Attention-based domain adaptive residual life prediction method
Through the Transformer model integrating multi-scale convolutional neural network and self-attention mechanism, combining efficient channel attention mechanism and joint distribution difference loss, a deep information fusion discriminant transfer learning network is built, which solves the scarcity of data and inconsistent distribution under rolling bearing conditions. The problem of the prediction accuracy is significantly improved.
Patent Information
- Application Number
- CN202411050914.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2024-11-19
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing deep learning models face the problems of scarcity and inconsistent distribution when processing variable working conditions of rolling bearings, resulting in reduced prediction accuracy and difficult to apply to actual production.
Adopting the domain adaptive residual life prediction method based on attention, local and global features are extracted through the Transformer model integrating multi-scale convolutional neural network and self-attention mechanism, combining high-efficiency channel attention mechanism and joint distribution difference loss method to improve the features Perform alignment and fusion, and build a deep information fusion discriminant transfer learning network for prediction.
The accuracy of feature extraction and the domain adaptability of the model are improved, and the accuracy of the remaining life prediction of rolling bearings is significantly improved, and the prediction accuracy reduction caused by data scarcity and inconsistent distribution is solved.
Smart Images

Figure CN118981679A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data prediction, and in particular to a domain adaptive remaining life prediction method based on attention. Background Art
[0002] Rolling bearings are core components in the machinery industry and are widely used in various types of mechanical equipment. Once a bearing fails, it will cause serious economic losses and unexpected downtime. In order to avoid these risks, it is very necessary to discover and identify various abnormalities and failures through prognosis and health management. Among them, the remaining useful life (RUL) prediction technology is particularly important in preventing accidents and early fault detection.
[0003] At present, RUL prediction methods are mainly divided into two categories: model-based methods and data-driven methods. Model-based methods establish bearing degradation models to evaluate the health status and predict degradation trends. Although they can better describe the degradation process, as the complexity of the equipment increases, it becomes challenging to establish an accurate physical model.
[0004] With the development of artificial intelligence technology, data-driven prediction methods have received widespread attention. These methods can predict RUL even in the case of unknown failure mechanisms by capturing the mapping relationship between condition monitoring data and bearing degradation. Deep learning, as a key data-driven prediction method, can learn from a large amount of high-quality data and has been widely used in RUL prediction, such as deep belief networks, convolutional neural networks, and long short-term memory networks. However, these methods face the problems of data scarcity and inconsistent distribution when dealing with variable operating condition data.
[0005] In order to improve the ability of deep learning models to process data features under different working conditions, transfer learning methods have been proposed and have achieved good results, but these methods usually require that the distribution of training data and test data is the same, which is difficult to meet in actual industrial scenarios. As a result, the accuracy of cross-domain prediction is reduced and it is difficult to apply to actual production.
[0006] To solve the above problems, the present invention proposes a domain adaptive remaining life prediction method based on attention. Summary of the invention
[0007] In view of the shortcomings of the prior art, the present invention provides an attention-based domain adaptive remaining life prediction method, which solves the problem of reduced prediction accuracy due to data scarcity and inconsistent distribution when predicting the remaining life of rolling bearings under variable working conditions.
[0008] To achieve the above objectives, the present invention is implemented by the following technical solutions: a domain adaptive remaining life prediction method based on attention, comprising the following steps:
[0009] S1. Obtain historical degradation data of rolling bearings under different working conditions;
[0010] S2, using the Transformer model integrating multi-scale convolutional neural network and self-attention mechanism to extract local and global features of historical degradation data;
[0011] S3, apply efficient channel attention mechanism to improve the weight of important features in each channel of feature extractor;
[0012] S4, preprocessing the extracted features to construct a pre-trained label predictor;
[0013] S5. Propose a joint distribution difference loss method to align features;
[0014] S6. Construct a deep information fusion discriminant transfer learning network for remaining life prediction.
[0015] Preferably, the step S2 specifically includes the following steps:
[0016] S2.1. Construct a Transformer model integrating multi-scale convolutional neural network and self-attention mechanism;
[0017] S2.2, input the historical degradation data into MS-CNN to extract local features;
[0018] S2.3. Use the Transformer’s self-attention mechanism to extract the global features of the data.
[0019] S2.4, combine local features and global features to extract multi-dimensional and multi-angle features of the signal;
[0020] S2.5. Use the Transformer self-attention mechanism to guide the model to focus on the dependencies between different positions in the sequence.
[0021] Preferably, the local features in step S2 include time domain features, frequency domain features, time-frequency domain features and vibration features, and the global features in step S2 include long-term dependence features, sequence features, global correlation features and overall dynamic features.
[0022] Preferably, the step S3 specifically includes the following steps:
[0023] S3.1. Construct a channel attention mechanism module to enhance key features by calculating the importance weight of each channel.
[0024] S3.2, input the fused feature vector into the channel attention mechanism module, and calculate the attention weight of each channel;
[0025] S3.3. According to the calculated attention weights, the feature vectors are weighted to improve the representation ability of important features;
[0026] S3.4, output the weighted feature vector as input for subsequent model training;
[0027] S3.5. Adaptively obtain the deep performance degradation characteristics of rolling bearings under different working conditions.
[0028] Preferably, the formula for calculating the channel weight in step S3.1 is:
[0029] Among them, w i is the importance weight of the ith channel, e i is the feature of the ith channel obtained after global average pooling, C is the total number of channels, and the fusion feature calculation formula in step S3.2 is:
[0030] Among them, F 融合 is the fused feature vector, w i is the importance weight of the ith channel, F i is the feature vector of the ith channel, and the formaldehyde feature calculation formula in step S3.3 is:
[0031] F 加权 =W·F, where F 加权 is the weighted eigenvector, W is the matrix containing all channel weights, and F is the input eigenvector.
[0032] Preferably, the historical degradation data under different working conditions in step S1 include vibration signals, temperature data and rotation speed.
[0033] Preferably, the step S5 specifically includes the following steps:
[0034] S5.1. Construct a joint distribution difference loss function to measure the difference between marginal distribution and conditional distribution respectively.
[0035] S5.2. Calculate the difference between marginal distribution and conditional distribution on the data of source domain and target domain respectively.
[0036] S5.3. Perform weighted summation of the marginal distribution difference and the conditional distribution difference to obtain the joint distribution difference loss.
[0037] S5.4. Align the features by minimizing the joint distribution difference loss through the back propagation algorithm;
[0038] S5.5. Align the data features at different degradation stages, align the marginal distribution and the conditional distribution, and retain the information in the target specific field.
[0039] The marginal distribution difference calculation formula in step S5.2 is:
[0040] Among them, P and D represent the marginal distribution of the source domain and the target domain respectively. is the mapping function, n and m are the number of samples in the source domain and the target domain respectively, and the conditional distribution difference calculation formula in step S5.2 is:
[0041] in, and They represent the conditional means of the source domain and the target domain under category c, respectively, and C is the number of categories.
[0042] Preferably, the calculation formula for the joint distribution difference loss in step S5.3 is:
[0043] JLCDA(P,Q)=λ·MMD 2 (P, Q) + (1-λ) Cond_DIff(P, Q), where
[0044] λ is a weight coefficient that controls the relative importance of the difference between the marginal distribution and the conditional distribution.
[0045] Preferably, the step S6 specifically includes the following steps:
[0046] S6.1. Use the feature fusion module and JLCDA mechanism to build a deep information fusion discriminant transfer learning network;
[0047] S6.2, build a deep information fusion module to fuse the features of the source domain and the target domain;
[0048] S6.3. Construct a discriminator to distinguish the features of the source domain and the target domain.
[0049] S6.4, input the fused features into the discriminant transfer learning network to perform model training;
[0050] S6.5, transfer the knowledge of the source domain to the target domain through transfer learning method;
[0051] S6.6. Evaluate the prediction results of the model and optimize the model structure.
[0052] The present invention provides a domain adaptive remaining life prediction method based on attention.
[0053] Beneficial effects:
[0054] 1. The present invention integrates a Transformer model with a multi-scale convolutional neural network and a self-attention mechanism, which can adaptively extract local and global features of the bearing under different working conditions, solves the problem of insufficient feature extraction due to receptive field limitations, and improves the accuracy of feature extraction.
[0055] 2. By utilizing an efficient channel attention mechanism, the present invention can adaptively improve the weights of important features in each channel of the feature extractor, solve the problem of insufficient feature extraction caused by data scarcity, and enhance the domain adaptation capability of the model under different working conditions.
[0056] 3. The present invention, through the proposed joint distribution difference loss method, can align the marginal distribution and the conditional distribution at the same time, solving the problem of inconsistent data distribution between the source domain and the target domain, thereby more effectively retaining the information of the target specific field and enhancing the generalization ability of the prediction model.
[0057] 4. The present invention realizes the deep fusion of source domain and target domain features by constructing a deep information fusion discriminant transfer learning network, solves the problem of reduced prediction accuracy caused by inconsistent data distribution and data scarcity, and significantly improves the accuracy of rolling bearing remaining life prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 is a flow chart of the present invention;
[0059] Figure 2 This is a diagram of the domain adaptive network architecture of the present invention. DETAILED DESCRIPTION
[0060] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0061] Please refer to the attached Figure 1 -Attached Figure 2 The embodiment of the present invention provides a domain adaptive remaining life prediction method based on attention, comprising the following steps:
[0062] S1. Obtain historical degradation data of rolling bearings under different working conditions;
[0063] The historical degradation data under different working conditions in step S1 include vibration signals, temperature data and rotation speed.
[0064] Specifically, step S1 includes the following steps:
[0065] The vibration signal, temperature data and speed of the rolling bearing under different working conditions are obtained from the sensor.
[0066] The acquired raw data is cleaned to remove noise and outliers to ensure data quality.
[0067] Sort the cleaned data in chronological order to ensure the continuity and timing characteristics of the data.
[0068] The processed data is stored in the database for easy subsequent calling and processing.
[0069] S2, using the Transformer model integrating multi-scale convolutional neural network and self-attention mechanism to extract local and global features of historical degradation data;
[0070] S3, apply efficient channel attention mechanism to improve the weight of important features in each channel of feature extractor;
[0071] S4, preprocessing the extracted features to construct a pre-trained label predictor;
[0072] S5. Propose a joint distribution difference loss method to align features;
[0073] S6. Construct a deep information fusion discriminant transfer learning network for remaining life prediction.
[0074] Specifically, the technical solution adopted by the present invention is to first construct a device degradation feature extraction model of the Transformer that inherits the multi-scale convolutional neural network and the self-attention mechanism, respectively perform local and global feature extraction and mining on the data, and use the transformer self-attention mechanism to guide the model to pay attention to the dependencies between different positions in the sequence, so as to achieve the complementarity of the local and global features of the signal, and adaptively extract the life state features of the bearing under different working conditions from multiple dimensions and angles, thus overcoming the problem of insufficient feature attention caused by the limitation of the receptive field; at the same time, use the efficient channel attention mechanism to improve the weight of important features in each channel of the feature extractor, adaptively obtain the deep performance degradation features of the rolling bearing under different working conditions, and pre-train the label predictor based on this; secondly, in order to enhance the distribution alignment of the two domains and match the marginal distribution and conditional distribution, an improved joint distribution difference loss is proposed, and the data features of different degradation stages are aligned respectively, so as to achieve marginal distribution alignment and conditional distribution alignment, while retaining the information of the target specific field. Using the feature fusion module and JLCDA mechanism, a deep information fusion discriminant transfer learning network is constructed to realize the remaining life prediction.
[0075] The S2 step specifically includes the following steps:
[0076] S2.1. Construct a Transformer model integrating multi-scale convolutional neural network and self-attention mechanism;
[0077] S2.2, input the historical degradation data into MS-CNN to extract local features;
[0078] S2.3. Use the Transformer’s self-attention mechanism to extract the global features of the data.
[0079] S2.4, combine local features and global features to extract multi-dimensional and multi-angle features of the signal;
[0080] S2.5. Use the Transformer self-attention mechanism to guide the model to focus on the dependencies between different positions in the sequence.
[0081] The local features in step S2 include time domain features, frequency domain features, time-frequency domain features and vibration features. The global features in step S2 include long-term dependency features, sequence features, global correlation features and overall dynamic features.
[0082] Specifically, in step S2.1:
[0083] Multi-scale convolutional neural network: extracts local features of different scales through multiple convolutional layers, enhancing the model's ability to capture different frequency and time features.
[0084] Transformer model with self-attention mechanism: The self-attention mechanism is used to capture the global features of the data and focus on the dependencies between different positions in the signal sequence.
[0085] The above two technologies solve the problem that traditional methods are difficult to capture local and global features at the same time due to limited receptive field, and the extracted features may contain a lot of redundant information and noise, affecting model performance.
[0086] Specific solutions:
[0087] In step S2.2,
[0088] Local feature extraction: Extract local features of rolling bearings from time domain, frequency domain, time-frequency domain and vibration features.
[0089] The problem that a single characteristic dimension is difficult to fully reflect the degradation of rolling bearings is solved.
[0090] Time Domain Features: Calculate the statistical characteristics of the signal, such as mean, standard deviation, skewness, and kurtosis.
[0091]
[0092] Frequency domain features: Spectral features such as main frequency, bandwidth, center frequency, etc. are calculated through Fourier transform.
[0093]
[0094] Time-frequency domain features: Time-frequency features such as energy distribution and instantaneous frequency are obtained through short-time Fourier transform or wavelet transform.
[0095]
[0096] Vibration characteristics: Extract acceleration, velocity and displacement characteristics to reflect the mechanical vibration state of the bearing.
[0097] In step S2.3,
[0098] Global feature extraction: Capture long-term dependencies and global patterns in signals, and enhance the model’s ability to perceive the overall trend of signals.
[0099] Sequence features: Through the multi-head self-attention mechanism, the model's attention to the key positions in the signal sequence is enhanced.
[0100] Global correlation feature: Calculate the correlation matrix between different time points to reveal the relationship between signals at different time points.
[0101] Overall dynamic characteristics: extract the dynamic changes of the signal over the entire time range, including the growth rate and attenuation rate of the signal.
[0102] In step S2.4,
[0103] Feature fusion: Combine local features and global features to fully reflect the multi-dimensional and multi-angle characteristics of the signal.
[0104] Feature fusion: The local features extracted by MS-CNN and the global features extracted by Transformer are fused to form a comprehensive feature vector.
[0105] Multi-dimensional and multi-angle feature extraction: Through feature fusion, the model's ability to perceive the signal's features in various dimensions and angles is enhanced.
[0106] In step S2.5,
[0107] Focus on key positions: Through the self-attention mechanism, the model is guided to focus on key positions in the signal sequence to improve the effect of feature extraction.
[0108] Self-attention mechanism: Through the self-attention mechanism, the correlation between different positions in the signal sequence is calculated to guide the model to focus on key positions.
[0109]
[0110] Position dependency: The self-attention mechanism is used to capture the position dependency in the signal sequence and improve the accuracy of feature extraction.
[0111] Through the above steps, the Transformer model integrating multi-scale convolutional neural network and self-attention mechanism is used to perform local and global feature extraction, which can effectively solve the problems of insufficient feature extraction, feature redundancy and noise, short-term dependency and feature singularity in the existing technology, and comprehensively improve the accuracy and reliability of rolling bearing remaining life prediction.
[0112] The S3 step specifically includes the following steps:
[0113] S3.1. Construct a channel attention mechanism module to enhance key features by calculating the importance weight of each channel.
[0114] S3.2, input the fused feature vector into the channel attention mechanism module, and calculate the attention weight of each channel;
[0115] S3.3. According to the calculated attention weights, the feature vectors are weighted to improve the representation ability of important features;
[0116] S3.4, output the weighted feature vector as input for subsequent model training;
[0117] S3.5. Adaptively obtain the deep performance degradation characteristics of rolling bearings under different working conditions.
[0118] The formula for calculating the channel weight in step S3.1 is:
[0119] Among them, w i is the importance weight of the ith channel, e i is the feature of the ith channel obtained after global average pooling, C is the total number of channels, and the fusion feature calculation formula in step S3.2 is:
[0120] Among them, F 融合 is the fused feature vector, w i is the importance weight of the ith channel, F i is the feature vector of the ith channel. The formaldehyde feature calculation formula in step S3.3 is:
[0121] F 加权 =W·F, where F 加权 is the weighted eigenvector, W is the matrix containing all channel weights, and F is the input eigenvector.
[0122] Step S5 specifically includes the following steps:
[0123] S5.1. Construct a joint distribution difference loss function to measure the difference between marginal distribution and conditional distribution respectively.
[0124] S5.2. Calculate the difference between marginal distribution and conditional distribution on the data of source domain and target domain respectively.
[0125] S5.3. Perform weighted summation of the marginal distribution difference and the conditional distribution difference to obtain the joint distribution difference loss.
[0126] S5.4. Align the features by minimizing the joint distribution difference loss through the back propagation algorithm;
[0127] S5.5. Align the data features at different degradation stages, align the marginal distribution and the conditional distribution, and retain the information in the target specific field.
[0128] The marginal distribution difference calculation formula in step S5.2 is:
[0129] Among them, P and D represent the marginal distribution of the source domain and the target domain respectively. is the mapping function, n and m are the number of samples in the source domain and the target domain respectively, and the formula for calculating the conditional distribution difference in step S5.2 is:
[0130] in, and They represent the conditional means of the source domain and the target domain under category c, respectively, and C is the number of categories.
[0131] The formula for calculating the joint distribution difference loss in step S5.3 is:
[0132] JLCDA(P,Q)=λ·MMD 2 (P, Q) + (1-λ) Cond_DIff(P, Q), where
[0133] λ is a weight coefficient that controls the relative importance of the difference between the marginal distribution and the conditional distribution.
[0134] Step S6 specifically includes the following steps:
[0135] S6.1. Use the feature fusion module and JLCDA mechanism to build a deep information fusion discriminant transfer learning network;
[0136] S6.2, build a deep information fusion module to fuse the features of the source domain and the target domain;
[0137] S6.3. Construct a discriminator to distinguish the features of the source domain and the target domain.
[0138] S6.4, input the fused features into the discriminant transfer learning network to perform model training;
[0139] S6.5, transfer the knowledge of the source domain to the target domain through transfer learning method;
[0140] S6.6. Evaluate the prediction results of the model and optimize the model structure.
[0141] Specifically, in the above steps, the feature fusion module: constructs a feature fusion module to fuse the features of the source domain and the target domain through linear transformation and nonlinear activation function.
[0142] F 融合 =σ(W s ·F s +W t ·F t )
[0143] Among them, F 融合 is the fused feature, W s and W t are the linear transformation matrices of the source domain and the target domain, respectively, and F s and F t are the features of the source domain and the target domain respectively, and σ is a nonlinear activation function.
[0144] Construct a discriminator to distinguish the features of the source domain and the target domain through a binary classification task.
[0145] D(F)=sigmoid(W d ·F+b d
[0146] Where D(F) is the output of the discriminator, W d is the weight matrix of the discriminator, b d is the bias term, F is the input feature, and \text{sigmoid} is the activation function.
[0147] Discriminative transfer learning network: The fused features are input into the discriminative transfer learning network, and effective knowledge transfer is achieved through joint training.
[0148] L total =L task +λL JLCDA +βL D
[0149] Among them, L total is the total loss function, L task is the mission loss, L JLCDA is the joint distribution difference loss, L D is the discriminative loss, λ and β are weight coefficients.
[0150] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A domain-adaptive remaining life prediction method based on attention, It is characterized in that The following steps are involved: S1. Obtain historical degradation data of rolling bearings under different working conditions; S2, using the Transformer model integrating multi-scale convolutional neural network and self-attention mechanism to extract local and global features of historical degradation data; S3, apply efficient channel attention mechanism to improve the weight of important features in each channel of feature extractor; S4, preprocessing the extracted features to construct a pre-trained label predictor; S5. Propose a joint distribution difference loss method to align features; S6. Construct a deep information fusion discriminant transfer learning network for remaining life prediction.
2. According to claim 1, a domain adaptive remaining life prediction method based on attention, It is characterized in that The S2 step specifically includes the following steps: S2.
1. Construct a Transformer model integrating multi-scale convolutional neural network and self-attention mechanism; S2.2, input the historical degradation data into MS-CNN to extract local features; S2.
3. Use the Transformer’s self-attention mechanism to extract the global features of the data. S2.4, combine local features and global features to extract multi-dimensional and multi-angle features of the signal; S2.
5. Use the Transformer self-attention mechanism to guide the model to focus on the dependencies between different positions in the sequence.
3. The method for predicting remaining life span based on domain adaptation based on attention according to claim 2, It is characterized in that The local features in step S2 include time domain features, frequency domain features, time-frequency domain features and vibration features, and the global features in step S2 include long-term dependence features, sequence features, global correlation features and overall dynamic features.
4. The method for predicting remaining life span based on domain adaptation based on attention according to claim 1, It is characterized in that The S3 step specifically includes the following steps: S3.
1. Construct a channel attention mechanism module to enhance key features by calculating the importance weight of each channel. S3.2, input the fused feature vector into the channel attention mechanism module, and calculate the attention weight of each channel; S3.
3. According to the calculated attention weights, the feature vectors are weighted to improve the representation ability of important features; S3.4, output the weighted feature vector as input for subsequent model training; S3.
5. Adaptively obtain the deep performance degradation characteristics of rolling bearings under different working conditions.
5. The method for predicting remaining life span based on domain adaptation based on attention according to claim 4, It is characterized in that The formula for calculating the channel weight in step S3.1 is: Among them, w i is the importance weight of the ith channel, e i is the feature of the ith channel obtained after global average pooling, C is the total number of channels, and the fusion feature calculation formula in step S3.2 is: Among them, F 融合 is the fused feature vector, w i is the importance weight of the ith channel, F i is the feature vector of the ith channel, and the formaldehyde feature calculation formula in step S3.3 is: F 加权 =W·F, where F 加权 is the weighted eigenvector, W is the matrix containing all channel weights, and F is the input eigenvector.
6. The method for predicting remaining life span based on domain adaptation based on attention according to claim 1, It is characterized in that The historical degradation data under different working conditions in step S1 include vibration signals, temperature data and rotation speed.
7. The method for predicting remaining life span based on domain adaptation based on attention according to claim 1, It is characterized in that The step S5 specifically includes the following steps: S5.
1. Construct a joint distribution difference loss function to measure the difference between marginal distribution and conditional distribution respectively. S5.
2. Calculate the difference between marginal distribution and conditional distribution on the data of source domain and target domain respectively. S5.
3. Perform weighted summation of the marginal distribution difference and the conditional distribution difference to obtain the joint distribution difference loss. S5.
4. Align the features by minimizing the joint distribution difference loss through the back propagation algorithm; S5.
5. Align the data features at different degradation stages, align the marginal distribution and the conditional distribution, and retain the information in the target specific field.
8. The method for predicting remaining life span based on domain adaptation based on attention according to claim 6, It is characterized in that The marginal distribution difference calculation formula in step S5.2 is: Among them, P and D represent the marginal distribution of the source domain and the target domain respectively. is the mapping function, n and m are the number of samples in the source domain and the target domain respectively, and the conditional distribution difference calculation formula in step S5.2 is: in, and They represent the conditional means of the source domain and the target domain under category c, respectively, and C is the number of categories.
9. The method for predicting remaining life span based on domain adaptation based on attention according to claim 6, It is characterized in that The calculation formula for the joint distribution difference loss in step S5.3 is: JLCDA(P, Q) = λ·MMD 2 (P, Q) + (1 - λ)·Cond_DIff(P, Q), where λ is a weight coefficient that controls the relative importance of the difference between the marginal distribution and the conditional distribution.
10. The method for predicting remaining life span based on domain adaptation based on attention according to claim 1, It is characterized in that The step S6 specifically includes the following steps: S6.
1. Use the feature fusion module and JLCDA mechanism to build a deep information fusion discriminant transfer learning network; S6.2, build a deep information fusion module to fuse the features of the source domain and the target domain; S6.
3. Construct a discriminator to distinguish the features of the source domain and the target domain. S6.4, input the fused features into the discriminant transfer learning network to perform model training; S6.5, transfer the knowledge of the source domain to the target domain through the transfer learning method; S6.
6. Evaluate the prediction results of the model and optimize the model structure.
Citation Information
Cited By
Gas cylinder management method and system
CN120354081A
A gas cylinder management method and system
CN120354081B
Method and system for evaluating service life of power electronic device based on electrical and thermal stress
CN120705798A
Method for predicting residual life of composite material structure based on domain self-adaption under digital-analog combined driving
CN120781692A
Battery charge state estimation method and system based on ultrasonic transmission signal
CN120847621A
