Continuous remaining service life prediction method based on multi-branch transverse amplification network
Through multi-branch lateral amplification network and memory weight constraint regularization terms, combined with time-frequency feature fusion technology, the problem of insufficient adaptability of equipment RUL prediction models in different components and environments in the existing technology is solved, and high-precision lifelong RUL prediction and knowledge retention.
Patent Information
- Application Number
- CN202410714779.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-04
- Publication Date
- 2024-11-08
AI Technical Summary
The existing equipment residual service life prediction model is mainly concentrated in a single component and a specific operating condition. It is difficult to effectively learn and predict in different components and dynamically changing environments, and lacks lifelong RUL prediction capabilities.
Using a multi-branch lateral amplification network (MBHAN) method, the device status monitoring data is continuously collected, and the fast Fourier transform and normalization process is carried out, and the MBHAN model is constructed, using memory weight to constrain regularization terms and time-frequency characteristics. Fusion technology to achieve continuous RUL prediction of mechanical equipment.
High-precision RUL prediction in different components and dynamic environments is achieved, reducing catastrophic forgetting, protecting the knowledge of previous tasks from interference from new tasks, and improving the adaptability and accuracy of the model.
Smart Images

Figure CN118917166A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of mechanical equipment remaining service life prediction, and relates to a continuous remaining service life prediction method based on a multi-branch lateral amplification network. Background Art
[0002] The safe and reliable operation of equipment is the core content to ensure the production activities of the factory. In order to ensure the operating status of the equipment and formulate a reasonable maintenance plan, it is very important to study a life prediction method with high accuracy and high robustness. As one of the core components, rolling bearings are widely used in aircraft, high-speed rail, etc. However, bearings are often in harsh working environments, which can easily cause bearings to fail, such as wear and pitting, and thus affect the normal operation of the equipment. Therefore, it is necessary to make a high-precision prediction of the remaining service life of rolling bearings.
[0003] Existing research mainly focuses on the remaining useful life (RUL) prediction task of a single component, aiming to improve the accuracy of RUL prediction of a single component under specific operating conditions. Some studies focus on the cross-operating condition RUL prediction task of the same component. In other words, these models are often only applicable to specific components and operating conditions. However, in actual engineering, it is common to continuously collect condition monitoring data of various components under different operating conditions. Current RUL prediction models are generally trained on data sets with specific distributions, which limits their ability to continuously and adaptively learn degradation information in dynamic environments such as different components and dynamically changing operating conditions. Therefore, there is an urgent need to develop a new RUL prediction method to effectively achieve lifelong RUL prediction task learning and RUL prediction. Summary of the invention
[0004] In view of this, the purpose of the present invention is to provide a continuous remaining useful life prediction method based on a multi-branch lateral amplification network, so as to realize the learning and RUL prediction of the lifelong RUL prediction task of mechanical equipment.
[0005] In order to achieve the above object, the present invention provides the following technical solutions:
[0006] A method for predicting the continuous remaining useful life based on a multi-branch lateral amplification network specifically comprises the following steps:
[0007] S1: Continuously collect condition monitoring data of mechanical equipment under different working conditions, and obtain frequency domain signals from them through fast Fourier transform; the frequency domain signals contain specific frequency components that indicate fault characteristics or degradation modes;
[0008] S2: Normalize the frequency domain signal;
[0009] S3: Divide the normalized data into normal data and degraded data;
[0010] S4: Construct the MBHAN model and use the loss function with the MWC regularization term to train and optimize the MBHAN model, where MBHAN represents a multi-branch horizontal expansion network and MWC represents a memory weight constraint;
[0011] S5: Use the optimized MBHAN model to predict the remaining useful life of mechanical equipment.
[0012] Further, in step S4, the constructed MBHAN model includes a feature extraction backbone network and a task branch network;
[0013] For the first RUL prediction task, the MBHAN network structure can be expressed as express After network ψ, RUL represents the remaining useful life; assuming that the feature extraction backbone network consists of N feature extraction networks, each feature extraction network ψ has a different number of RUL mapping networks MBHAN is represented by:
[0014]
[0015] in, It is MBHAN; i is the i-th feature extraction network; yes i The j-th RUL mapping network in; is the network combination operator;
[0016] On this basis, when a new RUL prediction task needs to be processed, the task branch network or feature extraction backbone network needs to be expanded.
[0017] Furthermore, in step S4, when a new RUL prediction task needs to be processed, the specific steps of expanding the task branch network or the feature extraction backbone network are as follows:
[0018] 1) Determine task similarity;
[0019] In the setting of continuous RUL prediction task, the data of previous tasks are not available, and the data of the current task consists of data and corresponding labels; therefore, the training data of the current task is directly input into the model Ψ, and then the predicted RUL of all RUL mapping networks can be obtained, which can be expressed as:
[0020]
[0021] Among them, p ij Indicated by The predicted RUL obtained; the average RUL of each feature extraction network is calculated and expressed as:
[0022]
[0023] Then, calculate P mean The mean absolute error (MAE) vector E between the actual RUL label:
[0024] E=|ARul-P mean |=[e1,e2…,e n ]
[0025] Where ARul is the actual URL label; the minimum error in E will be returned; if e i If the difference is smaller than the preset threshold, it means that the current task is different from the one using e i Otherwise, it is considered that there are large differences between all previous tasks and the current task.
[0026] 2) Extending the Task Branch Network: Typically, the feature extraction network extracts generic degradation features from the input monitoring data and does not have task-specific feature extraction and mapping capabilities. i The task is similar to that of i Add a new RUL mapping network later (new task branch), namely the feature extraction network ψ i By the network and Network Share; in this case, the structure of MBHAN is updated as:
[0027]
[0028] In order to make full use of the relevant information of previous tasks and speed up the network convergence, the new task branch network The weights of the previous network The weight parameters are initialized; In addition, considering the network ψ i When learning new task knowledge, it may be affected, and the MWC mechanism is used to consolidate the learned task knowledge; therefore, the loss function loss combines the mean square error (MSE) loss and the MWC regularization term loss MWC , used to update network parameters, can be expressed as:
[0029]
[0030] Where pRul is the predicted RUL, λ is the weight of the MWC regularization term, is the network i The importance of parameters, is the current task in the network ψ i The parameters on is the previous task in the network ψ i The optimal parameters on ;
[0031] During the training process, in addition to the network and network ψ i All other networks are frozen.
[0032] 3) Expanding the feature extraction backbone network: When the current task is dissimilar to all previous tasks, a new feature extraction network ψ needs to be expanded new At the same time, due to the new feature extraction network ψ new There is no branch network for any specific task, you need to add a task branch network Therefore, in this case, the structure of MBHAN is updated to:
[0033]
[0034] In order to focus on learning the knowledge representation of the current task, the newly expanded network Instead of using the parameters of the previous network for initialization, a random initialization method is used; at the same time, the loss function only uses the mean square error (MSE) loss for network update, which can be expressed as:
[0035] loss=(PRul-ARul) 2
[0036] Similar to 2), the parameters of the extended network are isolated from the parameters of other networks, that is, when training a new RUL prediction task, in addition to the network Except for , the parameters of all other networks are not updated.
[0037] Further, in step S4, the feature extraction network is composed of two parallel sub-networks, which are used to extract the degradation information hidden in the time domain signal and the frequency domain signal respectively; the sub-network includes a convolution sub-network for spatial feature extraction, four TCN blocks for capturing long-term dependencies, and HAS for important feature selection and multi-scale degradation information modeling; the convolution sub-network uses a large convolution kernel to expand the receptive field and retain more information from the original input data as much as possible; then, the deep degradation features obtained from the two sub-networks are fused using a connection operation; wherein TCN represents a temporal convolutional network and HAS represents hierarchical self-attention.
[0038] Further, in step S4, the HAS mechanism is as follows: first, define the attention level of HSA as n; each level corresponds to a different local attention area; the lower levels focus on smaller local areas and can capture fine-grained features and dependencies within these areas; the higher levels have a wider attention range, enabling them to capture dependencies across larger areas; then, determine the block size w = [w1, w2, …w n ]; in each level, divide the input sequence into different subsequences according to the block size w; in the i-th attention level and in each subsequence, use linear transformation to generate local query, key and value vectors; use cosine similarity to measure the similarity between the local query vector and the local key vector; then, divide the obtained similarity score by the scaling factor Where D corresponds to the input dimension of the subsequence; then, the attention weight is calculated using the Softmax function and assigned to the local value vector; finally, the attention values of all levels are integrated and output through linear transformation.
[0039] Further, in step S4, the RUL mapping network follows the feature extraction network, including a TCN block for task-related feature extraction, an adaptive average pooling layer, and a fully connected layer for RUL mapping; the network is responsible for mapping the fused degraded features to the RUL of the mechanical device.
[0040] Further, in step S4, the RUL label is prepared: the RUL value is normalized into a RUL percentage, and the RUL percentage label is defined as:
[0041]
[0042] Among them, ARul t is the actual RUL percentage label at time t, ARul t The value range of R is 0 to 1; T and R t Respectively represent the total life and actual RUL at time t, Indicates the RUL corresponding to the initial degradation point.
[0043] Further, in step S5, when predicting the remaining useful life of mechanical equipment, for the first RUL prediction task, the constructed prediction network is ψ1 Its initial weights are randomly generated; the loss function adopts the mean square error (MSE) loss; for the subsequent RUL prediction task, MBHAN is used to continuously learn task knowledge by using only the current task data; however, the test data includes the test sets of all previous tasks and the current test set.
[0044] The beneficial effects of the present invention are:
[0045] (1) The hierarchical self-attention (HSA) mechanism designed in this paper can model multi-scale local dependencies, which enables the RUL prediction model to capture degradation features and dependencies at different granularity levels, achieving a wider attention range and more comprehensive representation of input monitoring data.
[0046] (2) The time-frequency feature fusion temporal convolutional network (TFFTCN) designed in the present invention fully utilizes the advantages of time domain signals and frequency domain signals. By combining these two types of signals, TFFTCN can improve the information representation capability required for accurate RUL prediction and make full use of the complementarity between time domain and frequency domain features, which is responsible for mapping the fused degradation features to the RUL of mechanical equipment.
[0047] (3) The multi-branch horizontal expansion network (MBHAN) constructed by the present invention promotes the RUL prediction model to keep important parameters close to their original values during the training process of new tasks by adding a memory weight constraint (MWC) regularization term to the task loss function. This MWC helps to mitigate catastrophic forgetting and promotes continuous RUL prediction.
[0048] (4) The present invention uses task branch network enhancement based on semi-parameter sharing for tasks with smaller differences, and uses feature extraction backbone network enhancement based on parameter isolation for tasks with larger differences. This method effectively protects the acquired knowledge from being disturbed by new tasks.
[0049] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:
[0051] Figure 1 This is the HSA schematic diagram;
[0052] Figure 2 It is a structural diagram of DDC;
[0053] Figure 3 Schematic diagram of the TFFTCN structure;
[0054] Figure 4 This is the schematic diagram of MBHAN;
[0055] Figure 5RUL prediction results of TFFCNN, including (a) Task 1, (b) Task 3. DETAILED DESCRIPTION
[0056] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.
[0057] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.
[0058] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.
[0059] See also Figure 1 to Figure 5 , is a continuous remaining useful life prediction method based on a multi-branch horizontal expansion network, using TFFTCN (time-frequency feature fusion temporal convolutional network) and MBHAN (multi-branch horizontal expansion network), which can perform continuous RUL prediction tasks, and its process is summarized as follows:
[0060] 1) Frequency domain data conversion: Frequency domain signals contain specific frequency components that indicate fault characteristics or degradation modes. Using frequency domain signals can provide valuable insights into RUL prediction of mechanical equipment such as rolling bearings. Use Fast Fourier Transform (FFT) to obtain frequency domain signals:
[0061]
[0062] Where x(t) represents the original vibration signal, Represents a frequency domain signal.
[0063] 2) Normalization: Consistency and normalization of the input are important. Any significant differences or changes in the input data may negatively affect the performance of the network. Therefore, the input data is normalized using max-min normalization, calculated as:
[0064]
[0065] Among them, X norm and X are the normalized original input data, including time domain signals and frequency domain signals.
[0066] 3) Identification of Initial Degradation Point (IDP): Since the collected data covers the entire process of mechanical equipment from normal operation to final complete failure, it can be divided into normal data and degradation data. This embodiment uses the 3σ principle of RMS to identify IDP.
[0067] 4) RUL label production: In actual scenarios, the total life values of mechanical equipment often vary significantly. Therefore, the RUL value is usually normalized to a RUL percentage. For normal state data, the corresponding RUL percentage label is 1. For degraded state data, their corresponding RUL percentage labels decrease linearly from 1 to 0. Therefore, the RUL percentage label is defined as:
[0068]
[0069] Among them, ARul t is the actual RUL percentage label at time t, ARul t The value range of R is 0 to 1; T and R t Represent the total life and actual RUL at time t respectively.
[0070] 4) MBHAN model construction:
[0071] (1) Hierarchical Self-Attention (HSA) Mechanism
[0072] In order to reduce the computational complexity of SA, the block-local SA mechanism first divides the input sequence into multiple fixed-size blocks. Each block contains a portion of the elements of the input sequence. Then, within each block, SA is applied to calculate the attention weights of the elements within the block. This enables the prediction model to capture the relationships and dependencies between elements in the local sequence. Under the influence of this mechanism, although it effectively reduces the complexity of the model, the use of smaller block sizes may lead to information loss or incomplete representation of long-range dependencies. In addition, due to the use of fixed-size subsequences, the model's attention is limited to the information in these subsequences, and may not be able to model dependencies of different scales. This block-local SA mechanism will limit the overall performance of the RUL prediction model. To solve this problem, a new hierarchical self-attention (HSA) mechanism is proposed, such as Figure 1 As shown in Figure 2, the developed hierarchical self-attention mechanism is able to model multi-scale local dependencies. It enables the RUL prediction model to capture degradation features and dependencies at different levels of granularity, achieving a wider range of attention and a more comprehensive representation of input supervision data.
[0073] First, we define the attention level of HSA as n. Each level corresponds to a different local attention region. Lower levels focus on smaller local regions and can capture fine-grained features and dependencies within these regions. Higher levels have a wider attention range, enabling them to capture dependencies across larger regions. Then, the block size w = [w1, w2, …w n ]. In each level, the input sequence is divided into different subsequences according to the block size w. At the i-th attention level and within each subsequence, linear transformations are used to generate local query, key, and value vectors. Cosine similarity is used to measure the similarity between the local query vector and the local key vector. Then, the obtained similarity score is divided by the scaling factor Where D corresponds to the input dimension of the subsequence. Then, the attention weight is calculated using the Softmax function and assigned to the local value vector. Finally, the attention values of all levels are integrated and output through linear transformation.
[0074] The calculation process of HSA is as follows:
[0075]
[0076] Among them, reshape is the matrix reshape function, and F represents the input feature. The local query vector Local key vector Local value vector Calculated as
[0077]
[0078] in, They are The similarity score is calculated as follows:
[0079]
[0080] The attention coefficient is:
[0081]
[0082] At the i-th attention layer, the subsequence The output is calculated as follows:
[0083]
[0084] Then, the output of the i-th attention level can be written as:
[0085]
[0086] Then, the attention values of all levels are combined as follows:
[0087]
[0088] Finally, the output of HSA can be written as follows:
[0089]
[0090] Both traditional SA and block-local SA may ignore fine-grained local dependencies or fail to capture dependencies at other scales. Compared with traditional SA and block-local SA, the HSA mechanism introduces a hierarchical design that divides attention into multiple levels. Different levels are used to capture degradation information at different scales. This hierarchical design allows capturing local degradation features at different scales while maintaining computational efficiency.
[0091] (2) Structure of the Time-Frequency Fusion Temporal Convolutional Network (TFFTCN)
[0092] The Temporal Convolutional Network (TCN) uses dilated causal convolution (DCC) to capture the temporal dependencies in the input sequence data, making it very suitable for time series prediction tasks. The structure of DCC is Figure 2 Assume that the input time series data is represented by X = [x1, x2, ..., x n ], the corresponding output sequence is expressed as Y = [y1, y2, ..., y n ]. In DCC, y at a specific time point tThe value is only affected by the input data at the current time point and the previous time point, that is:
[0093] y t =f(x1,x2,…,x t )
[0094] This causal relationship ensures that future information does not leak into the forecast, thus maintaining the integrity of the temporal order. Figure 2 In the figure, DCC layers are stacked with expansion factors of d = 1, 2, and 4. As the value of d increases, the receptive field expands, allowing fewer network layers to understand more information. By incorporating larger receptive fields, the forecasting model can effectively capture and exploit temporal patterns over long periods of time, enhancing its ability to capture long-term dependencies in time series data.
[0095] In order to further enhance the signal understanding and modeling capabilities in remaining useful life (RUL) prediction, a time-frequency feature fusion temporal convolutional network (TFFTCN) is designed, which fully exploits the advantages of both time domain and frequency domain signals. By combining these two types of signals, TFFTCN can improve the information representation capabilities required for accurate RUL prediction and fully exploit the complementarity between time domain and frequency domain features. The structure of TFFTCN is shown in Figure 2. Figure 3 As shown in the figure, it includes a feature extraction network and a RUL mapping network. The feature extraction network consists of two parallel sub-networks, which are used to extract the degradation information hidden in the time domain signal and the frequency domain signal respectively. The sub-network includes a convolution sub-network for spatial feature extraction, four TCN blocks for capturing long-term dependencies, and HSA for important feature selection and multi-scale degradation information modeling. In particular, the convolution sub-network uses large convolution kernels to expand the receptive field and retain as much information as possible from the original input data. Then, the deep degradation features obtained from the two sub-networks are fused using a connection operation. The RUL mapping network follows the feature extraction network and includes a TCN block for task-related feature extraction, an adaptive average pooling layer, and a fully connected layer for RUL mapping. This network is responsible for mapping the fused degradation features to the RUL of the mechanical device.
[0096] (3) Memory Weight Constrained Loss
[0097] Memory Weight Constraint (MWC) is a regularization loss term that aims to retain important knowledge learned from previous tasks and adapt to new tasks in a continuous learning environment. Similar to EWC and MAS, MWC assigns importance values to parameters based on their contribution to the performance of the RUL prediction model on previous tasks. These importance values are used to constrain the learning process and prevent significant changes to important model parameters. For a trained model Γ, its parameters are denoted as θ * , which includes the weights and biases of the prediction network. These parameters play a key role in determining the prediction accuracy of the model, and any changes to these parameters will directly affect the performance of the model. When the network parameters change slightly due to learning new task knowledge, it will lead to a corresponding change in the model output. This change in model output can be approximated as linear and can be expressed as:
[0098]
[0099] Where D represents the data sample, is the Jacobian matrix, which can be calculated. Therefore, the Jacobian matrix reflects the sensitivity of the model output to parameter changes. Parameters with larger gradients (elements of the Jacobian matrix) are more sensitive to data changes. Reducing the update amplitude of sensitive parameters helps prevent catastrophic forgetting. Therefore, the importance value of the parameter can be estimated by the following formula:
[0100]
[0101] Considering the large difference in values leading to information imbalance, the maximum and minimum normalization operation is used. The importance value can be rewritten as:
[0102]
[0103] From this, the MWC regularization term can be obtained as:
[0104]
[0105] By adding the MWC regularization term to the task loss function, the RUL prediction model is forced to keep important parameters close to their original values during the training process of new tasks. This MWC helps to alleviate catastrophic forgetting and promotes continuous RUL prediction.
[0106] (4) Multi-branch horizontal expansion network (MBHAN)
[0107] Although the regularization method can help preserve task knowledge in continuous RUL prediction scenarios with small-difference tasks, it encounters challenges when faced with large differences between tasks and a considerable number of tasks. In this case, the performance of the RUL prediction model will rapidly degrade, affecting both the previous tasks and the current task. To solve this problem, the present invention proposes a new MBHAN that considers task similarity. The MBHAN principle diagram is shown in the figure. Figure 4 As shown in Figure 2. The core idea of this method revolves around solving rule prediction tasks with different degrees of task differences. For tasks with small differences, the task branch network enhancement strategy and MWC are used to reduce the storage requirements of the model and reduce the interference between tasks. On the other hand, for tasks with large differences, feature extraction backbone network enhancement is used to prevent interference with other tasks.
[0108] The feature extraction network in MBHAN is represented as ψ, and the RUL mapping network is represented as Therefore, for the first RUL prediction task, the network structure can be expressed as express After network ψ. Assume that the feature extraction backbone network in MBHAN consists of N feature extraction networks, each of which has a different number of RUL mapping networks. Therefore, MBHAN can be expressed as:
[0109]
[0110] in, It is MBHAN; i is the i-th feature extraction network; is i For the jth RUL mapping network in, according to the MBHAN principle, the differences between all tasks using the same feature extraction network are small; It is defined as a network combination operator. On this basis, when a new RUL prediction task needs to be processed, it is necessary to expand the task branch network or feature extraction backbone network. Its construction steps are as follows:
[0111] i) Task similarity judgment: In the setting of continuous RUL prediction tasks, the data of the previous task is not available, and the data of the current task consists of data and corresponding labels. Therefore, the present invention directly inputs the training data of the current task into the model Ψ, and then the predicted RUL of all RUL mapping networks can be obtained, which can be expressed as:
[0112]
[0113] Among them, p ij Indicated by The predicted RUL is obtained. The average RUL of each feature extraction network is calculated and expressed as:
[0114]
[0115] Then, calculate P mean The mean absolute error (MAE) vector between the actual RUL label:
[0116] E=|ARul-P mean |=[e1,e2…,e n ]
[0117] Where ARul is the actual URL label. The minimum error in E will be returned. If e i If the difference is smaller than the preset threshold, it means that the current task is different from the one using e i Otherwise, it is assumed that there are large differences between all previous tasks and the current task.
[0118] ii) Task branch network expansion: Usually, the feature extraction network extracts general degradation features from the input monitoring data and does not have task-specific feature extraction and mapping capabilities. i The task is similar to that of i Add a new RUL mapping network later (new task branch), namely the feature extraction network ψ i By the network and Network In this case, the structure of MBHAN is updated as follows:
[0119]
[0120] In order to make full use of the relevant information of previous tasks and speed up the network convergence, the new task branch network The weights of the previous network In addition, considering the network ψ i When learning new task knowledge, it may be affected, and the MWC mechanism is used to consolidate the learned task knowledge. Therefore, the loss function combines the mean square error (MSE) loss and loss MWC , used to update network parameters, can be expressed as:
[0121]
[0122] During the training process, in addition to the network and network ψ i All other networks are frozen.
[0123] iii) Feature extraction backbone network expansion: When the current task is dissimilar to all previous tasks, a new feature extraction network ψ needs to be expanded new At the same time, due to the new feature extraction network ψ new There is no branch network for any specific task, you need to add a task branch network Therefore, in this case, the structure of MBHAN is updated to:
[0124]
[0125] In order to focus on learning the knowledge representation of the current task, the newly expanded network Instead of using the parameters of the previous network for initialization, a random initialization method is used. At the same time, the loss function only uses the MSE loss for network update, which can be expressed as:
[0126] loss=(PRul-ARul) 2
[0127] Similar to ii), the parameters of the extended network are isolated from the parameters of other networks, that is, when training a new RUL prediction task, in addition to the network Except for , the parameters of all other networks are not updated.
[0128] The details of the feature extraction network and RUL mapping network are shown in Table 1. In Table 1, In_c and Out_c are the number of input channels and output channels respectively; In_f is the number of input features, and Out_f is the number of output features. D_s is the expansion coefficient of DDC; W_s is the window size vector in HSA, where different elements represent the window sizes of different attention levels.
[0129] Table 1 Detailed information of the prediction network
[0130]
[0131] 6) RUL prediction: For the first RUL prediction task, the prediction network constructed is ψ1 Its initial weights are randomly generated. The loss function adopts MSE loss. For the subsequent RUL prediction task, this embodiment uses the proposed MBHAN to continuously learn task knowledge by using only the current task data. However, the test data includes the test sets of all previous tasks and the current test set.
[0132] 7) Performance evaluation: Mean absolute error (MAE) is used to measure the average absolute difference between the predicted and actual RUL, which can quantify the overall degree of deviation of the model in RUL prediction. The lower the MAE value, the closer the model's prediction results are to the actual RUL value, and the higher the MAE value, the greater the prediction error. The Score function is used to evaluate overestimated or underestimated prediction results, and a larger penalty is imposed on overestimated results. This is because overestimating RUL predictions can lead to more serious consequences or losses than underestimating. Their formulas are defined as follows:
[0133]
[0134] Among them, Y i is defined as:
[0135]
[0136] Experimental verification:
[0137] 1) Data Introduction
[0138] (1) XJSY bearing dataset
[0139] The XJSY bearing dataset was collected through the accelerated bearing degradation experiment. It includes the life cycle data of 15 rolling bearings under three different working conditions. The working condition information is shown in Table 2. The experimental bearings used are of the LDK UER204 rolling bearing type. The sampling frequency during the experiment was set to 25.6kHz and the sampling interval was 1 minute. For each sampling instance, the duration of data collection was 1.28 seconds.
[0140] Table 2 XJSY bearing data set
[0141]
[0142] 2) Full life gear dataset
[0143] In this experiment, multiple fatigue life tests were carried out using the gear contact fatigue test bench at Newcastle University, UK. The test bench includes a gear test platform, a cooling and lubrication system, a measurement and control platform, and two gear pairs (experimental gearboxes). The experimental gearboxes are placed symmetrically on the test bench. Accelerometers are placed in the horizontal and vertical directions of the gearbox to collect vibration signals with a sampling frequency of 25kHz, a sampling time of 0.2 seconds, and a sampling interval of 1 minute. Four sets of vibration signal data sets from operation to failure were collected under two different working conditions. Table 3 provides further details of the experimental conditions.
[0144] Table 3 Full life gear data information
[0145]
[0146] (2) TFFTCN Verification
[0147] The TFFTCN model accurately predicts the RUL of the bearing. The predicted RULs curves for Task 1 and Task 3 are shown in Figure 2. Figure 5 (a) and Figure 5 (b). Figure 5 It can be seen that the predicted RUL in the normal stage has the smallest fluctuation and is very close to the actual RUL value. Importantly, in terms of the degradation stage, the predicted RUL curves of both tasks closely track the actual RUL trend. This observation provides evidence to support the effectiveness of the TFFTCN model in accurately predicting the bearing RUL.
[0148] To verify the superiority of the TFFTCN model, we conducted a comprehensive comparison using the MAE and Score evaluation indicators. The prediction results of the TFFTCN model were compared with the other 6 models, and the evaluation index values were calculated, as shown in Table 4. Table 4 clearly shows that the TFFTCN model outperforms the other models in terms of MAE and Score values for all three tasks. Obviously, TFFTCN presents the lowest MAE value and the highest Score value, indicating better prediction accuracy and overall performance on the XJSY bearing dataset. Comparing the prediction results of TFFTCN and SATCN, it is found that the prediction error of TFFTCN is significantly reduced. Specifically, for tasks 1, 2, and 3, the MAE of TFFTCN is reduced by 0.048, 0.029, and 0.058, respectively, and the Score is increased by 0.122, 0.097, and 0.042, respectively. These improvements clearly demonstrate the effectiveness of the HSA mechanism in improving prediction accuracy. This is because HSA can capture degenerate features and dependencies at different attention levels. In addition, the comparison between TCN1 and TCN2 shows that the prediction accuracy of TCN1 and TCN2 is improved, highlighting the effectiveness of integrating frequency domain signals. In summary, the proposed TFFTCN model has a strong ability to learn degradation features of different scales, thus significantly improving the prediction accuracy of RUL.
[0149] Table 4 Comparison results
[0150]
[0151] 2) Validation of the sustained RUL prediction method
[0152] MBHAN and MWC mechanisms are applied to continuously predict the remaining useful life (RUL) based on a given task sequence under a specified task sequence. The mean absolute error (MAE) values of the prediction results are calculated and presented in Table 6. Comparing Table 6 with Table 5, it can be found that when the MCA mechanism is adopted, the algorithm has achieved significant improvements in the prediction performance of the continuous prediction task. The MAE value of the previous task is greatly reduced, indicating that the continuous prediction is more accurate. The MAE value of the current task is very close to that of a single RUL prediction task. In addition, even if four tasks are learned consecutively, the knowledge gained from the first learning task is still retained. These improved results provide strong evidence for the effectiveness of the MWC mechanism in alleviating the catastrophic forgetting problem. By retaining the knowledge gained in the previous tasks while adapting to the new task, MBHAN is able to accurately predict and reduce the performance degradation of the previous RUL prediction task.
[0153] Table 5 MAE values of RULs predicted without using KWCMWC mechanism
[0154]
[0155] Table 6 MAE values of RULs predicted using KWCMWC mechanism
[0156]
[0157] Table 7 MAE values of RULs predicted by different CL methods
[0158]
[0159]
[0160] Table 7 shows the MAE values of the predicted RULs obtained by the above CL methods. In addition, the MAE values obtained for a single RUL prediction task are also included in Table 7 for a more direct comparison. By comparing the prediction results of the CL-free method and MBTCN1, it is found that the task branch network can effectively retain the degenerate information of a specific task. The comparative analysis of MBTCN1 and MBTCN2 shows that the MWC term helps to control the impact of new tasks on the shared feature extraction network and alleviate the catastrophic forgetting of the model when learning task knowledge in sequence. Among all CL methods, the MBHAN method proposed in this paper has the best prediction performance because it significantly reduces the MAE value. Compared with other methods, this highlights its more outstanding ability to retain knowledge in previous tasks while learning new tasks. In summary, for tasks with smaller differences, the task branch network enhancement based on semi-parameter sharing is adopted, and for tasks with larger differences, the feature extraction backbone network enhancement based on parameter isolation is adopted. This method effectively protects the acquired knowledge from interference from new tasks.
[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.
Claims
1. A method for predicting the continuous remaining useful life based on a multi-branch lateral amplification network. It is characterized in that The method specifically comprises the following steps: S1: Continuously collect condition monitoring data of mechanical equipment under different working conditions and obtain frequency domain signals through fast Fourier transform; frequency domain signals contain specific frequency components that indicate fault characteristics or degradation modes; S2: Normalize the frequency domain signal; S3: Divide the normalized data into normal data and degraded data; S4: Construct the MBHAN model and use the loss function with the MWC regularization term to train and optimize the MBHAN model, where MBHAN represents a multi-branch horizontal expansion network and MWC represents a memory weight constraint; S5: Use the optimized MBHAN model to predict the remaining useful life of mechanical equipment.
2. The method for predicting the continuous remaining useful life according to claim 1, It is characterized in that In step S4, the constructed MBHAN model includes a feature extraction backbone network and a task branch network; For the first RUL prediction task, the MBHAN network structure can be expressed as express After network ψ, RUL represents the remaining useful life; assuming that the feature extraction backbone network consists of N feature extraction networks, each feature extraction network ψ has a different number of RUL mapping networks MBHAN is represented by: Among them, Ψ is MBHAN; ψ i is the i-th feature extraction network; is i The j-th RUL mapping network in; is the network combination operator; On this basis, when a new RUL prediction task needs to be processed, the task branch network or feature extraction backbone network needs to be expanded.
3. The method for predicting the continuous remaining useful life according to claim 2, It is characterized in that In step S4, when a new RUL prediction task needs to be processed, the specific steps of expanding the task branch network or the feature extraction backbone network are as follows: 1) Determine task similarity; In the setting of continuous RUL prediction task, the data of previous tasks are not available, and the data of the current task consists of data and corresponding labels; therefore, the training data of the current task is directly input into the model Ψ, and then the predicted RUL of all RUL mapping networks is obtained, which is expressed as: Among them, p ij Indicated by The predicted RUL obtained; the average RUL of each feature extraction network is calculated and expressed as: Then, calculate P mean The mean absolute error vector E between the actual RUL label: E=|ARulP mean |=[and 1 ,and 2 …,and n ] Where ARul is the actual URL label; the minimum error in E will be returned; if e i If the difference is smaller than the preset threshold, it means that the current task is different from the one using e i The differences between the tasks are small; otherwise, it is considered that there are large differences between all previous tasks and the current task; 2) Expand the task branch network: If the current task is consistent with using ψ i The task is similar to that of i Add a new RUL mapping network later That is, the feature extraction network ψ i By the network and Network Share; in this case, the structure of MBHAN is updated as: In order to make full use of the relevant information of previous tasks and speed up the network convergence, the new task branch network The weights of the previous network The weight parameters are initialized; in addition, the MWC mechanism is used to consolidate the learned task knowledge; therefore, the loss function loss combines the mean square error loss and the MWC regularization term loss MWC , used to update network parameters, expressed as: Where pRul is the predicted RUL, λ is the weight of the MWC regularization term, is the network i The importance of parameters, is the current task in the network ψ i The parameters on is the previous task in the network ψ i The optimal parameters on ; During the training process, in addition to the network and network ψ i All other networks are frozen. 3) Expanding the feature extraction backbone network: When the current task is dissimilar to all previous tasks, a new feature extraction network ψ needs to be expanded new At the same time, due to the new feature extraction network ψ new There is no branch network for any specific task, you need to add a task branch network Therefore, in this case, the structure of MBHAN is updated to: In order to focus on learning the knowledge representation of the current task, the newly expanded network Instead of using the parameters of the previous network for initialization, a random initialization method is used; at the same time, the loss function only uses the mean square error loss for network update, which is expressed as: loss=(PRul-ARul) 2 Similar to 2), the parameters of the extended network are isolated from the parameters of other networks, that is, when training a new RUL prediction task, in addition to the network Except for , the parameters of all other networks are not updated.
4. The method for predicting the continuous remaining useful life according to claim 2, It is characterized in that In step S4, the feature extraction network consists of two parallel sub-networks, which are used to extract the degradation information hidden in the time domain signal and the frequency domain signal respectively; the sub-network includes a convolution sub-network for spatial feature extraction, four TCN blocks for capturing long-term dependencies, and HAS for important feature selection and multi-scale degradation information modeling; the convolution sub-network uses a large convolution kernel to expand the receptive field and retain more information from the original input data; then, the deep degradation features obtained from the two sub-networks are fused using a connection operation; wherein TCN represents a temporal convolutional network and HAS represents hierarchical self-attention.
5. The method for predicting the continuous remaining useful life according to claim 4, It is characterized in that In step S4, the HAS mechanism is as follows: first, define the attention level of HSA as n; each level corresponds to a different local attention area; the low level focuses on smaller local areas and captures the fine-grained features and dependencies within these areas; the high level has a wider attention range, enabling it to capture dependencies across larger areas; Then, the block size w = [w 1 ,w 2 ,…w n ]; in each level, divide the input sequence into different subsequences according to the block size w; in the i-th attention level and in each subsequence, use linear transformation to generate local query, key and value vectors; use cosine similarity to measure the similarity between the local query vector and the local key vector; then, divide the obtained similarity score by the scaling factor Where D corresponds to the input dimension of the subsequence; then, the attention weight is calculated using the Softmax function and assigned to the local value vector; finally, the attention values of all levels are integrated and output through linear transformation.
6. The method for predicting the continuous remaining useful life according to claim 2, It is characterized in that In step S4, the RUL mapping network follows the feature extraction network, including a TCN block for task-related feature extraction, an adaptive average pooling layer, and a fully connected layer for RUL mapping; the network is responsible for mapping the fused degradation features to the RUL of the mechanical device.
7. The method for predicting the continuous remaining useful life according to claim 3, It is characterized in that In step S4, the RUL label is prepared: the RUL value is normalized into a RUL percentage, and the RUL percentage label is defined as: Among them, ARul t is the actual RUL percentage label at time t, ARul t The value range of R is 0 to 1; T and R t Respectively represent the total life and actual RUL at time t, Indicates the RUL corresponding to the initial degradation point.
8. The method for predicting the continuous remaining useful life according to claim 1, It is characterized in that In step S5, when predicting the remaining useful life of mechanical equipment, for the first RUL prediction task, the prediction network constructed is Its initial weights are randomly generated; the loss function adopts mean square error loss; for the subsequent RUL prediction task, MBHAN is used to continuously learn task knowledge by using only the current task data; the test data includes the test sets of all previous tasks and the current test set.
Citation Information
Cited By
Turnout switch equipment mechanical fault diagnosis method based on complementary feature enhancement and consistency learning
CN119779655A
