A bearing remaining life prediction method based on BN-RBM and DA-BiLSTM

Abstract

This invention discloses a bearing remaining life prediction method based on BN-RBM and DA-BiLSTM. First, raw vibration signals from the bearing during operation are collected and preprocessed. Then, a bearing remaining life prediction model is constructed. This model includes a BN-RBM feature extraction network, a DA-BiLSTM feature fusion network, and a linear regression layer. The BN-RBM feature extraction network dynamically updates the prior weight distribution of the Restricted Boltzmann Machine (RBM) using posterior probabilities constructed from Bayesian BN, and extracts features using the trained RBM. The DA-BiLSTM feature fusion network performs bidirectional feature fusion, and the linear regression layer performs linear activation, outputting the predicted bearing remaining life. This invention effectively improves the accuracy of bearing remaining life prediction by providing timely and effective life prediction before failure, reducing the risk of downtime caused by failures.

Description

Technical Field

[0001] This invention relates to the field of mechanical data processing technology, specifically a bearing remaining life prediction method based on BN-RBM and DA-BiLSTM. Background Technology

[0002] Bearings are common mechanical components widely used in various industries. However, with increasing usage time, bearings are often subjected to extremely high loads during operation, and under such harsh conditions, bearing defects become increasingly apparent. Bearings frequently fail due to wear and fatigue, which can lead to machine downtime and production line interruptions. However, when a bearing first fails, the vibration signals are weak and difficult to detect. Therefore, predicting and assessing the remaining service life of bearings is crucial for scheduling planned maintenance before they approach failure, ensuring reliable equipment operation, and protecting personnel and equipment safety. Bearing remaining service life prediction is based on the current state and historical operating data of the bearing to predict its remaining life at a future point in time. Currently, significant progress has been made in the field of bearing life prediction, such as feature extraction methods based on multi-scale morphological decomposition spectral entropy, and pattern recognition methods using support vector machines and artificial neural networks. However, traditional RUL prediction methods have significant limitations in feature extraction capabilities and the ability to model dynamic changes in time-series data. RBMs excel at learning nonlinear features in high-dimensional data, but lack the ability to model time dependencies. BiLSTMs are good at capturing long-term dependency structures in time series, but they are sensitive to the quality of input features and easily affected by redundant information, thus impacting prediction stability and accuracy. Furthermore, traditional model parameter optimization relies on human experience or greedy search strategies, making them prone to getting trapped in local optima and lacking robustness. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a bearing remaining life prediction method based on BN-RBM and DA-BiLSTM, which effectively improves the accuracy of bearing remaining life prediction. By performing timely and effective life prediction before failure, the risk of downtime caused by failure is reduced, and the time, economic and energy losses are reduced.

[0004] The technical solution of this invention is as follows:

[0005] A bearing remaining life prediction method based on BN-RBM and DA-BiLSTM specifically includes the following steps:

[0006] (1) Collect the original vibration signals during the operation of the bearing and perform data preprocessing to construct the bearing dataset;

[0007] (2) Construct a bearing remaining life prediction model. The bearing remaining life prediction model includes a BN-RBM feature extraction network, a DA-BiLSTM feature fusion network, and a linear regression layer. The BN-RBM feature extraction network uses the posterior probability constructed by Bayesian BN to dynamically update the weight prior distribution of the Restricted Boltzmann Machine (RBM), obtaining a dynamic RBM weight prior distribution. Then, the RBM trained by the dynamic RBM weight prior distribution is used to extract the degradation features in the bearing data. The DA-BiLSTM feature fusion network uses the sliding window mechanism of the BiLSTM network to construct time series samples and performs bidirectional feature fusion on the extracted features. The sliding window length is determined by the Dragonfly Optimization (DA) algorithm. The linear regression layer linearly activates the fused features of the bearing and outputs the predicted value of the bearing remaining life.

[0008] (3) Construct a loss function to train the bearing remaining life prediction model, and use the trained bearing remaining life prediction model to predict the bearing remaining life and obtain the bearing remaining life.

[0009] The data preprocessing involves performing a fast Fourier transform on the original vibration signal to convert the time-domain signal into a frequency-domain signal, obtaining the frequency-domain amplitude spectrum information, and constructing a bearing remaining life label, thereby constructing a bearing dataset.

[0010] The BN-RBM feature extraction network performs feature extraction in the following steps:

[0011] S21. Construct a three-layer Bayesian BN network B=(F, D, R). The set of nodes representing the original degradation characteristics of the bearing, which are selected from historical degradation data and reflect the degradation pattern of the bearing; R represents the set of nodes in the degradation stage; R represents the set of nodes with remaining bearing life.

[0012] The conditional dependencies between nodes in a BN network are described using a conditional probability table, as shown in equation (1) below:

[0013] (1);

[0014] In equation (1), Its function is to encode prior knowledge of degradation feature changes, degradation stages, and remaining life, which is obtained by training based on the historical degradation data of the bearing;

[0015] S22. Posterior probability of the degenerate stage of inference: given the degenerate feature vector at the current time t. The posterior probability of the degradation stage is inferred using a Bayesian BN network, as shown in the following equation (2):

[0016] (2);

[0017] In equation (2), This represents the degenerate feature vector extracted at the current time t; Indicates the first A stage of degradation; Represents the stage of degradation The prior probability is obtained from the statistical analysis of historical degradation data of the bearing; Represents the stage of degradation Below, observed degradation characteristics The likelihood probability is obtained by looking up the conditional probability table; Represents the current degradation feature After that, it enters a degenerative stage. The posterior probability;

[0018] The output of the Bayesian BN network is the current degradation feature. Below, the posterior probability distribution composed of the posterior probabilities of each degradation stage. ;

[0019] S23. Construct the weight prior distribution of the Restricted Boltzmann Machine (RBM), as shown in the following equation (3):

[0020] (3);

[0021] In equation (3), Represents the stage of degradation Lower, RBM weight The prior distribution is a Gaussian distribution. ; Represents the stage of degradation The corresponding weighted average; Represents the stage of degradation The corresponding weighted covariance;

[0022] S24. The posterior probability constructed using Bayesian BN is used to dynamically update the weight prior distribution of the Restricted Boltzmann Machine (RBM), resulting in the dynamic RBM weight prior distribution. , see the following formula (4) for details:

[0023] (4);

[0024] In equation (4), Represents the current degradation feature After that, it enters a degenerative stage. The posterior probability; This represents the updated dynamic weight mean at the current time t. This represents the dynamic weighted covariance at the current time t after the update;

[0025] S25. A Restricted Boltzmann Machine (RBM) trained using a dynamic RBM weight prior distribution is used to extract degenerate features from bearing data. The extracted feature vector is denoted as... , This represents the number of nodes in the hidden layer of the RBM;

[0026] S26. Introduce a Dropout layer to randomly zero out the features extracted by the Restricted Boltzmann Machine (RBM). See equation (5) below for details:

[0027] (5);

[0028] In equation (5), Representing the The first sample The values ​​of the degenerate features after processing by the Dropout layer. Represents the first layer of the input Dropout layer The first sample A degenerate feature vector, This represents the set retention threshold;

[0029] The features extracted by the Restricted Boltzmann Machine (RBM) are randomly zeroed out to output a higher-order feature representation. , This represents a degenerate feature vector after being randomly set to zero. Represents all degenerate feature vectors A set of.

[0030] The specific steps for bidirectional feature fusion using the DA-BiLSTM feature fusion network are as follows:

[0031] The degenerate feature set output by the S31 and BN-RBM feature extraction networks is Organized into time series samples according to time sequence, the generated time series samples are shown in the following formula (6):

[0032] (6);

[0033] In equation (6), This represents the m-th time series sample; This represents the length of the sliding window, determined by the Dragonfly Optimization (DA) algorithm. This represents the set of degradation features at time t in the m-th time series sample;

[0034] S32. The BiLSTM network consists of multiple layers of BiLSTM structure. Each layer of BiLSTM structure includes a forward LSTM and a backward LSTM. The processing procedure of each layer of BiLSTM structure is shown in the following equation (7):

[0035] (7);

[0036] In equation (7), This represents the hidden state of the feedforward LSTM at time t. This represents the hidden state of the feedforward LSTM at time t-1. This represents the hidden state of the backward LSTM at time t. This represents the hidden state of the backward LSTM at time t+1. Represents a forward LSTM, Represents backward LSTM, This represents the bidirectional fusion characteristic at time t;

[0037] S33. A Dropout layer is introduced to randomly zero out the time-series samples processed by the BiLSTM network, and global average pooling is used to obtain a fixed vector representation, as shown in equation (8) below:

[0038] (8);

[0039] In equation (8), The Dropout layer is processed in the same way as formula (5). This represents the fused features output by the DA-BiLSTM feature fusion network.

[0040] The sliding window length The optimization is determined by the Dragonfly Optimization Algorithm (DA), and the specific steps are as follows:

[0041] S311. Construct the configuration parameters of the BiLSTM network and combine them to form a vector. , Parameters in Represents the number of layers in a BiLSTM network. Represents the hidden layer size of BiLSTM Represents learning rate, Represents the Dropout ratio, Represents the length of the sliding window;

[0042] Each For an individual, defined within a preset range, n are randomly generated. At the same time, initialize its velocity vector. ;

[0043] S312. In each iteration, The parameters in the formula are applied to the BiLSTM network, trained and evaluated on the validation set, and the prediction mean square error (MSE) is calculated. The fitness value is set to the validation MSE. The group with the smallest and largest MSE in the current group is labeled, corresponding to food source and enemy, respectively.

[0044] S313, the Dragonfly Optimization Algorithm (DA) includes the following five types of behavior vectors:

[0045] Separation Position avoidance between individuals;

[0046] Alignment Speed ​​is averaged compared to neighbors;

[0047] Cohesion : Move closer to the center of the group;

[0048] feeding : To move closer to the optimal individual;

[0049] Avoid the enemy Stay away from the worst individuals;

[0050] And combined with inertia weight The speed and parameter configuration are dynamically updated, as detailed in equations (9) and (10):

[0051] (9);

[0052] (10);

[0053] In equations (9) and (10), The velocity vector at time t+1 The velocity vector at time t. This represents the configuration parameters of the BiLSTM network at time t. The weights represent the configuration parameters of the BiLSTM network at time t+1. and inertia weight The process gradually transitions from exploration to utilization during iteration, stopping when the maximum number of iterations is reached or the fitness improvement falls below a threshold; the output result, the optimal individual, represents the optimal parameter configuration for the BiLSTM network.

[0054] The linear regression layer linearly activates the fusion characteristics of the bearing and outputs the predicted value of the remaining bearing life, as shown in the following formula (11):

[0055] (11);

[0056] In equation (11), The weight parameters represent the weights of the fully connected layer. Represents the bias of the fully connected layer. Represents the activation function. This represents the predicted value of the remaining life of the bearing.

[0057] The loss function is shown in equation (12):

[0058] (12);

[0059] In equation (12), Representing the A sample of real bearing remaining life labels. Representing the Predicted remaining life of a sample bearing. Represents the regularization coefficient. The parameter set representing the bearing remaining life prediction model. Represents the total number of samples. This represents the square of the L1 norm.

[0060] Advantages of this invention:

[0061] (1) The BN-RBM feature extraction network of the present invention uses the posterior probability constructed by Bayesian BN to dynamically update the weight prior distribution of the Restricted Boltzmann Machine (RBM), calculates the posterior probability of feature parameters based on historical fault data, constructs a probabilistic graphical model to quantify the uncertainty of the feature extraction process, and outputs confidence intervals instead of a single predicted value. This enables the RBM to automatically focus on the more sensitive features of the current stage during feature extraction (such as focusing on lubrication features in the mild degradation stage and focusing on metal wear features in the severe degradation stage), thereby enhancing the feature expression ability and the pertinence and accuracy of the remaining life prediction. It can more comprehensively reflect the uncertainty of bearing remaining life prediction and provide richer information for decision-making in industrial applications.

[0062] (2) The present invention inputs the high-order feature vector output by the BN-RBM feature extraction network into the BiLSTM network, and uses its bidirectional memory mechanism to capture the long-term dependency relationship between the bearing's previous and subsequent states during operation, thereby accurately depicting the fault development trend and outputting the life prediction value at the corresponding time step.

[0063] (3) The present invention can effectively avoid getting stuck in local optima by using the global search mechanism of the Dragonfly Optimization Algorithm (DA), effectively improve the prediction accuracy, accelerate the convergence process of the model, and significantly reduce the training time. By simulating the five behaviors of dragonflies in nature: separation, alignment, alliance, prey hunting and evasion of predators, it can effectively search for the optimal feature combination in the high-dimensional feature space and optimize the parameter configuration of BiLSTM to adapt to different data features, so that the BiLSTM network can more flexibly cope with data changes when processing complex time series data. Attached Figure Description

[0064] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] See Figure 1 A method for predicting the remaining life of bearings based on BN-RBM and DA-BiLSTM specifically includes the following steps:

[0067] (1) Collect the original vibration signal during the operation of the bearing and perform data preprocessing. The data preprocessing is to perform fast Fourier transform on the original vibration signal, convert the time domain signal into the frequency domain signal, obtain the frequency domain amplitude spectrum information, construct the bearing remaining life label, and thus construct the bearing dataset.

[0068] (2) Construct a bearing remaining life prediction model, which includes a BN-RBM feature extraction network, a DA-BiLSTM feature fusion network and a linear regression layer;

[0069] The S2 and BN-RBM feature extraction networks are used for feature extraction, specifically including the following steps:

[0070] S21. Construct a three-layer Bayesian BN network B=(F, D, R). The set of nodes representing the original degradation characteristics of bearings (including but not limited to frequency domain first octave amplitude, kurtosis, energy entropy, impact factor, envelope spectrum maximum value, wavelet packet energy entropy, and time domain statistics (mean, standard deviation, skewness, kurtosis)) is selected from historical degradation data. These degradation characteristics are highly sensitive to the remaining life of the bearing and can effectively reflect the degradation patterns of bearings such as lubrication failure and metal wear. The set of nodes representing the degradation stage , that is, the nodes in the degradation stage include the healthy stage , the mild degradation stage and the severe degradation stage ; R represents the set of nodes (continuous values or discrete intervals, such as "RUL>100h", "50h<RUL≤100h", etc.) of the remaining useful life (RUL) of the bearing;

[0071] The conditional dependencies between the nodes in the BN network are described by a conditional probability table, as shown in the following formula (1):

[0072] (1);

[0073] In formula (1), The role of is to encode the prior knowledge of the change of degradation characteristics, the degradation stage, and the remaining life, and it is obtained by training based on the historical degradation data of the bearing;

[0074] S₂₂, infer the posterior probability of the degradation stage: Given the degradation feature vector at the current time t , use the Bayesian BN network to infer the posterior probability of the degradation stage, as shown in the following formula (2):

[0075] (2);

[0076] In formula (2), represents the degradation feature vector extracted at the current time t; represents the th degradation stage; represents the prior probability of the degradation stage , which is obtained by statistically analyzing the historical degradation data of the bearing; represents the likelihood probability of observing the degradation feature under the degradation stage , which is obtained by querying the conditional probability table; represents the posterior probability of being in the degradation stage after given the current degradation feature ;

[0077] The output of the Bayesian BN network is the posterior probability distribution composed of the posterior probabilities of each degradation stage under the current degradation feature ; For example: ;

[0078] ;

[0079] S₂₃, construct the prior distribution of the weights of the restricted Boltzmann machine RBM, as shown in the following formula (3):

[0080] (3);

[0081] In equation (3), Represents the stage of degradation Lower, RBM weight The prior distribution is a Gaussian distribution. ; Represents the stage of degradation The corresponding weighted average; Represents the stage of degradation The corresponding weighted covariance;

[0082] S24. The posterior probability constructed using Bayesian BN is used to dynamically update the weight prior distribution of the Restricted Boltzmann Machine (RBM), resulting in the dynamic RBM weight prior distribution. , see the following formula (4) for details:

[0083] (4);

[0084] In equation (4), Represents the current degradation feature After that, it enters a degenerative stage. The posterior probability; This represents the updated dynamic weight mean at the current time t. This represents the dynamic weighted covariance at the current time t after the update;

[0085] S25. A Restricted Boltzmann Machine (RBM) trained using a dynamic RBM weight prior distribution is used to extract degenerate features from bearing data. The extracted feature vector is denoted as... , Represents the number of nodes in the hidden layer of the RBM; feature vector It is a high-order feature representation learned by an RBM network with the node set F of the original degradation features of the bearing as input. It reflects the deep potential feature structure of the current state of the bearing. Compared with the original degradation feature set F of the bearing, it has stronger discriminativeness and robustness, and is suitable for subsequent time series modeling and remaining life prediction tasks.

[0086] S26. Introduce a Dropout layer to randomly zero out the features extracted by the Restricted Boltzmann Machine (RBM). See equation (5) below for details:

[0087] (5);

[0088] In equation (5), Representing the The first sample The values ​​of the degenerate features after processing by the Dropout layer. Represents the first layer of the input Dropout layer The first sample A degenerate feature vector, This represents the set retention threshold, with a value between 0.5 and 0.8; scaling operations (divided by) This ensures consistency between the expected features during the training and inference phases, preventing the overall output from shrinking due to random neuron shutdowns.

[0089] The features extracted by the Restricted Boltzmann Machine (RBM) are randomly zeroed out to output a higher-order feature representation. , This represents a degenerate feature vector after being randomly set to zero. Represents all degenerate feature vectors A set;

[0090] Before inputting the data into the BiLSTM network, a sliding window mechanism is used to construct time-series samples: , The length of the sliding window. This represents the m-th time series sample; this process ensures a natural connection between feature extraction and temporal modeling, providing structured input for subsequent lifetime prediction;

[0091] The specific steps for bidirectional feature fusion using the S3 and DA-BiLSTM feature fusion networks are as follows:

[0092] The degenerate feature set output by the S31 and BN-RBM feature extraction networks is Organized into time series samples according to time sequence, the generated time series samples are shown in the following formula (6):

[0093] (6);

[0094] In equation (6), This represents the m-th time series sample; Represents the length of the sliding window; This represents the set of degradation features at time t in the m-th time series sample;

[0095] The sliding window length φ in equation (6) is determined by the Dragonfly Optimization Algorithm (DA), and the specific steps are as follows:

[0096] S311. Construct the configuration parameters of the BiLSTM network and combine them to form a vector. , Parameters in Represents the number of layers in a BiLSTM network. Represents the hidden layer size of BiLSTM Represents learning rate, Represents the Dropout ratio, Represents the length of the sliding window;

[0097] Each For an individual, defined within a preset range, n are randomly generated. At the same time, initialize its velocity vector. ;

[0098] S312. In each iteration, The parameters in the formula are applied to the BiLSTM network, trained and evaluated on the validation set, and the prediction mean square error (MSE) is calculated. The fitness value is set to the validation MSE. The group with the smallest and largest MSE in the current group is labeled, corresponding to food source and enemy, respectively.

[0099] S313, the Dragonfly Optimization Algorithm (DA) includes the following five types of behavior vectors:

[0100] Separation Position avoidance between individuals;

[0101] Alignment Speed ​​is averaged compared to neighbors;

[0102] Cohesion : Move closer to the center of the group;

[0103] feeding : To move closer to the optimal individual;

[0104] Avoid the enemy Stay away from the worst individuals;

[0105] And combined with inertia weight The speed and parameter configuration are dynamically updated, as detailed in equations (9) and (10):

[0106] (9);

[0107] (10);

[0108] In equations (9) and (10), The velocity vector at time t+1 The velocity vector at time t. This represents the configuration parameters of the BiLSTM network at time t. The weights represent the configuration parameters of the BiLSTM network at time t+1. and inertia weight During iteration, the process gradually transitions from exploration (diversity) to utilization (convergence), stopping when the maximum number of iterations is reached or the fitness improvement falls below a threshold; the output result is the optimal individual, which represents the optimal parameter configuration of the BiLSTM network.

[0109] S32. The BiLSTM network consists of multiple layers of BiLSTM structure. Each layer of BiLSTM structure includes a forward LSTM and a backward LSTM. The processing procedure of each layer of BiLSTM structure is shown in the following equation (7):

[0110] (7);

[0111] In equation (7), This represents the hidden state of the feedforward LSTM at time t. This represents the hidden state of the feedforward LSTM at time t-1. This represents the hidden state of the backward LSTM at time t. This represents the hidden state of the backward LSTM at time t+1. Represents a forward LSTM, Represents backward LSTM, This represents the bidirectional fusion characteristic at time t;

[0112] S33. A Dropout layer is introduced to randomly zero out the time-series samples processed by the BiLSTM network, randomly discarding some neuron activation values ​​to prevent overfitting. Global average pooling is then used to obtain a fixed vector representation, as shown in equation (8) below:

[0113] (8);

[0114] In equation (8), The Dropout layer is processed in the same way as formula (5). The fused features represent the output of the DA-BiLSTM feature fusion network;

[0115] The linear regression layer performs linear activation on the fusion characteristics of the bearing and outputs the predicted value of the remaining bearing life, as shown in the following formula (11):

[0116] (11);

[0117] In equation (11), The weight parameters represent the weights of the fully connected layer. Represents the bias of the fully connected layer. Represents the ReLU activation function. The predicted value representing the remaining life of the bearing;

[0118] (3) Construct a loss function to train the bearing remaining life prediction model. The loss function is shown in the following formula (12):

[0119] (12);

[0120] In equation (12), Representing the A sample of real bearing remaining life labels. Representing the Predicted remaining life of a sample bearing. Represents the regularization coefficient. The parameter set representing the bearing remaining life prediction model. Represents the total number of samples. Represents the square of the L1 norm;

[0121] Then, the trained bearing remaining life prediction model is used to predict the bearing remaining life and obtain the bearing remaining life.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A bearing residual life prediction method based on BN-RBM and DA-BiLSTM, characterized in that: Specifically comprising the following steps: (1) Collecting original vibration signals in the bearing running process and performing data preprocessing to construct a bearing dataset; (2) Constructing a bearing residual life prediction model, the bearing residual life prediction model comprising a BN-RBM feature extraction network, a DA-BiLSTM feature fusion network and a linear regression layer; the BN-RBM feature extraction network uses the posterior probability constructed by the Bayesian BN to dynamically update the weight prior distribution of the restricted Boltzmann machine RBM, to obtain a dynamic RBM weight prior distribution, and then uses the dynamic RBM weight prior distribution to guide the post-training restricted Boltzmann machine RBM to extract features of the degradation characteristics in the bearing data; the DA-BiLSTM feature fusion network is constructed by using the sliding window mechanism of the BiLSTM network to construct a time sequence sample, and performs bidirectional feature fusion on the extracted features, and the sliding window length is determined by the dragonfly optimization algorithm DA; the linear regression layer performs linear activation on the fusion features of the bearing, and outputs the predicted value of the bearing residual life; The specific steps of the DA-BiLSTM feature fusion network for bidirectional feature fusion are: S31, the degradation feature set output by the BN-RBM feature extraction network is organized in time sequence as time sequence samples, and the generated time sequence samples are specifically as follows (6): （6）； In formula (6), represents the mth time sequence sample; represents the sliding window length, which is determined by a dragonfly optimization algorithm DA optimization; represents the degradation feature set at time t in the mth time sequence sample; S32, the BiLSTM network comprises a plurality of BiLSTM structures, each BiLSTM structure comprises a forward LSTM and a backward LSTM, and the processing process of each BiLSTM structure is as shown in formula (7): （7）； In formula (7), denotes the hidden layer state of the forward LSTM at time t, denotes the hidden layer state of the forward LSTM at time t-1, denotes the hidden layer state of the backward LSTM at time t, denotes the hidden layer state of the backward LSTM at time t+1, denotes the forward LSTM, denotes the backward LSTM, denotes the bidirectional fusion feature at time t; S33, a Dropout layer is introduced to randomly zero-process the time sequence sample processed by the BiLSTM network, and a fixed vector representation is obtained using global average pooling, as shown in formula (8): （8）； In formula (8), represents a Dropout layer, and the processing process is the same as that in formula (5); represents the fusion feature output by the DA-BiLSTM feature fusion network. The sliding window length w is determined by the dragonfly optimization algorithm DA, and the specific steps are: S311, configuration parameters of the BiLSTM network are constructed, and are combined to form a vector , parameters in represent the number of layers of the BiLSTM network, represent the BiLSTM hidden layer size, represent the learning rate, represent the Dropout proportion, represent the sliding window length; Each For an individual, defined in the preset range, randomly generate n At the same time, initialize its speed vector ; S312. In each iteration, The parameters in the formula are applied to the BiLSTM network, trained and evaluated on the validation set, and the prediction mean square error (MSE) is calculated. The fitness value is set to the validation MSE. The group with the smallest and largest MSE in the current group is labeled, corresponding to food source and enemy, respectively. S313, the dragonfly optimization algorithm DA comprises the following five types of behavior vectors: separation : inter-individual position avoidance; Alignment : speed vs. neighbor average; coagulation : to gravitate toward the center of a group; taxis : approach to the fittest individual; avoiding enemies : moving away from the worst individual; combining the inertia weight , dynamically update the speed and parameter configuration, see the following formula (9) and formula (10): （9）； （10）； In formula (9) and formula (10), The velocity vector represents the t+1 moment, The velocity vector represents the t moment, The configuration parameter of the BiLSTM network at the t moment, The configuration parameter of the BiLSTM network at the t+1 moment; weight And the inertial weight Gradually transition from exploration to utilization in iterations, stop when reaching the maximum number of iterations or the fitness improvement is lower than the threshold value; the output result is the optimal individual, that is, the parameter configuration of the best BiLSTM network. (3) Constructing a loss function to train the bearing residual life prediction model, using the trained bearing residual life prediction model to predict the bearing residual life, and obtaining the residual life of the bearing.

2. The bearing residual life prediction method based on BN-RBM and DA-BiLSTM according to claim 1, characterized in that: The data preprocessing is a fast Fourier transform processing of the original vibration signal, which converts the time domain signal into a frequency domain signal and obtains frequency domain amplitude spectrum information, constructs a bearing residual life label, and thereby constructs a bearing dataset.

3. The bearing residual life prediction method based on BN-RBM and DA-BiLSTM according to claim 2, characterized in that: The specific steps of the BN-RBM feature extraction network for feature extraction are as follows: S21, constructing a three-layer Bayesian BN network B=(F, D, R), a node set representing original degradation characteristics of the bearing, the original degradation characteristics being screened from historical degradation data and reflecting a degradation mode of the bearing; a node set representing a degradation stage; and R represents a node set of a remaining life of the bearing. The conditional dependence between nodes in the BN network is described using a conditional probability table, as shown in formula (1): （1）； In formula (1), The role of is to encode the prior knowledge of degradation feature change, degradation stage and remaining life, which is trained based on the historical degradation data of the bearing. S22, infer the posterior probability of the degradation phase: given the degradation feature vector of the current time t The posterior probability of the degradation phase is inferred using the Bayesian BN network, as shown in the following equation (2): （2）； In formula (2), represents the degradation feature vector extracted at the current time t; represents the kth degradation stage; represents the degradation stage prior probability, obtained by statistical analysis of historical degradation data of the bearing; represents the degradation stage observed degradation feature likelihood probability, obtained by querying the conditional probability table; represents the posterior probability of being in the degradation stage after the given current degradation feature is observed. The output of the Bayesian BN network is the current degradation feature The posterior probability distribution is composed of the posterior probabilities of each degradation stage ; S23, constructing a weight prior distribution of the restricted Boltzmann machine RBM, as shown in formula (3): （3）； In formula (3), representing the degradation phase the prior distribution of the RBM weights is a Gaussian distribution ; representing the degradation phase the corresponding weight mean; representing the degradation phase the corresponding weight covariance; S24, the prior distribution of the weight of the restricted Boltzmann machine (RBM) is updated dynamically by using the posterior probability constructed by the Bayesian BN, to obtain a dynamic RBM weight prior distribution Specifically, see formula (4): （4）； In formula (4), representing a given current degradation feature after, in a degradation phase posterior probability; representing the updated dynamic weight mean at the current time t; representing the updated dynamic weight covariance at the current time t; S25, the restricted Boltzmann machine RBM guided by the dynamic RBM weight prior distribution is used to extract the degradation features in the bearing data, and the extracted feature vector is denoted as , the number of nodes representing the RBM hidden layer; S26, introducing a Dropout layer to randomly zero-process the features extracted by the restricted Boltzmann machine RBM, as shown in formula (5): （5）； In formula (5), represent the first sample the first degradation feature value after the Dropout layer processing, represent the first sample the first degradation feature vector input into the Dropout layer, represent the set retention threshold; high order feature representation output by the features extracted by the restricted boltzmann machine (RBM) after random zeroing processing , represents a degenerated feature vector after random zeroing processing, represents a set of all degenerated feature vectors .

4. The bearing residual life prediction method based on BN-RBM and DA-BiLSTM according to claim 1, characterized in that: The linear regression layer performs linear activation on the fusion features of the bearing, and outputs the predicted value of the bearing residual life, as shown in formula (11): （11）； In formula (11), weight parameters representing the fully connected layer; bias representing the fully connected layer, representing the activation function, representing the predicted value of the bearing remaining life.

5. The bearing residual life prediction method based on BN-RBM and DA-BiLSTM according to claim 4, characterized in that: The loss function is as shown in formula (12): （12）； In formula (12), represent the true bearing residual life label of the first sample, represent the predicted value of the bearing residual life of the first sample, represent the regularization coefficient, represent the parameter set of the bearing residual life prediction model, represent the total number of samples, represent the square of the L1 norm.

Images