A bearing residual service life prediction method based on feature migration
By constructing source and target networks based on feature transfer and using fault features to assist degradation features, the problems of high data requirements and unstable accuracy in bearing remaining service prediction are solved, achieving high-precision and highly interpretable prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ARTIFICIAL INTELLIGENCE RES INST OF HEFEI COMPREHENSIVE NAT SCI CENT (ANHUI ARTIFICIAL INTELLIGENCE LAB)
- Filing Date
- 2023-08-18
- Publication Date
- 2026-07-28
AI Technical Summary
Existing technologies for predicting the remaining service life of bearings suffer from limited processing capabilities for low-quality data and high requirements for domain knowledge. This is especially true when it is difficult to obtain complete equipment lifecycle degradation data, leading to unstable prediction accuracy and poor interpretability.
A feature transfer-based approach is adopted. By constructing a source network and a target network, and using equipment fault data to assist in the prediction of equipment life cycle degradation data, a fault feature transfer loss function is designed to realize the transfer of fault features to degradation features, thereby assisting in the prediction of remaining service life.
It reduces the need for remaining useful life data, improves prediction accuracy and stability, and enhances the interpretability of predictions.
Smart Images
Figure CN116893055B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of predictive maintenance technology, and more specifically to a method for predicting the remaining service life of bearings based on feature migration. Background Technology
[0002] Currently, intelligent manufacturing is receiving increasing attention, and one of the core aspects of intelligent manufacturing is predictive maintenance of equipment. Bearings, as crucial components widely used in machinery, primarily function to reduce friction and ensure smoother rotation, significantly impacting production safety. Furthermore, as wear parts, bearings have a high failure rate; therefore, predicting the remaining service life of bearings has significant application value.
[0003] Traditional methods for predicting remaining useful life can be broadly categorized into model-based and statistical methods. Model-based methods require multi-domain knowledge to construct physical models that accurately reflect machine degradation. However, acquiring domain knowledge is challenging, and constructing accurate physical models is difficult due to the complexity of system architecture and operating environment. Statistical methods focus on building stochastic models describing the degradation process to predict remaining useful life based on monitored equipment degradation variables. However, these methods have limited ability to handle low-quality data.
[0004] In recent years, the combination of deep learning and remaining useful life (UPS) prediction has become a trend due to its ability to improve prediction accuracy. Data-driven approaches utilize deep learning models to establish potential relationships between equipment monitoring data and UPS labels. These methods possess powerful degradation feature extraction capabilities, enabling them to effectively handle massive and complex structured data and reduce the need for domain knowledge. However, these methods rely on large amounts of UPS data, which is extremely difficult to obtain in reality.
[0005] Compared to complete equipment lifecycle degradation data, acquiring equipment fault data is less costly. Since fault characteristics are crucial for predicting remaining service life, this invention proposes a feature transfer-based method for predicting bearing remaining service life. This method aims to utilize fault characteristics to assist in remaining service life prediction, thereby reducing the need for remaining service life data. Summary of the Invention
[0006] To address the aforementioned technical problems, this invention provides a method for predicting the remaining service life of bearings based on feature migration.
[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0008] A method for predicting the remaining service life of bearings based on feature transfer includes the following steps:
[0009] Step 1: Acquire equipment fault data and equipment full-cycle degradation data, and perform data preprocessing; the amount of equipment fault data is greater than the amount of equipment full-cycle degradation data.
[0010] Step 2: Construct the source network, which includes a fault feature extraction module and a fault classifier; the fault feature extraction module outputs fault features. Fault classifiers classify faults based on fault characteristics. Equipment failure type Perform classification; design the fault classification loss function. class And use equipment failure data to train the source network;
[0011] Step 3: Construct the target network, which includes a degradation feature extractor and a remaining lifetime predictor; the degradation feature extractor E R Output degradation features Remaining useful life predictor C R Based on degradation characteristics For the remaining service life of the equipment Perform prediction; design the remaining useful life prediction loss function (Loss). rul And use the device's full-cycle degradation data to train the target network;
[0012] Step 4: Design the Fault Feature Migration Loss Function TL Through fault characteristics auxiliary degradation characteristics Learning, and then transferring fault features.
[0013] Furthermore, step one specifically includes:
[0014] Step S11, collect equipment fault data: the equipment vibration signal collected at time t is recorded as... 's' represents the number of sensors, and the sliding pane method is used to generate equipment fault training samples. T represents the length of the sliding pane, and the training samples for device failures. The corresponding fault type is denoted as y. F The final total number of equipment fault training samples is N;
[0015] Step S12, collect equipment full-cycle degradation data: the equipment vibration signal collected at time t will be recorded as... s represents the number of sensors, and the sliding pane method is used to generate equipment degradation training samples. T represents the length of the sliding pane; device degradation training samples The corresponding remaining useful life is denoted as y. RThe final total number of equipment degradation training samples is M; the total number of equipment degradation training samples M is less than one-tenth of the total number of equipment failure training samples N, that is, M < 0.1N.
[0016] Furthermore, step two specifically includes:
[0017] Step S21, construct the fault feature extraction module E of the source network. F Fault Feature Extraction Module E F It consists of three time attention modules, E F The temporal attention module in the code is formalized as follows:
[0018]
[0019]
[0020]
[0021]
[0022]
[0023] in All are trainable parameters; s is the number of sensors. It is the output feature of the (i+1)th layer temporal attention module. These are the output features of the i-th layer temporal attention module. When i = 0, For fault query characteristics at time t; The fault key value feature at time t, The fault matching features at time t, Let be the fault attention matrix at time t, softmax(·) be the nonlinear activation function, and d be the number of fault types.
[0024] The output of the previous temporal attention module is used as the input of the next temporal attention module, and the output of the last temporal attention module is used as the fault feature extractor E. F Output fault characteristics
[0025] Step S22, construct the fault classifier C for the source network. F Based on fault characteristics Equipment failure type Classification is performed; Fault classifier C F Formalized as follows:
[0026]
[0027] in Based on fault characteristics The one-dimensional vector obtained by unfolding All are trainable parameters; d represents the number of fault types.
[0028] Step S23: Design the fault classification loss function Loss class The source network is trained using equipment fault data.
[0029]
[0030] B is the size of the data batch during the training of the source network.
[0031] Furthermore, step three specifically includes:
[0032] Step S31: Construct the degenerate feature extractor E for the target network R Degradation Feature Extraction Module E R It consists of three time attention modules, E R The temporal attention module in the code is formalized as follows:
[0033]
[0034]
[0035]
[0036]
[0037]
[0038] in All are trainable parameters, where s is the number of sensors. These are the output features of the (i+1)th layer's temporal attention module. These are the output features of the i-th layer temporal attention module. When i = 0, Features for querying remaining lifespan at time t The remaining lifetime key value feature at time t Matching features for remaining lifetime at time t Let t be the remaining lifetime attention matrix, softmax(·) be the nonlinear activation function, and d be the number of fault types.
[0039] The output of the previous temporal attention module is used as the input of the next temporal attention module, and the output of the last temporal attention module is used as the degenerate feature extractor E. R Degradation characteristics of output
[0040] Step S32: Construct the remaining lifetime predictor C for the target network R According to the characteristics of degradation For the remaining service life of the equipment Make predictions; Remaining useful life predictor C R Formalized as follows:
[0041]
[0042] in Degenerative characteristics The one-dimensional vector obtained by unfolding; All of these are trainable parameters;
[0043] Step S33: Design the Loss function for predicting remaining useful life rul The target network was trained using data from the device's entire lifecycle degradation.
[0044]
[0045] B is the size of the data batch during the training of the target network.
[0046] Furthermore, step four specifically includes:
[0047] Fault Feature Transfer: Utilizing Fault Features auxiliary degradation characteristics The loss function for fault feature transfer learning. TL as follows:
[0048]
[0049] First, the source network is trained to classify device faults. Then, the parameters of the source network are fixed, and the loss function is applied. rul and Loss TL The parameters of the target network are trained together.
[0050] Compared with the prior art, the beneficial technical effects of the present invention are:
[0051] This invention first pre-trains a source network based on temporal attention for fault diagnosis. Then, it uses temporal attention to construct a target network and fits the relationship between time series and remaining useful life. Finally, it designs a feature transfer loss function to transfer features extracted from the source network to the target network, helping the target network learn fault features and better predict the remaining useful life of equipment. This invention utilizes fault features to assist in remaining useful life prediction, thereby reducing the need for remaining useful life data. It overcomes the shortcomings of existing methods, such as insufficient stability and poor interpretability. Attached Figure Description
[0052] Figure 1 This is a flowchart of the prediction method of the present invention. Detailed Implementation
[0053] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.
[0054] The device described in this embodiment is a bearing.
[0055] The bearing remaining service life prediction method based on feature transfer in this invention specifically includes:
[0056] Step 1: Acquire a large amount of equipment failure data and a small amount of equipment lifecycle degradation data, and perform data preprocessing.
[0057] Step one specifically includes the following sub-steps S11 to S22:
[0058] S11: Collect a large amount of equipment fault data. The equipment vibration signal collected at time t is denoted as... 's' represents the number of sensors. The sliding pane method is used to generate equipment fault training samples. T represents the length of the sliding pane, y F Indicates equipment failure training samples The corresponding fault type. Assume the total number of equipment fault training samples generated is N.
[0059] S12: Collect a small amount of equipment lifecycle degradation data. The equipment vibration signal collected at time t is denoted as... 's' represents the number of sensors. The sliding pane method is used to generate equipment degradation training samples. T represents the length of the sliding pane. y R Indicates equipment degradation training samples The corresponding remaining service life. Assume the total number of equipment degradation training samples generated is M. Since the degradation cycle is usually quite long, the total number of equipment degradation training samples M is usually much smaller than the total number of equipment failure training samples N.
[0060] Step 2: Train the source network using equipment fault data.
[0061] Step two specifically includes the following sub-steps S21 to S23:
[0062] S21: Constructing the source network fault feature extractor E F Fault Feature Extraction Module E F It consists of three time-attention modules. The time-attention module is formalized as follows:
[0063]
[0064]
[0065]
[0066]
[0067]
[0068] in All are trainable parameters, where s is the number of sensors. It is the output feature of the time attention module in the (i+1)th layer. These are the output features of the i-th layer temporal attention module. When i = 0,
[0069] Repeat this process three times to obtain the source network fault feature extractor E. F The final output fault characteristics
[0070] S22: Constructing the source network fault classifier C F Based on fault characteristics Classify equipment fault types. Source network fault classifier C F Formalized as follows:
[0071]
[0072] in Based on fault characteristics The one-dimensional vector obtained by expansion. All of these are trainable parameters. d represents the fault type. Since the health status of the equipment needs to be considered, the total number of categories is d+1.
[0073] S23: Design the fault classification loss function Loss class :
[0074]
[0075] B is the size of the data batch during the training of the source network.
[0076] Step 3: Train the target network using a small amount of data from the full lifecycle degradation of the device.
[0077] Step three specifically includes the following sub-steps S31 to S33:
[0078] S31: Construct the target network degenerate feature extractor E R Degradation Feature Extraction Module E RIt consists of three time-attention modules. The time-attention module is formalized as follows:
[0079]
[0080]
[0081]
[0082]
[0083]
[0084] in All are trainable parameters, where s is the number of sensors. It is the output feature of the time attention module in the (i+1)th layer. These are the output features of the i-th layer temporal attention module. When i = 0,
[0085] Repeat this process three times to obtain the target network degradation feature extractor E. R Final output degradation features
[0086] S32: Constructing a predictor of the remaining lifetime of the target network C R According to the characteristics of degradation Predicting the remaining lifespan of equipment. Target network remaining lifespan predictor C R Formalized as follows:
[0087]
[0088] in Degenerative characteristics The one-dimensional vector obtained by expansion. All of these are trainable parameters.
[0089] S33: Loss function for predicting remaining useful life rul :
[0090]
[0091] B is the size of the data batch during the training of the target network.
[0092] Step 4: Fault Feature Migration.
[0093] Step four specifically includes the following sub-steps S41 to S42:
[0094] S41: Due to the typically small number of training samples for device degradation, it is difficult to obtain a good degradation feature. Therefore, it is necessary to utilize fault characteristics. auxiliary degradation characteristics The learning process. The fault feature transfer loss function is as follows:
[0095]
[0096] The goal is to minimize the degradation characteristics. Approaching fault characteristics
[0097] S42: The specific training process is as follows: First, train the source network to classify device faults. Then, fix the parameters of the source network and apply the loss function. rul and Loss TL Together, we train the parameters of the target network.
[0098] It should be noted that the fault feature migration loss designed in this embodiment can effectively reduce the need for equipment lifecycle degradation data by leveraging equipment fault data.
[0099] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.
[0100] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A method for predicting the remaining service life of bearings based on feature transfer, comprising the following steps: Step 1: Acquire equipment fault data and equipment full life cycle degradation data, and perform data preprocessing; The number of equipment failure data is greater than the number of equipment full life cycle degradation data. Step 2: Construct the source network, which includes a fault feature extraction module and a fault classifier; the fault feature extraction module outputs fault features. The fault classifier classifies faults based on their characteristics. Equipment failure type Perform classification; design a fault classification loss function. And use equipment failure data to train the source network; Step 3: Construct the target network, which includes a degradation feature extractor and a remaining lifetime predictor; Degeneracy Feature Extractor Output degradation features Remaining useful life predictor Based on degradation characteristics For the remaining service life of the equipment Make predictions; Design a loss function for predicting remaining useful life. And use the device's full-cycle degradation data to train the target network; Step 4: Design the fault feature migration loss function Through fault characteristics auxiliary degradation characteristics Learning to transfer fault features; Fault feature transfer includes: leveraging fault features auxiliary degradation characteristics Learning, loss function for fault feature transfer as follows: First, train the source network to classify device faults. Then, fix the parameters of the source network and apply the loss function. and The parameters of the target network are trained together; Step one specifically includes: Step S11, collect equipment fault data: (The data will be collected in the following steps.) The vibration signals of the equipment collected at all times are recorded as follows: , The number of sensors is represented, and the sliding pane method is used to generate equipment fault training samples. , Indicates the length of the sliding pane, and the training samples for device failures. The corresponding fault type is denoted as The final total number of equipment fault training samples obtained is ; Step S12, collect equipment full-cycle degradation data: This will be... The vibration signals of the equipment collected at all times are recorded as follows: , The number of sensors is represented by a sliding pane method to generate device degradation training samples. Indicates the length of the sliding pane; device degradation training samples The corresponding remaining service life is denoted as The final total number of equipment degradation training samples is M; the total number of equipment degradation training samples M is less than the total number of equipment failure training samples. One-tenth, that is .
2. The bearing remaining service life prediction method based on feature transfer according to claim 1, characterized in that, Step two specifically includes: Step S21: Construct a fault feature extraction module for the source network. Fault Feature Extraction Module It consists of three time-attention modules. The temporal attention module in the code is formalized as follows: ； ； ； ； ； in , , , , , All of these are trainable parameters; It refers to the number of sensors. It is the output feature of the (i+1)th layer temporal attention module. It is the output feature of the i-th layer temporal attention module, when hour, ; For fault query characteristics at time t; The fault key value feature at time t, The fault matching features at time t, Let be the fault attention matrix at time t. It is a non-linear activation function. Number of fault types; The output of the previous temporal attention module is used as the input of the next temporal attention module, and the output of the last temporal attention module is used as the fault feature extractor. Output fault characteristics ; Step S22, construct the fault classifier for the source network. Based on fault characteristics Equipment failure type Classification; Fault Classifier Formalized as follows: ； in Based on fault characteristics The one-dimensional vector obtained by unfolding , All of these are trainable parameters; Number of fault types; Step S23: Design the fault classification loss function The source network is trained using equipment fault data. ； It is the size of the data batch during the training of the source network.
3. The bearing remaining service life prediction method based on feature transfer according to claim 1, characterized in that, Step three specifically includes: Step S31: Construct the degenerate feature extractor for the target network Degradation Feature Extraction Module It consists of three time-attention modules. The temporal attention module in the code is formalized as follows: ； ； ； ； ； in , , , , , All of these are trainable parameters. It refers to the number of sensors. These are the output features of the (i+1)th layer's temporal attention module. It is the output feature of the i-th layer temporal attention module, when hour, ; Features for querying remaining lifespan at time t The remaining lifetime key value feature at time t Matching features for remaining lifetime at time t The remaining lifetime attention matrix at time t It is a non-linear activation function. Number of fault types; The output of the previous temporal attention module is used as the input of the next temporal attention module, and the output of the last temporal attention module is used as the degenerate feature extractor. Degradation characteristics of output ; Step S32: Construct a predictor of the remaining lifetime of the target network According to the characteristics of degradation For the remaining service life of the equipment Make predictions; Remaining useful life predictor Formalized as follows: ； in Degenerative characteristics The one-dimensional vector obtained by unfolding; , All of these are trainable parameters; Step S33: Design the remaining useful life prediction loss function The target network was trained using data from the device's entire lifecycle degradation. ； It is the size of the data batch during the training of the target network.