Rolling bearing residual life prediction method based on heterogeneous dimension adversarial network
By employing the Nyquist sampling theorem and frequency resolution to select the dimension of the life point in bearing remaining life prediction, and combining it with a heterogeneous dimension adversarial network, the problem of asymmetric cross-domain data distribution is solved, and more accurate remaining life prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-18
- Publication Date
- 2026-04-07
AI Technical Summary
In the prediction of the remaining life of cross-domain bearings, the existing technology has problems. Because the life point dimension of the source domain and the target domain are the same, but the semantic information is not equivalent, the model mistakenly treats the difference in the number of cycles as the degradation mode, which undermines the stability of the knowledge transfer process. In particular, when there are differences in operating conditions such as speed, the data distribution measurement is distorted.
By adopting a lifetime point dimension selection rule based on the Nyquist sampling theorem and frequency resolution, and combining it with a heterogeneous dimensionality adversarial network (HDAN), the HDAN is trained through source and target domain feature extractors and a mutual information calculator to achieve cross-domain data distribution alignment and improve prediction accuracy.
By using heterogeneous dimensional adversarial networks, cross-domain invariant degradation trends can be accurately extracted, improving the accuracy and stability of bearing remaining life prediction, especially with significant data deviation alignment under changing operating conditions.
Smart Images
Figure CN121809209A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of rolling bearing remaining life prediction technology, and more specifically, relates to a rolling bearing remaining life prediction method based on heterogeneous dimensional adversarial networks. Background Technology
[0002] Due to the combined effects of harsh external operating environments and internal fatigue wear, bearings inevitably experience performance degradation during service, leading to abnormal operating conditions and ultimately functional failure. Therefore, research on bearing remaining life prediction methods is a crucial technical means to ensure the safe operation of mechanical equipment. Considering the current scarcity of industrial data on the entire lifespan of mechanical equipment, an effective solution for bearing remaining life prediction based on transfer learning is proposed.
[0003] Traditional cross-domain bearing remaining life prediction methods require the source and target domains to have the same life point dimension in order to effectively calculate and reduce the data distribution differences between the two domains. However, when operating conditions such as rotational speed differ, having the same life point dimension does not necessarily mean semantic equivalence. Figure 1 This is a vibration waveform diagram of an LDKUER204 bearing with an outer ring fault, operating at 1000 rpm and 2000 rpm. (The last sentence appears to be incomplete and possibly refers to a different product.) Figure 1 As shown, although the lifetime point dimensions are the same for both operating conditions (ds=dt), the fault information they contain is significantly different: the source domain signal contains approximately 6 BPFO cycles, while the target domain contains approximately 12. This results in an imbalance of fault information contained in the same sample; the source domain is relatively sparse, while the target domain sample contains denser impact components and more significant degradation features. Therefore, if direct alignment is performed based on the same lifetime point dimension, the model may mistakenly interpret the difference in the number of cycles as a cross-domain degradation pattern, thereby compromising the stability of the entire knowledge transfer process. It is evident that in practical applications, due to differences in operating speed and fault location, the assumption of equal lifetime point dimensions between the source and target domains inevitably leads to asymmetry in degradation information between the two domains, resulting in inherent distortion in the measurement of cross-domain data distribution. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting the remaining life of rolling bearings based on heterogeneous dimensional adversarial networks. It formulates a life point dimension selection rule based on the Nyquist sampling theorem and frequency resolution, and uses a heterogeneous dimensional adversarial network (HDAN) to make the distribution alignment focus on the deviation of the real data caused by the change of operating conditions, thereby improving the accuracy of remaining life migration prediction.
[0005] To achieve the above-mentioned objective, the rolling bearing remaining life prediction method based on heterogeneous dimensional adversarial networks of the present invention includes the following steps:
[0006] S1: Set the fault types of the rolling bearing according to actual needs, and record the number of fault types as follows: Based on actual needs, a specific operating condition is set as the source domain condition. Under this condition, raw vibration signals from several full-life cycles are collected from rolling bearings experiencing different faults. Several remaining life prediction samples are then extracted. The input samples for each remaining life prediction sample include... Vibration signal feature data at each lifetime point, labeled with the remaining lifetime of the last lifetime point of the input sample, constitute the source domain sample set, where the dimension of the lifetime points is... The following formula is used to determine it:
[0007] ,
[0008] in, This indicates the sampling frequency of the original vibration signal. This represents the preset coefficient. Represents the set of positive integers. , This indicates the rolling bearing under source domain operating conditions. Fault characteristic frequency The set of fault characteristic frequencies constituted ;
[0009] S2: For the target domain operating conditions requiring remaining life prediction, several raw vibration signals of the rolling bearing are collected under the target domain operating conditions, and then several signals containing... Vibration signal feature data at each lifetime point are used as input samples for the target domain, thus forming the target domain sample set; where the dimension of the lifetime points is... The following formula is used to determine it:
[0010] ,
[0011] in, , This indicates the rolling bearing under the target domain operating conditions. Fault characteristic frequency The set of fault characteristic frequencies constituted;
[0012] S3: Construct heterogeneous dimensional adversarial networks, including source domain feature extractors. Target domain feature extractor Remaining life predictor Mutual Information Calculator ,in:
[0013] Source Domain Feature Extractor Source domain data features are used to extract source domain input samples and sent to the remaining lifetime predictor. Mutual Information Calculator ;
[0014] Target Domain Extractor Used to extract target domain data features from input samples and send them to the mutual information calculator. ;
[0015] Remaining life predictor Used to predict remaining lifespan based on input data characteristics;
[0016] Mutual Information Calculator Used to generate mutual information estimates of source domain data features and target domain data features;
[0017] S4: Train a heterogeneous dimensional adversarial network using source domain sample sets and target domain sample sets. The specific method is as follows:
[0018] In each training round, samples are selected from the source domain sample set and the target domain sample set respectively. The input samples constitute input sample pairs , Then input the source domain sample and target domain input samples Input source domain feature extractor respectively and target domain feature extractor To obtain the source domain data features and target domain data features Remaining life predictor Based on source domain data characteristics Obtain the predicted remaining lifespan Then, the predicted loss of remaining lifetime is calculated using the following formula. :
[0019] ,
[0020] in, Represents the source domain input sample Corresponding real tags , This represents the preset remaining lifetime prediction loss function;
[0021] Mutual Information Calculator Calculate source domain data features and target domain data features mutual information estimate Then, the mutual information loss is calculated using the following formula. :
[0022] ,
[0023] in, This represents the preset mutual information loss function;
[0024] Loss prediction based on remaining lifespan For remaining lifetime predictor Perform parameter updates based on the negative of the mutual information loss. Source domain feature extractor and target domain feature extractor Perform parameter updates based on mutual information loss. Mutual Information Calculator Update the parameters;
[0025] S5: Extract the target domain extractor from the trained heterogeneous dimensional adversarial network. and remaining lifetime predictor The remaining lifetime prediction model constitutes the target domain;
[0026] S6: The time to be predicted under the target domain operating conditions and prior to... The characteristic data of the rolling bearing vibration signal at each life point constitute the input sample, which is then input into the remaining life prediction model of the target domain to obtain the predicted remaining life.
[0027] This invention presents a method for predicting the remaining life of rolling bearings based on heterogeneous dimensional adversarial networks. First, raw vibration signals from several full life cycles of rolling bearings under different fault types are collected under source domain operating conditions, from which a source domain sample set is extracted. Then, a target domain sample set is extracted under target domain operating conditions. The life point dimensions under both source and target domain operating conditions are determined based on the fault characteristic frequencies under the corresponding operating conditions. A heterogeneous dimensional adversarial network is constructed, comprising a source domain feature extractor, a target domain feature extractor, a remaining life predictor, and a mutual information calculator. This network is trained using both the source and target domain sample sets. From the trained heterogeneous dimensional adversarial network, the target domain feature extractor and the remaining life predictor are extracted to form a remaining life prediction model for the target domain. This model is then used to predict the remaining life of rolling bearings under the target domain operating conditions.
[0028] The present invention has the following beneficial effects:
[0029] 1) This invention formulates a lifetime point dimension selection rule based on the Nyquist sampling theorem and frequency resolution, and adopts a heterogeneous dimension adversarial network to make the distribution alignment focus on the real data deviation caused by the change of operating conditions, so that the final bearing remaining life migration prediction model can more accurately extract the cross-domain invariant degradation trend and improve the prediction accuracy.
[0030] 2) When calculating mutual information, this invention uses adaptive empirical mutual information to accurately assess the similarity of data distributions in two domains, thereby improving the performance of remaining lifetime migration prediction. Attached Figure Description
[0031] Figure 1 The waveform diagrams are of the outer ring of the LDKUER204 bearing under operating conditions of 1000rpm and 2000rpm.
[0032] Figure 2 This is a simplified model diagram of a rolling bearing;
[0033] Figure 3 This is a flowchart illustrating a specific implementation of the rolling bearing remaining life prediction method based on heterogeneous dimensional adversarial networks according to the present invention.
[0034] Figure 4 This is a structural diagram of the heterogeneous dimensional adversarial network in this invention. Detailed Implementation
[0035] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.
[0036] To better illustrate the technical solution of the present invention, the technical principles of the present invention will be briefly explained first.
[0037] The lifetime point dimension directly determines the input size of the network model. A reasonable lifetime point length should cover at least one complete failure cycle to ensure that each lifetime point contains sufficient degradation information and impact features. If the dimension is too short, the lifetime point may not be able to fully represent the degradation state of the bearing, resulting in incomplete semantic information; conversely, if the dimension is too long, although it can contain richer feature information, it will significantly increase computational complexity. Therefore, the appropriate selection of the lifetime point dimension is crucial to ensuring the accuracy and stability of cross-domain remaining lifetime migration prediction.
[0038] When rolling elements pass through localized defects in the raceway or cage, damage such as material spalling causes the contact surface to lose its smoothness. According to Hertzian contact theory, the elastic contact at the defect point will cause a momentary increase in contact force, thus generating an impact signal. During bearing operation, this impact repeats periodically with the rotational cycle as its period, which is determined by the rotational speed and the location of the defect. Ultimately, it manifests as a series of repetitive pulses in the vibration signal. Although the amplitude of the impact signal may differ in different fault cycles due to variations in local load distribution, its basic waveform pattern remains consistent. Therefore, the lifetime point dimension should be set to the number of data points required to cover a complete fault cycle.
[0039] Figure 2 This is a simplified model diagram of a rolling bearing. (For example...) Figure 2 As shown, the key geometric parameters of a rolling bearing include: the number of rolling elements Z, and the diameter of the rolling elements. and pitch circle diameter Frequency conversion and contact angle Different types of failures correspond to different impact frequencies. For example, outer ring failures generate outer ring pass frequency (BPFO), inner ring failures generate inner ring pass frequency (BPFI), rolling element failures generate rolling element spin frequency (BSF), and cage failures generate fundamental frequency (FTF). The formulas for calculating their characteristic failure frequencies are as follows:
[0040] (1)
[0041] (2)
[0042] (3)
[0043] (4)
[0044] It should be noted that the pitch diameter It can be approximated by the bearing outer diameter D and the bearing inner diameter d:
[0045] (5)
[0046] In the formula and These represent the inner raceway diameter and the outer raceway diameter, respectively.
[0047] Assume that the rolling bearing has If there are several fault types, then under a certain operating condition, there are... Fault characteristic frequency , This constitutes a fault characteristic frequency set. The fault characteristic frequency set in this embodiment .
[0048] Finally, considering the sampling factor defined by the Nyquist sampling theorem, at the th In the case of class-specific failures, the life point dimension d of a rolling bearing can be reformulated as:
[0049] (6)
[0050] in, This indicates the sampling frequency of the original vibration signal. This represents the preset coefficient. and These represent the rolling bearing's state under a certain operating condition. The fault impact cycle and fault characteristic frequency of each fault.
[0051] Furthermore, to ensure semantic separability between different fault types, the lifetime point dimension should be chosen sufficiently to distinguish adjacent fault frequency components. Specifically, as the lifetime point dimension increases, the time span covered by the time-domain signal extends, resulting in higher frequency resolution in spectral analysis and more accurate characterization of fault characteristic frequency distributions. Conversely, if the lifetime point dimension is too small, insufficient frequency resolution may prevent the full revelation of weaker frequency domain features in the early degradation stage. Therefore, the selection principle for the lifetime point dimension must also satisfy the following inequality:
[0052] (7)
[0053] Based on equation (7) and the consideration of minimizing computational resources, the selection rule for the final lifetime point dimension of this invention is as follows:
[0054] (8)
[0055] in, This indicates the sampling frequency of the original vibration signal. This represents the preset coefficient. Represents the set of positive integers. .
[0056] Based on the above analysis, this invention proposes a method for predicting the remaining life of rolling bearings based on heterogeneous dimensional adversarial networks. Figure 3 This is a flowchart illustrating a specific implementation of the rolling bearing remaining life prediction method based on heterogeneous dimensional adversarial networks according to the present invention. Figure 3 As shown, the rolling bearing remaining life prediction method based on heterogeneous dimensional adversarial networks of the present invention includes the following steps:
[0057] S301: Obtain the source domain sample set:
[0058] Set the fault types of the rolling bearing according to actual needs, and record the number of fault types as follows: Based on actual needs, a specific operating condition is set as the source domain condition. Under this condition, raw vibration signals from several full-life cycles are collected from rolling bearings experiencing different faults. Several remaining life prediction samples are then extracted. The input samples for each remaining life prediction sample include... Vibration signal feature data at each lifetime point, labeled with the remaining lifetime of the last lifetime point of the input sample, constitute the source domain sample set, where the dimension of the lifetime points is... The following formula is used to determine it:
[0059] (9)
[0060] in, This indicates the sampling frequency of the original vibration signal. This represents the preset coefficient. Represents the set of positive integers. , This indicates the rolling bearing under source domain operating conditions. Fault characteristic frequency The set of fault characteristic frequencies constituted .
[0061] S302: Obtain the target domain sample set:
[0062] For target domain operating conditions requiring remaining life prediction, several raw vibration signals of the rolling bearing are collected under the target domain operating conditions, and then several signals containing... Vibration signal feature data at each lifetime point are used as input samples for the target domain, thus forming the target domain sample set; where the dimension of the lifetime points is... The following formula is used to determine it:
[0063] (10)
[0064] in, , This indicates the rolling bearing under the target domain operating conditions. Fault characteristic frequency The set of fault characteristic frequencies constituted.
[0065] S303: Constructing Heterogeneous Dimensional Adversarial Networks
[0066] To achieve heterogeneous dimension alignment and better complete the remaining lifetime migration prediction task, this invention constructs a heterogeneous dimension adversarial network. Figure 4 This is a structural diagram of the heterogeneous dimensional adversarial network in this invention. For example... Figure 4 As shown, the source domain feature extractor in this invention Target domain feature extractor Remaining life predictor Mutual Information Calculator , , , ,and These represent the weight parameters for each module. Each module will then be explained in detail below.
[0067] Source Domain Feature Extractor Source domain data features are used to extract source domain input samples and sent to the remaining lifetime predictor. Mutual Information Calculator .
[0068] Target Domain Extractor Used to extract target domain data features from input samples and send them to the mutual information calculator. .
[0069] Remaining life predictor It is used to predict the remaining lifespan based on the characteristics of the input data.
[0070] Mutual Information Calculator Used to calculate mutual information between source domain data features and target domain data features.
[0071] As described above, this invention sets up two private feature extractors for the source and target domains respectively to extract latent representations. This design differs from the shared feature extractor structure commonly used in traditional models. This strategy enhances the model's sensitivity to differences in cross-domain feature distributions because the two private extractors amplify this difference, thereby improving the domain confusion capability of the HDAN model. In this embodiment, the source domain feature extractor... and target domain feature extractor Using the same structure, including four cascaded one-dimensional convolutional layers, data features are extracted layer by layer.
[0072] S104: Training Heterogeneous Dimensional Adversarial Networks:
[0073] A heterogeneous dimensional adversarial network is trained using source and target domain sample sets. Its optimization objectives mainly include two points: 1) minimizing the remaining lifetime prediction error of the source domain through supervised training; 2) maximizing the mutual information between the source and target domain features through an adversarial training mechanism. Therefore, the specific training method is as follows:
[0074] In each training batch, samples are selected from the source domain sample set and the target domain sample set respectively. The input samples constitute input sample pairs , Then input the source domain sample and target domain input samples Input source domain feature extractor respectively and target domain feature extractor To obtain the source domain data features and target domain data features Remaining life predictor Based on source domain data characteristics Obtain the predicted remaining lifespan Then, the predicted loss of remaining lifetime is calculated using the following formula. :
[0075] (11)
[0076] in, Represents the source domain input sample Corresponding real tags , This represents the preset remaining lifetime prediction loss function. In this embodiment, the calculation formula for the remaining lifetime prediction loss function is as follows:
[0077] (12)
[0078] Mutual Information Calculator Calculate source domain data features and target domain data features mutual information Then, the mutual information loss is calculated using the following formula. :
[0079] (13)
[0080] in, This represents the preset mutual information loss function.
[0081] In this invention, it is necessary to minimize the predicted loss of remaining lifetime; therefore, the predicted loss is based on the remaining lifetime. For remaining lifetime predictor Parameters are updated. The two feature extractors and the mutual information calculator employ an adversarial mechanism, requiring the maximization of the mutual information between the data features obtained by the two feature extractors while simultaneously ensuring the effectiveness of the mutual information calculator module. Therefore, the negative of the mutual information loss is used. Source domain feature extractor and target domain feature extractor Perform parameter updates based on mutual information loss. Mutual Information Calculator Update the parameters.
[0082] Regarding the mutual information loss function, due to the different lifetime point dimensions, the feature dimensions of the data feature spaces obtained from the source and target domains are also inconsistent. Commonly used distribution difference measurement methods (such as maximum mean difference and correlation alignment) are difficult to effectively characterize cross-domain distribution differences under heterogeneous feature dimensions. To address this, this embodiment designs an adaptive empirical mutual information calculation method. Its core idea is to quantify the information correlation between domains through mutual information, thereby reflecting the degree of similarity between cross-domain data distributions: the larger the mutual information value, the higher the similarity; the smaller the mutual information value, the lower the similarity.
[0083] Given two random variables and Mutual information between the source and target domains It can be defined as:
[0084] (14)
[0085] In the formula and Representing variables respectively and The marginal probability density; express and The joint probability density; and They represent and The sample space. According to the following definition of KL divergence:
[0086] (15)
[0087] Equation (11) can be redefined as:
[0088] (16)
[0089] Equation (13) shows that precise calculation is only feasible when the probability distribution of discrete or continuous variables is explicitly known. However, this condition is not met in practical engineering. To overcome this limitation, this embodiment uses the Donsker-Varadhan variational representation theorem to calculate the mutual information of unknown continuous variables. Equation (13) can be restated as follows:
[0090] (17)
[0091] in, This represents a mutual information estimation neural network.
[0092] From equation (14), it can be seen that in order to make the right term Obtaining the supremum, the function space It should include as many functions as possible. Given the powerful nonlinear approximation capabilities of neural networks, this embodiment directly uses them as the function hypothesis space, i.e. Figure 4 The mutual information calculator shown Therefore, the final computational form of mutual information can be expressed as:
[0093] (18)
[0094] Based on the above analysis, the calculation formula for the mutual information loss function in this embodiment can be obtained as follows:
[0095] (19)
[0096] S105: Constructing a target domain remaining lifetime prediction model:
[0097] Extract the target domain extractor from the trained heterogeneous dimensional adversarial network. and remaining lifetime predictor The remaining lifetime prediction model constitutes the target domain;
[0098] S106: Predict the remaining lifetime under target domain operating conditions:
[0099] The time to be predicted under the target domain operating conditions and before The vibration signal characteristic data of each lifetime point constitute the input sample, which is then input into the remaining lifetime prediction model of the target domain to obtain the predicted remaining lifetime.
[0100] Example
[0101] To better illustrate the technical solution of this invention, specific examples are used to experimentally verify the invention. The rolling bearing lifecycle dataset in this embodiment comes from Xi'an Jiaotong University, with a sampling frequency of 25.6 kHz, a sampling interval of 1 minute, and a sampling duration of 1.28 seconds. The bearing dataset includes three operating conditions, with load and speed information of 12 kN & 2100 rpm, 11 kN & 2250 rpm, and 10 kN & 2400 rpm, respectively. Five operational failure monitoring vibration signals were collected under each operating condition. This invention will select the operating condition data numbered B1_1, B1_2, and B1_3 for remaining life migration prediction experiments.
[0102] Based on the aforementioned rolling bearing full life cycle dataset, six remaining life migration prediction tasks across operating conditions were constructed, as shown in Table 1. To ensure the reliability of the experimental results, all models were executed five times in each migration prediction task.
[0103] To demonstrate the superiority of the remaining lifetime prediction method proposed in this invention, Score and Mean Absolute Error (MAE) are used as evaluation metrics. A higher Score indicates better prediction performance, and a lower MAE value indicates better prediction performance. Five remaining lifetime migration prediction network models are selected as comparison methods: WDCNN, AdaBN, ID-AE, DDC, and DANN. Table 1 compares the remaining lifetime prediction performance of the present invention and the comparison methods in this embodiment.
[0104]
[0105] Table 1
[0106] As shown in Table 1, by comparing the experimental data results (mean and standard deviation, with bold indicating the best result under this transfer task), it can be concluded that the rolling bearing life prediction method based on heterogeneous dimensional adversarial networks of this invention has higher transfer prediction accuracy, stronger robustness and generalization ability.
[0107] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.
Claims
1. A method for predicting the remaining life of rolling bearings based on heterogeneous dimensional adversarial networks, characterized in that... Includes the following steps: S1: Set the fault types of the rolling bearing according to actual needs, and record the number of fault types as follows: ; Based on actual needs, a specific operating condition is set as the source domain condition. Under this condition, raw vibration signals from several full-life cycles are collected from rolling bearings experiencing different fault conditions. Several remaining life prediction samples are then extracted. The input samples for each remaining life prediction sample include... Vibration signal feature data at each lifetime point, labeled with the remaining lifetime of the last lifetime point of the input sample, constitute the source domain sample set, where the dimension of the lifetime points is... The following formula is used to determine it: ， in, This indicates the sampling frequency of the original vibration signal. This represents the preset coefficient. Represents the set of positive integers. , This indicates the rolling bearing under source domain operating conditions. Fault characteristic frequency The fault characteristic frequency set constituted ; S2: For the target domain operating conditions requiring remaining life prediction, several raw vibration signals of the rolling bearing are collected under the target domain operating conditions, and then several signals containing... Vibration signal feature data at each lifetime point are used as input samples for the target domain, thus forming the target domain sample set; where the dimension of the lifetime points is... The following formula is used to determine it: ， in, , This indicates the rolling bearing under the target domain operating conditions. Fault characteristic frequency The set of fault characteristic frequencies constituted; S3: Construct heterogeneous dimensional adversarial networks, including source domain feature extractors. Target domain feature extractor Remaining life predictor Mutual Information Calculator ,in: Source Domain Feature Extractor Source domain data features are used to extract source domain input samples and sent to the remaining lifetime predictor. Mutual Information Calculator ; Target Domain Extractor Used to extract target domain data features from input samples and send them to the mutual information calculator. ; Remaining life predictor Used to predict remaining lifespan based on input data characteristics; Mutual Information Calculator Used to generate mutual information estimates of source domain data features and target domain data features; S4: Train a heterogeneous dimensional adversarial network using source domain sample sets and target domain sample sets. The specific method is as follows: In each training round, samples are selected from the source domain sample set and the target domain sample set respectively. The input samples constitute input sample pairs , Then input the source domain sample and target domain input samples Input source domain feature extractor respectively and target domain feature extractor To obtain the source domain data features and target domain data features Remaining life predictor Based on source domain data characteristics Obtain the predicted remaining lifespan Then, the predicted loss of remaining lifetime is calculated using the following formula. : ， in, Represents the source domain input sample Corresponding real tags , This represents the preset remaining lifetime prediction loss function; Mutual Information Calculator Calculate source domain data features and target domain data features mutual information estimate Then, the mutual information loss is calculated using the following formula. : ， in, This represents the preset mutual information loss function; Loss prediction based on remaining lifespan For remaining lifetime predictor Perform parameter updates based on the negative of the mutual information loss. Source domain feature extractor and target domain feature extractor Perform parameter updates based on mutual information loss. Mutual Information Calculator Update the parameters; S5: Extract the target domain extractor from the trained heterogeneous dimensional adversarial network. and remaining lifetime predictor The remaining lifetime prediction model constitutes the target domain; S6: The time to be predicted under the target domain operating conditions and prior to... The characteristic data of the rolling bearing vibration signal at each life point constitute the input sample, which is then input into the remaining life prediction model of the target domain to obtain the predicted remaining life.
2. The method for predicting the remaining life of rolling bearings according to claim 1, characterized in that, The failure types of the rolling bearing include outer ring failure, inner ring failure, rolling element failure, and cage failure.
3. The method for predicting the remaining life of rolling bearings according to claim 1, characterized in that, The source domain feature extractor and target domain feature extractor It adopts the same structure, including 4 cascaded one-dimensional convolutional layers.
4. The method for predicting the remaining life of rolling bearings according to claim 1, characterized in that, The formula for calculating the remaining lifetime prediction loss function is as follows: 。 5. The method for predicting the remaining life of rolling bearings according to claim 1, characterized in that, The formula for calculating the mutual information loss function is as follows: 。