Pulse timing dynamics network-based aero-engine remaining useful life prediction method and system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TAIHANG NATIONAL LABORATORY
- Filing Date
- 2026-03-25
- Publication Date
- 2026-06-02
Smart Images
Figure CN121902328B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of aero-engine technology, and in particular to an aero-engine RUL prediction method and system based on pulse time-series dynamics networks. Background Technology
[0002] In RUL (Remaining Useful Life) prediction, degradation signals are often extremely weak in the early stages of life, while noise, operational maneuvers, and external disturbances are significant. Traditional real-valued networks perform intensive multiplication and addition at each time step, resulting in high energy consumption; at the same time, continuous activation propagates irrelevant disturbances completely. Spiking Neural Networks (SNNs), utilizing threshold triggering and event sparsity, naturally possess the potential to suppress short-term noise and reduce computational redundancy under the mechanism of "only releasing when there is sufficient information."
[0003] However, the conversion from conventional ANN (Artificial Neural Network) to SNN is prone to the accumulation of quantization and timing errors during low-latency (few time steps) inference: membrane potential overshoot leads to output deviation; and with multi-layer stacking, the error is amplified layer by layer, thereby reducing the timing fidelity of the RUL curve. Existing spiking networks are mostly designed for vision / classification tasks and have not been specifically designed for the long-term, multi-condition, and low-latency requirements of PHM (Prognostics and Health Management) regression.
[0004] Existing technologies often employ standard LIF pulse networks for temporal regression, but these can only emit positive pulses. Overshooting of the membrane potential can only be eliminated through leakage and reset, making it difficult to accurately approximate continuous activations during low-time-step inference. When the degradation signal is weak and noise frequently triggers the threshold, the pulse activity becomes dominated by noise, affecting regression stability. Alternatively, lightweight variants of real-valued networks such as temporal convolutions and Transformers can be used. Even with fewer parameters, dense multiply-accumulate operations under continuous activations still exist, and compression often sacrifices robustness or requires complex retraining, making it difficult to maintain stability under strong noise and domain drift conditions. Summary of the Invention
[0005] In view of this, embodiments of this application provide a method and system for predicting the RUL of aero-engines based on pulse time-series dynamics networks, which at least partially solves the problem in the prior art of how to reduce the accumulation of timing errors and maintain regression accuracy in low-latency (few time steps) and multi-layer stacked pulse networks.
[0006] In a first aspect, embodiments of this application provide a method for predicting the RUL (Range Limiting) of aero-engines based on pulse-series dynamics networks, including:
[0007] A pulse temporal dynamics network is constructed and trained. The pulse temporal dynamics network includes multiple stacked temporal pulse convolutional modules, and residual temporal coupling connection channels are set between adjacent temporal pulse convolutional modules.
[0008] Acquire the timing data generated during the operation of the target device and convert it into a discrete pulse sequence;
[0009] Discrete pulse sequences are input into a pulse temporal dynamics network for temporal feature extraction and fusion, and a high-level fused pulse sequence is output.
[0010] Based on a high-level fusion pulse sequence, a predicted value for the remaining service life of the target device is generated.
[0011] According to a specific implementation of an embodiment of this application, the step of inputting the discrete pulse sequence into a pulse temporal dynamics network for temporal feature extraction and fusion includes:
[0012] In at least one layer of the pulse timing dynamics network, at each time step, signed integral firing neurons are used to sequentially integrate, leak, determine firing threshold, and reset the input discrete pulse sequence.
[0013] The pulse sequence output by the signed integral firing neuron is input into the temporal pulse convolution module. Each temporal pulse convolution module performs one-dimensional temporal convolution, gated normalization and pulse pooling in sequence.
[0014] Through the residual temporal coupling connection channel, the pulse sequence containing fine temporal information output by the shallow temporal pulse convolution module is additively fused with the pulse sequence containing abstract semantic information output by the deep temporal pulse convolution module step by step. The residual temporal coupling connection channel is equipped with a learnable attenuation coefficient, which adaptively adjusts the intensity of the temporal information transmitted from the shallow layer to the deep layer.
[0015] According to a specific implementation of an embodiment of this application, the step of using a signed integral firing neuron to sequentially integrate, leak, determine the firing threshold, and reset the input discrete pulse sequence includes:
[0016] The signed integral-emitting neuron receives a discrete pulse sequence, multiplies it by the corresponding weights, and adds it to the membrane potential to complete the integration.
[0017] The integrated membrane potential is multiplied by the leakage factor to update the membrane potential;
[0018] Determine whether the updated membrane potential exceeds the positive firing threshold or falls below the negative firing threshold;
[0019] If the updated membrane potential exceeds the positive firing threshold, a positive pulse is fired, and a subtraction reset is performed;
[0020] If the updated membrane potential is lower than the negative firing threshold, a negative pulse is fired and a subtraction reset is performed;
[0021] If the updated membrane potential is between the positive and negative firing thresholds, no pulse is fired, and the membrane potential remains unchanged.
[0022] According to a specific implementation of an embodiment of this application, each temporal pulse convolution module sequentially performs one-dimensional temporal convolution, gated normalization, and pulse pooling, including:
[0023] A one-dimensional temporal convolution operation is performed on the pulse sequence input to the current temporal pulse convolution module to extract the degradation trend and features within a local time period and obtain a continuous value feature map;
[0024] Perform gating normalization and adaptively adjust the firing threshold or gain of the signed integral firing neuron based on the continuous value feature map;
[0025] Pulse pooling is performed to downsample the gated and normalized feature map to expand the receptive field and capture long-term trends.
[0026] According to a specific implementation of an embodiment of this application, the execution of gating normalization, which adaptively adjusts the firing threshold or gain of neurons based on continuous value feature maps, includes:
[0027] Calculate the absolute average of all eigenvalues of the continuous value feature map at the current time step, and use this absolute average as the intensity value reflecting the continuous value feature map;
[0028] The intensity value is input into a learnable gating function to obtain an adjustment factor. The expression for the learnable gating function is: Where sigmoid is the activation function, β t Let w be the adjustment factor, b be the first learnable parameter, and g be the second learnable parameter. t This is the intensity value;
[0029] The adjustment factor can be used as a multiplicative gain and directly applied to the continuous feature map of the input to adjust the input gain; or, the firing threshold of the signed integral firing neurons in the next layer can be dynamically adjusted according to the adjustment factor.
[0030] The continuous value feature map after adjusting the input gain or the adjusted firing threshold is passed to the next layer of signed integral firing neurons.
[0031] According to a specific implementation of an embodiment of this application, the expression for the one-dimensional temporal convolution is:
[0032] ,
[0033] in, This represents the output value of the l-th convolutional filter at time t. Let S represent the weight parameters of the l-th convolutional filter at time delay k, where S represents the current time step and K represents the size of the convolutional kernel.
[0034] According to a specific implementation of an embodiment of this application, the training pulse timing dynamics network includes:
[0035] The temporal pulse convolution module is trained end-to-end using a backpropagation algorithm and a surrogate gradient method, and uses historical device operating data and its corresponding real remaining lifespan labels as supervision signals during the training process.
[0036] When the input to the sequential pulse convolution module comes from the quantization encoding front end, a training method combining quantization training and pulse network training is adopted.
[0037] According to a specific implementation of an embodiment of this application, the training pulse timing dynamics network includes:
[0038] The pulse timing dynamics network is trained in stages. First, the deep temporal pulse convolutional modules are frozen, and only the shallow temporal pulse convolutional modules and readout layers are trained, so that the network learns a stable basic pulse firing pattern.
[0039] The deep temporal pulse convolution module is gradually unfrozen, and all parameters are jointly adjusted to learn complex long-term degradation dependencies.
[0040] According to a specific implementation of an embodiment of this application, the method further includes:
[0041] When there is no pulse input at a certain time step or the membrane potential does not exceed the firing threshold, skip part of the one-dimensional temporal convolution or only perform addition accumulation.
[0042] Secondly, embodiments of this application also provide a RUL prediction system for aero-engines based on pulse-series dynamics networks, used to implement the RUL prediction method for aero-engines based on pulse-series dynamics networks as described in any embodiment of the first aspect, the system comprising:
[0043] The training module is used to build and train the pulse temporal dynamics network, which includes multiple stacked temporal pulse convolutional modules, with residual temporal coupling connection channels set between adjacent temporal pulse convolutional modules;
[0044] The data acquisition module is used to acquire time-series data generated during the operation of the target device and convert it into a discrete pulse sequence;
[0045] The feature extraction and fusion module is used to input discrete pulse sequences into a pulse temporal dynamics network for temporal feature extraction and fusion, and output a high-level fused pulse sequence.
[0046] The prediction value generation module is used to generate a prediction value of the remaining service life of the target device based on a high-level fused pulse sequence.
[0047] Beneficial effects:
[0048] The RUL prediction method and system for aero-engines based on a pulse temporal dynamics network in this application embodiment achieves effective processing of aero-engine time-series data by constructing a pulse temporal dynamics network containing multiple stacked temporal pulse convolution modules with residual temporal coupling connection channels between adjacent modules. During feature extraction and fusion, the network utilizes signed integral firing neurons to perform pulse sequence integration, leakage, firing threshold judgment, and resetting. Through one-dimensional temporal convolution, gated normalization, and pulse pooling operations of the temporal pulse convolution modules, combined with additive fusion of the residual temporal coupling connection channels, it can fully capture the fine temporal information and abstract semantic information during equipment degradation. This effectively improves the RUL prediction accuracy in low-latency, multi-layer stacked scenarios, reduces the accumulation of temporal errors, and provides reliable technical support for aero-engine health management. Attached Figure Description
[0049] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0050] Figure 1 A flowchart of an aero-engine RUL prediction method based on a pulse-time dynamics network according to an embodiment of the present invention;
[0051] Figure 2 This is a schematic diagram of the overall structure of a pulse timing dynamics network according to an embodiment of the present invention;
[0052] Figure 3 This is a schematic diagram of the timing of an S-IF neuron according to an embodiment of the present invention;
[0053] Figure 4 This is a low-latency inference timing diagram according to an embodiment of the present invention;
[0054] Figure 5 This is a reliability characteristic analysis diagram of the C-MAPSS dataset according to an embodiment of the present invention;
[0055] Figure 6 This is a complete lifecycle diagram of a single engine according to an embodiment of the present invention. Detailed Implementation
[0056] The embodiments of this application will now be described in detail with reference to the accompanying drawings.
[0057] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application 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 this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0058] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.
[0059] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The illustrations only show the components related to this application and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.
[0060] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.
[0061] The abbreviations and key terms used in this application are defined as follows:
[0062] SNN (Spiking Neural Network): A neural network that transmits information using discrete impulse events.
[0063] LIF (Leaky Integrate-and-Fire): A leaky integral firing model with membrane potential leakage and threshold triggering.
[0064] S-IF (Signed Integrate-and-Fire): A signed integral firing model that allows positive / negative impulses for error cancellation.
[0065] STDN (Spiking Temporal Dynamics Network): The core network structure of this invention.
[0066] TSCB (Temporal Spiking Convolution Block): This module includes temporal convolution, gated normalization, and pulse pooling.
[0067] RTC (Residual Temporal Coupling): Residual temporal coupling is used to transfer temporal memories across layers and suppress accumulated errors.
[0068] BPTT: Backpropagation via time.
[0069] Q-BPTT: BPTT training process combined with quantization-pulse.
[0070] C-MAPSS (Commercial Modular Aero-Propulsion System Simulation): Simulation of commercial modular aero-propulsion systems.
[0071] This application provides a core technical solution for "Simuli Temporal Dynamics Network STDN (S-IF / LIF+TSCB+Residual Temporal Coupling)," targeting low-power, low-latency online inference (RUL) for aero-engines. The temporal impulse network, composed of S-IF / LIF neurons, a hierarchical TSCB module, and RTC residual coupling, can receive inputs from a quantization-pulse coding front-end or any impulse encoder. An S-IF neuron mechanism capable of generating positive / negative impulses is introduced on top of LIF for low-latency error cancellation and stable firing. A TSCB module, incorporating temporal convolution, gated normalization, and impulse pooling, is constructed to extract degenerate dynamic features at multiple scales. Cross-layer residual temporal coupling with RTC preserves temporal memory and stabilizes long-range dependent regression outputs. A training and inference process compatible with surrogate gradient / BPTT is provided, supporting joint optimization with the quantization front-end. The following detailed description of this application is provided with reference to the accompanying figures.
[0072] Firstly, referring to Figure 1 and Figure 2 This application provides a method for predicting the RUL (Range Limiting) of aero-engines based on pulse-series dynamics networks, including:
[0073] Step S101: Construct and train a pulse temporal dynamics network. The pulse temporal dynamics network includes multiple stacked temporal pulse convolutional modules, and residual temporal coupling connection channels are set between adjacent temporal pulse convolutional modules.
[0074] Step S102: Obtain the timing data generated during the operation of the target device and convert it into a discrete pulse sequence;
[0075] Step S103: Input the discrete pulse sequence into the pulse temporal dynamics network for temporal feature extraction and fusion, and output a high-level fused pulse sequence;
[0076] Step S104: Based on the high-level fusion pulse sequence, generate the predicted value of the remaining service life of the target equipment.
[0077] In this embodiment, the pulse temporal dynamics network achieves deep mining of complex time-series data of aero-engines through stacked temporal pulse convolutional modules. Each temporal pulse convolutional module can capture dynamic features at different time scales, while the residual temporal coupling connection channel effectively alleviates the gradient vanishing problem during deep network training and enhances the stability of feature transfer. In step S102, the multidimensional time-series data generated during the operation of the target equipment, such as continuous monitoring data collected by sensors for temperature, pressure, and speed, is converted into discrete pulse sequences through preset pulse coding rules. This conversion method preserves the temporal correlation and key event information of the original data, laying the foundation for feature extraction in subsequent networks. When the discrete pulse sequence ( Figure 2 In The input pulse vector (representing the i-th sample at time step 1) is input into the pulse temporal dynamics network. Each temporal pulse convolution module sequentially performs convolution operations and pulse activation processing on it. Through multi-level feature extraction and fusion, low-level raw data features are gradually transformed into high-level fused pulse sequences that can characterize the evolution of equipment health status. In step S104, based on the high-level fused pulse sequence, combined with the degradation model and lifespan decay law of the aero-engine, mapping calculations are performed through a preset prediction head network to finally generate a specific predicted value for the remaining service life of the target equipment (e.g., ...). Figure 2 In , where represents the RUL prediction value of the i-th sample, and this prediction value can provide accurate decision support for condition-based maintenance and safe operation of aero engines.
[0078] In specific implementation, step S102 involves acquiring one-dimensional or multi-dimensional time-series monitoring data generated by the target device during operation. This data includes, but is not limited to, vibration signals, temperature, pressure, and rotational speed. The time-series data is preprocessed, including data cleaning, outlier handling, and normalization, to eliminate the influence of dimensions and improve data quality. A pulse encoder is used to convert the obtained continuous-value time-series data into discrete pulse sequences. The encoding method can employ rate encoding, delay encoding, or a quantization threshold-based encoding method. This step transforms the real-value calculation domain into the pulse event domain, providing input for subsequent spiking neural network processing.
[0079] In one embodiment, the pulse-time dynamics network employs LIF-based dynamics:
[0080] (1) Each neuron integrates the input pulse at time t and leaks it. When the threshold is crossed, it fires and updates the membrane potential by "subtraction reset";
[0081] (2) This dynamic provides a unified time-series state variable for subsequent S-IF extension and TSCB construction.
[0082] The dynamics of LIF can be expressed by the following equation:
[0083] ,
[0084] in, Let be the membrane potential at time t+1, and λ be the leakage factor. Let W be the membrane potential at time t. ij Let represent the synaptic connection weights from neuron j to neuron i. For pulse, θ i is the threshold, i is the index of the postsynaptic neuron, and j is the index of the presynaptic neuron.
[0085] In one embodiment, the step of inputting the discrete pulse sequence into a pulse temporal dynamics network for temporal feature extraction and fusion includes:
[0086] In at least one layer of the pulse timing dynamics network, at each time step, signed integral firing neurons are used to sequentially integrate, leak, determine firing threshold, and reset the input discrete pulse sequence.
[0087] The pulse sequence output by the signed integral firing neuron is input into the temporal pulse convolution module. Each temporal pulse convolution module performs one-dimensional temporal convolution, gated normalization and pulse pooling in sequence.
[0088] Through the residual temporal coupling connection channel, the pulse sequence containing fine temporal information output by the shallow temporal pulse convolution module is additively fused with the pulse sequence containing abstract semantic information output by the deep temporal pulse convolution module step by step. The residual temporal coupling connection channel is equipped with a learnable attenuation coefficient, which adaptively adjusts the intensity of the temporal information transmitted from the shallow layer to the deep layer.
[0089] In practice, during low-latency inference, to approximate quantized activation, the network is allowed to generate negative pulses (inhibitory events) to counteract membrane potential overshoot. By symmetrically distributing positive and negative pulses, the time-accumulated pulses are more accurately matched to the target simulated activation, thereby reducing interlayer accumulation errors.
[0090] The S-IF mechanism can be represented by the following formula:
[0091]
[0092] in, For positive issuance threshold, The negative issuance threshold This represents a state variable or counter associated with neuron i at time t.
[0093] In this embodiment, by setting a symmetrical structure with positive and negative firing thresholds, neurons can more accurately capture the fluctuation characteristics of signals when processing pulse sequences. Excitatory pulses are generated when the membrane potential reaches the positive firing threshold, while inhibitory pulses are generated when the membrane potential falls below the negative firing threshold. This bidirectional regulation mechanism effectively avoids signal distortion problems that may occur under a single threshold. In low-latency inference scenarios, the introduction of negative pulses can promptly offset the overshoot of the membrane potential, ensuring that the error between the time-accumulated pulse and the target simulated activation is controlled within a minimal range.
[0094] In this embodiment, a residual temporal path is maintained between adjacent TSCBs to additively fuse low-level temporal memories with high-level abstract features. Residual coupling reduces gradient vanishing and information forgetting, stabilizing regression curves under long-range dependencies. A learnable decay coefficient is introduced into the residual branch to adapt to different degradation rates and task profiles. This coefficient is dynamically adjusted through backpropagation to maintain the optimal weight ratio between shallow, fine-grained temporal features and deep semantic features during fusion. For example, in the early degradation stage of the device, the decay coefficient automatically increases to strengthen the transmission of shallow, original temporal fluctuations; while in the near-failure stage, the coefficient decreases to highlight the dominant role of the high-level abstract degradation trend. This adaptive residual temporal coupling mechanism not only solves the decay problem of cross-layer temporal information transmission in traditional deep SNNs but also achieves accurate matching of features at different degradation stages through dynamic weight allocation, further improving the robustness and generalization ability of RUL prediction.
[0095] Furthermore, the step of using a signed integral firing neuron to sequentially integrate, leak, determine the firing threshold, and reset the input discrete pulse sequence includes:
[0096] The signed integral-emitting neuron receives a discrete pulse sequence, multiplies it by the corresponding weights, and adds it to the membrane potential to complete the integration.
[0097] The integrated membrane potential is multiplied by the leakage factor to update the membrane potential;
[0098] Determine whether the updated membrane potential exceeds the positive firing threshold or falls below the negative firing threshold;
[0099] If the updated membrane potential exceeds the positive firing threshold, a positive pulse is fired, and a subtraction reset is performed;
[0100] If the updated membrane potential is lower than the negative firing threshold, a negative pulse is fired and a subtraction reset is performed;
[0101] If the updated membrane potential is between the positive and negative firing thresholds, no pulse is fired, and the membrane potential remains unchanged.
[0102] Specifically, refer to Figure 3 This demonstrates the process by which positive and negative pulses cancel out membrane potential overshoot. In low-timestep inference, to quickly approximate a real-valued target, neurons may need to accumulate a high input within a single timestep. This can cause the membrane potential to far exceed the positive threshold (e.g., the target is to output the equivalent value of "1.5", but a single pulse only represents "1"). If standard LIF (which can only emit positive pulses and has a hard reset) is used, the extra 0.5 will be directly discarded, leading to errors. However, with S-IF, the membrane potential can first exceed the positive threshold to emit a positive pulse (contribution +1), and the excess portion (e.g., 0.5) can be retained through subtraction reset. In the next timestep or even the same timestep (depending on the model design), if this positive excess still results in a high membrane potential, the network can use other inputs or mechanisms to make the neuron's membrane potential fall below the negative threshold, thereby emitting a negative pulse (contribution -1). The positive and negative pulses accumulate over time (+1 + (-1) = 0), and combined with the fine continuous value retained in the membrane potential (e.g., 0.5), non-integer values such as "1.5" or "0.5" can be represented more accurately. The core function of the negative pulse here is to "hedge" or "reverse" excessive positive excitation, achieving a more accurate approximation of continuous values within the discrete pulse event domain. In this embodiment, a subtraction reset strategy is adopted, which is an optimized state update rule that can retain residual information after pulse emission. Combined with the positive and negative dual-pulse emission capability of S-IF, it together constitutes a high-precision regression under low time step and low latency conditions.
[0103] In one embodiment, refer to Figure 4Let the length of the encoded pulse sequence be T (e.g., T = 4~10). At each time step, STDN performs an integration-leakage-threshold comparison on the input pulse. S-IF allows overshoot correction via negative pulses within the same layer, enabling outputs close to those of high-precision real-valued networks even with small T values. To improve stability, the firing threshold can be increased with layer depth, while the leakage coefficient can be decreased with layer depth to enhance memory; alternatively, grouped convolutions can be used in TSCB to reduce the propagation of coupling noise.
[0104] In one embodiment, each temporal pulse convolution module sequentially performs one-dimensional temporal convolution, gated normalization, and pulse pooling, including:
[0105] One-dimensional temporal convolution operation is performed on the pulse sequence input to the current temporal pulse convolution module to extract the degradation trend and features within a local time period and obtain a continuous value feature map to capture short-term drift and degradation local features.
[0106] Perform gate normalization and adaptively adjust the firing threshold or gain of the signed integral firing neuron based on the continuous value feature map to suppress impulse bursts when noise is strong and improve response during the degeneration acceleration phase.
[0107] Pulse pooling is performed to downsample the gated and normalized feature map to expand the receptive field and capture long-term trends.
[0108] In practice, one-dimensional temporal convolution operations, by setting convolution kernels of different sizes, can capture subtle changes at different time scales, such as short-term fluctuations and medium-term trend inflection points. The gating normalization mechanism, with the help of learnable parameters, dynamically balances the importance of each channel in the feature map. When the input signal contains a lot of background noise, the mechanism automatically increases the firing threshold to reduce the generation of invalid pulses. When key performance degradation indicators change significantly, the threshold is rapidly reduced and the gain is increased to ensure that effective features are fully activated. Pulse pooling adopts max pooling or average pooling strategies, which reduces the temporal dimension of the feature map while retaining key pulse information, enabling subsequent network layers to focus on the overall degradation trend over a longer time range, laying the foundation for accurately predicting the remaining service life of aero-engines.
[0109] Specifically, TSCB is divided into shallow layers (short kernel, long receptive field, focusing on local drift) and deep layers (long kernel, long receptive field or dilated convolution, focusing on long-term trends). By gradually expanding the receptive field through pulsed pooling, the network can maintain stable output during the healthy plateau period and quickly adjust the prediction slope at the degradation inflection point.
[0110] Furthermore, the execution of gating normalization, adaptively adjusting the firing threshold or gain of neurons based on continuous value feature maps, includes:
[0111] Calculate the absolute average of all eigenvalues of the continuous value feature map at the current time step, and use this absolute average as the intensity value reflecting the continuous value feature map;
[0112] The intensity value is input into a learnable gating function to obtain an adjustment factor. The expression for the learnable gating function is: Where sigmoid is the activation function, β t Let w be the adjustment factor, b be the first learnable parameter, and g be the second learnable parameter. t This is the intensity value;
[0113] The adjustment factor can be used as a multiplicative gain and directly applied to the continuous feature map of the input to adjust the input gain; or, the firing threshold of the signed integral firing neurons in the next layer can be dynamically adjusted according to the adjustment factor.
[0114] The continuous value feature map after adjusting the input gain or the adjusted firing threshold is passed to the next layer of signed integral firing neurons.
[0115] In this embodiment, the gating normalization mechanism introduces learnable parameters w and b, enabling the model to adaptively adjust the adjustment factor β according to the intensity of different input feature maps. t When the intensity value g of the continuous value feature map t When the value is large, the β output by the sigmoid function is... t A value closer to 1 indicates a weaker gain adjustment to the input feature map or a smaller adjustment to the firing threshold, thus preserving more original feature information; while when g... t When β is small, t Getting closer to 0 significantly reduces the input gain or firing threshold, preventing weak features from being overwhelmed by noise. This dynamic adjustment effectively enhances the network's adaptability to features of varying intensities, improves the model's robustness in processing multi-sensor time-series data from aero-engines, and lays a better input foundation for subsequent pulse coding and feature extraction by signed integral firing neurons.
[0116] Furthermore, the expression for the one-dimensional temporal convolution is:
[0117] ,
[0118] in, This represents the output value of the l-th convolutional filter at time t. Let S represent the weight parameters of the l-th convolutional filter at time delay k, where S represents the current time step and K represents the size of the convolutional kernel.
[0119] In one embodiment, the training pulse timing dynamics network includes:
[0120] The temporal pulse convolution module is trained end-to-end using a backpropagation algorithm and a surrogate gradient method, and uses historical device operating data and its corresponding real remaining lifespan labels as supervision signals during the training process.
[0121] When the input to the sequential pulse convolution module comes from the quantization encoding front end, a training method combining quantization training and pulse network training is adopted.
[0122] In this embodiment, a surrogate gradient is used to replace the non-differentiable points of the impulse function during the training phase, combined with backpropagation via time expansion using BPTT. This is because the impulse firing function (such as the step function of S-IF) is non-differentiable and cannot be directly used with gradient descent. This embodiment uses a combination of surrogate gradient and backpropagation via time (BPTT), which allows the error gradient to "bypass" the non-differentiable points and continue to propagate to shallower layers and earlier time steps in the network, thereby updating the weights and parameters. Specifically, BPTT propagates the gradient back along both the network layer and time step dimensions on the time-expanded computation graph, starting from the loss at the last time step T. This allows the network to learn dependencies across long time steps, which is crucial for capturing long-term patterns from early weak degradation to final failure. If the input to STDN comes from a quantized front-end, it can be combined with a quantized STE (Straight-Through Estimator) to form Q-BPTT (Quantized-BPTT) joint training.
[0123] This joint training approach allows for collaborative optimization of the quantization error of the quantization front-end and the prediction error of the pulse timing dynamics network within a unified optimization framework. This avoids the adverse effects of quantization distortion on subsequent network performance, a problem inherent in traditional staged training. Specifically, the STE (Synchronous Transformation) process performs the actual quantization operation during forward propagation, while ignoring the non-differentiability of the quantization function during backward propagation. It directly passes the gradient to the continuous-valued parameters before quantization, ensuring that the parameters of the quantization front-end are updated synchronously with the weights, thresholds, and other parameters of the pulse network. This end-to-end joint training mechanism effectively improves the accuracy and consistency of the entire process from raw sensor data to the final RUL prediction result, making it particularly suitable for complex systems like aero-engines, which have extremely high requirements for prediction accuracy and reliability.
[0124] In one embodiment, the training pulse timing dynamics network includes:
[0125] The pulse timing dynamics network is trained in stages. First, the deep temporal pulse convolutional modules are frozen, and only the shallow temporal pulse convolutional modules and readout layers are trained, so that the network learns a stable basic pulse firing pattern.
[0126] The deep temporal pulse convolution module is gradually unfrozen, and all parameters are jointly adjusted to learn complex long-term degradation dependencies.
[0127] This embodiment employs a "hierarchical freezing" strategy: first, shallow TSCBs are trained to emit stable signals, then deeper layers are gradually unfrozen to learn long-range dependencies. The parameters of all deep TSCB modules (e.g., the last 50% of layers) are frozen, and only the shallow TSCBs, the final readout layer, and any existing encoder parameters are trained. The goal is to allow the network to first learn to extract stable, fundamental features from input impulses and establish reasonable impulse emission patterns. During this stage, the loss decreases rapidly, indicating that the network has learned an initial mapping. The frozen deep TSCB modules are gradually unfrozen, either layer by layer or group by group. The aim is to allow the deep network to focus on learning more complex and longer-term temporal dependencies and abstract patterns, building upon the stable shallow features. At this point, the RTC mechanism becomes crucial, ensuring that the fine-grained temporal information extracted from the shallow layers can be effectively utilized by the deeper layers. This phased strategy is more stable and converges faster than direct end-to-end training, and effectively alleviates the instability problem of gradient propagation in deep temporal networks.
[0128] In one embodiment, the method further includes:
[0129] When there is no pulse input at a certain time step or the membrane potential does not exceed the firing threshold, skip part of the one-dimensional temporal convolution or only perform addition accumulation.
[0130] In practical implementation, in event-driven implementations, when there is no pulse input at a certain time step or the membrane potential does not exceed the threshold, part of the convolution can be skipped or only addition accumulation can be performed; TSCB can be implemented in hardware as a pulse sparse convolution kernel, significantly reducing the number of MACs. For airborne platforms, key operators can be mapped to integer accumulation and comparison operations, reducing power consumption and latency.
[0131] Figure 5 and Figure 6 The performance comparison of RUL prediction using multiple models is shown. Figure 5The horizontal axis represents True RUL, and the vertical axis represents Rredicted RUL. Ground Truth in the graph represents the true, known remaining useful life (RUL) change curve of the device, and is the gold standard for evaluating the prediction accuracy of all models. RQ-SNN is a quantized spiking neural network, representing the STDN of this application. Transformer is the Transformer model, a powerful deep learning model based on self-attention mechanisms, which performs well in long sequence modeling and is often used as a state-of-the-art baseline for time-series prediction tasks. TimeSeriesTrans is a temporal Transformer model, a variant of the Transformer model specifically designed for time-series data. CNN-LSTM is a hybrid model of convolutional neural network and long short-term memory network, combining the classic hybrid architecture of CNN (extracting local features) and LSTM (modeling long-term dependencies), and is a commonly used benchmark model in the PHM field. BiLSTM is a bidirectional long short-term memory network, a recurrent neural network capable of learning sequence dependencies simultaneously from forward and backward directions. TCN is a temporal convolutional network, a convolutional neural network that uses techniques such as dilated convolution to capture long-range dependencies. XGBoost is an extreme gradient boosting algorithm. RandomForest is a random forest algorithm that improves performance by constructing multiple decision trees and combining their results. LightGBM is a lightweight gradient boosting machine that trains quickly and consumes little memory. CatBoost is a categorical feature gradient boosting algorithm that handles categorical features well. SVR is Support Vector Regression, a regression algorithm based on Support Vector Machines (SVM), suitable for small sample datasets. Figure 5 and Figure 6 The data intuitively demonstrates the superiority of the STDN network in RUL prediction accuracy, especially in capturing complex degradation trends.
[0132] It should be noted that, regarding neuron model replacements: S-IF can be replaced with dual-threshold LIF, the Izhikevich simplified model (a simplified mathematical model proposed by biophysicist Eugene M. Izhikevich), or ALIF (Adaptive Leaky Integrate and Fire) with adaptive thresholding to enhance memory or suppress bursts. Convolutional structure replacements: One-dimensional convolutions in TSCB can be replaced with separable convolutions, dilated convolutions, or frequency-domain convolutions based on FFT (Fast Fourier Transform) to change the receptive field and computational cost. Correction mechanism replacements: Negative pulse correction can be replaced with "signed residual membrane potentials" or "error feedback channels" to achieve equivalent overshoot suppression. Coupling method replacements: RTC can be replaced with cross-layer gated fusion or skip-connection attention, but additive dominance should still be maintained for hardware implementation.
[0133] Secondly, embodiments of this application also provide a RUL prediction system for aero-engines based on pulse-series dynamics networks, used to implement the RUL prediction method for aero-engines based on pulse-series dynamics networks as described in any embodiment of the first aspect, the system comprising:
[0134] The training module is used to build and train the pulse temporal dynamics network, which includes multiple stacked temporal pulse convolutional modules, with residual temporal coupling connection channels set between adjacent temporal pulse convolutional modules;
[0135] The data acquisition module is used to acquire time-series data generated during the operation of the target device and convert it into a discrete pulse sequence;
[0136] The feature extraction and fusion module is used to input discrete pulse sequences into a pulse temporal dynamics network for temporal feature extraction and fusion, and output a high-level fused pulse sequence.
[0137] The prediction value generation module is used to generate a prediction value of the remaining service life of the target device based on a high-level fused pulse sequence.
[0138] The embodiments provided by the present invention have the following advantages:
[0139] Suppressing accumulated errors: S-IF provides a negative pulse correction channel, which can cancel out membrane potential overshoot in a short time, improving low-delay inference accuracy;
[0140] Multi-scale degradation capture: TSCB+pulse pooling forms a hierarchical representation from local drift to long-term trend, adapting to non-stationary, multi-stage degradation;
[0141] Robust noise reduction: Gating normalization reduces sensitivity during the noise phase, reduces irrelevant impulses caused by mission maneuvers, and improves the stability of the RUL curve;
[0142] Edge-friendly: Event sparsity replaces a large number of MACs with conditionally triggered accumulation / comparison, reducing energy consumption and latency.
[0143] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for predicting the RUL (Range Limit Intake) of aero-engines based on pulse temporal dynamics networks, characterized in that, include: A pulse temporal dynamics network is constructed and trained. The pulse temporal dynamics network includes multiple stacked temporal pulse convolutional modules, and residual temporal coupling connection channels are set between adjacent temporal pulse convolutional modules. Acquire the timing data generated during the operation of the target device and convert it into a discrete pulse sequence; Discrete pulse sequences are input into a pulse temporal dynamics network for temporal feature extraction and fusion, and a high-level fused pulse sequence is output. Based on a high-level fused pulse sequence, a predicted value of the remaining service life of the target equipment is generated. The step of inputting the discrete pulse sequence into the pulse temporal dynamics network for temporal feature extraction and fusion includes: In at least one layer of the pulse timing dynamics network, at each time step, signed integral firing neurons are used to sequentially integrate, leak, determine firing threshold, and reset the input discrete pulse sequence. The pulse sequence output by the signed integral firing neuron is input into the temporal pulse convolution module. Each temporal pulse convolution module performs one-dimensional temporal convolution, gated normalization and pulse pooling in sequence. Through the residual temporal coupling connection channel, the pulse sequence containing fine temporal information output by the shallow temporal pulse convolution module is additively fused with the pulse sequence containing abstract semantic information output by the deep temporal pulse convolution module step by step. The residual temporal coupling connection channel is equipped with a learnable attenuation coefficient, which adaptively adjusts the intensity of the temporal information transmitted from the shallow layer to the deep layer.
2. The method for predicting the RUL of aero-engines based on pulse temporal dynamics networks according to claim 1, characterized in that, The process of using signed integral firing neurons to sequentially integrate, leak, determine firing thresholds, and reset the input discrete pulse sequence includes: The signed integral-emitting neuron receives a discrete pulse sequence, multiplies it by the corresponding weights, and adds it to the membrane potential to complete the integration. The integrated membrane potential is multiplied by the leakage factor to update the membrane potential; Determine whether the updated membrane potential exceeds the positive firing threshold or falls below the negative firing threshold; If the updated membrane potential exceeds the positive firing threshold, a positive pulse is fired, and a subtraction reset is performed; If the updated membrane potential is lower than the negative firing threshold, a negative pulse is fired and a subtraction reset is performed; If the updated membrane potential is between the positive and negative firing thresholds, no pulse is fired, and the membrane potential remains unchanged.
3. The method for predicting the RUL of aero-engines based on pulse temporal dynamics networks according to claim 1, characterized in that, Each temporal pulse convolution module sequentially performs one-dimensional temporal convolution, gated normalization, and pulse pooling, including: A one-dimensional temporal convolution operation is performed on the pulse sequence input to the current temporal pulse convolution module to extract the degradation trend and features within a local time period and obtain a continuous value feature map; Perform gating normalization and adaptively adjust the firing threshold or gain of the signed integral firing neuron based on the continuous value feature map; Pulse pooling is performed to downsample the gated and normalized feature map to expand the receptive field and capture long-term trends.
4. The method for predicting the RUL of aero-engines based on pulse-time dynamics networks according to claim 3, characterized in that, The execution of gating normalization, which adaptively adjusts the firing threshold or gain of neurons based on continuous value feature maps, includes: Calculate the absolute average of all eigenvalues of the continuous value feature map at the current time step, and use this absolute average as the intensity value reflecting the continuous value feature map; The intensity value is input into a learnable gating function to obtain an adjustment factor. The expression for the learnable gating function is: Where sigmoid is the activation function, β t Let w be the adjustment factor, b be the first learnable parameter, and g be the second learnable parameter. t This is the intensity value; The adjustment factor can be used as a multiplicative gain and directly applied to the continuous feature map of the input to adjust the input gain; or, the firing threshold of the signed integral firing neurons in the next layer can be dynamically adjusted according to the adjustment factor. The continuous value feature map after adjusting the input gain or the adjusted firing threshold is passed to the next layer of signed integral firing neurons.
5. The method for predicting the RUL of aero-engines based on pulse-time dynamics networks according to claim 1, characterized in that, The expression for the one-dimensional temporal convolution is: ， in, This represents the output value of the l-th convolutional filter at time t. Let S represent the weight parameters of the l-th convolutional filter at time delay k, where S represents the current time step and K represents the size of the convolutional kernel.
6. The method for predicting the RUL of aero-engines based on pulse temporal dynamics networks according to claim 1, characterized in that, The training pulse timing dynamics network includes: The temporal pulse convolution module is trained end-to-end using a backpropagation algorithm and a surrogate gradient method, and uses historical device operating data and its corresponding real remaining lifespan labels as supervision signals during the training process. When the input to the sequential pulse convolution module comes from the quantization encoding front end, a training method combining quantization training and pulse network training is adopted.
7. The method for predicting the RUL of aero-engines based on pulse temporal dynamics networks according to claim 1, characterized in that, The training pulse timing dynamics network includes: The pulse timing dynamics network is trained in stages. First, the deep temporal pulse convolutional modules are frozen, and only the shallow temporal pulse convolutional modules and readout layers are trained, so that the network learns a stable basic pulse firing pattern. The deep temporal pulse convolution module is gradually unfrozen, and all parameters are jointly adjusted to learn complex long-term degradation dependencies.
8. The method for predicting the RUL of aero-engines based on pulse temporal dynamics networks according to claim 2, characterized in that, The method further includes: When there is no pulse input at a certain time step or the membrane potential does not exceed the firing threshold, skip part of the one-dimensional temporal convolution or only perform addition accumulation.
9. A RUL prediction system for aero-engines based on pulse-series dynamics networks, used to implement the RUL prediction method for aero-engines based on pulse-series dynamics networks as described in any one of claims 1-8, characterized in that, The system includes: The training module is used to build and train the pulse temporal dynamics network, which includes multiple stacked temporal pulse convolutional modules, with residual temporal coupling connection channels set between adjacent temporal pulse convolutional modules. The data acquisition module is used to acquire time-series data generated during the operation of the target device and convert it into a discrete pulse sequence; The feature extraction and fusion module is used to input discrete pulse sequences into a pulse temporal dynamics network for temporal feature extraction and fusion, and output a high-level fused pulse sequence. The prediction value generation module is used to generate a prediction value of the remaining service life of the target device based on a high-level fused pulse sequence.