Prediction method of remaining service life of bearings under migration condition based on contrast double confrontation mechanism

Abstract

The present invention discloses a method for predicting the remaining service life of a bearing under a migration condition based on a contrastive double-adversarial mechanism, which obtains a bearing vibration signal and a corresponding label of a single time step, decomposes the bearing vibration signal of a single time step, obtains the intrinsic mode IMF, reconstructs the bearing vibration signal of a single time step, and performs fault point detection; constructs a sample containing time series information; constructs a remaining service life prediction model, inputs the sample into the remaining service life prediction model, and outputs the remaining service life corresponding to the sample; the remaining service life prediction model includes a dynamic convolutional long short-term memory neural network feature extractor and a predictor; and constructs a migration training framework of a contrastive double-adversarial mechanism. The present invention combines multi-scale permutation entropy to perform noise reduction processing on the original vibration signal, and uses hidden Markov based on Bayesian reasoning to detect the fault occurrence point during the bearing degradation process. Through this technology, the influence of non-degraded data information on the remaining service life of the degradation process is reduced, and the workload and difficulty of the prediction task are reduced.

Description

Technical Field

[0001] The present invention relates to vibration signal analysis and engineering component life prediction technologies, and particularly to a method for predicting the remaining service life of bearings under migrating operating conditions based on a contrastive dual adversarial mechanism. Background Art

[0002] With the rapid development of industrial intelligence, mechanical industrial equipment has shown a trend of complexity, integration and intelligence, which has greatly increased the difficulty of mechanical equipment maintenance and had a huge impact on the safety and efficiency of equipment use. As a key component of rotating machinery, rolling bearings are widely used in core equipment in the fields of aerospace, transportation, wind power generation and shipping. Due to the complex and harsh working environment, bearings must maintain high precision and high load continuous operation and be able to withstand various impacts. Therefore, bearings are one of the most prone to failure parts of mechanical equipment. Once the bearing fails, it will often cause damage to the entire mechanical system over time, affecting the work efficiency of the equipment and increasing maintenance costs. The remaining service life prediction technology monitors and collects the operating data of the equipment, analyzes the collected signals, judges the status of the relevant parts, and predicts the remaining service life of the parts. Through this technology, the traditional passive or planned equipment maintenance plan can be converted into active maintenance and predictive maintenance, which will greatly reduce the maintenance cost of the equipment and improve maintenance efficiency. At present, the remaining service life prediction methods of bearings are mainly based on two types of methods: model-based and data-driven. Among them, the model-based method requires a lot of expert knowledge when establishing the prediction model, and the model is only applicable to a specific working environment, making this method difficult to be widely used; the data-driven remaining service life prediction technology of bearings is mainly based on deep learning methods. This method only needs to perform simple preprocessing on the collected historical monitoring data, and send the processed data to the deep learning model for training. Through the model, the deep features hidden in the historical data are mined, and the mapping relationship between the detection data and the remaining service life is constructed, so that an efficient bearing remaining service life prediction model can be trained. At present, the prediction models of deep learning mainly include CNN, RNN, LSTM, GRU, Transformer, TCN and other networks. These network models have made extraordinary achievements in the field of life prediction. At present, the main method for detecting the working state of bearings is to collect the vertical vibration data and horizontal vibration data of the bearings during the working process through acceleration sensors, and determine the working state and remaining service life of the bearings by analyzing the vibration data. However, in actual applications, the working environment of bearings is often complex and changeable, which makes it difficult for the trained deep learning prediction model to accurately predict the remaining service life of bearings under the actual working environment. In addition, the staged nature of the bearing degradation process and the complexity and variability of the vibration signals collected during the degradation process make it difficult for the model to accurately predict the remaining service life at different degradation stages. Summary of the invention

[0003] Purpose of the invention: The purpose of the present invention is to solve the deficiencies in the prior art and to provide a method for predicting the remaining useful life of bearings under migration conditions based on a comparative double-adversarial mechanism. The present invention determines the point of occurrence of bearing failure, i.e., the starting point of degradation, through an unsupervised detection method, and then trains the feature extraction module DTSConvLSTM through source domain data. The module fully extracts the degradation information contained in the bearing degradation process through a dynamic time scale, and introduces a migration training method of a double-adversarial mechanism. With the help of a small amount of target domain (i.e., actual working condition) data, the target domain prediction feature extraction model module DTSConvLSTM is subjected to comparative adversarial training, so that the model can extract domain-invariant features to improve the prediction model's ability to predict the remaining useful life of bearings under actual working conditions.

[0004] Technical solution: A method for predicting the remaining service life of a bearing under a migration condition based on a comparative double-antagonism mechanism of the present invention comprises the following steps:

[0005] Step S1: Collect and preprocess bearing vibration data to obtain the bearing vibration signal x of a single time step. t and the corresponding remaining useful life label Remaining useful life label The remaining service life of the corresponding time step is used to calculate the subsequent losses and the evaluation index of the prediction effect;

[0006] Step S2, using the adaptive noise method CEEMDAN to decompose the bearing vibration signal of a single time step obtained in step S1, obtain the intrinsic mode IMF, and reconstruct the bearing vibration signal of a single time step according to the permutation entropy size of each IMF; at the same time, using the hidden Markov model based on Bayesian reasoning to detect fault points;

[0007] Step S3, constructing samples containing time series information and dividing them into training sets, verification sets and test sets; using the data after the fault point detected in step S2, using the time window embedding method to reconstruct the bearing vibration signal of a single time step, splicing it according to a certain time step length, and obtaining a high-dimensional vector containing time series information, the high-dimensional vector is the feature of the training sample, and the label of the sample is the label corresponding to the last time step of the high-dimensional vector; at the same time, dividing the source domain bearing data into source domain training bearing data and source domain verification bearing data, and dividing the target domain bearing data into target domain training set bearing data and target domain test set bearing data;

[0008] Step S4, construct a remaining service life prediction model DTSConvLSTM-NN, input the sample in step S3 into the remaining service life prediction model DTSConvLSTM-NN, and output the remaining service life corresponding to the sample; the remaining service life prediction model includes a dynamic convolutional long short-term memory neural network feature extractor DTSConvLSTM and a predictor MLP; the hidden high-dimensional features of the sample are extracted by the feature extractor (for subsequent calculation of the remaining service life, and at the same time, in the contrast adversarial training stage, it is also used as the input of the discriminator and the contrast loss calculation module to participate in the calculation of the adversarial loss and the contrast loss); the predictor MLP includes a multi-layer full connection to obtain the predicted value; the fully connected layer is used for the H output of the above DTSConvLSTM t The hidden state is fully connected and the output value is the predicted value of the model;

[0009] Step S5: construct a transfer training framework of the contrast double adversarial mechanism, where the transfer training framework is divided into a pre-training stage and a contrast double adversarial training stage;

[0010] The specific method of the above-mentioned contrastive dual adversarial training stage is as follows: first, a target domain prediction model with the same structure as the pre-training stage is constructed and initialized with the pre-trained weights; then, the discriminators of the source domain and the target domain and a contrast loss calculation module are introduced to calculate the domain discrimination loss and the target domain information loss; finally, the target domain prediction model, the domain discriminator and the contrast loss calculation module are updated through the source domain data and the target domain data training set to complete the training of the target domain model;

[0011] Step S6: input the target domain test set data of step S3 into the target domain prediction model obtained in step S5. The output of the target domain prediction model is the predicted remaining useful life.

[0012] Furthermore, in step S1, for the bearing vibration signal x at time t t The acquisition method is to collect the vibration data of the bearing throughout its life cycle, and use the time window method to perform time stepping. The calculation formula is: t =n×tb×(t-1);

[0013] Where n is the number of bearing vibration signal data points in each time step;

[0014] The calculation formula for the corresponding label is as follows:

[0015]

[0016] Where RUL t is the current time step, RUL T is the total number of time steps over the entire life cycle of the bearing, It is the percentage of the remaining life of the current time step to the total life cycle.

[0017] Furthermore, the step S2 uses the adaptive noise method CEEMDAN to analyze the bearing vibration signal x at a single time step. t The specific method of decomposition is as follows:

[0018] First, the original single time step bearing vibration signal x t Add white noise w of different amplitudes i (t), generate different signal sets:

[0019] Where, ∈ is the noise amplitude, M is the number of times the noise is added, i represents the sequence number of the white noise, i = 1, 2, 3, ..., M;

[0020] Next, for each Perform empirical mode decomposition EMD() and calculate the mean to obtain the first eigenmode function

[0021] Then, calculate the residual signal r t 1 ,

[0022] In the residual signal Continue to repeat the above steps on the basis until the preset conditions are reached and the decomposition is completed;

[0023] The specific method of reconstructing the bearing vibration signal of a single time step in step S2 is as follows:

[0024] For a given eigenmode component Construct the embedding vector x according to the embedding dimension m and the time delay τ (that is, how many data points are taken to construct the permutation vector) e , denoted by x e ={x j ,x j+τ ,x j+2τ ,…,x j+(m-1)τ};

[0025] e=1,2,3,…; j=1,2,3,…,n-(m-1)τ;

[0026] By embedding the vector, we can get The order of arrangement of the corresponding elements is used to calculate the permutation entropy - PE of the intrinsic mode component through different arrangement orders. i , the calculation formula is as follows:

[0027]

[0028] In the formula, S p is the number of times the pth arrangement pattern appears, k is the total number of arrangement pattern types that appear in all arrangement patterns, M k It is the sum of the cumulative occurrences of all permutation patterns;

[0029] After obtaining the permutation entropies of all eigenmode components, the eigenmode components with smaller permutation entropies are selected for signal reconstruction.

[0030] Furthermore, the step S2 performs fault point detection based on a hidden Markov model of Bayesian reasoning, and the specific method is as follows:

[0031] (1) Extract the statistical features of the vibration signal at each reconstructed time step and construct a single time step feature vector s i =[F 1 ,F 2 ,F 3 ,…,F T ];

[0032] Where T is the number of statistical features; through the state of a single time step, the sequence set S = {s 1 ,s 2 ,…,s i};

[0033] (2) Initialize the hidden Markov model and set it to have two hidden states: normal stage and degradation stage; use the sequence S as observation data to train the parameters of the hidden Markov model;

[0034] (3) The parameters of the hidden state transition probability matrix and the observation probability are automatically estimated through the expectation maximization algorithm, and the sequence S is inferred using the trained hidden Markov model. The initial time step state of the sequence is assumed to be normal by default;

[0035] (4) Based on the given observation data, infer the posterior probability P(x t |s t ,S), when three consecutive time steps are judged as faulty state or 5 out of 10 consecutive time steps are faulty state, the bearing enters the fault stage by default.

[0036] Furthermore, in step S4, the operation of the sample entering the dynamic convolutional long short-term memory neural network feature extractor DTSConvLSTM is as follows:

[0037] Compared with the traditional convolutional long short-term memory neural network, DTSConvLSTM has multiple time scales and can automatically adjust the weights of information generated at different time scales to adapt to rapidly changing vibration signals;

[0038] First, the feature extractor generates feature information with different time scales through convolution kernels of different sizes in the gate and unit. The calculation formula is as follows:

[0039]

[0040]

[0041]

[0042]

[0043] In the formula, x t is the bearing vibration signal sample at the time step input at time t, H t-1 is the hidden state generated in the previous time step, is the weight matrix of the forget gate, is the forgotten information generated by the forget gate at time t, k = k 1 ,k 2 ,… is the size of the convolution kernel (i.e., time scale), * represents convolution calculation, σ is the activation function, is the input information generated by the input gate at time t, is the output information generated by the output gate at time t, is the temporary memory information generated at time t; k = k 1 ,k 2 ,…represents different convolution kernel sizes;

[0044] Next, after generating features with different time scales, each gate and unit adaptively adjusts the representation information generated at different scales through the selective attention mechanism, thereby generating the output information of the gate and unit; the specific formula is expressed as follows:

[0045] Information at different time scales is added element by element to generate mixed information:

[0046]

[0047] Embed f using global mean pooling u ,i u ,o u ,g u Global information of the channel is generated:

[0048]

[0049] In the formula, It is the mean pooling operation;

[0050] Then, for f c ,i c,o c ,g c Perform full connection calculation to generate guidance information z=[z f ,z i ,z o ,z g ];

[0051]

[0052] Where δ is the ReLU activation function, is batch normalization, W∈R 4×d×c ;

[0053] Under the guidance of the guidance information z, cross-information channel soft attention is used to adaptively select information of different time scales. The attention vector calculation formula is as follows:

[0054]

[0055] In the formula, A s ∈R 4×C×d is a learnable matrix; They are The attention weight vector of

[0056] S represents convolution kernels of different sizes, and each element of the attention weight corresponds to a feature channel one by one;

[0057] Next, we obtain the weighted residual information for each time scale:

[0058]

[0059]

[0060]

[0061]

[0062] In the formula, The residual information retained by information channels at different time scales;

[0063] Finally, the output n of the forget gate, input gate, output gate, and temporary memory unit is generated by element-by-element addition. t ,i t ,o t ,g t :

[0064]

[0065] Finally, we get the forget gate, input gate, output gate, and temporary memory, combined with the relevant information and the memory information C of the previous moment. t-1 , generate the memory information C of the current moment t With hidden state H t ; The calculation formula is as follows:

[0066] C t =f t ⊙C t-1 +i t ⊙g t ;

[0067] H t =sigmoid(o t )⊙tanh(C t );

[0068] In the formula, ⊙ represents element-by-element multiplication, and sigmoid and tanh are activation functions.

[0069] Furthermore, in step S4, the MLP obtains the predicted value; specifically, the fully connected layer obtains the H output of the above DTSConvLSTM t The hidden state is fully connected and the output value is the predicted value of the model.

[0070] Furthermore, in the pre-training stage of step S5, the source domain training bearing data obtained in step S3 is used to train the remaining service life prediction model DTSConvLSTM-NN constructed in step S4 to obtain a source domain prediction model; the specific operations are as follows:

[0071] First, the source domain training bearing data samples are input into the remaining service life prediction model DTSConvLSTM-NN in a certain batch size for forward reasoning. The final fully connected output is the predicted value of the sample.

[0072] Then, the model prediction output is calculated by the MSE loss function The lifespan label corresponding to the sample The loss between them, the MSE calculation formula is as follows:

[0073]

[0074] Where N is the batch size, is the model's predicted output value of the remaining useful life of the i-th sample, is the actual value of the remaining useful life of the i-th sample;

[0075] Through loss MSEUpdate the model until it achieves good prediction results on both the training set data and the test set data in the source domain, and save the pre-trained model.

[0076] Furthermore, the step S5 compares the dual adversarial training stage involving five parts: source domain prediction model, target domain prediction model, confusion domain discriminator D 1 , source domain discriminator D 2 And contrast loss calculation module C; the specific training process is as follows:

[0077] First, a target domain prediction model with the same structure as the source domain prediction model in the pre-training phase is constructed, and it is initialized using the weights saved from the pre-training phase to freeze the source domain prediction model.

[0078] Then, the confusion domain discriminator D is trained by the source domain data and the target domain data (training set). 1 , source domain discriminator D 2 And the contrast loss calculation module C updates the target domain feature extractor. The training principle is as follows:

[0079] Use the source domain prediction model feature extractor E s Extract source domain features from source domain samples, denoted as e s ; Use the target domain prediction model feature extractor E t At the same time, the source domain sample features and the target domain sample features are extracted and recorded as and e t ;

[0080] Obfuscation Domain Discriminator D 1 Acceptance Features and e t , and classify them. The loss calculation formula is as follows:

[0081]

[0082] Where N is the number of batch samples, y i is the true label of the i-th sample, is the probability of the discriminator identifying it as a positive class; by confusing the discriminator D 1 , so that the model can fully extract the domain-invariant features of the source domain and the target domain;

[0083] Domain Discriminator D 2 Accept feature e s , And classify the category to which it belongs, and use the cross entropy function to calculate the classification loss. The formula is as follows:

[0084]

[0085] Where N is the number of batch samples, yi is the true label of the i-th sample, The probability that the discriminator identifies the positive class is minimized by The target domain feature extractor ensures the stable extraction of common features in the source domain and avoids the loss of common features, thereby ensuring the effective use of source domain data by the model; when the domain discriminator D 2 ,When both are able to be accurately identified as source domain features, it is assumed that the target domain feature extraction module has effectively extracted the universal features of the source domain;

[0086] Contrastive loss calculation module C, accepting target domain feature extraction module E t The extracted target domain features t , and use the full connection operation to restore the original dimension of the feature input, and use the reconstruction error calculated by the mean square error MSE as the reconstruction loss. The formula is as follows:

[0087]

[0088] In the formula, N is the number of batch samples, n is the number of features of each sample, and x i,j is the true value of the jth feature of the ith sample, is the reconstructed value of the jth feature of the i-th sample;

[0089] Next, calculate the joint loss loss u :

[0090] In the formula, loss u is the discriminator D 1 , D 2 The joint loss of the contrast loss calculation module, α, β, and γ are the hyperparameter weights of the device, and the loss is minimized by gradient back propagation u , when the domain confusion discriminator D 1 Unable to distinguish s 、e t , the source domain discriminator can effectively identify e s , When all of them are source domain features and the contrast loss is extremely small, the model can simultaneously extract the invariant features of the source domain and the target domain, while avoiding the loss of common features of the source domain and specific information of the target domain;

[0091] After a certain number of iterations of the above training, a small amount of labeled target domain data is introduced to calculate the prediction loss of the target domain model for the target domain data. pre :

[0092]

[0093] Where N is the batch size, is the model's predicted output value of the remaining useful life of the i-th sample, is the actual value of the remaining useful life of the i-th sample;

[0094] Finally, freeze D 1 , D 2 Compare with the loss calculation module to calculate the total loss t :

[0095]

[0096] By minimizing the loss t Update the target domain feature extractor E t and the predictor pre t This is used to further calibrate the target domain model; repeat the above training steps multiple times, and stop training when the domain identification loss, contrast loss, and target domain test loss reach a certain threshold; save the model.

[0097] Furthermore, after completing step S6, step S7 lifelong learning is also included, that is, when facing unknown working condition samples, only a small amount of unknown working condition sample data with labels is used, the unknown working condition data is used as the target domain, the past historical data is used as the source domain, and the previous target domain prediction model is used as the source domain prediction model. The training process of step S5 is repeated to complete the prediction task of the unknown working condition.

[0098] Beneficial effects: Compared with the prior art, the present invention has the following technical effects:

[0099] 1. Combine multi-scale permutation entropy to reduce noise on the original vibration signal, and use hidden Markov based on Bayesian reasoning to detect fault occurrence points during bearing degradation. This technology can reduce the impact of non-degraded data information on the remaining service life of the degradation process, and reduce the workload and difficulty of prediction tasks.

[0100] 2. The present invention proposes a new dynamic time scale convolutional long short-term memory neural network (DTSC0nvLSTM-NN) that can achieve multi-scale representation of feature change information through convolution kernels of different sizes, describe the change information of signal features at different scales over time, and dynamically select memory information at different time scales. When faced with the rapidly changing vibration signals collected during the bearing degradation process, the model has a stronger ability to mine time series information, more fully captures the time dependency between samples, and increases the model's predictive performance and generalization ability.

[0101] 3. The present invention proposes a transfer learning method of contrasting double adversarial mechanism. The method predicts the remaining service life of bearings under cross-operating conditions through double adversarial mechanism and contrastive learning method. By contrasting the constraint of loss, the model can save the specific information of the target domain as much as possible while imitating the distribution of source domain data, and calibrate the prediction model through a small number of target domain samples. The method realizes the prediction of the remaining service life of bearings under cross-operating conditions under the condition of a small number of target domain samples.

[0102] 4. The present invention has the basic attribute of lifelong learning. Whenever facing unknown working conditions, the learning method can be transferred through the above-mentioned comparative double confrontation mechanism to improve the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 The overall flow chart of the prediction of the present invention.

[0104] Figure 2 This is the network structure of the feature extractor (DTSConvLSTM) in the present invention.

[0105] Figure 3 This is a schematic diagram of adversarial training of the present invention.

[0106] Figure 4 It is the prediction effect diagram of the embodiment. DETAILED DESCRIPTION

[0107] The technical solution of the present invention is described in detail below, but the protection scope of the present invention is not limited to the embodiments.

[0108] like Figure 1 As shown, the method for predicting the remaining service life of a bearing under a migration condition based on a comparative double antagonism mechanism of the present invention comprises the following steps:

[0109] Step S1: Collect and preprocess bearing vibration data to obtain the bearing vibration signal x of a single time step. t and the corresponding remaining useful life label

[0110] Step S2, decomposing the bearing vibration signal of a single time step obtained in step S1, obtaining the intrinsic mode IMF, and reconstructing the bearing vibration signal of a single time step according to the permutation entropy of each IMF; at the same time, using a hidden Markov model based on Bayesian reasoning to detect fault points;

[0111] Step S3, constructing a sample containing time series information, and at the same time, dividing the source domain bearing data into source domain training bearing data and source domain verification bearing data, and dividing the target domain bearing data into target domain training set bearing data and target domain test set bearing data;

[0112] Step S4, constructing a remaining useful life prediction model DTSConvLSTM-NN, inputting the sample in step 3 into the remaining useful life prediction model DTSConvLSTM-NN, and outputting the remaining useful life corresponding to the sample; the remaining useful life prediction model includes a dynamic convolutional long short-term memory neural network feature extractor DTSConvLSTM and a predictor MLP; the feature extractor extracts hidden high-dimensional features of the sample; the predictor MLP includes multiple fully connected layers to obtain the predicted value of the remaining useful life;

[0113] Step S5, constructing a transfer training framework of the contrastive double adversarial mechanism, the transfer training framework is divided into a pre-training stage and a contrastive double adversarial training stage; the specific method of the contrastive double adversarial training stage is: firstly constructing a target domain prediction model with the same structure as the pre-training stage, and initializing it with the pre-trained weights; then, introducing discriminators of the source domain and the target domain and a contrastive loss calculation module, and calculating the domain discrimination loss and the target domain information loss; finally, updating the target domain prediction model, the domain discriminator and the contrastive loss calculation module through the source domain data and the target domain data training set, and completing the training of the target domain model;

[0114] Step S6: input the target domain test set data of step S3 into the target domain prediction model obtained in step S5. The output of the target domain prediction model is the predicted remaining useful life.

[0115] In step S1 of this embodiment, for the bearing vibration signal x at time t t The acquisition method is to collect the vibration data of the bearing throughout its life cycle, and use the time window method to perform time stepping. The calculation formula is: t =n×tb×(t-1);

[0116] Where n is the number of bearing vibration signal data points in each time step;

[0117] The calculation formula for the corresponding label is as follows:

[0118]

[0119] Where RUL t is the current time step, RUL T is the time step of the entire life cycle of the bearing, In this embodiment, step S2 uses the adaptive noise method CEEMDAN to calculate the bearing vibration signal x at a single time step. t The specific method of decomposition is as follows:

[0120] First, the original single time step bearing vibration signal x t Add white noise w of different amplitudes i(t), generate different signal sets:

[0121] Among them, ∈ is the noise amplitude, M is the number of times the noise is added, i represents the sequence number of the white noise, i = 1, 2, 3, ..., M; then, for each Perform empirical mode decomposition and calculate the mean to obtain the first eigenmode function

[0122] Then, calculate the residual signal

[0123] In the residual signal Continue to repeat the above operations on the basis of until the preset conditions are met and the decomposition is completed;

[0124] The specific method of reconstructing the bearing vibration signal of a single time step in step S2 is as follows:

[0125] For a given eigenmode component Construct an embedding vector x based on the embedding dimension m and time delay τ e , the formula is: e ={x j ,x j+τ ,x j+2τ ,…,x j+(m-1)τ};

[0126] In the above formula, e=1,2,3,…;j=1,2,3,…,n-(m-1)τ;

[0127] By embedding the vector, we can get The order of arrangement of the corresponding elements is used to calculate the permutation entropy - PE of the intrinsic mode component through different arrangement orders. i , the calculation formula is as follows:

[0128]

[0129] In the formula, S p is the number of times the pth arrangement pattern appears, k is the total number of arrangement pattern types that appear in all arrangement patterns, M k It is the sum of the cumulative occurrences of all permutation patterns;

[0130] After obtaining the permutation entropies of all eigenmode components, the eigenmode components with smaller permutation entropies are selected for signal reconstruction.

[0131] In step S2 of this embodiment, fault point detection is performed based on a hidden Markov model based on Bayesian reasoning. The specific method is as follows:

[0132] (1) Extract the statistical features of the vibration signal at each reconstructed time step and construct a single time step feature vector s i =[F 1 ,F 2 ,F 3 ,…,F T ];

[0133] Where T is the number of statistical features; through the state of a single time step, the sequence set S = {s 1 ,s 2 ,…,s i};

[0134] (2) Initialize the hidden Markov model and set it to have two hidden states: normal stage and degradation stage; use the sequence S as observation data to train the parameters of the hidden Markov model;

[0135] (3) The parameters of the hidden state transition probability matrix and the observation probability are automatically estimated through the expectation maximization algorithm, and the sequence S is inferred using the trained hidden Markov model. The initial time step state of the sequence is assumed to be normal by default;

[0136] (4) Based on the given observation data, infer the posterior probability P(x t |s t ,S), when three consecutive time steps are judged as faulty state or 5 out of 10 consecutive time steps are faulty state, the bearing enters the fault stage by default.

[0137] Taking the dual-time-scale DTSConvLSTM feature extractor as an example, its structure diagram is as follows Figure 2 As shown, the operation of the sample entering the dynamic convolutional long short-term memory neural network feature extractor DTSConvLSTM in step S4 is as follows:

[0138] First, the feature extractor generates feature information with different time scales through convolution kernels of different sizes in the gate and unit. The calculation formula is as follows:

[0139]

[0140]

[0141]

[0142]

[0143] In the formula, x t is the bearing vibration signal sample at the time step input at time t, H t-1 is the hidden state generated in the previous time step, is the weight matrix of the forget gate, is the forgotten information generated by the forget gate at time t, k = k 1 ,k 2 is the size of the convolution kernel (i.e., time scale), * represents convolution calculation, σ is the activation function, is the input information generated by the input gate at time t, is the output information generated by the output gate at time t, is the temporary memory information generated at time t;

[0144] Then, each gate and unit adaptively adjusts the representation information generated at different scales through the selective attention mechanism to generate the output information of the gate and unit; the specific formula is as follows:

[0145] Information at different time scales is added element by element to generate mixed information:

[0146] Embed f using global mean pooling u ,i u ,o u ,g u Global information of the channel is generated:

[0147]

[0148] In the formula, It is the mean pooling operation;

[0149] Then, for f c ,i c ,o c ,g c Perform full connection calculation to generate guidance information z=[z f ,z i ,z o ,z g ];

[0150]

[0151] Where δ is the ReLU activation function, is batch normalization, W∈R 4×d×c ;

[0152] Under the guidance of the guidance information z, cross-information channel soft attention is used to adaptively select information of different time scales. The attention vector calculation formula is as follows:

[0153]

[0154] In the formula, A s ∈R 4×C×dis a learnable matrix; They are The attention weight vector of S represents convolution kernels of different sizes, and each element of the attention weight corresponds to a feature channel one by one.

[0155] Next, we obtain the weighted residual information for each time scale:

[0156]

[0157]

[0158]

[0159]

[0160] In the formula, The residual information retained by information channels at different time scales.

[0161] Finally, the output f of the forget gate, input gate, output gate, and temporary memory unit is generated by element-by-element addition. t ,i t ,o t ,g t ;

[0162]

[0163] Finally, we get the forget gate, input gate, output gate, and temporary memory, combined with the relevant information and the memory information C of the previous moment. t-1 , generate the memory information C of the current moment t With hidden state H t ; The calculation formula is as follows:

[0164] C t =f t ⊙C t-1 +i t ⊙g t ;

[0165] H t =sigmoid(o t )⊙tanh(C t );

[0166] In the formula, ⊙ represents element-by-element multiplication, and sigmoid and tanh are activation functions.

[0167] In order to obtain the hidden state output H at this time step t After that, use the full connection calculation to calculate, and the output value is the model's prediction value of the remaining service life of the bearing. The calculation formula is as follows:

[0168]

[0169] This is the remaining service life of the bearing at the current time step.

[0170] like Figure 3 As shown, in the pre-training stage of step S5, the source domain training bearing data obtained in step S3 is used to train the remaining service life prediction model DTSConvLSTM-NN constructed in step S4 to obtain a source domain prediction model; the specific operations are as follows:

[0171] First, the source domain training bearing data samples are input into the remaining service life prediction model DTSConvLSTM-NN in a certain batch size for forward reasoning. The final fully connected output is the predicted value of the sample.

[0172] Then, the model prediction output is calculated by the MSE loss function The lifespan label corresponding to the sample The loss between them, the MSE calculation formula is as follows:

[0173]

[0174] Where N is the batch size, is the model's predicted output value of the remaining useful life of the i-th sample, is the actual value of the remaining useful life of the i-th sample;

[0175] Through loss MSE Update the model until it achieves good prediction results on both the training set data and the test set data in the source domain, and save the pre-trained model.

[0176] The above step S5 compares the dual adversarial training phase, which involves five parts: source domain prediction model, target domain prediction model, confusion domain discriminator D 1 , source domain discriminator D 2 And contrast loss calculation module C; the specific training process is as follows:

[0177] First, a target domain prediction model with the same structure as the source domain prediction model in the pre-training phase is constructed, and it is initialized using the weights saved from the pre-training phase to freeze the source domain prediction model.

[0178] Then, the confusion domain discriminator D is trained by the source domain data and the target domain data. 1 , source domain discriminator D 2 And the contrast loss calculation module C updates the target domain feature extractor. The training principle is as follows:

[0179] Use the source domain prediction model feature extractor E s Extract source domain features from source domain samples, denoted as e s ; Use the target domain prediction model feature extractor E t At the same time, the source domain sample features and the target domain sample features are extracted and recorded as and e t ;

[0180] Obfuscated Domain Discriminator D 1 Acceptance Features and e t , and classify them. The loss calculation formula is as follows:

[0181]

[0182] Where N is the number of batch samples, y i is the true label of the i-th sample, is the probability that the discriminator identifies it as a positive class;

[0183] Source Domain Discriminator D 2 Accept feature e s , And classify the category to which it belongs, and use the cross entropy function to calculate the classification loss. The formula is as follows:

[0184]

[0185] Where N is the number of batch samples, y i is the true label of the i-th sample, The probability that the discriminator identifies the positive class

[0186] Contrastive loss calculation module C, accepting target domain prediction model feature extractor E t The extracted target domain features t , and use the full connection operation to restore the original dimension of the feature input, and use the reconstruction error calculated by the mean square error MSE as the reconstruction loss. The formula is as follows:

[0187]

[0188] In the formula, N is the number of batch samples, n is the number of features of each sample, and x i,j is the true value of the jth feature of the ith sample, is the reconstructed value of the jth feature of the i-th sample;

[0189] Next, calculate the joint loss loss u :

[0190] In the formula, loss u is the discriminator D1 , D 2 The joint loss with the contrast loss calculation module, α, β, and γ are the device hyperparameter weights respectively;

[0191] After a certain number of iterations of the above training, a small amount of labeled target domain data is introduced to calculate the prediction loss of the target domain model for the target domain data. pre :

[0192]

[0193] Where N is the batch size, is the model's predicted output value of the remaining useful life of the i-th sample, is the actual value of the remaining useful life of the i-th sample;

[0194] Finally, freeze D 1 , D 2 Compare with the loss calculation module to calculate the total loss t :

[0195]

[0196] By minimizing Loss t Update the target domain feature extractor E t and the predictor pre t .

[0197] In this embodiment, step S6, the target domain test set data of step S3 is input into the target domain prediction model obtained in step S5, and the output of the target domain prediction model is the predicted remaining service life; Figure 4 The model uses the data of working condition 1 as the source domain data. The bearings 2_1 and 2_2 in the working condition 2 data are used as the target domain training set, and the prediction effect of the target domain test set bearings 2-6 is obtained.

[0198] The main evaluation indicators include: MAE, RMSE, and Score.

[0199] The absolute mean error MAE is mainly used to measure the overall prediction performance of the model. The calculation formula is as follows:

[0200]

[0201] In the formula, represents the true remaining useful life of the sample, It represents the remaining useful life predicted by the model, and n is the number of samples.

[0202] The root mean square error RMSE is more sensitive than MAE and can better measure the stability of the prediction model. The calculation formula is as follows:

[0203]

[0204] In the formula, represents the true remaining useful life of the sample, It represents the remaining useful life predicted by the model, and n is the number of samples.

[0205] The prediction score Score gives different weights to the advance prediction and the lagged prediction to measure the prediction effect of the model. The specific calculation formula is as follows:

[0206]

[0207]

[0208]

[0209] Where, %Er i is the prediction error of the ith sample, A i is the prediction score of the sample, and Score is the overall score of the model's prediction of the remaining service life of the bearing.

[0210] Completing step S6 also includes step S7 to introduce a lifelong learning model. The specific method is: when facing unknown working condition samples, only a small amount of unknown working condition sample data with labels is used, the unknown working condition data is used as the target domain, the previous historical data is used as the source domain, and the previous target domain prediction model is used as the source domain prediction model. Repeating the training process of step S5 can complete the prediction task of the unknown working condition.

Claims

1. A method for predicting the remaining service life of a bearing under migration conditions based on a comparative double-antagonism mechanism, characterized in that: The following steps are involved: Step S1: Collect and preprocess bearing vibration data to obtain the bearing vibration signal x of a single time step. t and the corresponding remaining useful life label Step S2, decomposing the bearing vibration signal of a single time step obtained in step S1, obtaining the intrinsic mode IMF, and reconstructing the bearing vibration signal of a single time step according to the permutation entropy of each IMF; at the same time, using a hidden Markov model based on Bayesian reasoning to detect fault points; Step S3, constructing a sample containing time series information, and at the same time, dividing the source domain bearing data into source domain training bearing data and source domain verification bearing data, and dividing the target domain bearing data into target domain training set bearing data and target domain test set bearing data; Step S4, constructing a remaining service life prediction model DTSConvLSTM-NN, inputting the sample in step S3 into the remaining service life prediction model DTSConvLSTM-NN, and outputting the remaining service life corresponding to the sample; the remaining service life prediction model includes a dynamic convolutional long short-term memory neural network feature extractor DTSConvLSTM and a predictor MLP; the hidden high-dimensional features of the sample are extracted by the feature extractor; the predictor MLP includes a multi-layer fully connected layer to obtain the predicted value of the remaining service life; Step S5, constructing a transfer training framework of the contrastive double adversarial mechanism, the transfer training framework is divided into a pre-training stage and a contrastive double adversarial training stage; the specific method of the contrastive double adversarial training stage is: firstly constructing a target domain prediction model with the same structure as the pre-training stage, and initializing it with the pre-trained weights; then, introducing discriminators of the source domain and the target domain and a contrastive loss calculation module, and calculating the domain discrimination loss and the target domain information loss; finally, updating the target domain prediction model, the domain discriminator and the contrastive loss calculation module through the source domain data and the target domain data training set, and completing the training of the target domain model; Step S6: input the target domain test set data of step S3 into the target domain prediction model obtained in step S5. The output of the target domain prediction model is the predicted remaining useful life.

2. The method for predicting the remaining service life of a bearing under a migration condition based on a comparison double confrontation mechanism according to claim 1 is characterized in that: In step S1, for the bearing vibration signal x at time t t The acquisition method is to collect the vibration data of the bearing throughout its life cycle, and use the time window method to perform time stepping. The calculation formula is: t =n×tn×(t-1); Where n is the number of bearing vibration signal data points in each time step; The calculation formula for the corresponding label is as follows: Where RUL t is the current time step, RUL T is the total number of time steps over the entire life cycle of the bearing, It is the percentage of the remaining life of the current time step to the total life cycle.

3. The method for predicting the remaining service life of a bearing under a migration condition based on a comparison double confrontation mechanism according to claim 1 is characterized in that: The step S2 uses the adaptive noise method CEEMDAN to analyze the bearing vibration signal x at a single time step. t The specific method of decomposition is as follows: First, the original single time step bearing vibration signal x t Add white noise w of different amplitudes i (t), generating different signal sets Where, ∈ is the noise amplitude, M is the number of times the noise is added, i represents the sequence number of the white noise, i = 1, 2, 3, ..., M; Next, for each Perform the empirical mode decomposition (EMD) operation and calculate the mean to obtain the first intrinsic mode function. Then, calculate the residual signal In the residual signal Continue to repeat the above steps on the basis until the preset conditions are reached and the decomposition is completed; The specific method of reconstructing the bearing vibration signal of a single time step in step S2 is as follows: For a given eigenmode component Construct an embedding vector x based on the embedding dimension m and time delay τ e , the vector is represented as: x e ={x j ,x j++ ,x j+2τ ,…,x j+(m-1)τ }; e represents the number of decomposed IMFs, e = 1, 2, 3, ...; j = 1, 2, 3, ..., n-(m-1)τ; n is the number of bearing vibration signal data points in each time step; By embedding the vector, we can get The order of arrangement of the corresponding elements is used to calculate the permutation entropy - PE of the intrinsic mode component through different arrangement orders. i , the calculation formula is as follows: In the formula, S p is the number of times the pth arrangement pattern appears, k is the total number of arrangement pattern types that appear in all arrangement patterns, M k It is the sum of the cumulative occurrences of all permutation patterns; After obtaining the permutation entropies of all eigenmode components, the eigenmode components with smaller permutation entropies are selected for signal reconstruction.

4. The method for predicting the remaining service life of a bearing under a migration condition based on a comparison double confrontation mechanism according to claim 1 is characterized in that: The step S2 performs fault point detection based on the hidden Markov model of Bayesian reasoning, and the specific method is: (1) Extract the statistical features of the vibration signal at each reconstructed time step and construct a single time step feature vector s i =[F1,F2,F3,…,F T ]; Where T is the number of statistical features; through the state of a single time step, the sequence set S = {s1, s2, ..., s i }; (2) Initialize the hidden Markov model and set it to have two hidden states: normal stage and degradation stage; use sequence S as observation data to train the parameters of the hidden Markov model; (3) The parameters of the hidden state transition probability matrix and the observation probability are automatically estimated through the expectation maximization algorithm, and the sequence S is inferred using the trained hidden Markov model. The initial time step state of the sequence is assumed to be normal by default; (4) Based on the given observation data, infer the posterior probability P(x t |s t ,S), when three consecutive time steps are judged as faulty state or 5 out of 10 consecutive time steps are faulty state, the bearing enters the degradation stage by default.

5. According to the method for predicting the remaining service life of a bearing under a migration condition based on a comparison double confrontation mechanism according to claim 1, the operation of the sample entering the dynamic convolutional long short-term memory neural network feature extractor DTSConvLSTM in step S4 is as follows: First, the feature extractor generates feature information with different time scales through convolution kernels of different sizes in the gate and unit. The calculation formula is as follows: In the formula, x t is the bearing vibration signal sample at the time step input at time t, H t-1 is the hidden state generated in the previous time step, is the weight matrix of the forget gate, is the forget information generated by the forget gate at time t, k=k1,k2,… is the size of the convolution kernel, * represents the convolution calculation, σ is the activation function, is the input information generated by the input gate at time t, is the output information generated by the output gate at time t, is the temporary memory information generated at time t; k = k1, k2, ... represents different convolution kernel sizes; Then, each gate and unit adaptively adjusts the representation information generated at different scales through the selective attention mechanism to generate the output information of the gate and unit; the specific formula is as follows: Information at different time scales is added element by element to generate mixed information: Embed f using global mean pooling u ,i u ,o u ,g u Global information of the channel is generated: In the formula, It is the mean pooling operation; Then, for fc c ,i c ,o c ,g c Perform full connection calculation to generate guidance information z=[z f ,z i ,z o ,z g ]; Where δ is the ReLU activation function, is batch normalization, W∈R 4×d×c ; Under the guidance of the guidance information z, cross-information channel soft attention is used to adaptively select information of different time scales. The attention vector calculation formula is as follows: In the formula, A s ∈R 4×C×d is a learnable matrix; They are The attention weight vector of S represents convolution kernels of different sizes, and each element of the attention weight corresponds to a feature channel one by one; Next, we obtain the weighted residual information for each time scale: In the formula, The residual information retained by information channels at different time scales; Finally, the output f of the forget gate, input gate, output gate, and temporary memory unit is generated by element-by-element addition. t ,f t ,o t ,g t ; Finally, we get the forget gate, input gate, output gate, and temporary memory, combined with the relevant information and the memory information C of the previous moment. t-1 , generate the memory information C of the current moment t With hidden state H t ; The calculation formula is as follows: C t =f t ⊙C t-1 +i t ⊙g t ; H t =sigmoid(o t )⊙tanh(C t ); In the formula, ⊙ represents element-by-element multiplication, sigmoid and tanh are activation functions; Finally, the MLP fully connected network in step S4 is used to obtain the predicted value, that is, the H output by the above DTSConvLSTM is connected through the fully connected layer. t The hidden state is fully connected and the output value is the predicted value of the model.

6. According to the method for predicting the remaining service life of a bearing under a migration condition based on a contrast dual confrontation mechanism according to claim 1, in the pre-training stage of step S5, the remaining service life prediction model DTSConvLSTM-NN constructed in step S4 is trained using the source domain training bearing data obtained in step S3 to obtain a source domain prediction model; the specific operations are as follows: First, the source domain training bearing data samples are input into the remaining service life prediction model DTSConvLSTM-NN in a certain batch size for forward reasoning. The final fully connected output is the predicted value of the sample. Then, the model prediction output is calculated by the MSE loss function The lifespan label corresponding to the sample The loss between them, the MSE calculation formula is as follows: Where N is the batch size, is the model's predicted output value of the remaining useful life of the i-th sample, is the actual value of the remaining useful life of the i-th sample; Through loss MSE Update the model until it achieves good prediction results on both the training set data and the test set data in the source domain, and save the pre-trained model.

7. According to the method for predicting the remaining service life of a bearing under a migration condition based on a contrastive double-adversarial mechanism in claim 1, the step S5 of the contrastive double-adversarial training phase involves five parts: a source domain prediction model, a target domain prediction model, a confusion domain discriminator D1, a source domain discriminator D2, and a contrastive loss calculation module C; the specific training process is as follows: First, a target domain prediction model with the same structure as the source domain prediction model in the pre-training phase is constructed, and it is initialized using the weights saved from the pre-training phase to freeze the source domain prediction model. Then, the confusion domain discriminator D1, the source domain discriminator D2 and the contrast loss calculation module C are trained through the source domain data and the target domain data to update the target domain feature extractor. The training principle is as follows: Use the source domain prediction model feature extractor E s Extract source domain features from source domain samples, denoted as e s ; Use the target domain prediction model feature extractor E t At the same time, the source domain sample features and the target domain sample features are extracted and recorded as and e t ; Confusion domain discriminator D1 accepts features and e t , and classify them. The loss calculation formula is as follows: Where N is the number of batch samples, y i is the true label of the i-th sample, is the probability that the discriminator identifies it as a positive class; The source domain discriminator D2 accepts feature e s , And classify the category to which it belongs, and use the cross entropy function to calculate the classification loss. The formula is as follows: Where N is the number of batch samples, y i is the true label of the i-th sample, is the probability that the discriminator identifies it as a positive class; Contrastive loss calculation module C, accepting target domain prediction model feature extractor E t The extracted target domain features t , and use the full connection operation to restore the original dimension of the feature input, and use the reconstruction error calculated by the mean square error MSE as the reconstruction loss. The formula is as follows: In the formula, N is the number of batch samples, n is the number of features of each sample, and x i,j is the true value of the jth feature of the ith sample, is the reconstructed value of the jth feature of the i-th sample; Next, calculate the joint loss loss u : In the formula, loss u is the joint loss of the discriminator D1, D2 and the contrast loss calculation module, α, β, γ are their hyper-parameter weights respectively; After a certain number of iterations of the above training, a small amount of labeled target domain data is introduced to calculate the prediction loss of the target domain model for the target domain data. pre : Where N is the batch size, is the model's predicted output value of the remaining useful life of the i-th sample, is the actual value of the remaining useful life of the i-th sample; Finally, freeze D1, D2 and the contrast loss calculation module to calculate the total loss t : By minimizing the loss t Update the target domain feature extractor E t and predictors.

8. According to the method for predicting the remaining service life of a migration condition bearing based on a comparative double-antagonism mechanism according to claim 1, after completing step S6, it also includes step S7 introducing a lifelong learning mode. The specific method is: when facing unknown working condition samples, only a small amount of unknown working condition sample data with labels is used, and the unknown working condition data is used as the target domain, the previous historical data is used as the source domain, and the previous target domain prediction model is used as the source domain prediction model. The training process of step S5 is repeated to complete the prediction task of the unknown working condition.

Images