Cross-domain adaptive lightweight residual life prediction method for rolling bearing

Abstract

The invention provides a cross-domain self-adaptive lightweight residual life prediction method for a rolling bearing, and aims to solve the problems of complex model and poor cross-working-condition adaptability of the residual life prediction of the rolling bearing. The method comprises the following steps: firstly, performing angle alignment and band-pass filtering processing on a vibration signal, extracting multi-scale features through envelope analysis, and constructing a life label; the constructed hybrid model fuses one-dimensional convolution, a bidirectional circulation unit and an attention mechanism, the feature extraction capability is enhanced by adopting hierarchical attention, and the cross-working-condition performance is improved through an adjustable time sequence structure. Feature distribution alignment is introduced to reduce inter-domain differences, the model scale is compressed through structured pruning after training, and fine tuning optimization is performed in combination with target domain data. According to the method, high-precision prediction can be realized under finite computing power, cross-working-condition self-adaptive updating is supported, and an effective solution is provided for intelligent operation and maintenance of the bearing.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment condition monitoring and predictive maintenance technology, and in particular relates to a method for predicting the remaining life of rolling bearings based on lightweight deep learning and cross-domain transfer. Background Technology

[0002] In the context of mechanical equipment condition monitoring and predictive maintenance, while deep learning-based methods for predicting the remaining life of rolling bearings can automatically extract multi-dimensional degradation features and improve prediction accuracy, the deep networks used typically have a large number of parameters and high computational complexity, requiring high-performance servers and GPU environments, making them difficult to deploy on inexpensive terminals or field equipment with limited computing power. Although techniques such as lightweight pruning and model distillation can reduce model size to some extent, they often result in a significant decrease in prediction performance, failing to meet the requirements for high reliability and robustness. Furthermore, the data distribution varies significantly between different operating conditions or equipment, and existing methods are prone to accuracy degradation when transferring data across datasets or validating in multiple scenarios, making it difficult to achieve true cross-domain generalization. Therefore, there is an urgent need for a lightweight rolling bearing remaining life prediction method that can maintain high prediction accuracy, run in real-time on ordinary computing devices, and possess strong cross-domain transfer capabilities. Summary of the Invention

[0003] To address the above technical problems, this invention provides a cross-domain adaptive lightweight remaining lifetime prediction method for rolling bearings. The method includes the following steps: First, parameterized data preprocessing is performed on the original vibration signals from the source and target domains through operations such as coordinate alignment, sliding window segmentation, multi-scale time / frequency domain feature extraction, Z-Score normalization, and RUL label attachment to construct training, validation, and test sets. Second, a lightweight baseline model composed of a cascaded 1D convolutional neural network, a bidirectional GRU, and a self-attention layer is designed to achieve automatic extraction of degradation features and preliminary remaining lifetime prediction. Then, a hierarchical subspace pyramid attention module is set after the one-dimensional convolutional feature extraction layer of the baseline model to selectively weight and fuse the time-scale-channel information of the convolutional output; and time-series encoding is then performed. A bottleneck-based Adapter fine-tuner is integrated between the BiGRU output and self-attention layer. During the training / transfer phase, the backbone parameters are frozen, and only a small number of new parameters are updated in the Adapter (and the regression output layer). This maintains or improves the remaining lifetime prediction accuracy with very few trainable parameters under conditions of limited parameter quantity and computational cost. Furthermore, a cross-domain transfer module with grouped convolution and feature concatenation is constructed, and adversarial loss or maximum mean difference loss is combined to align the feature distributions of the source and target domains, so as to significantly improve the model's generalization ability on different operating conditions and datasets. Finally, the fine-tuned model is subjected to structured channel pruning to further compress the model size, and exported to ONNX or TorchScript format to realize real-time online prediction of the remaining lifetime of rolling bearings on resource-constrained ordinary equipment.

[0004] According to one aspect of this application, a cross-domain adaptive lightweight remaining life prediction method for rolling bearings is proposed:

[0005] S1. The original vibration signals of the source and target domains are aligned with coordinates, denoised and bandpass filtered, sliced ​​and standardized according to the set window, and the remaining lifetime labels are constructed and normalized according to the degradation start point to form training, verification and test datasets.

[0006] S2. Design a lightweight baseline model, which consists of a one-dimensional convolution, a bidirectional gated recurrent unit and a self-attention unit, used to automatically extract degradation features and complete preliminary remaining lifetime prediction.

[0007] S3. After the convolutional backbone, a hierarchical subspace pyramid attention module is embedded. In each subspace, the temporal scale-channel information is selectively weighted and fused through multi-scale pyramid convolution. A bottleneck adapter is set in the temporal coding layer (between BiGRU output and self-attention). With the backbone parameters frozen, only a few newly added parameters are fine-tuned, thereby maintaining or improving prediction accuracy under the premise of lightweight design, and enhancing cross-domain adaptive capability and deployment efficiency.

[0008] S4. Construct a cross-domain transfer module, simultaneously implement statistical alignment with the maximum mean difference and domain adversarial alignment combined with the gradient reversal layer in the mid-to-high-level feature space, and use joint objectives for end-to-end training to make the feature distributions of the source domain and the target domain tend to be consistent, thereby significantly improving the cross-domain generalization ability.

[0009] S5. Perform channel-level structured pruning on the model that has completed transfer training and perform a small amount of retraining. Export the model in a deployable format and deploy it to embedded terminals or ordinary industrial computing devices to achieve online inference and prediction alarms.

[0010] S6. Based on the deployment model, online inference and closed-loop updates are carried out. The remaining lifetime results are output in real time according to a unified preprocessing process. When a small number of target domain labels or weak labels are obtained, the adapter is periodically triggered for small-step fine-tuning and lightweight re-pruning to maintain accuracy and stability in the long term.

[0011] Beneficial effects of this invention: This invention proposes a cross-domain adaptive lightweight remaining lifetime prediction method for rolling bearings. By combining parameterized preprocessing with a lightweight CNN-BiGRU-self-attention backbone, it efficiently models the bearing degradation process without relying on high-performance hardware. Secondly, it introduces hierarchical subspace pyramid attention and spatial pyramid convolution (SPC) into the backbone network. The Convolution module, with an external Adapter bottleneck fine-tuner at the temporal layer, enhances multi-scale key information and enables rapid intra-domain adaptation with minimal parameter additions, balancing model size and prediction accuracy. This model can fully learn degradation patterns from source domain data and output stable remaining lifetime predictions, while significantly reducing parameter count and FLOPs, inference latency, and storage requirements. Furthermore, adversarial alignment using MMD statistical alignment and a domain classifier combined with a gradient inversion layer achieves cross-domain distribution consistency, effectively improving generalization and robustness across different datasets and operating conditions, and reducing performance fluctuations and annotation costs caused by domain shifts. Finally, structured pruning is applied to the transfer-tuned network, and ONNX / TorchScript is exported, enabling real-time online RUL prediction and early warning on inexpensive devices such as embedded terminals or ordinary industrial PCs, facilitating deployment and maintenance. Compared to existing technologies, this invention significantly reduces model complexity and computational overhead while maintaining prediction accuracy, possessing comprehensive advantages such as strong cross-domain consistency, low deployment cost, and rapid update iteration. This allows for timely adjustments to maintenance strategies, reducing unexpected downtime and repair costs, and improving production line availability and overall economic efficiency.

[0012] The above and other objects, advantages and features of the present invention will become more apparent to those skilled in the art from the following detailed description of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the overall process of one embodiment of this application. The diagram shows the complete process performed sequentially through six stages: data acquisition and preprocessing, baseline model construction, lightweight enhancement, cross-domain transfer alignment, structured pruning and model export, and deployment and online inference closed loop. The dashed arrow at the "deployment and online inference closed loop" indicates the closed-loop update mechanism of small-step fine-tuning and lightweight re-pruning triggered when a small number of target domain labels or weak labels are obtained.

[0014] Figure 2 This is a system and data flow diagram of one embodiment of this application. The diagram shows, from top to bottom: the original vibration signal input, followed by degradation initiation detection, sliding window segmentation, feature extraction and standardization, and remaining lifetime label construction to form source and target domain datasets; the left side shows the model training process, including lightweight baseline network, hierarchical subspace pyramid attention module enhancement, adapter bottleneck fine-tuning, statistical alignment of maximum mean difference and domain adversarial alignment combined with gradient inversion layer, and training and validation; the right side shows the evaluation and deployment process, including performance metric evaluation, inference mode selection, model export to a deployable format and deployment, achieving online remaining lifetime prediction.

[0015] Figure 3 This is a flowchart illustrating the lightweight and transfer-aligned training process according to one embodiment of this application. The flowchart describes the lightweight training process starting from constructing the backbone network, inserting the hierarchical subspace pyramid attention module, and attaching the adapter fine-tuner. Subsequently, evaluation and iterative optimization are performed based on a joint objective including task loss, maximum mean difference and domain adversarial loss, and parameter sparsity regularization. After a small amount of retraining, an accuracy target is determined. Once the target is met, a deployable model is exported and deployed online, enabling online inference and lightweight iterative updates.

[0016] Figure 4 This is a schematic diagram of a lightweight baseline network structure according to an embodiment of this application. The diagram shows a network structure consisting of convolutional layers, pooling layers, bidirectional gated recurrent unit layers, and self-attention layers in sequence; a hierarchical subspace pyramid attention module is embedded in the convolutional layer to fuse multi-scale spatial-channel information; an adapter bottleneck fine-tuner is attached to the temporal layer to fine-tune only a few newly added parameters while freezing the backbone parameters; after self-attention aggregation, a continuous remaining lifetime estimate is output by a fully connected layer. Detailed Implementation

[0017] S1. Data preprocessing and sample construction.

[0018] Specifically, step S1 is as follows:

[0019] S101. Collect bearing vibration signals under different loads, speeds, and environmental conditions to form a sample set. Marked with the corresponding runtime (or degradation process).

[0020] S102. Perform detrending and bandpass filtering on the original signal, using sliding window segmentation: Let the window length L and step size Δ be the sequence blocks obtained.

[0021] X k =[s(kΔ),s(kΔ+1),…,s(kΔ+L-1)],k=1,…,K.

[0022] In the formula, L is the number of sample points in each window, Δ is the step size between the starting points of adjacent windows, k is the index of k windows, and K is the total number of windows.

[0023] S103. Extract multi-scale time / frequency domain features (or directly retain time segments as end-to-end input), and perform Z-Score standardization on each feature dimension:

[0024]

[0025] In the formula, x is the original eigenvalue. Let μ and σ be the standardized feature values, representing the mean and standard deviation of the feature in the sample set, respectively, and N be the number of samples used to estimate μ and σ.

[0026] S104. Generate RUL tags. Using the degradation start point t0 and the maximum lifetime R... max Based on this, a piecewise linear (or exponential) labeling strategy is used, for example, linear:

[0027] y k =max(0,R) max -α·(t k -t0))

[0028] Normalizable to [0,1], Based on this, source domain datasets are constructed respectively. With the target domain dataset

[0029] S103. Based on this, the key processing parameter set X′={x1,x2,x3} is selected through correlation analysis. l ,…,x m} (1≤i≤m) serves as the data basis for subsequent parameter prediction and regression.

[0030] S2. After completing the data preprocessing and sample construction in S1, a lightweight baseline model consisting of one-dimensional convolution, bidirectional GRU, and self-attention is established and trained to obtain preliminary remaining lifetime prediction capability.

[0031] Specifically, step S2 is as follows:

[0032] S201. Establish a baseline network consisting of a one-dimensional convolutional layer, a bidirectional GRU, and a self-attention layer, to capture local degradation patterns and long-range temporal dependencies. The output of the convolutional layer is:

[0033] F = Conv1D(X; W) c ,b c )

[0034] In the formula, X is the input sequence / feature tensor, and W... c With b c Let F be the weights and biases of the one-dimensional convolution kernel, and F be the output feature of the convolution.

[0035] The bidirectional GRU recursion is as follows:

[0036] z t =σ(W z x t +U z h t-1 ),r t =σ(W r x t +U r h t-1 )

[0037]

[0038] In the formula, x t Let h be the input vector at time t. t-1 ,h t The current hidden state; z t With r t These are the update gate and the reset gate, respectively. For candidate hidden state W z W r W h with U z U r U h σ is the corresponding weight matrix; σ(·) is the Sigmoid function, and tanh(·) is the hyperbolic tangent.

[0039] Self-attention aggregates as follows:

[0040] Q = XW Q K = XW K V = XW V

[0041] In the formula, W Q W K W V Let Q, K, and V be the linear projection weight matrices, representing the query, key, and value representations, respectively. kThe dimension of the key vector is used for scaling, and softmax(·) is the softmax function for row normalization.

[0042] S202. Use mean squared error as the main loss to train the baseline model.

[0043]

[0044] In the formula, With y i Let N be the predicted RUL and the true RUL labels for i samples, and N be the number of samples used for loss calculation.

[0045] S3, based on the baseline network trained by S2, further introduces a hierarchical subspace pyramid attention and Adapter bottleneck fine-tuning module to enhance multi-scale feature representation and reduce transfer fine-tuning cost without significantly increasing the number of parameters.

[0046] Specifically, step S3 is as follows:

[0047] S301. Insert a hierarchical subspace pyramid attention module after the convolutional backbone. First, divide the feature F into g subspaces according to channels, and then perform a kernel-scale analysis on the set {k0,…,k...}. S-1 Parallel convolution yields branch features {F} i For each branch, the scale weight {α} is generated through a two-layer MLP (GAP→SE) process followed by Softmax. i}, and with

[0048]

[0049] Weighted fusion is performed. In the formula, S is the scale / number of branches, and k... i Let F be the kernel size of the i-th branch. i For the branch output features at the corresponding scale, α i The attention weights (SE weight generator) for this branch are used to learn the importance of each branch / channel. Within a single branch, SPC is used for channel grouping and multi-kernel parallel convolution before concatenation. Finally, the original number of channels is restored via 1×1 convolution / residual.

[0050] S302. Insert an Adapter bottleneck fine-tuner into the BiGRU (and / or attention) layer, training only a small number of newly added parameters to maintain accuracy while reducing transfer costs. Let the input be u, and the Adapter be a residual structure of "dimensionality reduction - nonlinearity - dimensionality increase":

[0051] Adapter(u) = u + Uδ(Wu + b1) + b2

[0052] In the formula: W is the dimensionality reduction mapping weight matrix, U is the dimensionality increase mapping weight matrix, b1 and b2 are bias vectors, and δ(·) is a nonlinear activation function (GELU or ReLU); during fine-tuning, the trunk is frozen and only (W,u,b1,b2) is updated.

[0053] S4. Based on the lightweight enhancement model obtained in S3, maximum mean difference (MMD) statistical alignment and adversarial alignment combining a domain classifier with a gradient inversion layer (GRL) are introduced simultaneously in the mid-to-high-level feature space. The model is trained with a joint loss including task MSE, MMD, domain adversarial, and L1 regularization to make the feature distributions of the source domain and the target domain more consistent, thereby improving the generalization and robustness of the model under different datasets and working conditions.

[0054] Specifically, step S4 is as follows:

[0055] S401. Perform MMD statistical alignment in the mid-to-high-level feature space of the model. Let the feature representations of the source domain and the target domain be respectively... and Source / target domain sample number n s n t definition:

[0056]

[0057] In the formula, k(·,·) is the kernel function, and the RBF kernel is used in this embodiment. σ k The kernel bandwidth is obtained through a validation set search.

[0058] δ x =x predicted -x actual

[0059] S402. In addition to feature alignment, domain discrimination D is introduced. dom Adversarial alignment is achieved with a gradient inversion layer (GRL). The domain discriminator outputs the probability of a particular f in its source domain, and the domain alignment loss is defined as follows:

[0060]

[0061] In the formula, d∈{s,t} is the domain label, and p(d=·|f) is the domain discrimination probability. During backpropagation, GRL... The gradient inversion forces the feature extraction subnet to learn a "domain-inseparable" representation, thereby reducing cross-domain differences.

[0062] S403, Cross-domain training uses joint loss:

[0063]

[0064] in, Let N be the mean squared error of the task and the number of samples in the batch. For the predicted / true RUL)||Θ||1, λ, γ, and β are the L1 norm regularization terms of the parameter set Θ, and the loss weights are set according to the performance of the validation set.

[0065] S5 performs channel / convolution kernel-level structured pruning on the model trained across domains in S4 and exports the deployment format to meet the low-latency online inference requirements of ordinary and inexpensive devices.

[0066] Specifically, step S5 is as follows:

[0067] S501 performs structured pruning based on channel / kernel importance metrics (such as BatchNorm scaling factor or L1 norm of convolution), preserving important channels and obtaining lightweight network parameters Θ.

[0068] S502 fine-tunes the simplified model on a small sample in the target domain until the accuracy threshold is met; then it is exported in ONNX / TorchScript format.

[0069] The S503 deploys online inference services on common computing devices such as embedded / industrial PCs, enabling low-latency RUL prediction and real-time alarms.

[0070] S6. Based on the deployment model exported from S5, online inference and closed-loop updates are carried out. When a small number of new labels (or weak labels) are obtained on site, the Adapter-level small-step fine-tuning and lightweight re-pruning are periodically triggered to maintain accuracy and stability in the long term.

[0071] Specifically, step S6 is as follows:

[0072] After S601 goes live, it uses the same strategy as S1 to window and standardize newly arrived data, and then performs forward computation and output. in These are the RUL prediction values ​​obtained through online inference.

[0073] When the S602 collects a small number of target domain labels (or weak labels), it periodically triggers small-step adjustments and lightweight re-pruning at the Adapter level to maintain high accuracy and stability under resource constraints.

[0074] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] The above description is only the most effective embodiment of the present invention. It should be noted that for those skilled in the art, appropriate improvements and modifications can be made without departing from the working principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A cross-domain adaptive lightweight remaining life prediction method for rolling bearings, characterized in that... The method includes: Bearing vibration signals, running time, and operating condition information are collected from the source and target domains. Key operating condition / processing parameters (including speed, load, temperature, lubrication / contamination, sensor position, and bearing structural parameters) are extracted. The dataset is then divided into training, validation, and test sets according to operating conditions and life stages to ensure the comparability of samples across equipment and operating conditions and to provide operating condition labels for transfer learning. The original signal is resampled in the angular or order domain to suppress non-stationarity caused by speed fluctuations. Detrending and bandpass filtering around the bearing characteristic frequency are performed. The signal is sliced ​​according to a preset window of fixed speed or fixed duration and conditional standardization is performed. RUL tags are generated and normalized based on degradation start point detection to stably characterize bearing degradation and obtain consistent life labeling under variable speed and noise conditions. A lightweight baseline network consisting of one-dimensional convolution, bidirectional gated recurrent units (BiGRU), and self-attention is constructed to automatically extract degradation features such as impulse modulation and long-range dependence and complete RUL regression prediction, thus adapting to bearing failure mechanisms with low parameter quantity and computational cost. A hierarchical subspace pyramid attention module is set between the one-dimensional convolutional feature extraction layer and the bidirectional gated recurrent unit (BiGRU) to selectively weight and fuse the subspace-multi-scale information of the convolutional output. In the temporal coding layer, a bottleneck adapter is inserted between the BiGRU output and the self-attention layer. During the training and transfer phases, the backbone parameters except for the adapter are frozen, and only a few newly added parameters of the adapter and the output regression layer are updated. This maintains or improves the remaining lifetime prediction accuracy and enhances cross-domain adaptive capability under the condition of limited parameter quantity and computational cost. A cross-domain transfer module is established, which simultaneously implements statistical alignment using maximum mean difference (MMD) and domain adversarial alignment using gradient inversion layer (GRL) in the mid-to-high-level feature space. End-to-end training is performed using joint loss to make the feature distributions of the source domain and the target domain more consistent, thereby improving the generalization ability of RUL prediction in the target domain. Perform channel / kernel-level structured pruning on the model that has completed transfer training, and retrain it on a small sample in the target domain until the accuracy threshold is met. Then export the model as an ONNX or TorchScript deployable format and deploy it to embedded terminals or industrial computing devices to achieve online inference and alarm under resource-constrained conditions. In actual operation, the new data is windowed and standardized according to the same process as training, and the RUL prediction is output online. When a small number of target domain labels or weak labels are obtained, small-step fine-tuning and light re-pruning are periodically triggered, which only affect the adapter and a small number of high-level parameters, in order to cope with the drift of operating conditions and achieve closed-loop update and long-term stability.

2. The method according to claim 1, characterized in that: The lightweight baseline network performs RUL regression for rolling bearings according to the sequence of "convolutional feature extraction - temporal modeling - attention aggregation". Its computational relationships and functions are as follows: F=Conv1D(X;W c ,b c ) Used to enhance degradation features such as impact and envelope modulation associated with bearing defects. Bidirectional gated cyclic recursion: z t =σ(W z x t +U z h t-1 ),r t =σ(W r x t +U r h t-1 ) It is used to preserve the degradation trend and suppress occasional noise interference over a long life span. Self-attention convergence: It is used to adaptively highlight time segments related to lifespan degradation and weaken irrelevant components caused by changes in operating conditions, thereby improving cross-domain generalization and prediction accuracy.

3. The method according to claim 1, characterized in that: The baseline network is trained using mean squared error as the main loss. This is used to directly reduce the deviation between predicted lifetime and actual lifetime on a unified scale after normalization according to the degradation start point, so that the model learning corresponds one-to-one with the degradation degree of rolling bearings, and maintains stable convergence and comparability under the difference in operating conditions between the source domain and the target domain, thereby supporting cross-domain and adaptive RUL prediction. The method according to claim 1 is characterized by the hierarchical subspace pyramid attention module. The hierarchical subspace pyramid attention module is inserted after the convolutional backbone. First, the feature F is divided into g subspaces according to channels, and then the kernel scale set {k0,…,k...} is considered. S-1 Parallel convolution yields branch features {F} i For each branch, the scale weight {α} is generated through a two-layer MLP (GAP→SE) process followed by Softmax. i }, and with Weighted fusion is performed, and a channel grouping and multi-core parallel strategy of spatial pyramid convolution is adopted within a single branch to effectively reduce the amount of computation while maintaining the representation of multiple receptive fields. The overall module assigns higher weights to branches that are more in line with the true scale of defects and suppresses branches that are more affected by operating condition disturbances. This results in more consistent feature representations under different speeds, loads and bearing specifications, improving cross-domain adaptive capability and enhancing the accuracy and stability of remaining life prediction.

4. The method according to claim 1, characterized in that: The adapter unit is a residual bottleneck structure of "dimensionality reduction-nonlinearity-dimensionality increase", which only trains newly added parameters and satisfies: Adapter(u) = u + Uδ(Wu + b1) + b2 Wherein, W is the dimensionality reduction mapping weight matrix, U is the dimensionality increase mapping weight matrix, b1 and b2 are bias vectors, and δ(·) is a nonlinear activation function (GELU or ReLU). The adapter unit is set in the temporal modeling layer and before and after it. During training and operation and maintenance, the backbone network parameters are frozen and only the adapter is fine-tuned in small steps. It is used to quickly align feature representations under the differences in rotational speed, load, lubrication and sensors between the source domain and the target domain. Thus, cross-domain adaptation is achieved in scenarios with few or weak labels and limited edge computing power, controlling the model size and computational load, and maintaining the accuracy and long-term stability of rolling bearing remaining life prediction.

5. The method according to claim 1, characterized in that: The cross-domain migration module includes both statistical alignment and adversarial alignment, which are used to solve the problem of excessively large differences in data distribution of rolling bearings under different equipment, speed, load and sensor conditions, thereby achieving adaptive remaining life prediction. Recently, to mitigate statistical bias caused by different operating conditions, the RBF kernel was developed. The adversarial alignment part employs a domain classifier based on a gradient inversion layer, and defines the domain alignment loss as follows: And through the gradient inversion layer during backpropagation The gradient acquisition enables the feature extraction layer to learn to facilitate the learning of "domain-inseparable" features, thereby further reducing cross-domain differences. The joint training objective is: Where Θ is the set of trainable parameters, and λ, γ, and β are the loss weights.

6. The method according to claim 1, characterized in that: To enable online prediction of the remaining life of rolling bearings on resource-constrained embedded or industrial terminals and reduce the dependence on source domain features after cross-domain migration, channel / convolution kernel-level structured pruning is implemented on the trained model. The channel scaling coefficient of batch normalization (BatchNorm) or the L1 norm of the convolution kernel is used as an importance index to prioritize the retention of channels and convolution kernels that contribute highly to degradation discrimination. Then, recovery training is performed on a small sample in the target domain until the accuracy threshold is reached, and the model is exported to a deployable format (ONNX or TorchScript) to obtain a low-latency, low-power model that can be ported across devices.

7. The method according to claim 1, characterized in that: The online inference after deployment adopts the same preprocessing and standardization strategy as the training phase. The newly added vibration data is sliced ​​into windows and forward calculations are performed to obtain RUL prediction. Alarm logic is set in combination with the predicted value threshold and its changing trend to stably indicate the bearing degradation stage under the changing operating conditions such as speed and load fluctuations, thereby improving the timeliness and reliability of the early warning.

8. The method according to claim 1, characterized in that: When only a small number of target domain labels or weak labels are obtained, the adapter and a small number of high-level parameters are periodically fine-tuned, and light re-pruning is performed according to the pruning importance index to cope with the distribution drift caused by changes in lubrication status, load disturbances, etc. Without increasing the model size, the target domain feature distribution is continuously aligned to maintain cross-domain adaptive capability and long-term stability of rolling bearing remaining life prediction.

Images