Rolling bearing residual life prediction method based on liquid time constant network

Abstract

The application discloses a rolling bearing residual service life prediction method based on a liquid time constant network, relates to the field of mechanical equipment predictive maintenance and fault diagnosis, and comprises the following steps: acquiring an original vibration signal of a rolling bearing in operation, extracting a multi-dimensional degradation feature tensor, inputting a nonlinear projection layer to obtain a hidden representation sequence; inputting the hidden representation sequence into a stacked liquid time constant network layer to obtain an updated hidden state sequence, inputting the updated hidden state sequence into a time attention module for weighted aggregation to generate a global degradation representation vector, inputting the global degradation representation vector into a prediction module, performing multiple independent random forward propagation to obtain multiple prediction values; calculating the mean and standard deviation of the multiple prediction values as a prediction expectation value and an uncertainty measure respectively, and outputting a residual service life prediction result and a confidence interval thereof. The application realizes high-precision life prediction with uncertainty quantification.

Description

Technical Field

[0001] This application relates to the field of predictive maintenance and fault diagnosis of mechanical equipment, and in particular to a method for predicting the remaining service life of rolling bearings based on a liquid time constant network. Background Technology

[0002] Rotating machinery is widely deployed in various basic industries. Rolling bearings, as core supporting components, operate under variable loads and harsh conditions for extended periods, making them highly susceptible to fatigue damage and even sudden failure. Therefore, implementing precise health monitoring and making reliable predictions of remaining life before failure are key technical approaches to achieving predictive maintenance, reducing operating costs, and ensuring equipment safety.

[0003] With the development of sensing technology and the Industrial Internet of Things (IIoT), data-driven deep learning methods have become the mainstream paradigm. However, existing sequence models, such as recurrent neural networks and their variants, mainly rely on discrete-time-step recurrent update mechanisms. At the physical level, bearing degradation is a continuously evolving and non-stationary dynamic process with fluctuating degradation rates. When degradation enters an accelerated phase and signal characteristics undergo drastic transient changes, discrete recurrent mechanisms based on fixed time steps often respond slowly, making it difficult to smoothly approximate the true continuous degradation trajectory and limiting further improvements in prediction accuracy.

[0004] Neural Ordinary Differential Equations (Neural ODEs) and Liquid Time-Constant Networks (LTCs) can relax the mandatory alignment constraints of discrete time steps, allowing the hidden state to evolve smoothly in the continuous time domain, providing a way to process highly non-stationary signals. However, directly applying them to bearing remaining life prediction still faces two major challenges: First, the feature shocks in the later stages of degradation are extremely severe, and without adaptive awareness of the local evolution rate, conventional continuous integral solvers are prone to computational lag in abrupt change intervals; Second, the long-range time hidden states generated by continuous evolution contain a large amount of redundant information from the early healthy stages, and there is a lack of effective mechanisms to isolate the core degradation focus and perform global feature aggregation. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as the mismatch between discrete-time modeling and continuous degradation processes, difficulty in handling abrupt changes in degradation rates, and lack of effective uncertainty quantification, this application provides a method for predicting the remaining service life of rolling bearings based on a liquid time constant network. By constructing a state-dependent adaptive gating mechanism and an additive time attention module, the method dynamically adjusts the integration step size and focuses on key degradation features. Combined with a Monte Carlo random deactivation mechanism, it achieves high-precision life prediction with uncertainty quantification.

[0006] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for predicting the remaining service life of rolling bearings based on a liquid time constant network, including: The original vibration signal of the rolling bearing is obtained, the multi-domain degradation features at each sampling time are extracted, and a multi-dimensional feature vector is constructed. The multi-dimensional feature vectors at each time within the window are stacked into a multi-dimensional degradation feature tensor in a sliding window manner. The multidimensional degenerate feature tensor is input into a nonlinear projection layer, and after nonlinear mapping and random deactivation, it is mapped to the hidden space to obtain a hidden representation sequence; each time step in the hidden representation sequence corresponds to a hidden representation vector. The hidden representation sequence is input into a stacked liquid time constant network layer; the liquid time constant network layer models the bearing degradation process as an ordinary differential equation in which the hidden state evolves in the continuous time domain; each layer in the stacked liquid time constant network layer processes the input sequence of the corresponding layer along the time axis time by time, calculates the discretization integration step size according to the hidden representation vector at the current time through a state-dependent adaptive gating mechanism, and uses a numerical integrator to perform numerical integration on the ordinary differential equation at the discretization integration step size to update the hidden state of the corresponding layer at the current time; after traversing all time points, the updated hidden states at each time point are arranged into the output sequence of the corresponding layer and used as the input sequence of the next layer; after the last layer is processed, the updated hidden state sequence is obtained; The updated hidden state sequence is input into the temporal attention module for weighted aggregation to generate a global degradation representation vector. The global degradation representation vector is input into the prediction module; during the inference phase, the random deactivation layer in the prediction module is kept active, and multiple independent random forward propagations are performed to obtain multiple prediction values; the mean and standard deviation of the multiple prediction values ​​are calculated and used as the expected value and uncertainty measure, respectively, and the remaining useful life prediction result and its confidence interval are output according to the expected value and the uncertainty measure.

[0007] Optionally, the multi-domain degradation features include time-domain statistical indices, frequency-domain energy representations, time-frequency domain multi-scale wavelet features, and nonlinear dynamic indices.

[0008] Optionally, the ordinary differential equation is specifically: ; in, For the first l Layer at time t The hidden state, For the first l Layer at time t-1 is a hidden state. For the first l Layer at time t Input features, For the first l Layer input weights, For the first l Layer recurrent kernel weights, For the first l The relaxation time constant for adaptive learning of the layer.

[0009] Optionally, the discretization integration step size is calculated based on the hidden representation at the processing time using a state-dependent adaptive gating mechanism, as shown in the following formula: ; in, To discretize the integration step size, For the first l Layer at time t Input features, It is the Sigmoid activation function. For learnable weights of the gated fully connected layer, For the bias of the gated fully connected layer, For the first gated boundary parameters, This is the second gating boundary parameter.

[0010] Optionally, the numerical integrator is a fourth-order Runge-Kutta solver, and its update formula is: ; ; ; ; ; in, This is the first derivative estimate. It is an ordinary differential equation. For a moment t Input features, For a moment t -1 is a hidden state. To discretize the integration step size, This is the estimate of the second derivative. This is the third derivative estimate. This is the fourth derivative estimate. For a moment t The hidden state.

[0011] Optionally, the formula for calculating the global degradation representation vector is: ; ; ; in, For the first i Attention response score at each time step For the first i The sequence of hidden states updated at each time step. For querying the matrix, The key mapping matrix, Assign weights to the normalized attention. W The time window length, C This is the global degradation representation vector. No. k Attention response score at each time step.

[0012] Optionally, during the inference phase, execution M In the next independent random forward propagation, the formulas for calculating the expected value of the prediction and the uncertainty measure are as follows: ; ; in, For the first a The remaining lifetime prediction after a random forward propagation. M This represents the total number of random forward propagations. To predict the expected value, It is a measure of uncertainty.

[0013] Optionally, the nonlinear mapping includes a fully connected mapping with Gaussian error linear unit activation.

[0014] Optionally, the number of layers in the stacked liquid time constant network layer is 2 to 5.

[0015] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method for predicting the remaining service life of rolling bearings based on a liquid time constant network. First, by constructing stacked liquid time constant network layers, the bearing degradation process is modeled as an ordinary differential equation in the continuous time domain, breaking through the fixed step size limitation of traditional discrete-time models. The integral step size is dynamically adjusted through a state-dependent adaptive gating mechanism. The step size is widened during the stable wear period to accelerate evolution, and the step size is rapidly narrowed during the severe degradation period to achieve high-resolution sampling of abrupt changes, thereby more smoothly and accurately approximating the real non-stationary degradation trajectory.

[0016] Secondly, by using an additive temporal attention module to perform differentiated weighted aggregation of the hidden states at each time point, the system autonomously assesses the differences in the contribution of different historical times to the prediction results, focuses attention on key degradation periods, and achieves accurate extraction of core degradation features.

[0017] Furthermore, by maintaining the activation of the randomly deactivated layer during the inference phase and performing multiple independent random forward propagations, the expected value of the remaining lifetime prediction and its confidence interval are output, breaking through the limitation of traditional models that only output deterministic point estimates, and providing a statistical safety boundary for maintaining decision-making. At the same time, the hidden state trajectory formed by the evolution of the liquid time constant network layer in the continuous time domain has a phase transition inflection point that highly coincides with the degradation initiation time based on physical signal statistics on the time axis, proving that this application can autonomously internalize the underlying degradation law of mechanical systems without human labeling intervention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the first process of a method for predicting the remaining service life of rolling bearings based on a liquid time constant network. Figure 2 This is a schematic diagram of the second process of a method for predicting the remaining service life of rolling bearings based on a liquid time constant network. Figure 3 This is a schematic diagram of the liquid time constant network layer in this application; Figure 4 A flowchart illustrating the process of global context aggregation and probabilistic remaining lifetime inference; Figure 5 for Figure 1 In the fourth step, the degenerate manifold inflection point is compared with the initial degenerate point for verification, and the manifold and degenerate thermodynamic scatter plots modeled by ordinary differential equations are used. Figure 6 for Figure 1 In the fourth step, the comparison and verification between the degenerate manifold inflection point and the initial degenerate point is shown in the vibration signal and initial degradation time diagram corresponding to the manifold inflection point region. Figure 7 The following are schematic diagrams of the PRONOSTIA experimental platform and bearing operating conditions: (a) is an overview diagram of PRONOSTIA; (b) is a force transmission diagram; (c) is a pneumatic jack diagram; (d) is a shaft support bearing diagram; and (e) is a bearing diagram under normal and deteriorated conditions. Figure 8 The evolution trajectory of the integration step size within the liquid time constant network layer is shown in the diagram. Figure 9 A visualization diagram of RUL prediction results and multidimensional errors; Figure 10 A comparison chart showing the remaining useful life prediction results and errors of different models on the Bearing1_3 dataset. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To make the objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] In the embodiments of this application, such as Figure 1 and Figure 2 As shown, a method for predicting the remaining service life of rolling bearings based on a liquid time constant network is provided. This method includes steps 101 to 105. Wherein: Step 101: Obtain the original vibration signal of the rolling bearing during operation, extract the multi-domain degradation features at each sampling time, construct a multi-dimensional feature vector, and stack the multi-dimensional feature vectors at each time within the window into a multi-dimensional degradation feature tensor in a sliding window manner.

[0023] In a preferred embodiment, the multi-domain degradation features comprise a 44-dimensional feature vector, specifically including: 15 time-domain statistical indices, 16 frequency-domain energy representations, 10 time-frequency domain multi-scale wavelet features, and 3 nonlinear dynamic indices. The stacking involves organizing the multi-dimensional feature vectors acquired continuously within a fixed time window W into a sliding sequence stream along the time axis, and reconstructing it into tensor form. To give the model a sense of history.

[0024] The 15 time-domain statistical indicators include 6 standard statistics and 9 predictive indicators. The 6 standard statistics are mean, standard deviation, median, interquartile range, signal entropy, and zero-crossing rate. The 9 predictive indicators are root mean square, skewness, kurtosis, peak value, peak-to-peak value, crest factor, impulse factor, waveform factor, and margin factor. The 16 frequency-domain energy characteristics include 10 spectral descriptors and 6 frequency band energy correlation features. The 10 spectral descriptors are power spectral density, spectral mean, spectral standard deviation, spectral centroid, spectral extension, spectral entropy, spectral flux, peak-to-frequency ratio, total power, and spectral roll-off. The 6 frequency band energy correlation features are the energy proportion of 5 specific frequency bands, the high-low frequency energy ratio, and the mid-low frequency energy ratio. The 5 specific frequency bands are 0-0.5kHz, 0.5-2kHz, 2-5kHz, 5-10kHz, and 10-12.8kHz. The 10 time-frequency domain multi-scale wavelet features are used to decompose the signal into four levels using the db4 wavelet basis. Wavelet energy and wavelet Shannon entropy are extracted from one approximation coefficient cA4 and four detail coefficients cD4, cD3, cD2, and cD1, respectively. The three nonlinear dynamic indices are detrended fluctuation analysis, Hearst exponent, and correlation dimension (embedding dimension m=3).

[0025] Step 102: Input the multidimensional degenerate feature tensor into the nonlinear projection layer, and map it to the hidden space through nonlinear mapping and random deactivation processing to obtain a hidden representation sequence; each time step in the hidden representation sequence corresponds to a hidden representation vector.

[0026] In a preferred embodiment, the nonlinear mapping includes a fully connected mapping and Gaussian Error Linear Unit (GELU) activation, and the random deactivation process is a dropout layer with a dropout rate set to 0.2; the dimension of the hidden space is set to 32 to 64 dimensions, preferably 48 dimensions. This operation effectively reduces information redundancy and potential dimensional coupling by nonlinearly projecting high-dimensional physical features into a low-dimensional hidden space, thereby reducing subsequent computational overhead and local noise interference.

[0027] Step 103, as follows Figure 3As shown, the hidden representation sequence is input into a stacked Liquid Time-Constant Networks (LTC) layer. The LTC layer models the bearing degradation process as an ordinary differential equation in which the hidden state evolves in the continuous time domain. Each layer in the stacked LTC layer processes the input sequence of its corresponding layer along the time axis, calculates the discretization integration step size based on the hidden representation vector at the current time using a state-dependent adaptive gating mechanism, and performs numerical integration on the ordinary differential equation at the discretization integration step size to update the hidden state of the corresponding layer at the current time. After traversing all time points, the updated hidden states at each time point are arranged into the output sequence of the corresponding layer and used as the input sequence of the next layer. After the last layer is processed, the updated hidden state sequence is obtained. The number of layers in the stacked LTC layer is 2 to 5, preferably 3.

[0028] Step 104: Input the updated hidden state sequence into the temporal attention module for weighted aggregation to generate a global degradation representation vector.

[0029] In a preferred embodiment, the temporal attention module is an additive attention module, employing an additive attention mechanism. This module receives the updated hidden state sequence from the stacked liquid time constant network layer. Then, the attention response scores at each time step are calculated and globally normalized using the softmax operator to obtain the attention allocation weights at each time step. Based on these attention allocation weights... The hidden states at each time step are weighted and aggregated into a global degeneracy representation vector. This enables adaptive focusing and global aggregation of key degraded nodes.

[0030] ; Where W is the length of the time window. For the first i The hidden state after the update at a certain moment.

[0031] Step 105, as follows Figure 4 As shown, the global degradation representation vector is input into the prediction module; during the inference phase, the random deactivation layer in the prediction module is kept active, and multiple independent random forward propagations are performed to obtain multiple prediction values; the mean and standard deviation of the multiple prediction values ​​are calculated and used as the expected value and uncertainty measure, respectively, and the remaining useful life prediction result and its confidence interval are output according to the expected value and the uncertainty measure.

[0032] In a preferred embodiment, the prediction module is a multilayer perceptron. After normal training, the multilayer perceptron acts as the regression prediction head to perform the final lifetime mapping; during the inference phase, the randomly deactivated layers in the multilayer perceptron are forced to remain active, with a dropout rate set to 0.2, and the same test input sequence is used for... M Each forward propagation is a series of independent random forward propagations. Each forward propagation, due to the random dropout effect of random inactivation, is equivalent to random sampling from the approximate posterior distribution of the network weights, thus obtaining a set of distributed lifetime predictions. By calculating the mean and standard deviation of this set of predictions, the expected value and uncertainty measure are obtained, and a statistical confidence interval for the prediction results is constructed.

[0033] By implementing steps 101 to 105 above, the bearing degradation process is modeled as an Ordinary Differential Equation (ODE) in the continuous time domain, abandoning the traditional fixed step size update paradigm. Through a state-dependent adaptive gating mechanism, the integration step size can be dynamically adjusted according to the hidden representation vector at the current moment. The step size is widened during the stable wear period to accelerate evolution and suppress local disturbances, and the step size is rapidly contracted during the severe degradation period to achieve high-resolution sampling and reconstruction of abrupt features, thereby more smoothly and accurately approximating the real non-stationary degradation trajectory.

[0034] Meanwhile, the time attention module performs differentiated weighted aggregation of the updated hidden states at each time point, and autonomously evaluates the differences in the contribution of different historical time points to the prediction results. As the device approaches failure, this application can concentrate the attention weight on the time period containing key degradation information, effectively filter out redundant information in the early stable stage, and achieve accurate extraction of core degradation features.

[0035] Furthermore, breaking through the limitations of traditional models that only output deterministic point estimates, this application uses Monte Carlo Dropout (MC Dropout) during the inference phase to maintain the activation of the stochastic deactivation layer and perform multiple independent stochastic forward propagations. The mean and standard deviation of the predicted values ​​are calculated and used as the expected value and uncertainty measure, respectively, outputting the remaining lifetime prediction result and its confidence interval, providing a statistically safe boundary for maintenance decisions. In addition, the hidden state trajectory formed by the evolution of the liquid time constant network layer in the continuous time domain in this application shows that its "V-shaped" phase transition inflection point highly coincides with the degradation initiation time based on physical signal statistics on the time axis. This proves that this application can autonomously mine and internalize the continuous phase transition physical laws at the bottom layer of the mechanical system without human labeling intervention. Figure 5 and Figure 6 As shown.

[0036] In a preferred embodiment, in step 103, the degradation process of the bearing is modeled as a continuous-time ordinary differential equation (ODE) of the hidden state. For the first... For a layered liquid time constant network, the continuous derivative of its hidden state is defined as: ; in, For the first l Layer at time t The hidden state, For the first l Layer at time t -1 is a hidden state. For the first l Layer at time t Input features, For the first l Layer input weights, For the first l The kernel weights of a layered recurrent neural network. For the first l The relaxation time constant of the layer's adaptive learning is strictly limited to the range [0.2, 0.8].

[0037] Furthermore, through a state-dependent adaptive gating mechanism, the discretization integration step size is calculated based on the hidden representation at the processing time, using the following formula: ; in, To discretize the integration step size, For the first l Layer at time t Input features, It is the Sigmoid activation function. For learnable weights of the gated fully connected layer, For the bias of the gated fully connected layer, For the first gated boundary parameters, This is the second gating boundary parameter. In specific implementation, it is set... =0.01, =0.19, thus setting the discretized integration time step to 0.19. Strictly constrained within the range of 0.01 to 0.20. The physical significance of using these fixed values ​​is that this mechanism endows the network with the ability to adaptively adjust the sensing rate according to different degradation stages. When the bearing is in the potential degradation stage or the stable wear stage, the slowly changing feature evolution dynamics drive the gating mechanism output to approach the upper limit ( With an integration step size of →0.20, the large-span state transition effectively suppresses the computational dissipation of early historical redundancy; when the system enters the accelerated degradation period accompanied by non-stationary transient fault impacts, local high-frequency mutations will trigger a rapid contraction of the gating threshold, forcing the discretization step size to approach the minimum lower limit (…). →0.01), thereby achieving high-resolution dense sampling and fine-grained reconstruction of nonlinear transient burst trajectories in the later stage of degradation.

[0038] Furthermore, the numerical integration solver is a fourth-order Runge-Kutta (RK4) solver, and its update formula is: ; ; ; ; ; in, This is the first derivative estimate. It is an ordinary differential equation. For a moment t Input features, For a moment t -1 is a hidden state. To discretize the integration step size, This is the estimate of the second derivative. This is the third derivative estimate. This is the fourth derivative estimate. For a moment t The hidden state.

[0039] In a preferred embodiment, in step 104, the formula for calculating the global degradation representation vector is: ; ; ; in, For the first i Attention response score at each time step For the first i The sequence of hidden states updated at each time step. For querying the matrix, The key mapping matrix, Assign weights to the normalized attention. W The time window length, C This is the global degradation representation vector. No. k Attention response score at each time step.

[0040] In a preferred embodiment, in step 105, during the reasoning phase, the following steps are performed: M In the next independent random forward propagation, the formulas for calculating the expected value of the prediction and the uncertainty measure are as follows: ; ; in, For the first a The remaining lifetime prediction after a random forward propagation. M This represents the total number of random forward propagations. To predict the expected value, It is a measure of uncertainty.

[0041] To further illustrate the technical solution and beneficial effects of this application, the following specific embodiments are used for verification.

[0042] (1) Data acquisition implementation method like Figure 7 As shown, this application uses the PRONOSTIA experimental platform to conduct accelerated degradation tests on rolling bearings. The system synchronously collects vibration signals in the horizontal and vertical orthogonal directions using accelerometers mounted on the test bearing housing. The sampling frequency is set to 25.6kHz, and the system fully records the entire process data of the bearing under real-world working conditions, from normal operation to severe spalling and eventual failure.

[0043] (2) Implementation of Model Training and Parameter Optimization To ensure that the stacked liquid time constant network layer and the overall prediction model constructed in this application can converge stably and have good generalization ability, the following training and optimization mechanism is adopted in this embodiment: Data partitioning and cross-validation: After standardizing the collected bearing life cycle data, the bearing data under the same working conditions are divided into training set, validation set and test set. During the training process, a cross-validation strategy is introduced to perform rolling evaluation of the data under different working conditions, so as to eliminate the evaluation bias caused by a single data partition and ensure the generalization ability of the model in unknown life scenarios.

[0044] Loss function: Mean Squared Error (MSE) is used as the overall loss function for model training to constrain the deviation between the predicted expected value of the model output and the true remaining life label.

[0045] Optimizer and learning rate decay: The Adam optimizer is used to iteratively update the network parameters, and the batch size is set to 128. In order to accelerate model convergence and avoid getting stuck in local optima, an adaptive learning rate decay strategy based on validation set loss is introduced. When the validation set loss does not decrease within 5 consecutive training rounds, the learning rate decay is triggered, and the decay factor is set to 0.5.

[0046] Overfitting prevention and gradient stabilization mechanisms: An early stopping mechanism is introduced, which terminates training early when the validation set error no longer decreases within a set tolerance number of rounds, and saves the current optimal model weights; at the same time, in order to alleviate the gradient explosion problem that may occur during the numerical integration of continuous-time domain ordinary differential equations, a gradient clipping mechanism is used in conjunction, with the maximum norm of the clipping threshold set to 1.0, to ensure the numerical stability of the model in complex dynamic evolution calculations.

[0047] (3) Verification of the interpretability of the internal mechanism like Figure 8 As shown, the physical rationality of the proposed mechanism is verified by tracking the evolution trajectory of the discretized integral step size output by the state-dependent adaptive gating mechanism at each time step. In the early stage of equipment operation and the steady wear stage, the dominant integral step size is maintained at a high level close to the upper limit to filter high-frequency redundant noise and accelerate state evolution. When the equipment enters the accelerated degradation period and features exhibit high-frequency abrupt changes, the gating mechanism drives the integral step size to undergo dense local adaptive contraction, approaching the minimum lower limit, thereby achieving high-resolution sampling and reconstruction of local core degradation features.

[0048] (4) Validation of prediction effect and advantages like Figure 9 As shown, the remaining useful life prediction expectation curve generated by this application can track the actual degradation baseline with high fidelity. Its 95% confidence interval converges rapidly and closely encloses the actual trend after entering the accelerated degradation stage, providing a quantitative reference for on-site fault-tolerant maintenance decisions. Furthermore, as... Figure 10 As shown, compared with traditional sequence models based on fixed discrete time steps, such as Temporal Convolutional Network (TCN) and Convolutional Neural Network-Bilateral Long Short-Term Memory Network-Attention Mechanism (CNN-BiLSTM-Attention), the method in this application can effectively overcome the perceptual lag phenomenon in the abrupt change interval, exhibiting lower mean absolute error (MAE) and root mean square error (RMSE), and achieving smooth approximation and high-precision lifetime inference for highly nonlinear dynamic processes.

[0049] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0050] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting the remaining service life of rolling bearings based on a liquid time constant network, characterized in that, The method includes: The original vibration signal of the rolling bearing is obtained, the multi-domain degradation features at each sampling time are extracted, and a multi-dimensional feature vector is constructed. The multi-dimensional feature vectors at each time within the window are stacked into a multi-dimensional degradation feature tensor in a sliding window manner. The multidimensional degenerate feature tensor is input into a nonlinear projection layer, and after nonlinear mapping and random deactivation, it is mapped to the hidden space to obtain a hidden representation sequence; each time step in the hidden representation sequence corresponds to a hidden representation vector. The hidden representation sequence is input into a stacked liquid time constant network layer; the liquid time constant network layer models the bearing degradation process as an ordinary differential equation in which the hidden state evolves in the continuous time domain; each layer in the stacked liquid time constant network layer processes the input sequence of the corresponding layer along the time axis time by time, calculates the discretization integration step size according to the hidden representation vector at the current time through a state-dependent adaptive gating mechanism, and uses a numerical integrator to perform numerical integration on the ordinary differential equation at the discretization integration step size to update the hidden state of the corresponding layer at the current time; after traversing all time points, the updated hidden states at each time point are arranged into the output sequence of the corresponding layer and used as the input sequence of the next layer; after the last layer is processed, the updated hidden state sequence is obtained; The updated hidden state sequence is input into the temporal attention module for weighted aggregation to generate a global degradation representation vector. The global degradation representation vector is input into the prediction module; during the inference phase, the random deactivation layer in the prediction module is kept active, and multiple independent random forward propagations are performed to obtain multiple prediction values; the mean and standard deviation of the multiple prediction values ​​are calculated and used as the expected value and uncertainty measure, respectively, and the remaining useful life prediction result and its confidence interval are output according to the expected value and the uncertainty measure.

2. The method for predicting the remaining service life of rolling bearings based on a liquid time constant network according to claim 1, characterized in that, The multi-domain degradation features include time-domain statistical indices, frequency-domain energy characterization, time-frequency domain multi-scale wavelet features, and nonlinear dynamic indices.

3. The method for predicting the remaining service life of rolling bearings based on a liquid time constant network according to claim 1, characterized in that, The ordinary differential equation is specifically as follows: ； in, For the first l Layer at time t The hidden state, For the first l Layer at time t -1 is a hidden state. For the first l Layer at time t Input features, For the first l Layer input weights, For the first l Layer recurrent kernel weights, For the first l The relaxation time constant for adaptive learning of the layer.

4. The method for predicting the remaining service life of rolling bearings based on a liquid time constant network according to claim 1, characterized in that, The discretization integration step size is calculated based on the hidden representation at the processing time using a state-dependent adaptive gating mechanism, as shown in the following formula: ； in, To discretize the integration step size, For the first l Layer at time t Input features, It is the Sigmoid activation function. For learnable weights of the gated fully connected layer, For the bias of the gated fully connected layer, For the first gated boundary parameters, This is the second gating boundary parameter.

5. The method for predicting the remaining service life of rolling bearings based on a liquid time constant network according to claim 1, characterized in that, The numerical integrator is a fourth-order Runge-Kutta solver, and its update formula is as follows: ； ； ； ； ； in, This is the first derivative estimate. It is an ordinary differential equation. For a moment t Input features, For a moment t -1 is a hidden state. To discretize the integration step size, This is the estimate of the second derivative. This is the third derivative estimate. This is the fourth derivative estimate. For a moment t The hidden state.

6. The method for predicting the remaining service life of rolling bearings based on a liquid time constant network according to claim 1, characterized in that, The formula for calculating the global degradation representation vector is: ； ； ； in, For the first i Attention response score at each time step For the first i The sequence of hidden states after each time step update. For querying the matrix, The key mapping matrix, Assign weights to the normalized attention. W The time window length, C This is the global degradation representation vector. No. k Attention response score at each time step.

7. The method for predicting the remaining service life of rolling bearings based on a liquid time constant network according to claim 1, characterized in that, During the reasoning phase, execution M In the next independent random forward propagation, the formulas for calculating the expected value of the prediction and the uncertainty measure are as follows: ; ; in, For the first a The remaining lifetime prediction after a random forward propagation. M This represents the total number of random forward propagations. To predict the expected value, It is a measure of uncertainty.

8. The method for predicting the remaining service life of rolling bearings based on a liquid time constant network according to claim 1, characterized in that, The nonlinear mapping includes fully connected mapping and Gaussian error linear unit activation.

9. The method for predicting the remaining service life of rolling bearings based on a liquid time constant network according to claim 1, characterized in that, The number of layers in the stacked liquid time constant network layer is 2 to 5.

Images