Bearing rul prediction method based on time-frequency domain features and weighted domain adversarial learning

Abstract

The application relates to a bearing RUL prediction method based on time-frequency domain features and a weighted domain adversarial learning, comprising a method for training a bearing residual life prediction model; a domain self-adaptive transfer learning training method, the domain self-adaptive transfer learning training method comprising the following steps; a target domain RUL prediction model is constructed based on a feature extractor and an RUL predictor, and the RUL prediction model further comprises a domain discriminator; a source domain sample input tensor and a target domain sample input tensor are input into the target domain RUL prediction model for training, and the training process comprises a preheating learning stage and an adversarial learning stage; after the training is completed, a target domain to-be-detected sample input tensor is input into the target domain bearing RUL prediction model, and the residual life of the bearing is predicted. The bearing RUL prediction method based on the time-frequency domain features and the weighted domain adversarial learning solves the problems of insufficient feature representation, weak cross-domain generalization ability, low adversarial training efficiency and slow training convergence in the prior art, and realizes accurate prediction of the bearing RUL.

Description

Technical Field

[0001] This invention relates to a method for predicting the remaining service life of bearings, and more particularly to a bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning. Background Technology

[0002] As a core component of rotating machinery, the operating status of bearings directly determines the reliability and safety of the equipment. Accurate prediction of the remaining service life (RUL) of bearings is one of the core tasks of predictive health management (PHM), which can effectively avoid unplanned downtime and reduce equipment maintenance costs.

[0003] Existing bearing RUL prediction technologies suffer from the following shortcomings: First, traditional methods rely heavily on shallow features of one-dimensional time-series vibration signals, making it difficult to uncover deep correlations in the time-frequency joint domain of the signal. This limited feature representation capability leads to insufficient prediction accuracy. Second, most models are trained for a single equipment domain, such as bearings of a single model or operating condition. When faced with cross-equipment domains, i.e., different models or operating conditions, the data distribution in the target domain deviates significantly from the source domain, causing a sharp decline in model generalization ability. Third, existing domain adaptation methods often employ uniform weight adversarial training, failing to consider the difference in prediction errors between source and target domain samples. This makes it difficult to specifically optimize the feature alignment of easily confused samples, resulting in low adversarial training efficiency. Fourth, model training lacks a standardized pre-training-cross-domain adaptation process. Initial weight randomization leads to slow convergence and a tendency to get trapped in local optima during cross-domain training.

[0004] Therefore, there is a need for a bearing RUL prediction method that can fully exploit deep time-frequency domain features, efficiently mitigate domain offset, and balance training efficiency and prediction accuracy, so as to meet the RUL prediction needs across equipment domains in actual industrial scenarios. Summary of the Invention

[0005] To address the technical problems of low model generalization ability, randomized initial weights, and low training efficiency, this invention provides a bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning, which features strong cross-domain generalization ability, high adversarial training efficiency, fast training convergence, and accurate remaining lifetime prediction.

[0006] The present invention provides a bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning, including a method for training a bearing remaining life prediction model; and a step of using the remaining life prediction model to predict the remaining life of the bearing under test. Methods for training bearing remaining life prediction models include source domain model pre-training methods, which include the following steps; Collect vibration time-series signals of source domain samples and construct the source domain sample input tensor; The source domain sample input tensor is input into the source domain RUL prediction model to perform supervised pre-training, resulting in a converged source domain RUL prediction model. The RUL prediction model includes a feature extractor and a RUL predictor; The method for training the bearing remaining life prediction model also includes the domain adaptive transfer learning training method, which includes the following steps; A target domain RUL prediction model is constructed based on the feature extractor and RUL predictor of the converged source domain RUL prediction model. The target domain RUL prediction model also includes a domain discriminator. Collect vibration time-series signals of samples in the target domain and construct the input tensor of the samples in the target domain; The source domain sample input tensor and the target domain sample input tensor are input into the target domain RUL prediction model for training. The training process includes a warm-up learning stage and a domain adversarial learning stage, with the domain adversarial learning stage following the warm-up learning stage. During the warm-up learning phase, the feature extractor of the target domain bearing RUL prediction model is fixed, and the RUL predictor is trained on the basis. During the domain adversarial learning phase, the domain discriminator and feature extractor are trained alternately. During the adversarial learning phase, when training the domain discriminator, the feature extractor is fixed, and the source domain features and target domain features with real domain labels are used to train the domain discriminator and update the model parameters of the domain discriminator. During the adversarial learning phase, when training the feature extractor, the domain discriminator is fixed, and the feature extractor is trained using source domain features with true domain labels and target domain features with incorrect domain labels. The model parameters of the feature extractor and the RUL predictor are then updated. During training, the target domain RUL prediction model is evaluated using the target domain test set at predetermined intervals, and the evaluation value is obtained. The model training ends when the evaluation value is less than the predetermined value. After training, the test samples in the target domain are input into the tensor and then into the target domain RUL prediction model to predict the remaining life of the bearing.

[0007] The advantage of the bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning is that it pre-trains the source domain RUL prediction model before training the target domain RUL prediction model, thereby obtaining a converged source domain RUL prediction model. This provides initial weight parameters for the feature extractor and RUL predictor of the target domain RUL prediction model. When the system performs cross-domain detection of bearings in the target domain, it does not need to re-pre-train the model, avoiding the training inefficiency caused by random weights. This greatly improves the training efficiency of the target domain RUL prediction model and is suitable for the rapid processing of large-scale industrial data. Meanwhile, in the domain adaptive transfer learning training method, during the adversarial learning stage, the method adopts the step of alternating training of the domain discriminator and the feature extractor. When training the domain discriminator, source domain features and target domain features with real domain labels are used to train the domain discriminator, so that the domain discriminator can accurately identify the domain to which the sample belongs. When training the feature extractor, the labels of the target domain features are flipped to confuse the discriminator, thereby improving the feature extraction effect of the feature extractor on the alignment of sample features of source domain samples and target domain samples, improving the model's cross-domain generalization ability and the accuracy of RUL prediction, which is suitable for the actual needs of industrial scenarios and has good engineering application value.

[0008] The vibration time series signal of the sample may include the lateral and / or longitudinal vibration signal of the bearing, and the source domain sample input tensor and the target domain sample input tensor are tensors transformed from the original vibration time series signal.

[0009] The source domain features and target domain features mentioned above are the outputs of the source domain sample input tensor and the target domain sample input tensor after passing through the feature extractor, respectively.

[0010] Furthermore, in the bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning of the present invention, both the source domain sample input tensor and the target domain sample input tensor are time-frequency feature map tensors formed by converting sample vibration time-series signals.

[0011] This scheme uses time-frequency feature map tensors as model inputs. Compared with traditional time-series signals, time-frequency feature maps have strong feature representation capabilities. The system can divide the time-frequency feature map into sub-bands and extract features such as energy entropy, multi-scale energy accumulation, and feature time-frequency entropy from the sub-bands. At the same time, the feature extractor network based on deep time-frequency residuals can fully explore the deep energy features in the joint time-frequency domain. Compared with traditional one-dimensional time-series features, it can effectively improve the feature support capability of RUL prediction.

[0012] During training, the system can collect bidirectional vibration time-series signals of rolling bearings under different working conditions, and generate normalized RUL labels through time-series reversal, dimension reshaping and normalization processing. It then divides and constructs the original source domain dataset and the original target domain dataset, and uses steps such as synchronous squeezing wavelet transform to convert the original source domain dataset and the original target domain dataset into the corresponding time-frequency feature map tensor.

[0013] Furthermore, in the bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning of the present invention, during the domain adversarial learning stage, the domain discriminator loss L is obtained using the following formula. D ; ; When training the feature extractor, the feature extractor loss is obtained using the following formula. ; Among them, L WAL This is called weighted adversarial loss, w k For the first The weight coefficients of each sample, For smoothing coefficients, For the domain discriminator to the first The predicted probability of each sample; For the first The domain label of each sample, L BCE (p k ,d k ) represents the binary cross-entropy loss, L RUL For the remaining lifetime prediction loss of the source domain samples, N s N is the number of samples in the source domain. t The number of samples in the target domain; Among them, w k Obtained by the following formula; ; Where exp is the natural exponential function. For the sample predicted value; For source domain sample data, y s For the source domain sample data, y represents the true label; for the target domain sample data, y represents the true label. s This is the global average lifetime label for the source domain.

[0014] This scheme introduces a dynamic weighted domain adversarial learning mechanism, which dynamically assigns weighting coefficients to the loss of the domain discriminator based on the sample prediction error, and specifically optimizes the feature alignment effect of easily distinguishable samples, effectively alleviating the distribution offset between the source domain and the target domain, and improving the prediction accuracy of the prediction model across device domains.

[0015] Furthermore, the bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning of the present invention includes a feature extractor comprising a stacked two-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer and a max pooling layer, several cascaded Bottleneck residual modules, an adaptive average pooling layer, a flattening layer and a bottleneck layer.

[0016] Several Bottleneck residual modules are stacked in series, leveraging the bottleneck convolutional structure to achieve multi-dimensional feature refinement and deep semantic mining with low parameter count and low computational cost. Through multi-level progressive feature extraction, the local details, degradation patterns, and implicit correlations of time-frequency images are fully explored, enhancing the model's ability to represent complex nonlinearities and improving the network's accuracy in identifying features and degradation states of different bearing domains. The bottleneck layer at the end of the network outputs the feature vectors.

[0017] Furthermore, in the bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning of the present invention, the Bottleneck residual module includes a 1×1 convolutional layer, a 3×3 convolutional layer, and a 1×1 convolutional layer connected in series, and the input and output of the Bottleneck residual module are connected in a skip connection.

[0018] A series of 1×1, 3×3, and 1×1 convolutional layers are sequentially connected to form the Bottleneck residual module, which has a bottleneck structure. This reduces the number of model parameters and computational overhead while ensuring the ability to extract deep features. The skip connections between input and output constitute the residual mapping structure of the module, alleviating the gradient vanishing problem in deep networks and ensuring the training stability and feature representation performance of deep models.

[0019] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the following describes the embodiments of the present invention in detail. Attached Figure Description

[0020] Figure 1 This is a network structure diagram of the feature extractor.

[0021] Figure 2 This is a schematic diagram of the framework of the target domain RUL prediction model.

[0022] Figure 3 This is a time-series comparison diagram of the actual RUL and the predicted RUL in the embodiment. Detailed Implementation

[0023] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0024] See Figures 1 to 3 The bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning in this embodiment includes the following steps.

[0025] S1. Data Acquisition and Preprocessing 1.1 Acquire the horizontal and vertical vibration timing signals of the rolling bearing under different operating conditions (OC). The acquired data includes the full life cycle vibration signals of the bearing under three different operating conditions OC1, OC2, and OC3, and includes vibration data from both horizontal and vertical channels. The specific operating parameters for the three operating conditions are shown in the table below.

[0026] OC1 1800 4000 Bearing1_1~Bearing1_7 1800~2803 OC2 1650 4200 Bearing2_1~Bearing2_7 911~1708 OC3 1500 5000 Bearing3_1~Bearing3_3 1000~2152

[0027] 1.2 Preprocess the vibration time series data, including time series reversal, dimension reshaping, and normalization.

[0028] The timing reversal makes the bearing health status change from "healthy → fault" to "0 → N" in time step, in order to match the decreasing logic of RUL; Dimensional reshaping transforms time-series data into a standardized format of [number of samples, time step, 2]. Normalization employs min-max normalization to eliminate the influence of dimensions.

[0029] 1.3 Generate standardized RUL labels and perform normalization processing.

[0030] The normalized remaining service life label for each bearing sample is obtained using the following formula to adapt to regression task training.

[0031] ;

[0032] in, For the first The actual remaining lifetime of each sample; This refers to the total lifespan of the bearing from the start of operation to failure. The normalized RUL tag has a value range of 1. .

[0033] 1.4 Construct the original source domain and target domain datasets, dividing the data into the original source domain dataset. and the original target domain dataset .in, For the source domain Vibration signal of each sample For time step, Number of channels; For the source domain Normalized remaining useful life label corresponding to each sample; For the target domain Vibration signals of individual samples; This represents the total number of samples in the source domain. This represents the total number of samples in the target domain. The original source domain dataset consists of labeled data with sufficient RUL labels. The original target domain dataset consists of sparsely labeled or unlabeled data, with each device domain containing vibration signals from at least one bearing.

[0034] Step S2: Generate time-frequency feature maps of the source and target domains. 2.1 The preprocessed one-dimensional vibration time-series signal is converted into a two-dimensional time-frequency domain feature map matrix using Synchronous Squeezed Wavelet Transform (SSQ-CWT). This allows for the full exploitation of deep features in the joint time and frequency domains. The specific formula for SSQ-CWT is as follows: ;

[0035] in, These are the wavelet coefficients after synchronous compression. For scale parameters, These are translation parameters; These are the original continuous wavelet transform coefficients; This is an estimate of the instantaneous frequency. This is the Dirac function.

[0036] For a dual-channel signal, the time-frequency plots of the two channels are... and Take the absolute value, concatenate along the channel dimension, and unify the size. The concatenation transformation process is represented as: ;

[0037] in, The final input is the time-frequency image tensor; To obtain the modulus of the complex wavelet coefficients; This involves stitching together the time-frequency feature maps of two channels along the channel dimension. This is a bilinear interpolation scaling operation.

[0038] 2.2 Divide the time-frequency feature map into sub-bands and extract time-frequency features. Time-frequency features include energy entropy. Multi-scale energy accumulation Feature time-frequency entropy These features are used as inputs to the subsequent network, and the formulas for each feature are as follows: ; ; ; in For sub-band energy, the probability is , This refers to the energy of the sub-band unit.

[0039] Step S3: Pre-training of the source domain RUL prediction model A feature extractor based on a deep time-frequency residual network (DFT-ResNet) is constructed, and supervised pre-training is performed on the source domain training set to obtain a converged pre-trained model, which serves as the initial weights for subsequent cross-domain training. The specific process is as follows: 3.1 Network Architecture Construction. A feature extraction network based on DFT-ResNet is constructed, with the network structure as follows: Figure 1 As shown, the network input first passes through an initial layer containing 2D convolutions (Conv2d), batch normalization (BN), ReLU activation function, and max pooling (MaxPool2d) to extract shallow features. Then, four Bottleneck residual modules are concatenated, each containing 1×1, 3×3, and 1×1 convolutions (Conv) and skip connections to deepen feature representation. These are then processed by adaptive average pooling (AdaptiveAvgPool2d) and flattening operations. To reduce feature dimensionality and improve representational ability, a bottleneck layer (Bottleneck) is added after flattening. Finally, the feature vector is output after passing through a classifier.

[0040] The feature vectors output by the feature extractor are connected to a fully connected layer to serve as the normalized RUL values ​​output by the RUL predictor. Its network forward propagation process is represented as: ;

[0041] in, These are the learnable parameters of the feature extractor; These are the learnable parameters for the RUL predictor; This is the extracted feature vector.

[0042] 3.2 Source domain model pre-training.

[0043] The pre-training process uses only labeled data from the source domain. The network is trained under supervised supervision, and the loss function is the mean squared error (MSE). ;

[0044] in, This represents the loss function value during the pre-training phase. For the source domain Input time-frequency images of one sample; For the source domain The actual RUL label of the sample.

[0045] The pre-training process uses the AdamW optimizer with L2 regularization to prevent overfitting, and the initial learning rate is [value missing]. The learning rate decays in steps using the StepLR scheduler, decreasing by 50% every 50 epochs. After training, the model weights are saved. , which serve as the initialization parameters for domain adaptive transfer learning training. Iterative training is performed on the source domain training set, and the model's generalization ability is evaluated on the source domain test set every 5 rounds. After training to the preset number of rounds, the model weights and optimizer state are saved to obtain a converged source domain RUL prediction model, which has learned the basic mapping relationship between vibration features and RUL in the source domain.

[0046] Step S4: Domain adaptive transfer learning training.

[0047] Based on the source domain RUL prediction model, a target domain RUL prediction model is constructed, which includes a feature extractor, a RUL predictor, and a domain discriminator, such as... Figure 2 As shown, a sample-level weighting mechanism is introduced to optimize the adversarial process and achieve feature alignment between the source and target domains.

[0048] 4.1 Network architecture composition of the target domain RUL prediction model.

[0049] The target domain RUL prediction model mainly consists of the following three parts: Feature extractor Load the pre-trained weights in step S3 This parameter is used to extract domain-invariant features. During the transfer learning training phase, this parameter can be further fine-tuned.

[0050] RUL Predictor : Fully connected layer, used to predict RUL values, jointly optimized with the feature extractor.

[0051] Domain discriminator A binary classification neural network is used to determine whether input features originate from the source domain or the target domain. The discriminator structure is a multilayer perceptron (MLP), containing two hidden layers, each with 128 neurons, and an output layer with single-neuron sigmoid activation. The domain discriminator output is represented as follows: ;

[0052] in, To predict the probability that a sample belongs to the source domain for the discriminator; These are the learnable parameters of the discriminator.

[0053] 4.2 Dynamic weighted loss mitigation mechanism.

[0054] To reduce the distributional discrepancy between the source and target domains, domain adversarial training is introduced. This scheme proposes a weighted algorithm based on prediction consistency (WAL). First, the weight coefficients are calculated. The weight depends on the actual tags of the source domain. Compared with the target domain predicted value The smaller the difference, the greater the weight, indicating that the sample is more important in domain alignment. Dynamic weight calculation is based on the absolute error between the true source domain label and the predicted value, using the following formula: ;

[0055] in, The weight coefficient is 1 when the difference is 0 and the weight is 1 when the difference is 1. .

[0056] Weight It decreases as the error increases.

[0057] In actual batch training, the weights of the source domain and the target domain are concatenated to obtain the batch weight vector. : ;

[0058] 4.3 Definition of loss function.

[0059] Define a weighted binary cross-entropy loss function. Let the discriminator output be... Domain tags The source domain is 1, and the target domain is 0. A warm-up mechanism is introduced, with a total of 1 training rounds. The preheating ratio is Then the preheating cycle In the current round When, calculate the lifetime prediction loss; when At that time, add weighted averages to counteract losses.

[0060] Total loss function The total loss for cross-domain training is defined as the weighted sum of the RUL prediction loss and the weighted adversarial loss, as shown in the following formula: ;

[0061] in, To predict MSE loss for lifetime over the source domain, To weighted countermeasures against losses, To counteract the loss weighting coefficient. and The calculation formula is as follows: ;

[0062] ;

[0063] in, For the first The weighting coefficients of each sample; This is a smoothing coefficient to prevent gradient vanishing due to excessively small weights; For the discriminator to the first The predicted probability of each sample; For the first The true domain labels of each sample; The binary cross-entropy loss is expressed by the following formula: ;

[0064] 4.4 The parameters of the feature extractor and the domain discriminator are updated using an alternating optimization strategy.

[0065] During the warm-up phase, the feature extractor is fixed, and only the RUL prediction loss and the basic adversarial loss are optimized without introducing dynamic weights to ensure model convergence. During the weighting phase, the domain discriminator and feature extractor are trained alternately.

[0066] 4.4.1 In the alternating training phase, the feature extractor is first fixed. Update the domain discriminator To minimize the weighted adversarial loss and improve domain discriminative ability. Fixed feature extractor The parameters. Input source domain features. and target domain features .

[0067] Construct the discriminator label, defining the source domain label as 1 and the target domain label as 0, i.e. ; Calculate the discriminator loss And update via backpropagation parameters : ;

[0068] in, The learning rate of the discriminator; This represents the gradient of the loss function with respect to the discriminator parameters.

[0069] 4.4.2 Subsequently, the fixed-domain discriminator Update the feature extractor The classification error of the domain discriminator is maximized by a gradient inversion layer while minimizing the RUL prediction loss, thus achieving feature distribution alignment. Model performance is evaluated after each training round, and the optimal model parameters are saved. Fixed domain discriminator. The parameters are such that, in order to confuse the discriminator, the discriminator label of the target domain sample is flipped to 1. Calculate the feature extractor loss, which includes source domain lifetime prediction loss and adversarial loss. .

[0070] Updated via backpropagation and parameters : ;

[0071] in, is the learning rate for the feature extractor. Gradient clipping is used during training to prevent gradient explosion.

[0072] Step S5: Model Evaluation and Remaining Useful Life Prediction During Cross-Condition Migration During training, evaluation is performed at each preset interval using the target domain test set. Evaluation metrics include root mean square error (RMSE), mean absolute error (MAE), and a score function, as shown in the following formulas; ;

[0073] ;

[0074] ;

[0075] in, The number of test set samples in the target domain; For the first Predicted RUL values ​​for each test sample; For the first The true RUL value of each test sample. The scoring function imposes a greater penalty on overestimation, i.e., the predicted lifetime is greater than the actual lifetime, which meets the needs of practical engineering. If the current RMSE is less than the historical best RMSE, the current model state is saved. After training, load the optimal model state. Input the target domain test data and output the final predicted RUL sequence.

[0076] To verify the effectiveness of this invention, an analysis of the cross-condition transfer task for rolling bearings was conducted. This invention employs a leave-one-out cross-condition partitioning strategy. Compared with existing traditional transfer learning methods on the OC2→OC1 cross-condition transfer task, the bearing RUL prediction method of this embodiment achieves significant results, such as... Figure 3As shown in the figure, RMSE is the root mean square error of the normalized RUL prediction; a smaller value indicates higher prediction accuracy. Experimental results show that, compared with traditional transfer learning methods such as MMD (Maximum Mean Difference), DANN (Domain Adversarial Neural Network), and CORAL (Correlation Alignment), this method achieves the lowest RMSE index of 0.0491 on the OC2→OC1 task, significantly outperforming MMD (0.0840), DANN (0.0655), and CORAL (0.0904). The prediction error is reduced by approximately 25.0% compared to the second-best performing DANN and by approximately 45.7% compared to the largest error, CORAL. From the curve trends, CORAL and MMD show obvious overestimation of lifetime at most time steps, with the predicted curves deviating significantly from the actual RUL, and the deviation widening sharply in the later stages of lifetime degradation. In contrast, the predicted trajectory of our method closely follows the actual degradation trend throughout the entire lifetime, especially in the later stages, where it accurately tracks the rapid decline of RUL. This fully demonstrates that the proposed weighted adversarial mechanism can effectively alleviate the negative migration problem caused by operating condition differences and significantly improve the prediction robustness of cross-domain remaining useful life.

[0077] The above embodiments demonstrate that the present invention effectively solves the generalization problem of bearing life prediction models under different operating conditions through refined dataset partitioning strategies, deep time-frequency residual feature extraction, operating condition adaptive hyperparameter configuration, and dynamic weighted adversarial migration mechanism, providing a reliable technical solution for cross-operating condition health management of industrial equipment.

[0078] The above description is merely a preferred embodiment of the present invention, used to assist those skilled in the art in implementing the corresponding technical solutions, and is not intended to limit the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims. It should be noted that, for those skilled in the art, several equivalent improvements and modifications can be made based on the technical solutions of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Furthermore, it should be understood that although this specification describes the embodiments as described above, not every embodiment contains only one independent technical solution. This descriptive method is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions of each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning, including a method for training a bearing remaining life prediction model; and a step of using the remaining life prediction model to predict the remaining life of the bearing under test. Methods for training bearing remaining life prediction models include source domain model pre-training methods, which include the following steps; Collect vibration time-series signals of source domain samples and construct the source domain sample input tensor; The source domain sample input tensor is input into the source domain RUL prediction model to perform supervised pre-training, resulting in a converged source domain RUL prediction model. The RUL prediction model includes a feature extractor and a RUL predictor; Its features are: The method for training the bearing remaining life prediction model also includes the domain adaptive transfer learning training method, which includes the following steps; A target domain RUL prediction model is constructed based on the feature extractor and RUL predictor of the converged source domain RUL prediction model. The target domain RUL prediction model also includes a domain discriminator. Collect vibration time-series signals of samples in the target domain and construct the input tensor of the samples in the target domain; The source domain sample input tensor and the target domain sample input tensor are input into the target domain RUL prediction model for training. The training process includes a warm-up learning stage and a domain adversarial learning stage, with the domain adversarial learning stage following the warm-up learning stage. During the warm-up learning phase, the feature extractor of the target domain bearing RUL prediction model is fixed, and the RUL predictor is trained on the basis. During the domain adversarial learning phase, the domain discriminator and feature extractor are trained alternately. During the adversarial learning phase, when training the domain discriminator, the feature extractor is fixed, and the source domain features and target domain features with real domain labels are used to train the domain discriminator and update the model parameters of the domain discriminator. During the adversarial learning phase, when training the feature extractor, the domain discriminator is fixed, and the feature extractor is trained using source domain features with true domain labels and target domain features with incorrect domain labels. The model parameters of the feature extractor and the RUL predictor are then updated. During training, the target domain RUL prediction model is evaluated using the target domain test set at predetermined intervals, and the evaluation value is obtained. The model training ends when the evaluation value is less than the predetermined value. After training, the test samples in the target domain are input into the tensor and then into the RUL prediction model in the target domain to predict the remaining life of the bearing. Both the source domain sample input tensor and the target domain sample input tensor are time-frequency feature map tensors formed by converting sample vibration time-series signals. During the domain adversarial learning phase, the domain discriminator loss L is obtained using the following formula when training the domain discriminator. D ; ； When training the feature extractor, the feature extractor loss is obtained using the following formula. ; Among them, L WAL This is called weighted adversarial loss, w k For the first The weight coefficients of each sample, For smoothing coefficients, For the domain discriminator to the first The predicted probability of each sample; For the first The domain label of each sample, L BCE (p k ,d k ) represents the binary cross-entropy loss, L RUL For the remaining lifetime prediction loss of the source domain samples, N s N is the number of samples in the source domain. t The number of samples in the target domain; Among them, w k Obtained by the following formula; ； Where exp is the natural exponential function. For the sample predicted value; For source domain sample data, y s For the source domain sample data, y represents the true label; for the target domain sample data, y represents the true label. s This is the global average lifetime label for the source domain.

2. The bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning according to claim 1, characterized in that: The feature extractor includes a series of stacked 2D convolutional layers, batch normalization layers, ReLU activation layers and max pooling layers, several cascaded Bottleneck residual modules, adaptive average pooling layers, flattening layers and bottleneck layers.

3. The bearing RUL prediction method based on time-frequency domain features and weighted domain adversarial learning according to claim 2, characterized in that: The Bottleneck residual module includes a 1×1 convolutional layer, a 3×3 convolutional layer, and a 1×1 convolutional layer connected in series, and the input and output of the Bottleneck residual module are connected in a skip connection.

Images