VMD-BILSTM regression model are improved.

[0030] The embodiment of the application provides a large double-drive mill sliding shoe bearing life prediction method based on a self-adaptive enhanced VMD-BILSTM model, and the method comprises the following steps: Obtaining an original vibration signal, and performing noise reduction processing on the original vibration signal to obtain a pretreated vibration signal.

[0031] In this step, a variational mode decomposition (VMD) is used, and the VMD is a non-recursive signal filtering method and has the characteristics of self-adaptation and quasi-orthogonality. The method assumes that any signal is composed of IMFs with limited bandwidth and amplitude modulation frequency modulation, effectively avoiding the end effect and false component problem in the solving process, and the definition of the amplitude modulation frequency modulation signal is as follows.

[0032] (1) (2) In formulas (1) and (2), 、 is the instantaneous amplitude and frequency of , 、 is slowly varying relative to , , can be simplified into a harmonic signal with an amplitude of and a frequency of .

[0033] As an improved EMD method, the core mechanism of VMD is to convert the signal frequency band division problem into a variational optimization problem. Unlike the EMD method based on the processing mechanism of cyclic screening intrinsic mode function, VMD realizes the adaptive decomposition of the signal in the framework of functional analysis by constructing a constrained variational model. In the model solving process, the center frequency and bandwidth parameters of each modal component are dynamically adjusted through the alternating direction multiplier method, so as to accurately match the spectral distribution characteristics of the target signal. For a given signal to be decomposed, the algorithm finally obtains K IMF (intrinsic mode component) with physical significance through the optimization process of minimizing the sum of the bandwidths of all modal components.

[0034] The process can be described as a functional extremum problem with constraints, which is mathematically expressed as (4-3).

[0035] (3) wherein: is the filtered K Intrinsic Mode Function (IMF), is the center angular frequency of the k th IMF extracted by VMD, is the Dirac delta function, is the impulse response of Hilbert transform, t denotes the sampling time of the vibration signal, which is the basic variable for constructing time-domain or time-frequency domain features; The augmented Lagrange function (4) is introduced to solve the optimal solution of the above formula (3).

[0036] (4) In formula (4), is the penalty parameter; is the Lagrange multiplier; is the input original signal to be decomposed.

[0037] In the solving process, the Alternating Direction Method of Multipliers (ADMM) is used to solve the saddle point of the augmented function Lagrange, so as to decompose the original signal, and the process is as follows.

[0038] (1) initialization n = 0, , and ; (2) n = n + 1, execute the entire loop; (3) execute the first inner loop, update by ; (4) let k = k + 1, repeat step (3), when k = K , exit the first iteration; (5) execute the second inner loop, update by ; (6) let k = k + 1, repeat step (5), when k = K , end the second loop; (7) with updating ; (8) Perform steps (2-7) until the iteration stopping condition is met, output K IMFs.

[0039] wherein: n represents the number of iterations of the outer loop, controlling the number of iterations of the entire decomposition process; k represents the index number of the modal component, used to traverse and update all modal components; respectively represent the values of the n th modal component, center frequency and Lagrange multiplier in the k th iteration; K represents the preset total number of modal components, i.e. the number of IMFs; represents the step size parameter of the Lagrange multiplier update; represents the iteration stopping threshold.

[0040] It should be noted that when VMD processes a signal, the number of IMF components needs to be preset, and different K values will result in differences in the processing results. The penalty parameter in the VMD algorithm has a significant impact on the decomposition results, a small parameter will result in a larger bandwidth of the IMF component. a too large parameter will result in a smaller bandwidth of the component. Since the vibration signal is complex, K the value and the penalty parameter are difficult to determine, therefore, selecting the appropriate parameter combination is the key to extracting the signal fault degradation features.

[0041] An adaptive enhanced VMD-BILSTM model is constructed to extract degradation features from two dimensions of time domain and frequency domain and compress the feature dimension, and a degradation feature matrix is established as the input of the adaptive enhanced VMD-BILSTM model.

[0042] In this step, the whale optimization algorithm (WOA) is developed by imitating the social behavior of humpback whales when hunting, aiming to solve some optimization problems. WOA can better cope with randomness than other traditional group optimization algorithms, and its initial parameters do not need to be set too much, and the convergence speed is fast. The WOA running process contains two mutually coordinated stages of exploration and development to solve the imbalance between exploration and development during the search.

[0043] Suppose there are N whales hunting in space, and the position of each whale in space can be recorded as The whales performing the optimization task are defined as "search agents". The local optima found by these agents in the solution space are represented as local extrema. WOA utilizes these "search agents" with random initial positions to find the position of the most fish in the bubble net strategy to obtain the global optimum solution. The process is shown as follows.

[0044] (a) The parameters of the whale individual, population, position, and iteration number are initialized, and the first iteration is performed. i The individual position is shown as follows: (5) In the formula: is a random number, is the search space size , and are the minimum and maximum values of the boundary parameters, respectively.

[0045] (b) WOA selects the path of searching prey according to the coefficient vector A , and the probability of using the predation mechanism p determines the final predation mechanism, when and , the best agent performs a shrinkage surrounding as shown in formula (6): (6) In formula (6), , and p are random numbers with a value range of , t and t max are the current iteration and the maximum iteration number, and are the vector coefficients, linearly decrease to 0 during the iteration process, and represent the current solution position and the optimal solution position vector.

[0046] When and , a random search agent is selected for iterative update as shown in formula (7), and the spiral shrinkage method is used for iteration as shown in formula (4-8).

[0047] (7) In the formula: is a random individual position vector.

[0048] (8) In the formula, characterizing the distance between the whale group and the prey, l is a random number between b is the helical coefficient.

[0049] (c) determining whether a termination condition is met, i.e. ( depending on the specific problem iteration accuracy), a maximum number of iterations is reached, and the convergence factor in equation (7) is reduced to 0. If not, then step (b) is returned to, and if so, then the optimal search agent is output. The fitness value is calculated once per position update of the whale, and is updated by comparing the fitness value of the new whale individual. Then the convergence factor size is determined to select the iteration formula for iteration update until the termination condition is met, and the optimal parameters are output.

[0050] When optimizing the VMD parameters, the envelope entropy is set as the fitness function, and the number of iterations is set to 50 times. The envelope entropy can well evaluate the sparsity of the signal. When the entropy value is small, it indicates that the signal has obvious sparse characteristics, and the richer the degradation characteristics contained, the more the zero mean signal its envelope entropy is calculated as follows: (9) In equation (9), , is the signal after Hilbert envelope demodulation, is a normalized expression of

[0051] The Bayesian network is used to adaptively determine the optimal parameters of the adaptive enhancement VMD-BILSTM model, and a large-scale double-drive mill sliding shoe bearing residual service life prediction model is obtained. The bearing life is predicted through the large-scale double-drive mill sliding shoe bearing residual service life prediction model.

[0052] In this step, BO is proposed as an active optimization algorithm in 2012. This algorithm is mainly used to solve the extreme value problem of unknown expression function. A large number of experiments prove that Bayesian optimization has better advantages than traditional grid search and random search, and can largely avoid falling into local optimum.

[0053] The core architecture of Bayesian optimization consists of two key components: a proxy model and a collection function. The proxy model is responsible for building the probability distribution estimate of the objective function, and the commonly used modeling methods include Gaussian process and random forest regression. The collection function determines the optimal sampling point sequence to optimize the model fitting accuracy by balancing the relationship between local development and global exploration. Common collection strategies include "expected improvement", "upper confidence bound", and "probability improvement".

[0054] In the present application, Gaussian process is used as a prior distribution modeling tool, and the "probability boost" function is selected as the sampling strategy to achieve the global optimization of the objective function. In the optimization process, the search space of the hyperparameters is optimized by the optimizer through limited iterative experiments on the objective function to obtain the optimal hyperparameter combination.

[0055] (10) In formula (10), is the final optimal hyperparameter combination, is the acquisition function.

[0056] The probability distribution of the optimization process is constructed by using the BO formula as shown below.

[0057] (11) In formula (11), P ( E ) is a Gaussian distribution, and the Gaussian regression process P(D|E) is determined by the kernel matrix , and the expression is as follows: (12) The final goal of optimization is to continuously iterate to obtain the best hyperparameters, and through the model accuracy is optimized.

[0058] Bidirectional long short-term memory network (BILSTM): Large double-drive mill sliding shoe bearing RUL prediction is a long time series prediction problem, which needs to infer future changes based on current features. The degradation process involves fault accumulation, state change, and both current and historical monitoring values.

[0059] LSTM, as an improved RNN, can store current and historical data, capture long-term dependencies, and control selective flow of information by increasing gate unit structure, and process long time series data. It effectively solves the problems of difficult training, gradient dispersion and gradient explosion existing in traditional RNN. The RNN network cycle structure diagram and the LSTM unit structure diagram are as follows Figure 1 , Figure 2 .

[0060] In Figure 1 , is the input sequence information, is the state vector corresponding to the moment, and the memory cell in each layer of network is represented by A. One layer of RNN is composed of input layer, hidden layer and state layer. Figure 2 In , is tThe output of the memory cell at time t, is t The input of the memory cell at time t, denotes the sigmoid activation function, is the activation value of the forget gate, is the activation value of the input gate, is the candidate state of the memory cell, is the activation value of the output gate, is t The state of the memory cell at time t. The hidden state of the current LSTM cell together with the input value of the sample at the current time The forget gate decides which information in the memory cell should be forgotten, the input gate decides whether the current input information is written into the memory cell, and the output gate decides whether the information in the memory cell is used as the output of the current LSTM. However, LSTM can only learn forward information of time series and cannot use backward information. BiLSTM improves it by using forward and reverse LSTM, and the network structure is as shown in Figure 3 .

[0061] Figure 3 The relationship derivation in formula (13) to (17).

[0062] (13) (14) (15) (16) (17) (18) Wherein: is the weight vector and bias of the forget gate, is the convolution operation, is the weight vector and bias of the input gate, is the candidate weight vector and bias of the output gate, is the weight vector and bias of the output gate.

[0063] The network uses the forward layer and the reverse layer to independently process time series data, thereby fully utilizing the information contained in the past and future time data, and then transmits the processing results of the two layers to the output layer at the same time to obtain the final output .

[0064] (19) wherein, and are t the forward and reverse output of the momentary memory unit.

[0065] The construction of the adaptive enhanced VMD-BILSTM model mainly includes two processes, signal denoising and extracting degradation features and RUL prediction. Specifically, it can be subdivided into: denoising filtering of the original signal by VMD; calculation of the envelope entropy of the IMF, selection of the component with the minimum envelope entropy for time domain and frequency domain feature extraction and screening; model building, Bayesian training and optimization of the BILSTM model; analysis of bearing life test results. The specific steps are as follows.

[0066] (1) WOA optimizes the K and parameters of VMD to obtain the optimal penalty factor and value for denoising filtering to obtain the intrinsic mode function and residual subsequence.

[0067] (2) Calculate the envelope entropy value of each order modal function, and reconstruct the signal of the component with the minimum envelope entropy.

[0068] (3) Preprocess the reconstructed signal and extract various time domain and frequency domain features. Use monotonicity, trendiness and robustness indicators to evaluate and screen the features, and construct a feature parameter set as the feature input for subsequent model training.

[0069] (4) Build a BILSTM network, with the number of hidden layer nodes , learning rate and regularization coefficient as the variables for Bayesian network optimization, and set the fitness function as the root mean square error. The parameter search range is set to: [10, 100], [0.001, 1], [10 -11 , 10 -3 ], the maximum number of iterations is 20, the Adam optimizer is used for gradient optimization, and the optimal value of the prediction model parameters is obtained. At the same time, the optimal model is obtained by training the model, and the final effect is shown on the test set. The process is shown in Figure 4 .

[0070] In order to verify the prediction results of the model, the prediction results are verified:

[0071] In engineering applications, compared with lag prediction (prediction value > true value), lead prediction (prediction value < true value) is more valuable and can provide reliable basis for maintenance decision-making. In order to quantitatively evaluate the effectiveness and advantages of the prediction model, the root mean square error RMSE and the determination coefficient R 2 are used as quantitative indicators to evaluate the performance of the model, and the prediction value Real value y = {y1, y2, …, y n}, and the error evaluation is shown in formulas (20) and (21). In formulas (20) and (21), n is the number of prediction points, is the prediction value of the i-th sample, i is the serial number of the prediction point, is the average value of y i . The smaller the RMSE is, the better the prediction effect of the model is, and the smaller the error value between the prediction value and the real value is. R 2 measures the fitting degree of the regression model, and the value range is 0-1. The closer to 1, the better the fitting effect of the model is.

[0072] The present application aims at the problems that it is difficult to extract effective feature information of bearing signals under the disturbance of background noise, and a large amount of prior knowledge is needed to obtain a better regression result. If the noise signal is not processed or a simple noise reduction method is used, it will have a certain influence on the prediction result, resulting in a decrease in prediction accuracy. Therefore, a new RUL scheme of an adaptive enhanced VMD-BILSTM model is proposed. The method uses WOA to adaptively determine the K , ] parameters to decompose the original vibration signal, extracts features from the best components, and compresses the dimension of the input features through three indexes. Finally, a BO enhanced BILSTM VMD-BILSTM prediction model is constructed, and the accuracy and robustness of the model are verified through the XJTU-SY data set and the ablation experiment. The RUL research method based on the adaptive enhanced VMD-BILSTM model has the following advantages: (1) the original vibration signal is filtered by WOA-VMD, which reduces the influence of noise interference in the signal, proves the feasibility of noise reduction, and solves the problem of K , ] parameter selection; (2) the time domain and frequency domain features are selected independently, and the data dimension is compressed using a comprehensive degradation function to reduce redundant information and improve the training speed of the BO-BILSTM prediction model. BO optimizes the BILTSM network to adaptively determine the model parameters, reducing the cost of manual debugging. At the same time, the effectiveness of the enhanced VMD-BILSTM combination is verified through the XJTU-SY data set. Compared with BILSTM, CNN, and VMD-GRU regression models, the root mean square error is the smallest, and the determination coefficient is the largest, which shows that the adaptive enhanced VMD-BILSTM model can provide more accurate real-time maintenance strategies for equipment.

[0073] The above merely describes preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement and improvement within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for predicting the life of sliding bearings in large dual-drive mills based on an adaptive enhanced VMD-BILSTM model, characterized in that, The method includes: The original vibration signal is acquired, and noise reduction processing is performed on the original vibration signal to obtain the preprocessed vibration signal; An adaptive augmented VMD-BILSTM model is constructed, and degradation features are extracted from both the time and frequency domains and the feature dimensions are compressed. The degradation feature matrix is ​​then used as the input to the adaptive augmented VMD-BILSTM model. The optimal parameters of the adaptively enhanced VMD-BILSTM model are determined by using a Bayesian network to obtain a prediction model for the remaining service life of the sliding bearing of a large dual-drive mill. The life of the sliding bearing of the large dual-drive mill is then predicted using the prediction model for the remaining service life of the bearing.

2. The method for predicting the life of sliding bearings in large dual-drive mills based on the adaptive enhanced VMD-BILSTM model according to claim 1, characterized in that, In the steps of acquiring the original vibration signal and performing noise reduction processing on the original vibration signal to obtain the preprocessed vibration signal, the variational mode filtering method is used for noise reduction processing, and the mathematical expression is: in: It is obtained through filtering. K One intrinsic mode function (IMF). The first one extracted by VMD k The central angular frequency of each IMF It is the Dirac delta function. It is the impulse response of the Hilbert transform. t The sampling time of the vibration signal is the fundamental variable for constructing time-domain or time-frequency-domain features; The optimal solution is calculated by introducing an augmented Lagrange function, which is expressed as follows: In the formula, It is a penalty parameter. For Lagrange multipliers, It is the input raw signal to be decomposed.

3. The method for predicting the life of sliding bearings in large dual-drive mills based on the adaptive enhanced VMD-BILSTM model according to claim 1, characterized in that, The steps to calculate the optimal solution include: (1) Initialization n =0、 , and ; (2) n = n +1, execute the entire loop; (3) Execute the first inner loop, by renew ; (4) Order k = k +1, repeat step (3), when k = K At that time, exit the first iteration; (5) Execute the second inner loop, by renew ; (6) Order k = k +1, repeat step (5), when k = K When the second loop ends; (7) with renew ; (8) Execute step (2-7) when the iteration stopping condition is met. Output K One IMF; in: n Represents the number of iterations in the outer loop, controlling the total number of iteration rounds in the entire decomposition process; k The index number represents the modal component, used to iterate and update all modal components; Representing respectively in the n In the first iteration, the... k Each modal component, center frequency, and value of the Lagrange multiplier; K This represents the preset total number of modal components, i.e., the number of IMFs; The step size parameter represents the Lagrange multiplier update; This represents the iteration stopping threshold.

4. The method for predicting the life of sliding bearings in large dual-drive mills based on the adaptive enhanced VMD-BILSTM model according to claim 2, characterized in that, In the noise reduction step using variational mode filtering, the whale optimization algorithm is used to calculate the global optimum. (a) Initialization of individual whales, population, location, and iteration number, the first iteration... i The locations of the individuals are shown below: In the formula: random numbers, The search space size is , and These are the minimum and maximum values ​​of the boundary parameters, respectively. (b) Based on the coefficient vector A Choosing a path to search for prey, utilizing predation mechanisms and probabilistic strategies. p The final predation mechanism is determined when and The best agent's contraction and encirclement is represented as: In equation (4-6), , and p The range of values ​​is random numbers, t and t max These are the current iteration and the maximum number of iterations. and These are vector coefficients. It decreases linearly to 0 during the iteration process. and This represents the vectors representing the current solution position and the optimal solution position; when and Select a random search agent for iterative updates, the expression is: In the formula: To randomly select individual position vectors; The iteration uses a spiral contraction method, and the expression is: In the formula, Indicating the distance between whale pods and their prey, l yes Random numbers between b It is the helical coefficient; (c) Determine whether the termination condition is met, i.e. The maximum number of iterations is reached, and the convergence factor is... Reduce the value to 0. If the condition is not met, return to step (b). If the condition is met, output the best search agent.

5. The method for predicting the life of sliding bearings in large dual-drive mills based on the adaptive enhanced VMD-BILSTM model according to claim 4, characterized in that, Each time a whale updates its position, it calculates its fitness value. This fitness value is then compared to that of a new individual whale. An iterative update formula is selected based on the convergence factor, and the process continues until a termination condition is met. The optimal parameters are then output. In the step of calculating the global optimum using the whale optimization algorithm, the envelope entropy is used as the fitness function, and the number of iterations is set to 50 for zero-mean signals. Its envelope entropy The calculation formula is: In the formula, , yes The signal after Hilbert envelope demodulation yes The normalized expression.

6. The method for predicting the life of sliding bearings in large dual-drive mills based on the adaptive enhanced VMD-BILSTM model according to claim 5, characterized in that, In the step of adaptively determining the optimal parameters of the adaptively enhanced VMD-BILSTM model using Bayesian networks, during the optimization process, the search space for hyperparameters is considered. To conduct limited iterative experiments on the objective function Optimize to obtain the optimal combination of hyperparameters; In the formula, For the final optimal combination of hyperparameters, This is the data acquisition function; The probability distribution of the optimization process is represented as follows: In the formula, P ( E The distribution is Gaussian, and the Gaussian regression process P(D|E) is determined by the kernel matrix. To determine, the expression is: 。 7. The method for predicting the life of sliding bearings in large dual-drive mills based on the adaptive enhanced VMD-BILSTM model according to claim 1, characterized in that, The adaptive augmented VMD-BILSTM model uses an LSTM cell structure.

8. The method for predicting the life of sliding bearings in large dual-drive mills based on the adaptive enhanced VMD-BILSTM model according to claim 1, characterized in that, In the noise reduction process using variational mode filtering, the signal is set to consist of a finite bandwidth and an amplitude-modulated frequency-modulated (AM-FM) signal. The AM-FM signal is defined as follows: ; In the formula, , yes The instantaneous amplitude and frequency within a time interval , Simplified to amplitude The frequency is Harmonic signals.

9. The method for predicting the life of sliding bearings in large dual-drive mills based on the adaptive enhanced VMD-BILSTM model according to claim 4, characterized in that, In the steps of calculating the global optimum using the whale optimization algorithm, there are: N A whale hunts in a space, and the position of each whale in the space is recorded as: 。

Images