A wavelet enhanced multi-branch time series network fusion system and method for aero-engine remaining useful life prediction

Abstract

This invention provides a wavelet-enhanced multi-branch temporal network fusion system and method for predicting the remaining service life of aero-engines. It includes introducing a Chirplet-based wavelet channel attention module at the input, constructing a cross-branch interactive multi-branch temporal backbone network during the feature extraction stage, achieving bidirectional information correction through two-level cross-branch attention, designing a dual-branch prediction head, employing a lightweight temporal and channel hybridization module for efficient temporal mixing in the main path, and incorporating slow-fast dual-speed hybridization, ALiBi, and lightweight state-space gating into the bypass path based on Informer. Finally, the two outputs are fused through learnable gating. This invention effectively improves the accuracy, stability, and generalization ability of aero-engine remaining service life prediction.

Description

Technical Field

[0001] This invention belongs to the field of equipment health management and predictive maintenance technology, specifically relating to a wavelet-enhanced multi-branch temporal network fusion system and method for predicting the remaining service life of aero-engines. Background Technology

[0002] During long-term operation, aircraft engines undergo complex degradation accumulation and operational condition switching, and their health status directly affects flight safety and maintenance costs. Remaining service life prediction is a key step in monitoring their operational status and making maintenance decisions, and the prediction results play an important role in determining the timing of maintenance, spare parts scheduling, and task allocation.

[0003] In actual operation, sensor data often contains noise, operating condition disturbances, and differences in degradation stages, leading to fluctuations, lags, or failures across operating conditions in the prediction results, thereby weakening the reliability of maintenance decisions based on remaining service life prediction. Therefore, how to stably extract degradation evidence and achieve robust remaining service life prediction in the context of multiple sensors, multiple operating conditions, and strong noise remains a core challenge for the practical application of equipment health management and predictive maintenance technologies.

[0004] In recent years, with the rapid development of deep learning technology, the mainstream approach for building remaining useful life (RUL) prediction models is based on deep neural networks. Notable methods include Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), Gated Recurrent Units (GRUs), and Temporal Convolutional Networks (TCNs), as well as the recent Transformer series of solutions based on attention mechanisms. However, while traditional CNNs excel at local pattern extraction, in multi-sensor RULs, the sensitivity of different channels to degradation varies with operating conditions and stages. Meanwhile, although RNNs, GRUs, and LSTMs can process temporal information sequentially, they suffer from the problem of dilution of key stage information in long sequences, easily leading to delayed responses or unstable predictions at degradation inflection points. To address the distributional differences caused by multiple operating conditions, some works combine operating condition identification, decomposition, or multi-branch structures, attempting to characterize the coupling relationship between operating condition modulation and degradation evolution within the model. However, some "multi-branch" structures often remain at the level of simple parallel connection and post-fusion, lacking interpretable division of labor and mutual verification mechanisms, resulting in unstable fusion gains.

[0005] Currently, there is a lack of a system and method in the technical field that can simultaneously take into account the effective enhancement and selection of degradation information from multiple sensors, the joint modeling of global temporal dependence and local evolution details, and achieve high-precision and high-stability prediction of the remaining service life of aero-engines under multiple operating conditions.

[0006] Definitions: Bidirectional TCN: Bidirectional TCN (Bidirectional Temporal Convolutional Network) is an extended version of the standard Temporal Convolutional Network (TCN). By introducing a bidirectional structure, it models the past and future dependencies of time series simultaneously, thereby improving the ability to capture dynamic features of time series.

[0007] CBA1: First Cross-Branch Attention Module. This module enables the first cross-branch information exchange between the main branch and auxiliary branches. It uses main branch features as queries and auxiliary branch features as keys and values. Through an attention mechanism, it selects the most relevant supplementary information from the auxiliary branches for the current time step, updating the main branch features. Its purpose is to enable the main branch to maintain its ability to model the overall degradation trend while absorbing local dynamics and fine-grained supplementary information from the GRU and CNN branches.

[0008] MSCA: Multi-scale Channel Convolution Module. This module performs multi-scale convolution processing on the backbone features after the first cross-branch attention update. By setting convolution operators with different kernel sizes or different dilation rates, this module performs parallel filtering and enhancement of the backbone features at multiple scales to further extract local degradation patterns and stable trend information at different time scales, thus obtaining multi-scale enhanced backbone features.

[0009] CBA2: Second Cross-Branch Attention Module. Used to implement reverse correction of auxiliary branches by the main branch. This module uses the main branch features after MSCA processing as the correction center and performs a second cross-branch attention interaction on the auxiliary branch features, aligning the auxiliary branch features with the core degenerate feature subspace represented by the main branch. This reduces the sensitivity of the auxiliary branches to operating noise and spurious correlation patterns, resulting in corrected auxiliary branch features.

[0010] HorizonHead: In this invention, HorizonHead uses linear mapping and a lightweight multilayer perceptron to directly map the high-level temporal features extracted from the main or bypass paths into remaining lifetime prediction results with a length equal to the prediction step size. Its function is to complete the regression output from high-dimensional fused features to the final remaining lifetime sequence. Summary of the Invention

[0011] To address the aforementioned issues, this invention proposes a wavelet-enhanced multi-branch temporal network fusion system and method for predicting the remaining service life of aero-engines.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of aero-engines includes the following steps: S1. Construct a wavelet-enhanced multi-branch temporal network fusion framework, which includes a wavelet channel attention module, a multi-branch network with cross-branch interaction, and a dual-branch prediction head; the dual-branch prediction head includes a lightweight temporal and channel hybrid module and a dual-speed hybrid temporal encoder. S2. Receive time-series monitoring data from multiple sensors of the aero-engine to form an original degraded sequence, and perform normalization preprocessing and sample construction on the original degraded sequence to obtain a training dataset. Train the wavelet-enhanced multi-branch time-series network fusion framework using the training dataset to obtain the trained wavelet-enhanced multi-branch time-series network fusion framework; the loss function used to train the wavelet-enhanced multi-branch time-series network fusion framework is the mean square error loss function. S3. Collect and process time-series monitoring data from multiple sensors of the aero-engine to obtain the input sample sequence. , input sample sequence The remaining service life prediction results of aero-engines are obtained by inputting a pre-trained wavelet-enhanced multi-branch temporal network fusion framework.

[0013] In a further improvement, S3, the input sample sequence The method to obtain it is as follows: Step 1: Receive the time-series monitoring data from multiple sensors of the aero-engine and sort them according to the engine number and sampling time sequence to form the original degradation sequence corresponding to each engine; Step 2: Perform feature filtering and normalization preprocessing on the original degraded sequence to obtain a multivariate degraded sequence with uniform dimensions; Step 3: Slice the multivariate degenerate sequence using a sliding time window method to construct the input sample sequence. : in, Indicates the number of input channels. Indicates the length of the sample sequence. Indicates the first Each input channel at time position The value of , It represents the set of real numbers.

[0014] Further improvements include the following data processing method for the trained wavelet-enhanced multi-branch temporal network fusion framework: S31, Input Sample Sequence Wavelet transform is performed through the wavelet channel attention module to achieve multi-scale decomposition and enhancement, resulting in an enhanced sequence. The enhanced sequence is then mapped to the network feature space to obtain the enhanced input sample sequence. S32. Input the enhanced input sample sequence into a multi-branch network with cross-branch interaction, extract the degradation time-series features at different levels, and use cross-branch attention to interact and fuse the degradation time-series features at different levels to obtain shared fused features. S33. Feed the shared fusion features into the dual-branch prediction head: S33.1: Lightweight time-series and channel hybrid module, which decouples and hybridizes shared fusion features in the time dimension and channel dimension to obtain the main path prediction result; S33.2: Dual-speed hybrid temporal encoder, through fast and slow dual-path temporal resampling, multi-head self-attention branch with ALiBi bias and lightweight state space branch, jointly models the long-range dependence, gradual trend and local dynamics of shared fusion features to obtain bypass prediction results; S33.3: Adaptively fuse the main path prediction results and the bypass prediction results to output the remaining service life prediction results of the aero-engine: concatenate the main path prediction results and the bypass prediction results in the prediction dimension, generate gating coefficients through linear transformation and Sigmoid activation function, and perform weighted summation of the main path prediction results and the bypass prediction results based on the gating coefficients to output the remaining service life prediction results of the aero-engine.

[0015] Further improvements are made, and the specific steps of step S31 are as follows: S31.1, Wavelet channel attention module on input sample sequence Each sensor channel is subjected to a set of complex-valued Chirplet convolution kernels for one-dimensional convolution, yielding multi-scale, tunable time-frequency response and energy representation. Subsequently, using the time-frequency response and energy representation as "evidence," channel dependencies are explicitly modeled and sensor channels are recalibrated, providing a more discriminative and robust input representation for the subsequent time-series backbone. Among these, the wavelet channel attention module's... Chirplet cores Defined as: in, , , , These are the learnable scale, center frequency, frequency modulation, and initial phase; the relative time variable in the Chirplet kernel function. The imaginary unit satisfies ; S31.2, in time position Chirplet amplitude characteristics were obtained at [location]. After responding to the real and imaginary parts, the following is calculated: : in, and respectively in time position The real and imaginary parts of the response; S31.3, Chirplet amplitude characteristics Energy pooling is performed in the time dimension to obtain the energy of each input channel in the first dimension. Average energy over the Chirplet band: For the first The input channel in the first Average energy over a Chirplet band; The average energy in the Chirplet band of each input channel is stacked and combined according to the scale to form the energy vector of each input channel; For the first Energy vectors of each input channel; The attention weights for each input channel are generated using SE-MLP: For the first Each input channel weight; The input sample sequence X is recalibrated using the attention weights of each input channel to obtain the enhanced input sample sequence. , The element-level representation in the code is: in, Indicates the first Each input channel at time position The recalibrated value; all Reassemble according to the original structure to obtain the enhanced input sample sequence. .

[0016] Further improvements are made, and the specific steps of step S32 are as follows: S32.1, the backbone of the multi-branch network with cross-branch interaction is a bidirectional TCN backbone. The bidirectional TCN backbone performs forward and backward modeling on the enhanced input sample sequence F to obtain forward temporal features. and reverse time series features The bidirectional TCN backbone introduces a global gating coefficient. Forward temporal features and reverse time series features Perform weighted fusion to obtain the final main branch output. for: ; S32.2, a bidirectional TCN backbone with a GRU branch and a one-dimensional convolutional network CNN branch connected in parallel; The data processing methods for the GRU branch and the one-dimensional convolutional network (CNN) branch are as follows: The output of the GRU branch focuses on modeling sequence dependencies and short-term memory. For the GRU branch, The output of a 1D convolutional network (CNN) branch captures local smoothness and detailed textures through shallow multi-scale convolutions; and Mapping to the latent space dimension of the bidirectional TCN backbone, dimension alignment is performed in the intermediate layer to obtain the alignment projection matrix between the GRU branch and the one-dimensional convolutional network CNN branch. and : The features are GRU branches after alignment and projection in stage 0. The features of the one-dimensional convolutional network (CNN) branches after alignment and projection in stage 0; S32.3, design a three-stage structure "CBA1→MSCA→CBA2". In the intermediate layer, this three-stage structure progressively aligns and mutually corrects the three branches: the bidirectional TCN backbone, the GRU branch, and the one-dimensional convolutional network (CNN) branch. The first cross-branch attention step uses the bidirectional TCN backbone output as the query and the concatenated features of the GRU branch and the one-dimensional CNN branch as the key. The aligned and projected GRU branch features... Recorded as One-dimensional convolutional network (CNN) branch features Record , obtain memory characteristics As an auxiliary branch input in the first interaction phase: Then, multi-head scaling dot product attention is used, from... Select the most useful auxiliary information for each time step; The first time you cross a branch, note that CBA1's design only updates the output of the main branch. The first interaction phase is performed, and the output features from the GRU branch in the first interaction phase are... Output features from the CNN branch in the first interaction phase Remain unchanged: The core features are those following the first interaction; After the first cross-branch attention CBA1, multi-scale channel convolution (MSCA) is introduced for the second interaction stage. Multi-scale filtering is performed to obtain : in This represents a depthwise convolution operator with different kernel sizes and dilation rates along the channel dimension. is the scale number; s is the scale index of the multi-scale convolution. This is a multi-scale enhanced backbone feature output by multi-scale channel convolutional MSCA; The second cross-branch attention CBA2 backcorrection stage involves the GRU branch and the one-dimensional convolutional network CNN branch undergoing a third interaction. Through the second cross-branch attention CBA2, the GRU and CNN are pulled back to the original state while retaining their respective advantages. The core degenerate feature subspace reduces overfitting to operating condition noise and spurious correlation patterns. in, The GRU branch features are after the second cross-branch attention correction. These are the CNN branch features after the second cross-branch attention correction. This represents the submodule in the second cross-branch attention module used to correct GRU branch features. This represents the submodule in the second cross-branch attention module used to correct the branch features of the one-dimensional convolutional network (CNN). S32.4, will , and Concatenate according to channel dimension, and then... Convolution is used for unified mapping, and residual connections are introduced to obtain shared fused features. : ; in, This is the output projection matrix used for unified mapping of spliced ​​features. This is the residual projection matrix for residual mapping of the enhanced input sample sequence F.

[0017] Further improvements, the specific steps of step S33.1 are as follows: S33.1.1: Lightweight timing and channel hybrid module for shared fusion features in the time dimension Linear mixing is performed, with learnable weights applied to different time steps within each feature channel to reconstruct the most useful temporal combinations for prediction. The lightweight temporal and channel mixing module uses a shared temporal transformation matrix and bias across all channels to obtain the temporal mixing result. : in, Indicates shared and integrated features The A sequence of sensor channels, Indicates shared and integrated features Number of feature channels, Represents the learnable weight matrix for temporal mixing. Represents a learnable bias vector for temporal mixing; The time blending results corresponding to each feature channel Combining into temporal mixed feature tensors : ; Indicates the characteristics of sharing and integration Each feature channel undergoes a time-mixing operation with linear transformation in the time dimension; S33.1.2: A small feedforward network is used for feature mixing in the channel dimension, for any time step Temporal Mixed Feature Tensor Channel MLP first expands the dimensions to Throw it back This enables nonlinear channel interaction; then, after parallel computation across all time steps, the result is obtained. And then Apply layer normalization to stabilize the output obtained during training. ; after obtaining the output Subsequently, the lightweight timing and channel hybrid module... Global pooling is performed and concatenated along the channel dimension. Then, a global representation is obtained by linear mapping and a lightweight MLP-based HorizonHead. The global representation is then mapped to a baseline prediction sequence as the main path prediction result. in, To represent the temporal mixture feature tensor Perform a channel blending operation on the channel-dimensional nonlinear mapping at each time step. Let represent the channel blending feature tensor obtained after channel blending.

[0018] Further improvements are made, and the specific steps in step S33.2 are as follows: S33.2.1: Incorporate shared and integrated features Model dimensions mapped to the encoder To obtain sequence characterization : S33.2.2: Slow-Fast Dual-Speed ​​Temporal Resampling; The dual-speed hybrid temporal encoder includes a fast branch and a slow branch, where the fast branch is a fast branch encoder and the slow branch is a slow branch encoder; the fast branch maintains the original temporal resolution and shares fused features. Perform shallow modeling to capture rapidly changing local fine-grained dynamics. ; For fast branch encoders; The slow branch then follows the time axis to share fused features. Downsampling is performed to extract stable global semantic information and long-term dependencies at a lower temporal resolution. The downsampled input... ; The downsampling factor; To indicate after downsampling The feature vector at the k-th time step, The sequence representation E is in the first... The feature vectors of each time step; k represents the index of the time step in the downsampled sequence. Indicates the time length of sequence representation of E; Will Inputting the slow branch encoder yields: This is the output of the slow branch encoder. It is a slow branch encoder; Because the fast and slow branches have different time resolutions, the slow branch output needs to be upsampled back by a certain length. Record the result after upsampling as : For upsampling; Then, at each time step, gating fusion is performed on the fast and slow branches; set up , They represent and In the The feature vector at each time step, then the gate vector The definition is as follows: in and These are learnable parameters; The dual-speed hybrid timing encoder in the first The fusion features at each time step are written as: ; S33.2.3: A lightweight state space branch is introduced within each coding layer of the dual-speed hybrid temporal encoder and set in parallel with a multi-head self-attention branch with ALiBi bias to simultaneously model local content dependency and long-range trend memory; after combining the state sequences of all channels, the output of the lightweight state space branch is obtained; the output of the multi-head self-attention branch with ALiBi bias and the output of the lightweight state space branch are adaptively fused at the feature level through learnable gating to obtain the fusion result. The fusion result is residually connected with the corresponding coding layer input and transformed through a feedforward network to obtain the corresponding coding layer output, that is, the bypass prediction result.

[0019] A wavelet-enhanced multi-branch temporal network fusion system for predicting the remaining service life of an aero-engine includes an interconnected microprocessor and a memory, wherein the microprocessor is programmed or configured to execute the aforementioned wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of an aero-engine.

[0020] A computer-readable storage medium storing a computer program or instructions programmed or configured to execute, via a processor, the aforementioned wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of an aero-engine.

[0021] A computer program product includes a computer program or instructions that are programmed or configured to execute, via a processor, the aforementioned wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of an aero-engine.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: By introducing a wavelet enhancement mechanism, the ability to extract key features from multi-sensor degradation signals was improved. Simultaneously, a multi-branch temporal network and a cross-branch attention interaction mechanism were constructed to enhance the representation capabilities of local degradation details, global temporal dependencies, and feature fusion. A lightweight temporal and channel hybrid module and a dual-speed hybrid temporal encoder were designed to achieve collaborative modeling of long-term and short-term degradation information. Furthermore, by paralleling and gating fusion of lightweight state-space branches and self-attention branches, the model's ability to jointly characterize long-term trends and local dynamics under complex operating conditions was further enhanced, thereby effectively improving the accuracy, stability, and generalization ability of predicting the remaining service life of aero-engines. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0024] Figure 2 This is an algorithmic framework diagram of a wavelet-enhanced multi-branch temporal network fusion system and method for predicting the remaining service life of aero-engines, provided in an embodiment of the present invention.

[0025] Figure 3 This is a framework diagram of the wavelet channel attention module in an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of a multi-branch network process with cross-branch interaction in an embodiment of the present invention.

[0027] Figure 5 This is a diagram of the internal structure of the lightweight timing and channel hybrid module in an embodiment of the present invention.

[0028] Figure 6 Internal structure diagram of encoder layer with multi-head ALiBi self-attention and lightweight state space branch gating fusion.

[0029] Figure 7 This invention presents a performance comparison of models on four subsets of the CMAPSS dataset in this embodiment.

[0030] Figure 8 This is a comparison chart of predicted and actual values ​​on the FD001 subset in an embodiment of the present invention.

[0031] Figure 9 This is a comparison chart of predicted and actual values ​​on the FD002 subset in an embodiment of the present invention.

[0032] Figure 10 This is a comparison chart of predicted and actual values ​​on the FD003 subset in an embodiment of the present invention.

[0033] Figure 11 This is a comparison chart of predicted and actual values ​​on the FD004 subset in an embodiment of the present invention.

[0034] Figure 12 This is a prediction error distribution diagram on the FD001 subset in an embodiment of the present invention.

[0035] Figure 13 This is a prediction error distribution diagram on the FD002 subset in an embodiment of the present invention.

[0036] Figure 14 This is a prediction error distribution diagram on the FD003 subset in an embodiment of the present invention.

[0037] Figure 15 This is a prediction error distribution diagram on the FD004 subset in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0039] The core of this invention is to provide a wavelet-enhanced multi-branch temporal network fusion system and method for predicting the remaining service life of aero-engines, in order to solve the problems existing in the existing methods and technologies.

[0040] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] Figure 1 This is a basic flowchart illustrating a wavelet-enhanced multi-branch temporal network fusion system and method for predicting the remaining service life of an aero-engine, as provided in an embodiment of the present invention. Figure 1 As shown, it includes the following steps: S1: Receives time-series monitoring data from multiple sensors of the aero-engine and preprocesses and constructs samples from the raw degradation sequences. It performs data preprocessing such as normalization on the acquired multivariate sensor signals to form an input sample sequence for predicting remaining service life, providing a unified data representation for subsequent feature enhancement and time-series modeling. S2: The original multi-sensor degradation signal is decomposed and enhanced using wavelet transform on the input degradation sequence at multiple scales. Key frequency band information related to performance degradation is highlighted while noise interference is suppressed. The enhanced sequence is then mapped to the network feature space to obtain the basic input features. S3: The enhanced features are input into the multi-branch backbone network to extract degraded temporal features at different levels, and feature interaction and fusion are achieved through cross-branch attention. The bidirectional TCN selectively absorbs evidence from the auxiliary branch CNN and GRU, and the bidirectional TCN features reverse-correct the auxiliary branch CNN and GRU. S4: The shared fused features are fed into the dual-branch prediction head to construct a collaborative prediction between the main path and the bypass path. The main path uses a lightweight temporal and channel hybrid module to decouple and hybridize the shared features in the temporal and channel dimensions; the bypass path uses a dual-speed hybrid temporal encoder to jointly model long-range dependencies, gradual trends, and local dynamics through fast and slow dual-path temporal resampling, multi-head self-attention with ALiBi bias, and lightweight state space branching. S5: Adaptively fuse the main path prediction results and the bypass path prediction results to output the remaining service life prediction result of the aero-engine. The outputs of the two prediction branches are weighted and fused, enabling the model to adaptively select more effective prediction evidence according to different operating conditions and degradation stages, thereby outputting the final remaining service life prediction value.

[0042] Figure 2 This invention provides an algorithmic framework diagram for a wavelet-enhanced multi-branch temporal network fusion system and method for predicting the remaining service life of aero-engines. To effectively address the problems of insufficient utilization of multi-sensor degradation information, inadequate long-term dependency modeling capabilities, weak interaction and fusion effects of different branch features, and low prediction accuracy and stability under complex multi-condition working conditions in existing aero-engine remaining service life prediction methods, this study proposes a wavelet-enhanced multi-branch temporal network fusion framework for predicting the remaining service life of aero-engines. The core design improves the extraction capability of key features from multi-sensor degradation signals by introducing a wavelet enhancement mechanism. Simultaneously, a multi-branch temporal network and cross-branch attention interaction mechanism are constructed to improve the representation capabilities of local degradation details, global temporal dependencies, and feature fusion. A lightweight temporal and channel hybrid module and a dual-speed hybrid temporal encoder are designed to achieve collaborative modeling of long-term and short-term degradation information. Furthermore, the parallel connection and gated fusion of lightweight state-space branches and self-attention branches further enhance the model's ability to jointly characterize long-term trends and local dynamics under complex conditions, thereby effectively improving the accuracy, stability, and generalization ability of aero-engine remaining service life prediction. The overall framework consists of four core components: First, there is the wavelet channel attention module. This module explicitly introduces multi-scale frequency domain priors, which helps to suppress channels and frequency bands with unstable statistical behavior and high noise levels under different operating conditions. It also provides a cleaner and more structured input representation for subsequent cross-branch interactive backbones and dual-branch prediction heads. Second, a multi-branch network with cross-branch interaction. This part constructs a two-level cross-branch attention to achieve information exchange and correction, and combines multi-scale channel convolution to enhance the temporal feature representation, which is conducive to enhancing the ability to capture the common degradation characteristics of multiple working conditions, while preserving and utilizing diverse local dynamic information; Third is the lightweight time-series and channel hybrid module. This module is designed to extract time-dependent information and multi-channel interaction information of degraded sequences in a lightweight manner, reducing the number of parameters and computational burden while ensuring the expressive power of the prediction head. Fourthly, a dual-speed hybrid temporal encoder is constructed. This part establishes a dual-path temporal coding structure. The slow path encodes the input sequence at a lower sampling rate and lower temporal resolution to extract stable global semantic information and long-term degenerate dependencies; the fast path encodes the input sequence at a higher sampling rate and higher temporal resolution to capture rapidly changing local fine-grained dynamic features. Through fast and slow dual-path temporal resampling, multi-head self-attention with ALiBi bias, and lightweight state space branching, long-range dependencies, gradually changing trends, and local dynamics are jointly modeled. During the encoding process, the two paths are laterally connected to achieve feature interaction and information fusion in the intermediate layer, thereby enhancing the complementary representation ability between features at different time scales.

[0043] like Figure 3 As shown, this invention designs a wavelet channel attention module. This module applies a set of complex-valued Chirplet convolution kernels to each sensor channel for one-dimensional convolution, obtaining multi-scale, frequency-tunable time-frequency responses and energy representations. Then, using these time-frequency energies as "evidence," and drawing on the idea of ​​"explicitly modeling inter-channel dependencies and recalibrating channels" in channel attention, adaptive weighting is applied to each channel. This suppresses noise-dominated channel contributions, highlights degradation-sensitive channels, and provides a more discriminative and robust input representation for the subsequent time-series backbone.

[0044] like Figure 4 As shown, this invention designs a multi-branch network with cross-branch interaction. The backbone of this module uses a bidirectional TCN to model the input features forward and backward. To achieve an adaptive balance between forward and backward information, a global gating coefficient is introduced. Furthermore, the global statistics obtained by averaging over the time dimension, through gating fusion, allow the backbone to adaptively adjust the dependence ratio on "past information" and "future context" according to specific sequences, thus maintaining the causal structure while enhancing the ability to perceive long-range patterns.

[0045] To complement the representational bias of the main branch, this study uses a bidirectional TCN main branch with a parallel GRU branch and a CNN branch to provide local dynamic information under different inductive biases. GRU achieves an adaptive balance between preserving historical memory and absorbing new information through update and reset gates, thus alleviating the gradient vanishing problem of traditional RNNs and demonstrating comparable modeling capabilities and higher parameter efficiency to LSTM in various sequence modeling tasks. Therefore, the GRU branch can more sensitively capture short-term dependencies and nonlinear dynamic changes in continuous degradation processes, serving as supplementary evidence to the convolutional main branch. One-dimensional convolutional networks are often used in time series tasks to extract local morphological features and can form hierarchical representations of local patterns through multi-scale convolutional kernel stacking structures. Therefore, the CNN branch can provide the main branch with more sensitive feature supplements to local texture and fine-grained changes. For the input sample sequence F, the outputs of the two branches are as follows: in The focus is on modeling sequence dependencies and short-term memory. Local smoothness and detailed texture are captured through shallow multi-scale convolution.

[0046] Since the feature dimensions of the three branches may be different, in order to facilitate alignment and mutual correction between the subsequent cross-branch attention modules in the same representation space, linear projection is used to map the auxiliary branches to the latent space dimension of the main branch. ,use and Representing the alignment projection matrices of the GRU branch and the CNN branch respectively, we obtain: In this way, the GRU and CNN branches provide complementary evidence from their respective perspectives, enabling subsequent explicit cross-branch interactions to be gradually aligned and mutually corrected in the intermediate layers, rather than simply splicing and fusion at the output end.

[0047] To enable directional, explicit information exchange between the main branch and auxiliary branches, this paper designs a three-stage structure: "CBA1→MSCA→CBA2". Unlike the traditional approach of simply concatenating multi-branch features at the output end, this structure performs progressive alignment and mutual correction of the three branches in the intermediate layer.

[0048] The first cross-branch observation of CBA1 uses the bidirectional TCN backbone output as the query and the concatenated features of the auxiliary branches as the key. First, construct the memory features. Then, multi-head scaling dot product attention is used to... The most useful auxiliary information is selected for each time step. CBA1's design updates only the backbone features, specifically the output features from the GNN branch in the first interaction phase. Output features from the CNN branch in the first interaction phase Maintaining this consistency is equivalent to selectively absorbing supplementary evidence from GRU and CNN on top of BiTCN, allowing the backbone to incorporate more fine-grained temporal information without sacrificing stability. After CBA1, the backbone features have already incorporated information from the auxiliary branches, but local fluctuations may still exist at different time scales. Therefore, this paper introduces multi-scale channel convolution within the backbone to... Multi-scale filtering is performed. MSCA is equivalent to performing a multi-scale resampling and smoothing of the backbone features, reducing occasional spikes. This results in stable backbone features processed by MSCA. Then, the second cross-branch attention (CBA2) is responsible for reverse correcting the two auxiliary branches. Through CBA2, GRU and CNN are pulled back to the mean while retaining their respective advantages. The core degenerate feature subspace reduces overfitting to operating noise and spurious correlation patterns.

[0049] After completing the two rounds of interaction described above, this paper concatenates the final features of the three branches along the channel dimension, and then... Convolution performs a unified mapping and introduces residual connections to obtain the backbone output, represented as: in , , The output of the multi-branch backbone network that interacts across branches will serve as a shared representation for subsequent dual-branch prediction heads.

[0050] From an overall perspective, this backbone network, through structured multi-branch modeling and a direction-controllable cross-branch attention mechanism, is conducive to enhancing the ability to capture the commonalities of degradation under multiple operating conditions, while retaining and utilizing diverse local dynamic information.

[0051] like Figure 5 As shown, this invention designs a lightweight time-series and channel mixing module. First, linear mixing is performed in the time dimension. Within each channel, different time steps are learnedably weighted to reconstruct the most useful time combination for prediction. A time transformation matrix shared across all channels is used. and bias The result obtained after time mixing is: in, Indicates shared features The A channel sequence, This indicates the corresponding time-mixing result.

[0052] Restore all channels to tensor form, denoted as Subsequently, a small feedforward network is used for feature mixing in the channel dimension. For any time step... The feature vectors, channel MLP first expands the dimension to... Throw it back This enables nonlinear channel interaction. After parallel computation across all time steps, the result is obtained. And then Apply layer normalization to stabilize the output obtained during training. After obtaining the output Then, the main path undergoes global pooling and concatenation across the channel dimensions. Subsequently, through linear mapping and a lightweight MLP-based HorizonHead, this global representation is mapped in one step to a length of [length missing]. The RUL baseline prediction sequence.

[0053] like Figure 6 As shown, this invention designs an encoder layer with multi-head ALiBi self-attention and lightweight state space branch gating fusion. In the dual-branch prediction head, the bypass is responsible for modeling long-range dependencies and gradually changing trends, complementing the main path's lightweight temporal and channel mixing module. Its overall structure first performs feature embedding and slow-fast temporal resampling on the main output, then feeds it into the encoder-like branch based on Informer modification. Within each encoding layer, multi-head self-attention with ALiBi bias is connected in parallel with the lightweight state space branch, and then the output is fused through gating. The final bypass prediction and the main path prediction are fused at each time step using learnable gating.

[0054] First, the shared fusion feature H is mapped to the model dimension of the encoder. To obtain sequence characterization Fast branching at original resolution... Shallow modeling is performed, while the slow branch downsamples the time axis to perform deeper modeling at low-frequency scales: Slow branches should be checked first. Perform time downsampling. Let... If is the downsampling factor, then the downsampled input is denoted as . Its definition is: Then, Inputting the Slow branch encoder yields: Because the two branches have different time resolutions, the output of the slower branch needs to be upsampled back to the specified length first. The result after upsampling is recorded as follows: Then, at each time step, gating fusion is performed on the fast and slow branches. Let... , They represent and In the The feature vector at each time step, then the gate vector The definition is as follows, where and Learnable parameters: Based on this, the bypass is in the first The fusion features at each time step are written as: We employ this slow-fast structure to enable the bypass to model both gradual trends over long time spans and high-frequency local fluctuations while maintaining low computational cost. Meanwhile, although our branch is based on the Informer encoder framework, we have completely removed the original Informer decoder, retaining only the encoder and directly outputting the predicted features.

[0055] A single encoder layer employs a parallel structure of ALiBi self-attention branches and lightweight state-space branches to simultaneously model local content dependencies and long-range trend memories. Given the input from the previous layer... The encoder layer in this layer first calculates the multi-head self-attention output with ALiBi linear bias. : in For layer normalization, By incorporating a bias linearly proportional to the time interval into the scaled dot product attention, explicit positional encoding is avoided, and the extrapolation capability for different sequence lengths is improved. ALiBi no longer explicitly adds absolute positional encoding to the input; instead, it directly adds a linear bias term proportional to the temporal distance to the attention, thus introducing an inductive bias of recency bias while maintaining the standard attention form. The figure also shows its interaction with "Local-band + causalmask," which, while limiting the local bandwidth attention range, maintains the characterization of temporal structure through relative distance bias.

[0056] Meanwhile, this layer also recursively models the slowly varying trend through a lightweight state-space branch. The lightweight state-space branch adopts a diagonal, first-order discrete linear state-space form, and its recursion is equivalent to EMA / AR(1) filtering on each channel. for At time step ,aisle Input components on, For the corresponding state variables, For the first The learnable attenuation coefficients for each channel are then defined by the recursive process as follows: By combining the state sequences of all channels, a lightweight state-space branch output can be obtained. Subsequently, the state-space branch output is obtained recursively based on the lightweight state space. Then, output the results for both branches. and Perform feature-level adaptive fusion to obtain intermediate fused features. : in, For learnable parameters, This is the Sigmoid activation function.

[0057] Obtaining intermediate fusion characteristics Then, after random deactivation, it is input to the coding layer. Perform residual connections to obtain residual enhancement features. : Then, the residual enhancement features The input is fed forward into the network, and after layer normalization, the output of the coding layer is obtained. : Where Dropout() represents random deactivation, FFN() represents feedforward network, and LN() represents layer normalization. This represents the output feature of the coding layer, i.e., the side-channel prediction result.

[0058] Subsequently, adaptive gating fusion is performed on the main path prediction results and the bypass prediction results. Specifically, the main path prediction results and the bypass prediction results are concatenated in the prediction dimension, and gating coefficients are generated through linear transformation and Sigmoid activation function. Based on the gating coefficients, the main path prediction results and the bypass prediction results are weighted and summed to output the remaining service life prediction results of the aero-engine.

[0059] See Figure 7It can be seen that the wavelet-enhanced multi-branch temporal network fusion system and method for predicting the remaining service life of aero-engines in this embodiment outperforms existing methods for predicting the remaining service life of aero-engines in all aspects. To further illustrate the effectiveness of the wavelet-enhanced multi-branch temporal network fusion system and method for predicting the remaining service life of aero-engines in this embodiment, Figures 8 to 11 This is a comparison chart of predicted and actual values ​​on four subsets of the CMAPSS dataset in this embodiment of the invention. From the engine-by-engine endpoint prediction curve of FD001, it can be observed that the predicted sequence generally follows the changing trend of the actual RUL (Remaining Life). The two maintain a relatively consistent amplitude and phase relationship across most engine samples, indicating that the model can stably learn the mapping relationship between degradation degree and remaining life in a single-operation scenario. In the multi-operation scenarios of FD002 and FD004, the actual values ​​of the endpoint RUL show greater differences across engines and stronger dispersion between samples. Consequently, the prediction curves exhibit more frequent fluctuations and intersections, which is consistent with the representational heterogeneity caused by the coupling of operating condition migration and degradation in multi-operation scenarios. Nevertheless, the predicted points still maintain a similar range of variation as the actual points overall, and achieve relatively close endpoint estimates across a large number of engines. The engine-by-engine endpoint prediction results of FD003 show a good correspondence between the prediction and the actual values ​​across most engines.

[0060] Furthermore, the effectiveness of the wavelet-enhanced multi-branch temporal network fusion system and method for predicting the remaining service life of aero-engines is investigated. Figures 12 to 15This is a distribution of prediction errors on four subsets of the CMAPSS dataset in this embodiment of the invention. Error is defined as the difference between the predicted and actual values. The error histogram corresponding to FD001 shows that the errors are mainly concentrated near zero, with a relatively compact distribution and only a small number of long-tailed samples on both sides, reflecting that the overall model bias is small, but there are still some individual engines with significant overestimation / underestimation. Regarding the error distribution, the histogram of FD002 shows a wider distribution span and a more pronounced tail compared to FD001, indicating that under complex operating conditions, the model's uncertainty and the proportion of extreme error samples will increase. However, the main body of the errors still clusters around zero, indicating that the model does not exhibit systematic distortion on most samples. The engine-by-engine endpoint prediction results of FD003 show a good correspondence between the prediction and the actual values ​​on most engines, and the error distribution also shows a concentrated structure centered around zero, indicating that the model's degenerate representation of this subset has good transferability and stability. The error distribution of FD004 also exhibits a heavier tail feature, suggesting that in scenarios with richer operating conditions and more complex degradation patterns, a few "hard samples" from a few engines may still produce significant endpoint bias. Overall, the proposed model demonstrates a more concentrated and stable error pattern on single-operating-condition data. On multi-operating-condition data, although the error distribution inevitably widens and more obvious tail samples appear, the predicted main body remains concentrated near zero, demonstrating the model's robustness and usability under complex operating condition perturbations.

[0061] In summary, the wavelet-enhanced multi-branch temporal network fusion system and method for predicting the remaining service life of aero-engines in this embodiment processes the input using a wavelet-enhanced multi-branch temporal network architecture to obtain accurate remaining service life prediction results. At the input end, a Chirplet-based wavelet channel attention module is designed, driving adaptive channel selection through time-frequency multi-scale energy representation, thereby suppressing noise and operating condition disturbances and highlighting degradation-related information. A multi-branch temporal backbone network with cross-branch interaction is designed, using a bidirectional TCN as the core and combining GRU and lightweight CNN branches. Two-level cross-branch attention enables bidirectional information exchange between the main branch and auxiliary branch corrections, enhancing feature representation. A dual-branch prediction head is constructed. The main branch designs a lightweight temporal and channel mixing module for efficient temporal mixing, while the bypass is based on an improved Informer, incorporating a slow-fast dual-speed approach, ALiBi, and lightweight state-space branch gating to strengthen long-range dependency modeling. Learnable gating adaptively fuses the two outputs to achieve a more robust performance trade-off. This invention aims to improve the accuracy and robustness of predictions through in-depth collaborative analysis of global information and local details, and can be applied to fields such as equipment health management and predictive maintenance technology.

[0062] Furthermore, the present invention also provides a wavelet-enhanced multi-branch time-series network fusion system for predicting the remaining service life of an aero-engine, comprising a processor and a memory interconnected thereon, wherein the memory stores computer-executable instructions, and when the computer-executable instructions are executed by the at least one processor, the microprocessor is programmed or configured with any of the aforementioned wavelet-enhanced multi-branch time-series network fusion methods for predicting the remaining service life of an aero-engine.

[0063] Furthermore, the present invention also provides a computer-readable storage medium having program code loaded thereon, the program code containing a set of instructions that, when executed on a computing device, cause the computing device to perform the steps defined in any of the aforementioned wavelet-enhanced multi-branch temporal network fusion methods for predicting the remaining service life of aero-engines.

[0064] Furthermore, the present invention also provides a computer program product comprising program instructions stored on a computer-readable medium, which, when loaded and executed by a processor, are intended to implement any of the aforementioned wavelet-enhanced multi-branch temporal network fusion methods for predicting the remaining service life of aero-engines.

[0065] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (systems), or computer program products. Therefore, the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of computer program products embodied on one or more computer-readable storage media (including but not limited to disk storage, optical storage, etc.) containing computer-executable program code. The present invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It should be understood that each flowchart illustration and / or block, and combinations of 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 create a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce an implementation for the process. Figure 1 One or more processes and / or frames Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or frames Figure 1The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or frames Figure 1 The steps of the function specified in one or more boxes.

[0066] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of aero-engines, characterized in that, Includes the following steps: S1. Construct a wavelet-enhanced multi-branch temporal network fusion framework, which includes a wavelet channel attention module, a multi-branch network with cross-branch interaction, and a dual-branch prediction head; the dual-branch prediction head includes a lightweight temporal and channel hybrid module and a dual-speed hybrid temporal encoder. S2. Receive time-series monitoring data from multiple sensors of the aero-engine to form an original degraded sequence, and perform normalization preprocessing and sample construction on the original degraded sequence to obtain a training dataset. Train the wavelet-enhanced multi-branch time-series network fusion framework using the training dataset to obtain the trained wavelet-enhanced multi-branch time-series network fusion framework. S3. Collect and process time-series monitoring data from multiple sensors of the aero-engine to obtain the input sample sequence. , input sample sequence The remaining service life prediction results of aero-engines are obtained by inputting a pre-trained wavelet-enhanced multi-branch temporal network fusion framework.

2. The wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of aero-engines as described in claim 1, characterized in that, In S3, the input sample sequence The method to obtain it is as follows: Step 1: Receive the time-series monitoring data from multiple sensors of the aero-engine and sort them according to the engine number and sampling time sequence to form the original degradation sequence corresponding to each engine; Step 2: Perform feature filtering and normalization preprocessing on the original degraded sequence to obtain a multivariate degraded sequence with uniform dimensions; Step 3: Slice the multivariate degenerate sequence using a sliding time window method to construct the input sample sequence. : in, Indicates the number of input channels. Indicates the length of the sample sequence. Indicates the first Each input channel at time position The value of , It represents the set of real numbers.

3. The wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of aero-engines as described in claim 1, characterized in that, The data processing method for the trained wavelet-enhanced multi-branch temporal network fusion framework is as follows: S31, Input Sample Sequence Wavelet transform is performed through the wavelet channel attention module to achieve multi-scale decomposition and enhancement, resulting in an enhanced sequence. The enhanced sequence is then mapped to the network feature space to obtain the enhanced input sample sequence. S32. Input the enhanced input sample sequence into a multi-branch network with cross-branch interaction, extract the degradation time-series features at different levels, and use cross-branch attention to interact and fuse the degradation time-series features at different levels to obtain shared fused features. S33. Feed the shared fusion features into the dual-branch prediction head: S33.1: Lightweight time-series and channel hybrid module, which decouples and hybridizes shared fusion features in the time dimension and channel dimension to obtain the main path prediction result; S33.2: Dual-speed hybrid temporal encoder, through fast and slow dual-path temporal resampling, multi-head self-attention branch with ALiBi bias and lightweight state space branch, jointly models the long-range dependence, gradual trend and local dynamics of shared fusion features to obtain bypass prediction results; S33.3: Adaptively fuse the main path prediction results and the bypass prediction results to output the remaining service life prediction results of the aero-engine: concatenate the main path prediction results and the bypass prediction results in the prediction dimension, generate gating coefficients through linear transformation and Sigmoid activation function, and perform weighted summation of the main path prediction results and the bypass prediction results based on the gating coefficients to output the remaining service life prediction results of the aero-engine.

4. The wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of aero-engines as described in claim 3, characterized in that, The specific steps of step S31 are as follows: S31.1, Wavelet channel attention module on input sample sequence Each sensor channel is subjected to a set of complex-valued Chirplet convolution kernels for one-dimensional convolution, yielding multi-scale, tunable time-frequency response and energy representation. Subsequently, using the time-frequency response and energy representation as "evidence," inter-channel dependencies are explicitly modeled and sensor channel recalibration is performed, providing a more discriminative and robust input representation for the subsequent time-series backbone. Among these, the wavelet channel attention module's... Chirplet cores Defined as: in, , , , These are the learnable scale, center frequency, frequency modulation, and initial phase; the relative time variable in the Chirplet kernel function. The imaginary unit satisfies ; S31.2, in time position Chirplet amplitude characteristics were obtained at [location]. After responding to the real and imaginary parts, the following is calculated: : in, and respectively in time position The real and imaginary parts of the response; S31.3, Chirplet amplitude characteristics Energy pooling is performed in the time dimension to obtain the energy of each input channel in the first dimension. Average energy over the Chirplet band: For the first The input channel in the first Average energy over a Chirplet band; The average energy in the Chirplet band of each input channel is stacked and combined according to the scale to form the energy vector of each input channel; For the first Energy vectors of each input channel; The attention weights for each input channel are generated using SE-MLP: For the first Each input channel weight; The input sample sequence X is recalibrated using the attention weights of each input channel to obtain the enhanced input sample sequence. , The element-level representation in the code is: in, Indicates the first Each input channel at time position The recalibrated value; all Reassemble according to the original structure to obtain the enhanced input sample sequence. .

5. The wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of aero-engines as described in claim 3, characterized in that, The specific steps of step S32 are as follows: S32.1, the backbone of the multi-branch network with cross-branch interaction is a bidirectional TCN backbone. The bidirectional TCN backbone performs forward and backward modeling on the enhanced input sample sequence F to obtain forward temporal features. and reverse time series features The bidirectional TCN backbone introduces a global gating coefficient. Forward temporal features and reverse time series features Perform weighted fusion to obtain the final main branch output. for: ； S32.2, a bidirectional TCN backbone with a GRU branch and a one-dimensional convolutional network CNN branch connected in parallel; The data processing methods for the GRU branch and the one-dimensional convolutional network (CNN) branch are as follows: The output of the GRU branch focuses on modeling sequence dependencies and short-term memory. For the GRU branch, The output of a 1D convolutional network (CNN) branch captures local smoothness and detailed textures through shallow multi-scale convolutions; and Mapping to the latent space dimension of the bidirectional TCN backbone, dimension alignment is performed in the intermediate layer to obtain the alignment projection matrix between the GRU branch and the one-dimensional convolutional network CNN branch. and : The features are GRU branches after alignment and projection in stage 0. The features of the one-dimensional convolutional network (CNN) branches after alignment and projection in stage 0; S32.3, design a three-stage structure "CBA1→MSCA→CBA2". In the middle layer, the three-stage structure performs step-by-step alignment and mutual correction on the three branches of the bidirectional TCN backbone, GRU branch and one-dimensional convolutional network CNN branch in three stages: the first cross-branch attention CBA1 uses the output of the bidirectional TCN backbone as the query and the concatenation features of the GRU branch and the one-dimensional convolutional network CNN branch as the key value. Align and project the GRU branch features Recorded as One-dimensional convolutional network (CNN) branch features Record , obtain memory characteristics As an auxiliary branch input in the first interaction phase: Then, multi-head scaling dot product attention is used, from... Select the most useful auxiliary information for each time step; The first time you cross a branch, note that CBA1's design only updates the output of the main branch. The first interaction phase is performed, and the output features from the GRU branch in the first interaction phase are... Output features from the CNN branch in the first interaction phase Remain unchanged: The core features are those following the first interaction; After the first cross-branch attention CBA1, multi-scale channel convolution (MSCA) is introduced for the second interaction stage. Multi-scale filtering is performed to obtain : in This represents a depthwise convolution operator with different kernel sizes and dilation rates along the channel dimension. is the scale number; s is the scale index of the multi-scale convolution. This is a multi-scale enhanced backbone feature output by multi-scale channel convolutional MSCA; The second cross-branch attention CBA2 backcorrection stage involves the GRU branch and the one-dimensional convolutional network CNN branch undergoing a third interaction. Through the second cross-branch attention CBA2, the GRU and CNN are pulled back to the original state while retaining their respective advantages. The core degenerate feature subspace reduces overfitting to operating condition noise and spurious correlation patterns. in, The GRU branch features are after the second cross-branch attention correction. These are the CNN branch features after the second cross-branch attention correction. This represents the submodule in the second cross-branch attention module used to correct GRU branch features. This represents the submodule in the second cross-branch attention module used to correct the branch features of the one-dimensional convolutional network (CNN). S32.4, will , and Concatenate according to channel dimension, and then... Convolution is used for unified mapping, and residual connections are introduced to obtain shared fused features. : ； in, This is the output projection matrix used for unified mapping of spliced ​​features. This is the residual projection matrix for residual mapping of the enhanced input sample sequence F.

6. The wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of aero-engines according to claim 1, characterized in that, The specific steps of step S33.1 are as follows: S33.1.1: Lightweight timing and channel hybrid module for shared fusion features in the time dimension Linear mixing is performed, with learnable weights applied to different time steps within each feature channel to reconstruct the most useful temporal combinations for prediction. The lightweight temporal and channel mixing module uses a shared temporal transformation matrix and bias across all channels to obtain the temporal mixing result. : in, Indicates shared and integrated features The A sequence of sensor channels, Indicates shared and integrated features Number of feature channels Represents the learnable weight matrix for temporal mixing. Represents a learnable bias vector for temporal mixing; The time blending results corresponding to each feature channel Combining into temporal mixed feature tensors : ； Indicates the characteristics of sharing and integration Each feature channel undergoes a time-mixing operation with linear transformation in the time dimension; S33.1.2: A small feedforward network is used for feature mixing in the channel dimension, for any time step Temporal Mixed Feature Tensor Channel MLP first expands the dimensions to Throw it back This enables nonlinear channel interaction; then, after parallel computation across all time steps, the result is obtained. And then Apply layer normalization to stabilize the output obtained during training. ; after obtaining the output Subsequently, the lightweight timing and channel hybrid module... Global pooling is performed and concatenated along the channel dimension. Then, a global representation is obtained by linear mapping and a lightweight MLP-based HorizonHead. The global representation is then mapped to a baseline prediction sequence as the main path prediction result. in, To represent the temporal mixture feature tensor Perform a channel blending operation on the channel-dimensional nonlinear mapping at each time step. Let represent the channel blending feature tensor obtained after channel blending.

7. The wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of aero-engines according to claim 1, characterized in that, The specific steps in step S33.2 are as follows: S33.2.1: Incorporate shared and integrated features Model dimensions mapped to the encoder To obtain sequence characterization : S33.2.2: Slow-Fast Dual-Speed ​​Temporal Resampling; The dual-speed hybrid temporal encoder includes a fast branch and a slow branch, where the fast branch is a fast branch encoder and the slow branch is a slow branch encoder; the fast branch maintains the original temporal resolution and shares fused features. Perform shallow modeling to capture rapidly changing local fine-grained dynamics. ; For fast branch encoders; The slow branch then follows the time axis to share fused features. Downsampling is performed to extract stable global semantic information and long-term dependencies at a lower temporal resolution. The downsampled input... ; The downsampling factor; To indicate after downsampling The feature vector at the k-th time step, The sequence representation E is in the first... The feature vectors of each time step; k represents the index of the time step in the downsampled sequence. Indicates the time length of sequence representation of E; Will Inputting the slow branch encoder yields: This is the output of the slow branch encoder. It is a slow branch encoder; Because the fast and slow branches have different time resolutions, the slow branch output needs to be upsampled back by a certain length. Record the result after upsampling as : For upsampling; Then, at each time step, gating fusion is performed on the fast and slow branches; set up , They represent and In the The feature vector at each time step, then the gate vector The definition is as follows: in and These are learnable parameters; The dual-speed hybrid timing encoder in the first The fusion features at each time step are written as: ； S33.2.3: A lightweight state space branch is introduced within each coding layer of the dual-speed hybrid temporal encoder and set in parallel with a multi-head self-attention branch with ALiBi bias to simultaneously model local content dependency and long-range trend memory; after combining the state sequences of all channels, the output of the lightweight state space branch is obtained; the output of the multi-head self-attention branch with ALiBi bias and the output of the lightweight state space branch are adaptively fused at the feature level through learnable gating to obtain the fusion result. The fusion result is residually connected with the corresponding coding layer input and transformed through a feedforward network to obtain the corresponding coding layer output, that is, the bypass prediction result.

8. A wavelet-enhanced multi-branch temporal network fusion system for predicting the remaining service life of aero-engines, comprising interconnected microprocessors and memory, characterized in that, The microprocessor is programmed or configured to execute a wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of an aero-engine, as described in any one of claims 1-7.

9. A computer-readable storage medium storing a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, a wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of an aero-engine, as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, any one of claims 1 to 7, a wavelet-enhanced multi-branch temporal network fusion method for predicting the remaining service life of an aero-engine.

Images