Method and system for predicting residual service life of aero-engine rolling bearing based on coordinate neural field

Abstract

This invention discloses a method and system for predicting the remaining service life of rolling bearings in aero-engines based on coordinate neural fields, belonging to the field of fault prediction and health management technology. It includes: constructing and training a conditionally modulated coordinate neural field network, representing the time-frequency observations of vibration signals as a continuous field, and applying a random mask during training to learn missing measurement repair; aggregating the features extracted by the network into a latent representation, and training a conditional diffusion model to learn its evolutionary priors; during prediction, obtaining the current latent representation based on the latest observations, and introducing a gradient based on observation likelihood for posterior guidance during the backsampling process of the diffusion model to generate future trajectories; finally, decoding the trajectory and calculating a health index to obtain the probability distribution of the remaining service life. This invention achieves high-precision, calibrable, and highly adaptive probabilistic prediction of remaining service life by uniformly handling data missingness through continuous representation and combining generative models with posterior guidance.

Description

Technical Field

[0001] This invention relates to the field of aero-engine health management technology, specifically to a technology for predicting the failure and remaining service life of rolling bearings, and particularly to a method and system for predicting the remaining service life of aero-engine rolling bearings based on coordinate neural fields. Background Technology

[0002] As a critical rotating component, the health of the main bearing of an aero-engine directly determines flight safety. Accurate prediction of its Remaining Useful Life (RUL) is central to predictive maintenance. The degradation process of the bearing is reflected in vibration signals, exhibiting significant non-stationary characteristics. However, in practical engineering monitoring scenarios, acquiring vibration data often faces multiple challenges: 1) Limited by transmission bandwidth and edge computing capabilities, sparse sampling or intermittent acquisition is often used, leading to discontinuous data; 2) Sensors may experience intermittent failures or communication interference, resulting in complete data loss over long periods; 3) The complex and variable operating conditions of the engine (speed, load) cause data distribution drift.

[0003] Currently, mainstream RUL prediction methods are mainly based on deep learning models, such as Long Short-Term Memory networks and Transformers. These methods typically require fixed-length, well-sampled time series or time-frequency images as input. When the input data is missing or the sampling rate is inconsistent, preprocessing techniques such as interpolation and resampling are often needed to fill in and normalize the data. These operations introduce noise and distortion, destroying the integrity of fault features and severely affecting prediction accuracy. More importantly, most of these methods output point estimates or simple interval estimates, whose uncertainty is difficult to calibrate, making it impossible to provide a reliable probabilistic risk measure for high-risk maintenance decisions.

[0004] Coordinate neural fields (CNF) are an emerging implicit continuous representation technique, adept at learning continuous physical fields from discrete observations. Diffusion models, as a cutting-edge generative model, possess powerful capabilities in multimodal data generation and uncertainty quantification. However, currently, there is no technical solution to organically combine the two to construct a unified framework for predicting the remaining service life of rolling bearings, capable of handling missing data, performing probabilistic predictions, and possessing online adaptive capabilities. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for predicting the remaining service life of rolling bearings in aero-engines based on coordinate neural fields. It aims to fundamentally address data gaps through continuous representation and utilize a combination of generative models and posterior assimilation to achieve high-precision, calibrable, and highly adaptive probabilistic remaining service life prediction.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for predicting the remaining service life of rolling bearings in aero-engines based on coordinate neural fields; the method includes the following steps:

[0008] The continuous time-frequency field construction steps are as follows: Vibration signals from the rolling bearings of an aero-engine are acquired, and time-frequency transformation is performed to obtain discrete time-frequency observation data. Using the discrete time-frequency observation data as the training target, a conditionally modulated coordinate neural network is constructed and trained. This network takes as input continuous coordinates containing time, frequency, and channels, and a condition vector consisting of rotor speed, load, and engine exhaust temperature, and outputs a continuous time-frequency field representation. During training, a random mask is applied to the discrete time-frequency observation data.

[0009] The modeling steps for the latent space diffusion model are as follows: the features extracted by the coordinate neural network are aggregated into a latent representation of fixed dimensions through weighted average pooling; and a diffusion model conditioned on the working condition vector is trained in the latent representation space.

[0010] Posterior-guided prediction steps: In the prediction phase, the current latent representation is obtained based on the latest observation data; the current latent representation is used as the initial latent variable, and backsampling of the diffusion model is performed to generate multiple future latent variable trajectories; during the backsampling process, the likelihood gradient is calculated based on the difference between the latest observation data and the time-frequency field data reconstructed from the current latent variable, and the update of the latent variable is guided by the likelihood gradient.

[0011] Lifetime distribution decision steps: decode the multiple future latent variable trajectories into corresponding future time-frequency fields, and calculate the health index curve of each time-frequency field; based on the time when each health index curve first exceeds the preset failure threshold, obtain the remaining lifetime sample set; perform statistical analysis on the remaining lifetime sample set, and output the probability distribution information of the remaining lifetime.

[0012] Secondly, the present invention provides a system for predicting the remaining service life of rolling bearings in aero-engines based on coordinate neural fields, for implementing the above-mentioned method; the system includes:

[0013] The preprocessing module is used to acquire the vibration signal of the rolling bearing of the aero-engine and perform time-frequency transformation to obtain discrete time-frequency observation data;

[0014] A continuous time-frequency field modeling module, connected to the preprocessing module, is used to run a conditionally modulated coordinate neural field network. The coordinate neural field network takes as input continuous coordinates containing time, frequency, and channels, and a working condition vector consisting of rotor speed, load, and engine exhaust temperature, and outputs a continuous time-frequency field representation.

[0015] A latent diffusion and assimilation module, connected to the continuous time-frequency field modeling module, includes:

[0016] The latent representation aggregation unit is used to aggregate the features extracted by the continuous time-frequency field modeling module into a latent representation through weighted average pooling.

[0017] The conditional diffusion model unit is used to perform prior modeling and backsampling in the latent representation space with the conditional vector as a condition;

[0018] The posterior guidance unit is used to calculate the likelihood gradient based on the difference between the latest observation data and the time-frequency field data reconstructed from the current latent variables during the backsampling process, and to guide the latent variable update with the likelihood gradient.

[0019] The latent diffusion and assimilation module is configured to generate multiple future latent variable trajectories based on the latent representation, through the collaboration of the conditional diffusion model unit and the posterior guidance unit.

[0020] The decision output module, connected to the latent diffusion and assimilation module, is used to decode the generated future latent variable trajectory into a time-frequency field, calculate the health index and remaining lifespan distribution, and output the prediction results.

[0021] Beneficial effects:

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] This invention is based on the continuous function representation of coordinate neural fields. The model can directly handle observations with arbitrary missing values ​​and output results at arbitrary resolution, fundamentally solving the dependence on regular grid data and improving robustness. The generative framework based on the diffusion model naturally generates future distributions through sampling, and combined with posterior-guided online correction, makes the output confidence intervals and risk probabilities realistic and reliable.

[0024] This invention achieves online model updates through a posterior-guided mechanism, allowing it to adapt to new operating conditions or individual differences without retraining, greatly improving the model's practicality and deployment flexibility. By outputting the complete probability distribution and risk curve of remaining useful life, maintenance decisions can be upgraded from point estimation to scientific decision-making based on risk probability. Attached Figure Description

[0025] Figure 1 A flowchart illustrating the remaining useful life prediction method provided in this embodiment of the invention;

[0026] Figure 2 This is a schematic diagram of the probability distribution of remaining useful life (RUL) in an embodiment of the present invention;

[0027] Figure 3 This is a module structure diagram of the remaining useful life prediction system provided in an embodiment of the present invention. Detailed Implementation

[0028] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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 based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0030] Example 1

[0031] See Figure 1 This embodiment details the specific implementation steps of the method for predicting the remaining service life of aero-engine rolling bearings based on coordinate neural fields provided by the present invention.

[0032] Step S101: Data Acquisition and Preprocessing

[0033] The vibration signal a(t) of the rolling bearing is acquired by an accelerometer mounted on the engine casing or bearing housing. Simultaneously, engine operating parameters, such as high-pressure rotor speed N2, load, and engine exhaust temperature, are recorded to form the operating condition vector u. The original vibration signal a(t) is bandpass filtered (e.g., retaining the frequency band from 100Hz to 5000Hz to include the bearing's characteristic frequencies) to remove high-frequency noise and extremely low-frequency interference. The filtered signal is divided into windows of fixed duration (e.g., 2 seconds), with adjacent windows overlapping by 50%. A short-time Fourier transform is performed on the signal within each window, using a Hanning window as the window function, to obtain the complex time-frequency spectrum X(t, f). Its logarithmic magnitude log(|X(t, f)|+ε) (ε is a very small constant to prevent numerical underflow) is taken as the physical field value y to be modeled in subsequent steps. Meanwhile, based on the actual data collection records, the continuous vibration signal is converted into discrete time-frequency observation data, generating a binary mask matrix M with the same size as the observation data, where 1 indicates that there is a valid observation at that time-frequency point, and 0 indicates that it is missing.

[0034] Step S102: Constructing a coordinate neural field modulated with training conditions

[0035] The core of this step is to train a conditionally modulated coordinate neural network capable of representing discrete time-frequency observations as continuous fields. Specifically, a sinusoidal representation network (SIREN) structure based on characteristic linear modulation (FiLM) is employed.

[0036] Input and Output: The network takes normalized continuous coordinates s = (t, f, c) and a conditional embedding vector u encoded by a multilayer perceptron as input, and directly outputs the field value at that coordinate. In this context, time t and frequency f are both normalized to the [0, 1] interval; channel c is an embedded representation of the sensor index. Coordinates s define the query location, and u explicitly modulates network behavior to model the impact of operating conditions.

[0037] Network Structure: The coordinate neural network uses a sine wave representation of its structure. The main body is a fully connected network containing L layers; for example, L is 5-8 layers, preferably L=6, with each layer having a width of 256. The calculation formula for the l-th layer (l ranges from 0 to L-1) is typically expressed as:

[0038] h_{l+1} = sin(ω0 (W_l h_l + b_l))

[0039] The meaning of each symbol is as follows:

[0040] ω0: This is the base frequency or frequency scaling factor, a hyperparameter. It is not learned through training but is a constant set during network initialization. Its function is to scale the linear transformation result of the input to the sine function. The value of ω0 directly determines the range of signal frequencies that the network can effectively represent and fit. A larger ω0 value (e.g., 30) allows the network to capture higher frequency details in the signal, which is crucial for accurately representing the sharp spectral peaks and rapidly changing textures caused by impact faults in the vibration time-frequency plot. If ω0 is too small, the network will struggle to express high-frequency information, resulting in an overly smooth fitted time-frequency field that loses fault characteristics.

[0041] W_l: This is the weight matrix for the l-th layer, a trainable parameter. Its dimensions are (d_out, d_in), where d_in is the dimension of the input feature h_l, and d_out is the dimension of the output feature (i.e., the input to the next layer). The role of W_l is to perform linear transformations and feature space rotations on the input feature h_l, learning how to combine different coordinates and conditional information to generate meaningful intermediate representations.

[0042] h_l: The feature vector (or activation value) input to the l-th layer. For the input layer (l=0), h_0 is the total input of the network, i.e., the concatenated or fused continuous coordinates s and the conditional information modulated by the FiLM subnetwork. For the hidden layer, h_l is the result of the output of the previous layer after sinusoidal activation.

[0043] b_l: This is the bias vector for the l-th layer, also a trainable parameter. Its dimension is d_out. The role of the bias is to shift the result of the linear transformation W_l·h_l before applying the sine function. This provides the network with additional flexibility to adjust the phase of the sine wave, thereby better fitting complex functions. W_l·h_l + b_l completes an affine transformation (linear transformation + translation), mapping features from one space to another. Subsequently, ω0 globally scales this result, expanding its numerical range. Finally, the sin(·) function is applied to the scaled result. The periodicity, boundedness, and non-zero higher-order derivatives of the sine function give the SIREN network a significant advantage over traditional networks using activation functions such as ReLU and Tanh when fitting complex signals containing high-frequency components (such as vibration time-frequency maps). It avoids spectral bias and more accurately reconstructs the details and edges of the signal.

[0044] To incorporate the influence of operating conditions, a characteristic linear modulation technique is employed: an independent FiLM subnetwork (a 3-layer MLP) takes the conditional embedding as input and dynamically generates a pair of modulation parameters (γ_l, β_l) for each layer l of the SIREN based on the operating condition vector u. γ_l and β_l are the element-wise scaling factor and element-wise bias term, respectively. The modulated layer calculation becomes: h{l+1} = sin(ω0·(W_l) h_l + b_l)) ⊙γ_l+β_l, where ⊙ denotes element-wise multiplication. γ_l and β_l explicitly and finely modulate the amplitude (scaling) and baseline (bias) of the activation features in layer l. This allows the same network to dynamically adjust its internal feature representation according to different operating condition inputs (u), thereby accurately modeling the modulation effects of operating conditions such as spectral peak shift to higher frequencies under high speeds and overall energy enhancement under high loads on the vibration spectrum. This is one of the key mechanisms for achieving cross-operating condition adaptive prediction.

[0045] Training Objective and Masking Strategy: Discrete time-frequency observation data is used as the training objective. The training data consists of a large number of windows with time-frequency observations y and a mask M. In each training iteration, K=2048 coordinate points {s_i} are randomly sampled from the time-frequency grid of each window in the current batch, and the corresponding field values ​​{y_i} and mask values ​​{m_i} are obtained. To endow the network with strong missing data repair capabilities, a random mask M_train is applied to the input training objective (i.e., discrete time-frequency observation data) during training. The occlusion mode of the random mask includes at least one of the following: random point missing mode, block missing mode, and periodic missing mode. This multi-mode masking strategy simulates various real missing scenarios, forcing the network to learn generalization ability. For example, observation points are randomly set to missing (m_i changes from 1 to 0) with a probability of p=40% to simulate various data missing scenarios. The output of the coordinate neural field network is the predicted value {y_i} at these coordinate points. _i}.

[0046] Loss Function: The loss function used to train the coordinate neural network includes field value reconstruction loss at observed coordinate points and completion consistency loss at occluded coordinate points. The reconstruction loss ensures the network fits existing data; the completion consistency loss forces the network to generate outputs in the masked region that are physically consistent with the surrounding observations and operating conditions, which is crucial for achieving high-quality missing data restoration and super-resolution. By minimizing the aforementioned losses, the network learns to reconstruct a complete and continuous time-frequency field representation from missing observations.

[0047] This step yields an intelligent function Φ(s; u). For any given coordinates and operating conditions, it can provide reasonable time-frequency amplitude values, thereby transforming discrete, incomplete observations into a complete and continuous time-frequency physical field representation, providing a unified and high-quality data foundation for subsequent steps.

[0048] Step S103: Modeling the latent space diffusion model

[0049] The goal of this step is to model the degenerate dynamics in a low-dimensional, compact space.

[0050] Latent Representation Aggregation: Data is processed using a trained coordinate neural network Φ. For an input window, its output field value is obtained, and feature vectors of its intermediate layers at all sampled coordinate points are extracted. These features are then aggregated into a fixed-dimensional (e.g., 128-dimensional) latent representation z through an attention pooling layer. Specifically, the latent representation aggregation method is as follows: weighted average pooling is performed on the feature vectors of the intermediate layers of the coordinate neural network at all sampled coordinate points. The pooling weights are generated based on the sensitivity of the coordinates to faults through an attention mechanism. Specifically, the attention mechanism is implemented through a lightweight fully connected network that takes the frequency information of the coordinates as the main input and considers the temporal context to predict the importance score of each coordinate point to the current bearing health status representation, and uses this as the weight for the weighted average pooling. This means that coordinate points near the fault feature frequency will receive higher weights, thus aggregating a fixed-dimensional latent representation z with higher information density. This operation condenses the degradation state information at the current moment, allowing the aggregation process to focus on the time-frequency region that is more sensitive to faults.

[0051] Diffusion Model Training: Obtain the latent representation sequence {z} for all time windows and its corresponding condition u. Train a diffusion model conditioned on the operating condition vector in the latent representation space. This model includes a forward noise-adding process (gradually transforming the data into Gaussian noise) and a backward denoising process (reconstructing the data from the noise). The core of the model is a U-Net-structured noise predictor ε_θ(z_t, t, u), where z_t is the noisy latent variable at step t, t is the time-step embedding, and u is the condition. The training objective is to predict the noise added during the forward process. After training, the model learns the probability distribution p(z_{t+1}|z_t, u) of the latent representation z evolving over time. Its role is to establish a powerful prior model of the degradation process, capable of generating multiple possible future degradation paths under a given operating condition.

[0052] Step S104: Online posterior guided prediction

[0053] This step is central to online prediction and adaptation.

[0054] When the system receives a new observation window (corresponding to time-frequency data y_obs and mask M_obs) and the current operating condition u_now:

[0055] Obtain the current state: Input (s, u_now) into the trained coordinate neural field Φ, ignore the output, and directly extract the current latent representation z_now after attention pooling through step S102. Even if y_obs is missing, Φ can still provide a complete latent representation.

[0056] Set up the forecast task: Set the future time step H to be forecasted (e.g., 50 future windows).

[0057] Perform post-hoc guided sampling:

[0058] Backsampling to generate trajectories: Using the current latent representation z_now as the initial latent variable, a backsampling process of the diffusion model is executed using a trained noise predictor ε_θ and a deterministic or stochastic sampler such as the DDIM sampler to gradually generate multiple (H) future latent variable trajectories. For example, in the i-th sampling step, the current latent variable is z_i. This is the process of gradually denoising and generating data from noise. Specifically, for each latent time step i in the sampling process, the corresponding time-frequency field data is reconstructed using the decoding capability of the trained coordinate neural network Φ (or by inputting the latent variable z_i into a lightweight decoder network D). _i. Based on the latest actual observation data Y_obs (only at observed coordinate points), calculate the Gaussian observation likelihood function, for example:

[0059] L(z_i) ∝ exp(-‖M_obs ⊙ (Y_obs - _i)‖² / (2σ²))

[0060] Where M_obs is the observation mask matrix, ⊙ represents element-wise multiplication, and σ² is the preset observation noise variance.

[0061] Posterior guidance: During backsampling, the likelihood gradient of the latest observation data with respect to the current latent variable is calculated, and this likelihood gradient guides the update of the latent variable. Specifically, the likelihood gradient is calculated as follows: a Gaussian observation likelihood function is constructed based on the difference between the latest observation data and the time-frequency field data reconstructed from the current latent variable, and the gradient of this likelihood function with respect to the current latent variable is calculated. This likelihood gradient indicates how to fine-tune the currently generated latent variable so that the reconstructed data is more like the latest real observation. The diffusion model backsampler Sampler(z_i, ...) calculates the latent variable update amount that conforms to the prior distribution based on the noise predicted by the noise predictor ε_θ, the current latent variable z_i, and time step information. For example, under the DDIM sampling rule, this update amount can be specifically represented as a specific function based on ε_θ. The likelihood gradient guidance determines the next update direction of the latent variable by multiplying the calculated likelihood gradient by a step size coefficient η and then combining it with the noise update amount predicted in the diffusion model backsampling step.

[0062] Specifically, calculate the gradient of the observed likelihood L(z_i) with respect to the current latent variable z_i. _{z_i} log L(z_i). This likelihood gradient indicates how to adjust z_i to make its reconstruction more consistent with the latest observations. In each update step of the diffusion model backsampling, the above likelihood gradient is fused into the original sampling update direction with a certain weight. One specific implementation is as follows:

[0063] z_{i-1} = Sampler_ε_θ(z_i, t, u) +η· _{z_i} log L(z_i)

[0064] Where Sampler_ε_θ(...) represents the prior update calculated by the backsampler of the diffusion model of the noise predictor ε_θ, and η is an adjustable guide step size coefficient used to control the intensity of the introduction of observation information.

[0065] Repeat the above process to generate an observation-guided future latent variable trajectory. By changing the random seed or the randomness of the sampling process, and performing multiple independent samplings, N possible future trajectories {z_{1:H}^{(k)}}, k=1,...,N can be obtained.

[0066] The posterior guidance mechanism enables online posterior assimilation. It ensures that the generation process not only follows priors learned from historical data (diffusion model) but is also constrained by real-time observations. This allows the model to adapt to new operating conditions or sudden changes in circumstances reflected in the observations with zero samples, significantly improving the real-time performance and accuracy of predictions.

[0067] Step S105: Calculation and Output of Remaining Lifetime Distribution

[0068] The uncertainty generated in this step is transformed into decision information.

[0069] Trajectory Decoding and Health Index Calculation: Multiple future latent variable trajectories are decoded into corresponding future time-frequency fields. Specifically, for each generated future latent variable trajectory, the decoding part of the Φ network (or a separately trained lightweight decoder) is used to map it back to the time-frequency field sequence {Y_t^(k)}. For the time-frequency field Y_t^(k) at each future time t, its health index HI_t^(k) is calculated. The health index can be calculated in any of the following ways: (1) calculating the ratio of fault-sensitive frequency band energy to full-band energy from the future time-frequency field; (2) calculating spectral kurtosis or spectral entropy from the future time-frequency field; (3) mapping it through a pre-trained regression network with latent representation as input and health index as output. In this embodiment, spectral kurtosis is used as the health index because it is very sensitive to the impact characteristics caused by early bearing failures.

[0070] Failure Assessment and RUL Sample Generation: A failure threshold τ is set (e.g., spectral kurtosis greater than 8). For the k-th trajectory, the time point T_fail^(k) when its health index sequence {HI_t^(k)} first exceeds the threshold τ is found. Then, the remaining lifetime sample corresponding to this trajectory is RUL^(k) = T_fail^(k) - T_now, where T_now is the current time. Based on the time when each health index (HI) curve first exceeds the preset failure threshold, a RUL sample is generated for each trajectory, resulting in the remaining lifetime sample set {RUL^(k)}.

[0071] Statistical Analysis: Perform statistical analysis on the above sample set of remaining useful life, and output the probability distribution information of the remaining useful life, such as... Figure 2 As shown. Specifically, calculate the mean, median, and standard deviation of the sample set {RUL^(k)}; calculate its empirical distribution function, and read the 5% and 95% quantiles from it as the 90% confidence interval [RUL_{5%}, RUL_{95%}].

[0072] Output: This step directly produces actionable decision support information. Specifically, it calculates the empirical cumulative distribution function CDF(t) = P(RUL≤t), whose curve is the early failure probability curve (risk curve), visually representing the probability of failure occurring before any future time t. The system's final output includes: the expected value of RUL, the quantile interval at a specified confidence level (e.g., the 90% confidence interval), and the risk curve. When the risk curve exceeds a preset alarm threshold (e.g., 10%) in the short term (e.g., within the next 24 hours), the system automatically triggers an alert, thus achieving true risk warning.

[0073] Example 2

[0074] See Figure 3 This embodiment provides a coordinate neural field-based system for predicting the remaining service life of rolling bearings in aero-engines, used to implement the above-mentioned method. The system includes:

[0075] The preprocessing module 410 is used to acquire the vibration signal of the rolling bearing of the aero-engine and perform time-frequency transformation to obtain discrete time-frequency observation data. This module is responsible for receiving data from vibration sensors and operating condition sensors, performing signal filtering, segmentation, time-frequency transformation (such as STFT), and outputting discrete time-frequency observation matrix and corresponding mask matrix.

[0076] A continuous time-frequency field modeling module 420 is connected to a preprocessing module 410. This module internally runs a conditionally modulated coordinate neural network as described in step S102 of Embodiment 1. This coordinate neural network takes as input continuous coordinates containing time, frequency, and channels, and a condition vector consisting of rotor speed, load, and engine exhaust temperature, and outputs a continuous time-frequency field representation. Preferably, the conditionally modulated coordinate neural network is a sinusoidal representation network with characteristic linear modulation, and the training process of the coordinate neural network includes a step of applying a random mask to the input data.

[0077] The latent diffusion and assimilation module 430 is connected to the continuous time-frequency field modeling module 420. This module is the intelligent core of the system and specifically includes:

[0078] The latent representation aggregation unit 431 receives features from the continuous time-frequency field modeling module 420 and performs the aggregation operation as described in step S103, aggregating the features extracted by the continuous time-frequency field modeling module into a latent representation through weighted average pooling.

[0079] Conditional diffusion model unit 432 is used to perform the model function in step S103, and to perform prior modeling and backsampling in the latent representation space.

[0080] The posterior guidance unit 433 is used to perform the guidance function in step S104. During the backsampling process, it calculates the likelihood gradient based on the difference between the latest observation data and the time-frequency field data reconstructed from the current latent variables, and uses the likelihood gradient to guide the latent variable update.

[0081] The latent diffusion and assimilation module 430 is configured to generate multiple future latent variable trajectories based on the latent representation, through the collaboration of the conditional diffusion model unit and the posterior guidance unit. This clarifies the final output of the module.

[0082] The decision output module 440, connected to the latent diffusion and assimilation module 430, receives multiple future latent variable trajectories generated by the receiving module 430. This module is used to execute step S105 of Embodiment 1, decoding the generated future latent variable trajectories into a time-frequency field, calculating the health index and remaining lifetime distribution, and outputting the prediction results. The prediction results output by the decision output module include at least one of the following: the expected value of the remaining lifetime, the quantile interval at a specified confidence level, and the premature failure probability curve.

[0083] To balance computational efficiency, real-time requirements, and model complexity, the system in this embodiment of the invention can adopt an edge-cloud collaborative deployment architecture.

[0084] On an edge computing device (such as an airborne data acquisition unit), a preprocessing module 410 and a lightweight continuous time-frequency field modeling module 420 are deployed. Their responsibility is to process the raw signal in real time, complete the continuous field reconstruction and latent representation extraction, and upload the low-dimensional latent representation z and operating condition u, which greatly saves uplink bandwidth.

[0085] On the cloud server, a complete latent diffusion and assimilation module 430 and a decision output module 440 are deployed. The cloud leverages its powerful computing capabilities to perform computationally intensive diffusion model backsampling, posterior guidance, and trajectory decoding analysis to generate the final risk prediction report. This architecture balances the real-time performance of the edge side with the powerful computing capabilities of the cloud side. The cloud also includes a model management module responsible for version management, security updates, and incremental learning using anonymized runtime data for all AI models.

[0086] The technical solution provided by this invention can be directly integrated into the engine health management system of airlines and the remote monitoring platform of manufacturers for online life assessment and risk warning of bearings of wing engines. It provides key technical support for formulating condition-based maintenance plans and optimizing spare parts management, and has significant economic and safety value.

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining service life of rolling bearings in aero-engines based on coordinate neural fields, characterized in that, Includes the following steps: Vibration signals of rolling bearings in aero-engines are acquired, and time-frequency transformation is performed to obtain discrete time-frequency observation data; Using the discrete time-frequency observation data as the training target, a conditionally modulated coordinate neural network is constructed and trained. The coordinate neural network takes continuous coordinates containing time, frequency, and channel, as well as a working condition vector consisting of rotor speed, load, and engine exhaust temperature as input, and outputs a continuous time-frequency field representation. During training, a random mask is applied to the discrete time-frequency observation data. The features extracted by the coordinate neural network are aggregated into a fixed-dimensional latent representation through weighted average pooling; a diffusion model conditioned on the working condition vector is trained in the latent representation space. In the prediction phase, the current potential representation is obtained based on the latest observational data; Using the current latent representation as the initial latent variable, backsampling of the diffusion model is performed to generate multiple future latent variable trajectories; during the backsampling process, the likelihood gradient is calculated based on the difference between the latest observation data and the time-frequency field data reconstructed from the current latent variable, and the update of the latent variable is guided by the likelihood gradient; The multiple future latent variable trajectories are decoded into corresponding future time-frequency fields, and the health index curve of each time-frequency field is calculated. Based on the time when each health index curve first exceeds the preset failure threshold, a sample set of remaining service life is obtained. The sample set of remaining service life is statistically analyzed, and the probability distribution information of the remaining service life is output.

2. The method according to claim 1, characterized in that, The occlusion mode of the random mask includes at least one of the following: random point missing mode, block missing mode, and periodic missing mode.

3. The method according to claim 1, characterized in that, The conditionally modulated coordinate neural network is a sinusoidal representation network with characteristic linear modulation; the loss function for training the coordinate neural network includes field value reconstruction loss at observed coordinate points and completion consistency loss at occluded coordinate points.

4. The method according to claim 1, characterized in that, The weights for the weighted average pooling are generated using an attention mechanism based on the sensitivity of the coordinates to faults.

5. The method according to claim 1, characterized in that, The method of guiding the update of latent variables with the likelihood gradient is as follows: the likelihood gradient is multiplied by a step size coefficient and then added to the update amount calculated by the diffusion model backsampler to jointly determine the next update of the latent variables.

6. The method according to claim 1, characterized in that, The health index is calculated using any of the following methods: Calculate the ratio of fault-sensitive frequency band energy to full-band energy from the future time-frequency field; Calculate the spectral kurtosis or spectral entropy from the future time-frequency field; It is obtained by mapping through a pre-trained regression network that takes latent representations as input and health indices as outputs.

7. A system for predicting the remaining service life of rolling bearings in aero-engines based on coordinate neural fields, characterized in that, include: The preprocessing module is used to acquire the vibration signal of the rolling bearing of the aero-engine and perform time-frequency transformation to obtain discrete time-frequency observation data; A continuous time-frequency field modeling module, connected to the preprocessing module, is used to run a conditionally modulated coordinate neural field network. The coordinate neural field network takes as input continuous coordinates containing time, frequency, and channels, and a working condition vector consisting of rotor speed, load, and engine exhaust temperature, and outputs a continuous time-frequency field representation. A latent diffusion and assimilation module, connected to the continuous time-frequency field modeling module, includes: The latent representation aggregation unit is used to aggregate the features extracted by the continuous time-frequency field modeling module into a latent representation through weighted average pooling. The conditional diffusion model unit is used to perform prior modeling and backsampling in the latent representation space with the conditional vector as a condition; The posterior guidance unit is used to calculate the likelihood gradient based on the difference between the latest observation data and the time-frequency field data reconstructed from the current latent variables during the backsampling process, and to guide the latent variable update with the likelihood gradient. The latent diffusion and assimilation module is configured to generate multiple future latent variable trajectories based on the latent representation, through the collaboration of the conditional diffusion model unit and the posterior guidance unit. The decision output module, connected to the latent diffusion and assimilation module, is used to decode the generated future latent variable trajectory into a time-frequency field, calculate the health index and remaining lifespan distribution, and output the prediction results.

8. The system according to claim 7, characterized in that, The system adopts an edge-cloud collaborative deployment architecture: the preprocessing module and the continuous time-frequency field modeling module are deployed on edge computing devices; the latent diffusion and assimilation module and the decision output module are deployed on cloud servers.

9. The system according to claim 7, characterized in that, In the continuous time-frequency field modeling module, the conditionally modulated coordinate neural field network is a sinusoidal representation network with characteristic linear modulation, and the training process of the coordinate neural field network includes the step of applying a random mask to the input data.

10. The system according to claim 7, characterized in that, The prediction results output by the decision output module include at least one of the following: the expected value of remaining useful life, the quantile interval at a specified confidence level, and the premature failure probability curve.

Images