Aero-engine residual life prediction method and device based on feature quantization coding and quantized pulse mapping

Abstract

The embodiment of the application provides a kind of based on feature quantization encoding and quantization pulse mapping aero-engine residual life prediction method and device, it is related to prediction and health management technical field, wherein, the method comprises the following steps: after pre-processing and labelling residual life label to sensor data, generate pre-processed long sequence multi-channel sensor data;Time sequence sample construction is carried out to pre-processed long sequence multi-channel sensor data, generate the sample tensor of continuous real value sensor data;Time series deep model including feature quantization encoder, quantization to time mapper and prediction network is constructed and trained;Sample tensor is input to the feature quantization encoder of training and is converted into the pulse sequence in time domain for feature quantization encoding, the pulse sequence in time domain is input to the prediction network of training, and the predicted value of aero-engine residual life is output.The scheme improves the efficiency of deployment on the edge device with limited resources through the prediction network.

Description

Technical Field

[0001] This invention relates to the field of prediction and health management technology, and in particular to a method and apparatus for predicting the remaining life of an aero-engine based on feature quantization coding and quantization pulse mapping. Background Technology

[0002] As complex thermomechanical equipment, the operating status of aero engines directly affects flight safety and operating costs. To achieve predictive health management (PHM) for aero engines, accurately predicting their Remaining Useful Life (RUL) is crucial.

[0003] Aero engines typically operate under varying conditions, including altitude, Mach number, and thrust command. The data sequences monitored by their sensors exhibit complex characteristics such as strong non-stationarity, multi-source noise interference, and long-term dependencies. This presents a significant challenge to Restricted Length Undergoing (RUL) prediction. Existing RUL prediction methods, particularly those based on deep learning models (such as Long Short-Term Memory (LSTM), Temporal Convolutional Networks (TCN), or Transformers), while achieving significant progress in nonlinear modeling and long-range dependency capture, largely rely on high-precision continuous real-valued computation. These methods suffer from two main drawbacks: First, in the multi-layer propagation of deep networks, minute noise in the sensor data or disturbances caused by changes in operating conditions can easily accumulate and amplify layer by layer, affecting the stability and robustness of the prediction. Second, these models typically involve a large number of intensive multiply-accumulate operations, requiring high computational resources and storage bandwidth, resulting in high energy consumption and inference latency, making it difficult to meet the deployment requirements of airborne or edge computing platforms with stringent power consumption, size, and real-time performance requirements.

[0004] To address the challenges of model deployment, several model quantization and compression techniques have been proposed, which convert the weights and / or activation values ​​of the model into low-bit (e.g., 8-bit or 4-bit) fixed-point numbers after or during training. However, most existing quantization schemes are designed for general computer vision or natural language processing tasks, and their quantization strategies do not address the characteristics of "long-term weak signals" and "multi-condition drift" in aero-engine RUL prediction scenarios. These methods merely pursue numerical approximations, failing to explicitly utilize the noise suppression effect of discretization, and lacking mechanisms to further convert the discretized amplitude information into more energy-efficient event-driven temporal codes (e.g., pulse sequences). Therefore, existing technologies struggle to achieve an effective balance between realizing model robustness and meeting the low-power constraints of edge deployment. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method for predicting the remaining life of aero-engines based on feature quantization encoding and quantization pulse mapping, to solve the technical problems of existing deep learning models being sensitive to noise disturbances, having high computational overhead, and being difficult to deploy efficiently on resource-constrained edge devices. The method includes:

[0006] The sensor data generated during the operation of the aero-engine is acquired, and the sensor data is preprocessed and labeled with the remaining life tag to generate preprocessed long sequence multi-channel sensor data, wherein the sensor data is multi-channel and continuous time series data;

[0007] Temporal sample construction is performed on the preprocessed long sequence multi-channel sensor data to generate a sample tensor of continuous real-value sensor data.

[0008] A temporal deep model including a feature quantization encoder, a quantization-to-temporal mapper, and a prediction network is constructed, and the temporal deep model is trained to generate a trained temporal deep model.

[0009] The sample tensor of the continuous real-value sensor data is input into the trained feature quantization encoder for feature quantization encoding to generate a low-bit discrete amplitude sequence. The low-bit discrete amplitude sequence is then converted into a pulse sequence in the time domain by the trained quantization-to-time mapper. The pulse sequence in the time domain is input into the trained prediction network to output the predicted value of the remaining life of the aero-engine.

[0010] This invention also provides a device for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping, to address the technical problems of existing deep learning models being sensitive to noise disturbances, having high computational overhead, and being difficult to deploy efficiently on resource-constrained edge devices. The device includes:

[0011] The preprocessing module is used to acquire sensor data generated by the aero-engine during operation, preprocess the sensor data and construct a remaining life tag, and generate preprocessed long-sequence multi-channel sensor data, wherein the sensor data is multi-channel and continuous time series data.

[0012] A sample tensor construction module is used to construct time-series samples from the preprocessed long-sequence multi-channel sensor data to generate a sample tensor of continuous real-value sensor data.

[0013] The model training module is used to construct a temporal deep model including a feature quantization encoder, a quantization-to-temporal mapper, and a prediction network, and to train the temporal deep model to generate a trained temporal deep model.

[0014] The prediction output module is used to input the sample tensor of the continuous real-value sensor data into the trained feature quantization encoder for feature quantization encoding to generate a low-bit discrete amplitude sequence, and convert the low-bit discrete amplitude sequence into a pulse sequence in the time domain through the trained quantization-to-time mapper. The pulse sequence in the time domain is then input into the trained prediction network to output the predicted value of the remaining life of the aero-engine.

[0015] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least:

[0016] The deep learning model in this invention can be efficiently deployed on resource-constrained edge devices. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping, provided by an embodiment of the present invention.

[0019] Figure 2 This is a structural block diagram of an aero-engine remaining life prediction device based on feature quantization encoding and quantization pulse mapping provided in an embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] In this embodiment of the invention, a method for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping is provided, such as... Figure 1 As shown, the method includes:

[0023] Step S101: Acquire sensor data generated during the operation of the aero-engine, preprocess the sensor data and label it with the remaining life tag, and generate preprocessed long sequence multi-channel sensor data, wherein the sensor data is multi-channel and continuous time series data;

[0024] Step S102: Construct time-series samples from the preprocessed long-sequence multi-channel sensor data to generate a sample tensor of continuous real-value sensor data.

[0025] Step S103: Construct a temporal deep model including a feature quantization encoder, a quantization-to-temporal mapper, and a prediction network, and train the temporal deep model to generate a trained temporal deep model.

[0026] Step S104: Input the sample tensor of the continuous real-value sensor data into the trained feature quantization encoder for feature quantization encoding to generate a low-bit discrete amplitude sequence. Then, convert the low-bit discrete amplitude sequence into a pulse sequence in the time domain through the trained quantization-to-time mapper. Input the pulse sequence in the time domain into the trained prediction network to output the predicted value of the remaining life of the aero-engine.

[0027] In specific implementation, the following steps are used to acquire sensor data generated during the operation of the aero-engine, preprocess the sensor data and label it with remaining life tags, and then generate preprocessed long-sequence multi-channel sensor data:

[0028] For the sensor data of each sensor channel, wavelet transform and adaptive median filtering algorithms are used for smoothing, denoising, and correction to generate denoised sensor data. The mean and standard deviation of each sensor channel in the training dataset are calculated, and Z-score normalization is performed on the denoised sensor data in the training dataset using the mean and standard deviation to generate normalized sensor data. The end-of-life time of the aero-engine is determined based on the equipment failure definition and / or maintenance threshold. For any moment during the operation of the aero-engine... t A piecewise linear decay model is used to evaluate the remaining lifetime of the tag. RUL ( t ) is used for calculation, where, , R max This is a preset maximum remaining lifespan limit based on the stated end-of-life time. t thThe degradation start time is determined by arranging the normalized sensor data in chronological order and assigning the remaining lifetime tag to each timestamp. RUL ( t )and t The sensor data at different times are correlated to generate a long sequence of preprocessed multi-channel sensor data.

[0029] In specific implementation, the following steps are used to construct time-series samples from the preprocessed long-sequence multi-channel sensor data, generating a sample tensor of continuous real-value sensor data:

[0030] The parameters of the sliding window are set, including the window length L and the sliding step S. Starting from the beginning of the preprocessed long-sequence multi-channel sensor data, the sliding window with the specified length L and sliding step S slides and truncates data along the time dimension. Each slide and truncation captures a continuous data segment of length L containing all M sensor channels, forming an independent window sample. This sliding and truncation operation is repeated until the entire preprocessed long-sequence multi-channel sensor data is traversed, generating N window samples arranged in chronological order. These N window samples are then stacked in a batch dimension to generate a three-dimensional continuous real-valued sensor data sample tensor. Set the remaining lifetime label y corresponding to the i-th window sample. i = RUL ( ),in, RUL ( t )for t The remaining lifespan label at any given moment. The end time of the window is defined; the labels of all window samples are collected to generate the remaining lifetime label vector Y corresponding to the sample tensor X; the sample data pair consisting of the sample tensor X and the remaining lifetime label vector Y is output as a continuous real-valued sensor data sample tensor.

[0031] In practice, the following steps are used to construct a temporal deep model, including a feature quantization encoder, a quantization-to-temporal mapper, and a prediction network:

[0032] A feature quantization encoder is constructed using a learnable scaling factor and a low-bit quantization operator. The feature quantization encoder outputs a low-bit discrete amplitude sequence, wherein the feature quantization encoder is used to map continuous inputs to discrete amplitudes. A quantization-to-time mapper is constructed, and the low-bit discrete amplitude sequence is input into the quantization-to-time mapper to convert the low-bit discrete amplitude sequence into a sparse pulse sequence in the time domain and output it. A prediction network is constructed, and the sparse pulse sequence in the time domain is input into the prediction network to output a predicted value of the remaining life of the aero-engine.

[0033] In practice, the following steps are used to construct a feature quantization encoder using a learnable scaling factor and a low-bit quantization operator:

[0034] Configure the learnable scaling factor of the feature quantization encoder a i And configure the bit width of the low-bit quantization operator. b The learnable scaling factor is used to adaptively adjust the dynamic range of the input features. i For sensor channel index, b The sample tensor of the continuous real-valued sensor data is a positive integer less than 8. x As input, the sample tensor x The first in i Each input of each channel x i Computational quantization operator ,in, i For sensor channel index, a i For the first i Each channel corresponds to a learnable scaling factor, `round()` is the rounding function, and `clip()` is the truncation function; based on the quantization operator... Q i Generated low-bit discrete amplitude sequence Q ( x ) is the output of the feature quantization encoder.

[0035] In specific implementation, the quantization-to-time mapper is constructed through the following steps:

[0036] Configure a pulse coding strategy for the quantization-to-time mapper, the pulse coding strategy being used to convert discrete amplitude information into pulse events in the time domain, wherein the pulse coding strategy includes level-frequency coding, level-delay coding, and threshold crossover coding; configure a dynamic threshold unit for the quantization-to-time mapper, through which a time-varying threshold θ is generated. t Wherein, the threshold that changes over time i t The adjustment rules are one or more combinations of the following: changes in firing time, calculation based on statistics of the input discrete amplitude sequence within a local time window, and pulse coding strategy adaptive adjustment based on the state of neuron membrane potential or historical firing rate during threshold crossing coding; through the low-bit discrete amplitude sequence Q ( x Based on the dynamic threshold that changes over time i tAnd the pulse coding strategy, to generate a pulse sequence. S ( t ).

[0037] In specific implementation, the following steps are used to achieve the effect of the low-bit discrete amplitude sequence. Q ( x Based on the dynamic threshold that changes over time i t And the pulse coding strategy, to generate a pulse sequence. S ( t ):

[0038] pulse sequence S ( t Set the sequence to empty, initialize the time step counter, and initialize the neuron membrane potential state V.

[0039] Traversing the low-bit discrete amplitude sequence Q ( x All time steps t and their corresponding discrete amplitudes in ) q t This process continues until all input data has been processed, generating the final pulse sequence. S ( t ):

[0040] If the pulse coding strategy is the level-frequency coding, calculate the current time step. t The number of pulses n generated by the internal response t = F( q t ), in time step t Within, generate n according to the preset pattern. t A pulse is generated, and the timestamp t of the pulse is recorded in the pulse sequence. S ( t In ), where F( q t ) is a predefined amplitude-frequency mapping function, and F() is a non-decreasing function; if the pulse coding strategy is the level-delay coding, calculate the first pulse relative to the time step. t Delay in issuing the starting point t t = D( q t ), at the point in time t + t t A pulse is generated at the specified time point. t + t t Recorded as a timestamp in the pulse sequence S ( t In ), where D(q t ) is a predefined amplitude-delay mapping function, and D() is a non-increasing function; if the pulse coding strategy is the threshold crossing coding, the discrete amplitude q t As an input current, it is accumulated to the neuron membrane potential V( t )=V( t+ 1)+ q t Through the neuronal membrane potential V( t Determine whether a pulse is generated, and record the timestamp t of the pulse into the pulse sequence. S ( t In the process, after the pulse is fired, the neuron membrane potential V( t Reset to the preset reset potential, wherein the pulse sequence , For dynamic threshold i t The upper threshold, For dynamic threshold i t The lower threshold.

[0041] In specific implementation, the time-series deep model is trained through the following steps to generate the trained time-series deep model:

[0042] The time-series deep model is trained and iterated until a preset convergence condition is met: the prediction loss of the current batch is calculated based on the loss function L, and the quantization operator of the time-series deep model is calculated using the prediction loss. Q i Operator gradient and learnable scaling factor a i The scaling factor gradient; according to the optimization algorithm, the learnable scaling factor is updated using the scaling factor gradient. a i ; for the quantization operator Q i Configure a gradient approximation mechanism and use it as the training strategy during end-to-end training of the temporal deep model; optimize the parameters of the quantization-to-time mapper, wherein the parameters include the coefficients of the dynamic threshold adjustment function, the mapping relationship between the pulse firing rate and the amplitude, and the membrane potential reset value.

[0043] In practice, the prediction network is constructed through the following steps:

[0044] A feature extraction layer of the prediction network is constructed. This feature extraction layer is used to extract temporal features related to engine degradation from the input pulse sequence. The feature extraction layer includes a spiking neural network layer, a recurrent neural network layer, and an attention network layer. The spiking neural network layer is used to perform asynchronous event-driven processing on the pulse events. The recurrent neural network layer is used to convert the pulse sequence into a dense vector and then input it into the recurrent neural network layer of the long short-term memory network. The attention network layer is used to convert the pulse sequence into a binary or rate-coded form. A regression output layer of the prediction network is constructed. This regression output layer is used to map the extracted high-dimensional features into a continuous scalar value as a predicted value of the remaining life of the aero-engine.

[0045] Before providing a further detailed description of the embodiments of the present invention, the nouns and terms involved in the embodiments of the present invention will be explained, and the nouns and terms involved in the embodiments of the present invention shall be interpreted as follows.

[0046] (1) Remaining Useful Life (RUL): refers to the remaining operating time or number of cycles of a device, such as an aircraft engine, from the current moment until it can no longer perform its intended function (i.e., a functional failure occurs or a preset maintenance threshold is reached). In the context of this invention, it is a key health indicator that needs to be predicted.

[0047] (2) Prognostics and Health Management (PHM): is a system engineering approach that aims to achieve comprehensive management of equipment health status through monitoring, diagnosis and prediction of equipment status, thereby optimizing maintenance strategies, improving equipment reliability and availability and reducing life cycle costs.

[0048] (3) Feature Quantization Encoder (FQE): This refers to a processing unit or algorithm that maps continuous, high-precision input features (such as sensor data) to a discrete, low-bit-width set of values. Its core function is to compress information and suppress noise through discretization, and to provide a more robust and efficient data representation for subsequent processing.

[0049] (4) Quantization-to-Time Mapping (Q→T): This refers to an encoding mechanism that converts the discrete amplitude information output by the feature quantization encoder into a series of events in the time domain, usually represented as a pulse sequence. This mapping carries the original amplitude information through the timing, frequency, or number of pulses, thereby converting static values ​​into dynamic time-series events.

[0050] (5) Straight-Through Estimator (STE): This is a gradient approximation technique used in neural network training to handle non-differentiable operations (such as rounding or truncation functions in quantization). During backpropagation, it approximates the gradient of non-differentiable operations as 1 or a simple differentiable function, thus enabling end-to-end gradient descent training of models containing these operations.

[0051] (6) Edge Deployment: This refers to deploying trained models or algorithms on hardware devices (i.e., "edge devices") close to the data source for inference and computation, rather than on remote cloud servers. In the context of aero-engines, edge devices typically refer to onboard computers or near-end ground support equipment, characterized by limited computing resources, power consumption, and storage space.

[0052] This invention provides a method for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping.

[0053] In a basic implementation, the method first requires acquiring monitoring data from multiple sensors collected by the aero-engine during actual operation. This sensor data is multi-channel, potentially including multiple physical quantities such as temperature, pressure, speed, and vibration, and is recorded in continuous time series, reflecting the engine's state evolution under different flight phases and operating conditions. The acquired raw sensor data often contains noise, outliers, and drift caused by changes in operating conditions, therefore it needs to be preprocessed, and residual lifetime (RUL) labels for model training are constructed based on the engine's maintenance records or failure definitions. After this series of operations, a preprocessed long-sequence multi-channel sensor data is generated, which forms the basis for subsequent analysis.

[0054] Next, to adapt this continuous long-sequence data to the input format of the deep learning model, it is necessary to construct time-series samples from the preprocessed long-sequence multi-channel sensor data. This process typically involves splitting the long sequence into a series of short, overlapping or non-overlapping segments of fixed length, each segment constituting an independent training sample. In this way, the original one-dimensional time-series data is reconstructed into a high-dimensional sample tensor, which contains multiple dimensions such as time, channel, and sample batch, while its internal values ​​remain continuous real values. This continuous real-valued sensor data sample tensor will serve as the core input for the subsequent deep learning model.

[0055] Subsequently, a specially designed temporal deep model was constructed. The core architecture of this model consists of three key components connected in series: a feature quantization encoder, a quantization-to-temporal mapper, and a final prediction network. These three components work together to form a complete processing chain from raw sensor data to the final RUL prediction value. Before the model is deployed, the entire temporal deep model needs to be trained end-to-end using the aforementioned constructed sample tensor and corresponding RUL labels. The training objective is to make the model's prediction output as close as possible to the true RUL labels. After training, a post-trained temporal deep model is obtained, whose internal parameters have learned the complex mapping relationship from sensor data patterns to remaining lifetime.

[0056] During the prediction phase, the sample tensor of the continuous real-valued sensor data to be predicted is input into the entry point of the trained temporal deep model, namely the feature quantization encoder. The function of the feature quantization encoder is to convert the input continuous, high-precision floating-point numbers into low-bit discrete integer values, forming a low-bit discrete amplitude sequence. This process can effectively filter out small noise in the data and compress the data representation, because small-range numerical fluctuations are mapped to the same discrete value.

[0057] This low-bit discrete amplitude sequence is then passed to a trained quantized-to-time mapper. This mapper performs a novel encoding transformation that, instead of directly processing these discrete values, converts them into a series of impulse events in the time domain, generating a sparse impulse sequence, based on their magnitude. For example, a large discrete value might be mapped to a dense series of impulses, or an earlier-appearing impulse. This event-driven representation, similar to the information processing methods of biological neural systems, has the potential for extremely high computational efficiency.

[0058] Finally, the pulse sequence containing the original data information is input into the trained prediction network. The prediction network is designed to understand and process this time-series pulse signal, extracting deep-level features related to the engine's degradation state from the pulse sequence's pattern, frequency, and time intervals. Based on these extracted features, the prediction network ultimately outputs a continuous scalar value, which is the predicted remaining lifespan of the current aero-engine. Through this series of transformations from continuous to discrete, and then to time-series events, the entire prediction process not only enhances its resistance to noise but also paves the way for efficient computation on low-power hardware.

[0059] In a preferred embodiment, the steps of preprocessing sensor data and constructing remaining lifetime tags can be further refined. Specifically, for the raw data acquired by each sensor channel, a wavelet transform algorithm can first be used for multi-scale decomposition to separate signal components of different frequencies, and then combined with an adaptive median filtering algorithm to effectively smooth the data and correct sudden spike noise, thereby generating denoised sensor data. Wavelet transform is particularly good at handling non-stationary signals, while adaptive median filtering can well preserve the edge details of the signal; the combination of the two can significantly improve data quality.

[0060] During the data normalization phase, to ensure good consistency and generalization ability of the model across different individual engines or different datasets, the mean and standard deviation of each sensor channel in the training dataset can be calculated first. Then, using these statistics, Z-score normalization is performed on the denoised sensor data in all data (including training, validation, and test data). This strategy of using training set statistics for normalization prevents information from the test set from leaking into the training process, ensuring the fairness of the evaluation, and enabling the model to handle inputs with different numerical ranges.

[0061] The construction of the remaining life tag first requires a clear definition of the end-of-life time. This can be determined based on the equipment's functional failure criteria or preventative maintenance thresholds set by the manufacturer or operator. Once the end-of-life time is determined, for any time t in the aircraft engine's operating history, its true remaining life tag... RUL ( t The degradation rate (RUL) can be calculated using a piecewise linear degradation model. A typical form of this model is that, in the early stages of engine operation, when degradation is not significant, the RUL value is set to a preset maximum upper limit. R max When the process reaches a certain degenerate initiation point... t th Afterward, the RUL value begins to decay linearly or non-linearly until it reaches 0 at the end of its lifespan. This piecewise model, especially when setting... R max This avoids overfitting the model to minor, irregular sensor fluctuations during the engine's healthy state phase, allowing it to focus more on learning significant degradation patterns in the later stages. The calculation formula can be expressed as:

[0062]

[0063] in, R max This is a preset maximum remaining lifespan limit based on the stated end-of-life time. t thAt the onset of degradation, f() is a function describing the decay pattern, such as a linear function. This design makes the labels more consistent with the physical degradation process, thereby guiding the model to learn more meaningful features.

[0064] Finally, the normalized multi-channel sensor data is arranged in chronological order, and the remaining lifetime label RUL(t) calculated for each timestamp t is associated with it, forming a complete, labeled, preprocessed long sequence of multi-channel sensor data. This series of refined preprocessing and label construction steps provides high-quality input for subsequent model training and is the foundation for achieving high-precision predictions. Its technical effect lies in significantly improving the signal-to-noise ratio and label effectiveness of the input data by combining domain knowledge (such as piecewise decay models) and advanced signal processing techniques (such as wavelet transform), thereby reducing the difficulty of subsequent model learning and improving the accuracy and reliability of the final prediction.

[0065] Furthermore, the process of constructing time-series samples from preprocessed long-sequence multi-channel sensor data can be further specified. First, two key parameters of the sliding window need to be set: the window length L and the sliding step size S. The window length L determines the duration of historical data the model can "see" when making a prediction; a suitable L should be sufficient to capture a complete degradation-related dynamic process. The sliding step size S controls the time interval between two adjacent samples; a smaller S can generate more samples, which helps the model learn, but also increases data redundancy and computational cost.

[0066] After setting the parameters, a sliding window of size L will slide and truncate along the time dimension, starting from the beginning of the preprocessed long sequence of multi-channel sensor data. Each slide extracts a continuous data segment of length L containing all M sensor channels; this segment forms an independent window sample. For example, if L=30, S=1, and M=24, the first sample would be the data from the 24 channels from time 1 to time 30, the second sample would be the data from time 2 to time 31, and so on. This process is repeated until the sliding window has traversed the entire long sequence, ultimately generating N window samples arranged in chronological order.

[0067] To enable batch training, these N window samples are stacked in a new dimension (batch dimension) to generate a three-dimensional continuous real-valued sensor data sample tensor X. This tensor typically has dimensions (N, L, M), i.e., (number of samples, window length, number of channels). Its mathematical representation is:

[0068]

[0069] Where N is the total number of generated window samples, L is the window length, and M is the number of sensor channels. Simultaneously, a supervision label needs to be determined for each sample. The remaining lifetime label y corresponding to the i-th window sample is set. i The end time of this window The true remaining lifetime value, i.e., y i = RUL ( By aggregating the labels of all N window samples, a remaining lifetime label vector Y is formed, corresponding one-to-one with the sample tensor X. Finally, the sample data pairs (X,Y) consisting of the sample tensor X and the label vector Y are used as the final output for subsequent training and evaluation of the temporal deep model. This sliding window-based sample construction method successfully transforms time-series data of unequal length into structured data with a uniform format suitable for deep learning models, serving as a crucial bridge connecting data preprocessing and model training.

[0070] In another preferred embodiment, the steps of constructing a temporal deep model can be decomposed into the separate construction of three core components. First, a feature quantization encoder is constructed. The core function of this encoder is to map the continuous real-valued features of the input into discrete amplitudes using a learnable scaling factor and a low-bit quantization operator. This process is lossy compression, but its key lies in learning to control the degree of compression, preserving information useful for the prediction task while discarding irrelevant noise. The output is a low-bit discrete amplitude sequence.

[0071] Secondly, a quantization-to-time mapper is constructed. This mapper receives a discrete amplitude sequence from the feature quantization encoder and, according to a preset encoding strategy, converts it into a sparse pulse sequence in the time domain. This means that a sequence that originally had a value at each time step is transformed into a sequence in which an "event" (i.e., a pulse) occurs only at a specific moment. This representation greatly reduces data redundancy and makes it suitable for event-driven computational models.

[0072] Finally, a prediction network is constructed. This network is the "brain" of the entire model, receiving the pulse sequence generated by the quantization-to-time mapper. Since the input is in pulse form, the prediction network typically employs a special structure, such as a spiking neural network (SNN) or other architectures capable of handling temporal events. The network extracts high-dimensional temporal features from the input pulse patterns and ultimately maps these features to a single continuous scalar value through a regression output layer. This value is the model's prediction of the remaining lifespan of the aero-engine. This cascaded design of the three components forms a complete process from continuous signal to discrete encoding and then to event-driven processing. Its technical advantage lies in achieving a progressive abstraction and compression of information representation, which not only improves robustness to noise but also lays the foundation for low-power, low-latency edge computing.

[0073] Specifically, the construction of the feature quantization encoder involves: First, configuring two core elements for the encoder: a learnable scaling factor. a i And the bit width b of the low-bit quantization operator. Where, the scaling factor... a i This parameter is set independently for the i-th sensor channel. It is automatically optimized during model training via gradient descent, adaptively adjusting the dynamic range of the input features for that channel to better match the finite representation range of the quantizer. The bit width *b* is a hyperparameter, typically set to a positive integer less than 8, such as 2, 3, or 4. A smaller *b* results in higher quantization, better data compression, and stronger noise suppression, but may also lead to greater information loss. Choosing an appropriate *b* involves a trade-off between robustness and fidelity.

[0074] During forward propagation, when a sample tensor x of continuous real-valued sensor data is input, the encoder performs a quantization operation on each input value xi of the i-th channel. This operation is performed by the following quantization operator. Q i definition:

[0075]

[0076] Where i is the index of the sensor channel, a i It is the learnable scaling factor corresponding to the i-th channel. x i This is the input value for that channel. First, the input value xi is multiplied by its corresponding scaling factor. a i Then, the `round()` function rounds the result to the nearest integer. Finally, the `clip()` function truncates the rounded value to [...]. , Within the range of [-8, 7], this range is determined by the bit width b. For example, when b=4, the discrete values ​​of the output will be limited to [-8, 7]. After performing this operation on all channels and all time steps, the resulting low-bit discrete amplitude sequence... Q ( x This is then used as the final output of the feature quantization encoder. The technical advantage of this design lies in introducing a learnable scaling factor. a i This makes the quantization process no longer fixed and data-independent, but adaptively finds the optimal quantization "scale" for each feature channel, thereby achieving low-bit representation while preserving as much information as possible that is crucial for RUL prediction.

[0077] Furthermore, the construction of the quantization-to-time mapper also involves specific strategies and components. First, a pulse coding strategy needs to be configured for the mapper, defining how discrete amplitude information is converted into pulse events in the time domain. Optional strategies include: level-to-frequency coding, where a larger amplitude results in more pulses generated per unit time; level-to-delay coding, where a larger amplitude results in a shorter delay for the first pulse; and threshold-crossing coding, which mimics the working mechanism of biological neurons, accumulating membrane potential as input current and firing a pulse when the membrane potential exceeds a threshold. These strategies provide different ways to dynamically transform static amplitude information.

[0078] To enhance the adaptability of the mapping process, a dynamic thresholding unit can be configured for the quantization-to-time mapper. This unit is responsible for generating a threshold that varies over time. i t Instead of using a fixed threshold, this dynamic threshold... i t The adjustment rules can be very flexible. For example, it can be a function of time, changing with the increase of the running cycle; or it can be calculated based on the statistics (such as mean and variance) of the input discrete amplitude sequence within a certain local time window, increasing the threshold to suppress noise when the input signal fluctuates drastically and decreasing the threshold to improve sensitivity when the signal is stable; in particular, when using threshold crossing coding, the dynamic threshold can also be adaptively adjusted according to the neuron's own membrane potential state or historical firing rate, forming a feedback control. This dynamic threshold mechanism allows the pulse generation process to better adapt to the dynamic characteristics of the input signal and the degeneration state of the motor.

[0079] Ultimately, by integrating a pulse coding strategy and a dynamic threshold, quantization to a time mapper becomes functional. It receives low-bit discrete amplitude sequences. Q ( x As input, based on a dynamic threshold that changes over time. i t Using the selected pulse coding strategy, the final output is a sparse pulse sequence. S ( t Different input amplitudes (y-axis) are converted into pulses of different patterns on the time axis (x-axis). The technical advantage of this construction method is that it not only realizes the conversion from the amplitude domain to the time domain, but also endows the conversion process with high flexibility and adaptability by introducing dynamic thresholds and diverse coding strategies. This allows the generated pulse sequences to represent the key information in the original signal more intelligently and efficiently, providing high-quality input for subsequent event-driven processing.

[0080] Based on low-bit discrete magnitude sequence Q ( x ), dynamic threshold i t Pulse sequence generation using pulse coding strategies S ( t The specific process includes: First, before processing begins, the target pulse sequence is... S ( t The system is initialized with an empty sequence, along with a time-step counter and a neuron membrane potential state V for threshold crossing encoding. Then, the system iterates through the input discrete amplitude sequence. Q ( x Each time step t and its corresponding discrete amplitude in ) q t The system performs corresponding operations according to the selected pulse coding strategy until all input data has been processed, generating the final pulse sequence. S ( t ).

[0081] If the selected pulse coding strategy is level-frequency coding, for the input amplitude qt at each time step t, the number of pulses n to be generated in that time step is first calculated using a predefined, non-decreasing amplitude-frequency mapping function F(). t = F( q t The larger the amplitude, the better. t The larger it is, the more. Then, within the duration of the current time step t, n is generated according to a pre-defined pattern (such as a uniform distribution or a Poisson distribution). t Each pulse is recorded, and the timestamps of these pulses are recorded in the pulse sequence. S ( t)middle.

[0082] If the selected pulse coding strategy is level-delay coding, for the input amplitude qt at each time step t, the firing delay of the first pulse relative to the start point of the current time step will be calculated through a predefined, non-increasing amplitude-delay mapping function D(). t t = D( q t The larger the amplitude, the longer the delay. t t The smaller the value. Then, the system will [do something] at a specific time point. t + t t A pulse is generated at each time step, and this precise time point is recorded as a timestamp in the pulse sequence S(t). At most one pulse is generated at each time step, and its information is entirely carried by its precise occurrence time.

[0083] If the chosen pulse coding strategy is threshold crossing coding, the process simulates a simplified Leaky Integrate-and-Fire neuron model. At each time step t, the discrete amplitude qt is treated as the input current and accumulated onto the neuron's membrane potential V, i.e., V(t) = V(t-1) + qt (this is a simplified representation; in practice, a leakage term may be included). Then, it is determined whether the updated membrane potential V(t) exceeds the dynamic threshold. i t Specifically, an upper threshold can be set. and a lower threshold When V(t) ≥ At time t, the neuron fires a pulse, i.e., S(t) = 1, and records the timestamp t of the pulse; after firing the pulse, the membrane potential V(t) is reset to a preset reset potential (usually the lower threshold). Or a lower value), to prepare to receive new input. If V(t) < If no pulse is emitted, S(t) = 0. The pulse emission logic can be expressed as:

[0084]

[0085]

[0086] in, For dynamic threshold i t The upper threshold, For dynamic threshold i tThe lower threshold. These three specific implementation methods provide a clear operational path for converting discrete amplitudes into time-series pulses, making the entire Q→T mapping process clearly defined and feasible.

[0087] In a specific embodiment, the training process of a temporal deep model can be carried out using iterative training until a preset convergence condition is met (such as the training loss no longer decreasing or the maximum number of training epochs being reached). In each iteration, for the current batch of training samples, forward propagation of the model is first performed to obtain the predicted RUL values. Then, the prediction loss for the current batch is calculated based on a predefined loss function L (such as mean squared error MSE or Huber loss). This loss measures the difference between the model's predicted value and the true RUL label.

[0088] The next crucial steps are backpropagation and parameter update. Since the model includes a feature quantization encoder, its quantization operator... Q i The `round()` function in the code is non-differentiable, so regular gradient descent cannot be applied directly. To solve this problem, we need to quantize the operator. Q i The most common gradient approximation mechanism is the Straight-Through Estimator (STE). During backpropagation, the STE "bypasses" the non-differentiable points of the `round()` function, approximating its gradient as 1. This allows the lost gradient to propagate smoothly back to the layers before the quantization operator, enabling the computation of the gradient with respect to the learnable scaling factor `ai` (the scaling factor gradient).

[0089] Once the gradient of the scaling factor is obtained, an optimization algorithm (such as Adam or SGD) can be used to update the learnable scaling factor. a i The value of the loss is adjusted to reduce the prediction loss. Simultaneously, the gradient of the loss continues to propagate backward, updating other learnable parameters in the quantization-to-temporal mapper and the prediction network. For example, optimizations can be made to the coefficients of the dynamic threshold adjustment function in the quantization-to-temporal mapper, parameters in the amplitude-frequency / delay mapping, or the membrane potential reset value in the threshold crossing encoding.

[0090] In this way, the entire temporal deep model, including its unconventional quantization and impulse generation components, can be jointly optimized within a unified end-to-end training framework. The technical advantage of this training strategy lies in its ability to enable feature quantization, impulse mapping, and the final prediction task to learn collaboratively. Quantization parameters (such as ai) and parameters from other parts of the model evolve together to achieve optimal overall performance. This is far superior to the separate approach of fixing the quantization scheme before training subsequent networks, significantly reducing the accuracy loss caused by quantization and achieving Quantization-Aware Training (QAT).

[0091] In one embodiment, the prediction network can be constructed in two layers. The first layer is a feature extraction layer, whose core task is to extract deep temporal features related to engine health and degradation trends from the input pulse sequence. Considering that the input is sparse pulse events, this feature extraction layer can be implemented in several ways. One approach is to use a spiking neural network (SNN) layer, such as a network composed of leaky integral-fire (LIF) neurons or more complex Izhikevich neurons. SNNs naturally process pulse events asynchronously and event-driven, with computation occurring only when a pulse arrives, thus exhibiting extremely high energy efficiency. Another approach is to first convert the pulse sequence into a dense vector or image form through some means (such as time binning or rate encoding), and then input it into a traditional deep learning model. For example, recurrent neural network layers such as Long Short-Term Memory (LSTM) networks or Gated Recurrent Units (GRUs) can be used to capture temporal dependencies, or Transformer network layers based on self-attention mechanisms can be used to capture long-distance dependencies.

[0092] The second layer is the regression output layer, connected after the feature extraction layer. Its function is to map the high-dimensional feature vector output by the feature extraction layer to a continuous scalar value. Typically, this can be achieved using one or more fully connected layers (also called linear or dense layers). The final layer usually has only one output neuron and uses little or no linear activation function to directly output the regression's RUL prediction. This hierarchical structure decouples the complex feature extraction from the final task output, making the network design more modular. Its technical advantage lies in providing multiple optional feature extraction layer implementations, allowing the framework of this invention to be flexibly combined with various advanced temporal modeling techniques. Whether it's emerging SNNs or mature RNNs / Transformers, they can all utilize the robust and efficient impulse representation provided by the front end of this invention to find the optimal solution under different hardware platforms and performance requirements.

[0093] Now, we describe a minimum protection range embodiment that integrates the aforementioned preferred solutions, demonstrating a complete and efficient aircraft engine remaining life prediction system. From raw data acquisition to the final RUL prediction, the system employs specific and optimized technology choices at every step. Assume our goal is to monitor an aircraft engine equipped with 24 different sensors.

[0094] First, in the data preprocessing stage, the system acquires 24 channels of sensor data with a 1Hz sampling rate from the engine. For each channel, wavelet transform is applied for initial denoising, followed by an adaptive median filter with a window size of 5 to correct for potential transient spikes. Next, Z-score normalization is performed on the denoised data using the mean and standard deviation of each channel calculated across the entire training fleet data. For the RUL label, we set the maximum remaining lifetime upper limit Rmax to 125 flight cycles and constructed it using a piecewise linear decay model.

[0095] Secondly, in the time-series sample construction stage, we set the sliding window length L to 40 time steps (i.e., 40 seconds of data) and the sliding step size S to 1. This means that each prediction by the model is based on continuous data from the past 40 seconds, and there is a high degree of overlap between the samples to capture subtle state changes. Through this operation, the long sequence data is transformed into a sample tensor X with dimensions (N, 40, 24), where N is the total number of samples.

[0096] Next, the sample tensor is fed into a carefully configured temporal deep model. The model's Feature Quantization Encoder (FQE) is configured to use an independent learnable scaling factor ai (i=1 to 24) for each channel, and the quantization bit width b is chosen to be 4 bits. This means that each consecutive input value is mapped to a set of 16 levels of integers in the range [-8, 7]. This 4-bit quantization achieves a good balance between noise suppression and information fidelity.

[0097] The 4-bit discrete amplitude sequence is then fed into a time mapper (Q→T Mapper). In this embodiment, we choose the most efficient threshold crossing coding strategy. Internally, the mapper simulates a leaky integral-fire (LIF) neuron for each channel. The discrete amplitude qt acts as the input current, driving changes in the membrane potential V(t). Crucially, the firing threshold θt is dynamic, adaptively adjusting based on the standard deviation of the input amplitude sequence qt over the past 10 time steps: when the standard deviation increases (indicating signal fluctuations or increased noise), the threshold θt increases accordingly, making the neuron less likely to fire a pulse; conversely, when the signal is stable, the threshold decreases to maintain sensitivity to weak, degenerate signals. A pulse is generated when the membrane potential V(t) exceeds the dynamic threshold θt.

[0098] The generated sparse pulse sequence (a 24-channel pulse stream) is input into the prediction network. In this embodiment, the feature extraction layer of the prediction network employs a spiking neural network (SNN) composed of multiple layers of LIF neurons. The SNN processes the input pulses in an event-driven manner, performing computation only when a pulse arrives, greatly reducing computational power consumption. The output firing rate or membrane potential state of the last layer of spiking neurons in the SNN is aggregated to form a high-dimensional feature vector. Finally, this feature vector is passed through a fully connected regression layer with a single output unit to obtain the final RUL prediction value.

[0099] During the training phase, the entire model (including 24 learnable scaling factors ai, parameters of the dynamic threshold adjustment function, and synaptic weights of the SNN) is trained end-to-end using the Adam optimizer. The non-differentiable 4-bit quantization operator in the FQE is approximated by a pass-through estimator (STE), ensuring smooth gradient propagation.

[0100] The complete workflow of this embodiment is as follows: Real-time multi-channel sensor data streams from an aero-engine are acquired, denoised by wavelet and median filtering, and then segmented into 40x24 sample segments through a sliding window. Each sample segment enters the FQE, where its continuous values ​​are efficiently converted into 4-bit discrete integers by a learnable scaling factor. These integers then drive LIF neurons in the Q→T mapper, and a dynamic thresholding mechanism intelligently converts these discrete amplitudes into sparse, information-rich pulse sequences. Finally, this pulse sequence flows through a deep SNN, which extracts temporal patterns highly correlated with the engine's health status through its internal pulse dynamics, and these patterns are decoded into accurate RUL prediction values ​​by the final regression layer.

[0101] By synergistically combining all the aforementioned specific technical features, this embodiment achieves significant comprehensive technical benefits. Firstly, in terms of efficiency, due to the adoption of 4-bit quantization and event-driven SNN, compared to traditional 32-bit floating-point and computationally intensive LSTM or Transformer models, the computational load (in terms of multiply-accumulate operations) and memory usage of this scheme are reduced by more than an order of magnitude, making real-time online prediction possible on airborne edge devices with limited power consumption and computing power. Secondly, in terms of robustness, the quantization effect of FQE and the dynamic threshold mechanism of the Q→T mapper form a dual "noise filter," effectively suppressing data drift caused by sensor noise and changes in flight conditions. This makes the prediction results more stable and reliable in complex and variable real-world operating environments. Experiments show that in tests injecting Gaussian white noise, the prediction error (RMSE) of this scheme is reduced by approximately 15% compared to the baseline real-valued model. Finally, regarding accuracy, the end-to-end joint training strategy enables perfect coordination between quantization, pulse coding, and the prediction network. The quantization process itself becomes part of the model learning process, thereby minimizing information loss and ensuring prediction accuracy comparable to or even higher than advanced real-valued models. In summary, this embodiment achieves an efficient, robust, and accurate aero-engine remaining life prediction system through the deep integration and synergistic effect of various technical features.

[0102] The technical solution proposed in this invention has a wide range of applications. Its most direct and core application is in the aviation field, used for predictive health management (PHM) of various aero engines (such as turbofan, turboshaft, and turboprop engines). By deploying the methods or devices described in this invention in an onboard computer or a dedicated health monitoring unit, it is possible to predict the remaining life of key engine components (such as fans, compressors, turbines, and bearings) in real time and online. This enables airlines and maintenance departments to shift from the traditional time-based maintenance model based on fixed flight hours or cycles to a more advanced and economical condition-based maintenance model based on actual health status.

[0103] Specifically, when the predicted RUL value falls below a certain warning threshold, the system can automatically send an alert to the ground maintenance center, providing detailed sensor data and predictive justification. This allows maintenance teams to plan maintenance activities in advance and prepare necessary spare parts and tools, significantly reducing aircraft on-ground (AOG) time and improving fleet availability. Furthermore, accurate RUL prediction is the foundation for achieving a higher level of predictive maintenance, which helps optimize spare parts inventory management, avoids unnecessary over-maintenance, and thus significantly reduces total lifecycle operating costs.

[0104] Beyond aero engines, the technical concepts and framework of this invention are also applicable to other complex industrial equipment undergoing long-term, slow degradation processes. For example, in the energy sector, it can be used to predict the remaining lifespan of critical components (such as blades, gearboxes, and generators) in large gas turbines, steam turbines, or wind turbine generator sets. In the rail transportation sector, it can be used to monitor the health status of systems such as bearings, wheelsets, and traction motors in high-speed trains. In manufacturing, it can be used to predict the health evolution trends of critical production equipment such as precision CNC machine tools and large presses. In all these scenarios, the equipment typically operates under varying conditions, and sensor data is also subject to noise interference, while high requirements are placed on real-time performance and deployment costs. This invention, through its unique feature quantization and pulse mapping mechanism, provides a general, efficient, and robust front-end processing solution for PHM problems in these fields.

[0105] Based on the same inventive concept, this invention also provides an aero-engine remaining life prediction device based on feature quantization coding and quantization pulse mapping, as described in the following embodiments. Since the principle of the aero-engine remaining life prediction device based on feature quantization coding and quantization pulse mapping is similar to that of the aero-engine remaining life prediction method based on feature quantization coding and quantization pulse mapping, the implementation of the aero-engine remaining life prediction device based on feature quantization coding and quantization pulse mapping can refer to the implementation of the aero-engine remaining life prediction method based on feature quantization coding and quantization pulse mapping, and will not be repeated. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0106] Figure 2 This is a structural block diagram of an aero-engine remaining life prediction device based on feature quantization encoding and quantization pulse mapping according to an embodiment of the present invention, as shown below. Figure 2As shown, it includes: a preprocessing module 201, a sample tensor construction module 202, a model training module 203, and a prediction output module 204. The structure is described below.

[0107] The preprocessing module 201 is used to acquire sensor data generated by the aero-engine during operation, preprocess the sensor data and construct a remaining life tag, and generate preprocessed long-sequence multi-channel sensor data, wherein the sensor data is multi-channel and continuous time series data.

[0108] A sample tensor construction module 202 is used to construct time-series samples from the preprocessed long sequence multi-channel sensor data to generate a sample tensor of continuous real-value sensor data.

[0109] The model training module 203 is used to construct a temporal deep model including a feature quantization encoder, a quantization-to-temporal mapper and a prediction network, and to train the temporal deep model to generate a trained temporal deep model.

[0110] The prediction output module 204 is used to input the sample tensor of the continuous real-value sensor data into the trained feature quantization encoder for feature quantization encoding to generate a low-bit discrete amplitude sequence, and convert the low-bit discrete amplitude sequence into a pulse sequence in the time domain through the trained quantization-to-time mapper, and input the pulse sequence in the time domain into the trained prediction network to output the predicted value of the remaining life of the aero-engine.

[0111] In one embodiment, the preprocessing module includes:

[0112] The denoising unit is used to smooth and denoise the sensor data of each sensor channel using wavelet transform algorithm and adaptive median filtering algorithm to generate denoised sensor data.

[0113] The normalization unit is used to calculate the mean and standard deviation of each sensor channel in the training dataset, and to perform Z-score normalization on the denoised sensor data in the training dataset using the mean and standard deviation to generate normalized sensor data.

[0114] Define a life end time unit to determine the life end time of the aero engine based on the equipment failure definition and / or maintenance threshold;

[0115] The remaining life calculation unit is used to calculate the remaining life at any point during the operation of the aero-engine. t A piecewise linear decay model is used to evaluate the remaining lifetime of the tag. RUL ( t ) is used for calculation, where, ,R max This is a preset maximum remaining lifespan limit based on the stated end-of-life time. t th This is the moment when degradation begins;

[0116] The preprocessing unit is used to arrange the normalized sensor data in chronological order and assign the remaining lifetime tag corresponding to each timestamp. RUL ( t )and t The sensor data at different times are correlated to generate a long sequence of preprocessed multi-channel sensor data.

[0117] In one embodiment, constructing a sample tensor module includes:

[0118] The parameter setting unit is used to set the parameters of the sliding window, wherein the parameters of the sliding window include the window length L and the sliding step size S;

[0119] The window sample generation unit is used to slide and truncate the preprocessed long sequence multichannel sensor data in the time dimension using the sliding window with the window length L and the sliding step size S. Each time, a continuous data segment of length L containing all M sensor channels is slid and truncated to form an independent window sample. The sliding and truncating operation is repeated until the entire preprocessed long sequence multichannel sensor data is traversed to generate N window samples arranged in time order.

[0120] The sample tensor construction unit is used to stack N window samples in the batch dimension to generate a sample tensor of three-dimensional continuous real-valued sensor data. ;

[0121] The remaining lifetime label unit is used to set the remaining lifetime label y corresponding to the i-th window sample. i = RUL ( ),in, RUL ( t )for t The remaining lifespan label at any given moment. The end time of the window;

[0122] The label vector generation unit is used to collect the labels of all window samples and generate the remaining lifetime label vector Y corresponding to the sample tensor X.

[0123] The sample tensor output unit is used to output the sample data pair consisting of the sample tensor X and the remaining lifetime label vector Y as a continuous real-valued sensor data sample tensor.

[0124] In one embodiment, the model training module includes:

[0125] Output discrete amplitude sequence unit, used to construct a feature quantization encoder through a learnable scaling factor and a low-bit quantization operator, and output a low-bit discrete amplitude sequence through the feature quantization encoder, wherein the feature quantization encoder is used to map continuous input to discrete amplitude;

[0126] A time-mapping unit is constructed to construct the quantization-to-time mapper, input the low-bit discrete amplitude sequence into the quantization-to-time mapper, convert the low-bit discrete amplitude sequence into a sparse pulse sequence in the time domain, and output it.

[0127] A pulse sequence generation unit is used to construct a prediction network. The sparse pulse sequence in the time domain is input into the prediction network, and the predicted value of the remaining life of the aero-engine is output.

[0128] In one embodiment, the time mapper unit is further configured to configure a pulse coding strategy for the quantization-to-time mapper, the pulse coding strategy being used to convert discrete amplitude information into pulse events in the time domain, wherein the pulse coding strategy includes level-frequency coding, level-delay coding, and threshold crossing coding; and to configure a dynamic threshold unit for the quantization-to-time mapper, through which a threshold that varies with time is generated. i t Wherein, the threshold that changes over time i t The adjustment rules are one or more combinations of the following: changes in firing time, calculation based on statistics of the input discrete amplitude sequence within a local time window, and pulse coding strategy adaptive adjustment based on the state of neuron membrane potential or historical firing rate during threshold crossing coding; through the low-bit discrete amplitude sequence Q ( x Based on the dynamic threshold that changes over time i t And the pulse coding strategy, to generate a pulse sequence. S ( t ).

[0129] In one embodiment, the pulse sequence generating unit is further configured to generate the pulse sequence. S ( t Set the sequence to empty, initialize the time step counter and initialize the neuron membrane potential state V; traverse the low-bit discrete amplitude sequence. Q ( x All time steps t and their corresponding discrete amplitudes in ) q t This process continues until all input data has been processed, generating the final pulse sequence.S ( t If the pulse coding strategy is the level-frequency coding, calculate the current time step. t The number of pulses n generated by the internal response t = F( q t ), in time step t Within, generate n according to the preset pattern. t A pulse is generated, and the timestamp t of the pulse is recorded in the pulse sequence. S ( t In ), where F( q t ) is a predefined amplitude-frequency mapping function, and F() is a non-decreasing function; if the pulse coding strategy is the level-delay coding, calculate the first pulse relative to the time step. t Delay in issuing the starting point t t = D( q t ), at the point in time t + t t A pulse is generated at the specified time point. t + t t Recorded as a timestamp in the pulse sequence S ( t In ), where D( q t ) is a predefined amplitude-delay mapping function, and D() is a non-increasing function; if the pulse coding strategy is the threshold crossing coding, the discrete amplitude q t As an input current, it is accumulated to the neuron membrane potential V( t )=V( t+ 1)+ q t Through the neuronal membrane potential V( t Determine whether a pulse is generated, and record the timestamp t of the pulse into the pulse sequence. S ( t In the process, after the pulse is fired, the neuron membrane potential V( t Reset to the preset reset potential, wherein the pulse sequence , For dynamic threshold i t The upper threshold, For dynamic threshold i t The lower threshold.

[0130] In one embodiment, the model training module further includes:

[0131] An iterative unit is used to perform the following training and iteration on the temporal deep model until a preset convergence condition is met:

[0132] The gradient calculation unit is used to calculate the prediction loss of the current batch based on the loss function L, and to calculate the quantization operator of the temporal deep model using the prediction loss. Q i Operator gradient and learnable scaling factor a i The scaling factor gradient;

[0133] The scaling factor update unit is used to update the learnable scaling factor according to the scaling factor gradient based on the optimization algorithm. a i ;

[0134] Training strategy formulation unit, used for the quantization operator Q i Configure a gradient approximation mechanism and use the gradient approximation mechanism as the training strategy during end-to-end training of the temporal deep model;

[0135] The parameter optimization unit is used to optimize the parameters of the quantization-to-time mapper, wherein the parameters include the coefficients of the dynamic threshold adjustment function, the mapping relationship between pulse firing rate and amplitude, and the membrane potential reset value.

[0136] In one embodiment, the predicted value output module includes:

[0137] An extraction layer unit is constructed to build a feature extraction layer for the prediction network. The feature extraction layer is used to extract temporal features related to engine degradation from the input pulse sequence. The feature extraction layer includes a spiking neural network layer, a recurrent neural network layer, and an attention network layer. The spiking neural network layer is used to perform asynchronous event-driven processing on the pulse events. The recurrent neural network layer is used to convert the pulse sequence into a dense vector and then input it into the recurrent neural network layer of the long short-term memory network. The attention network layer is used to convert the pulse sequence into a binary or rate-encoded form.

[0138] An output layer unit is constructed to build the regression output layer of the prediction network. The regression output layer is used to map the extracted high-dimensional features into a continuous scalar value as a predicted value of the remaining life of the aero-engine.

[0139] The embodiments of the present invention achieve the following technical effects:

[0140] The robustness of the prediction model is improved by mapping continuous sensor data to a finite low-bit discrete interval through a feature quantization encoder. This effectively suppresses small-amplitude random noise and disturbances caused by changes in operating conditions in the input signal, as these minute changes are insufficient to alter the quantization interval to which the data belongs. This prevents noise from propagating and amplifying in deeper layers of the network, enhancing the model's prediction stability in complex operating environments. Computational complexity and power consumption are reduced by converting discrete amplitude information into a sparse pulse sequence in the time domain through a quantization-to-time mapper. This allows the subsequent prediction network to adopt an event-driven computation mode based on addition and comparison, replacing the dense, high-precision multiply-accumulate operations in traditional deep networks. This significantly reduces the model's computational load, memory access bandwidth, and power consumption, making it more suitable for online deployment on edge devices with limited computing resources and energy, such as airborne systems. Consistency between model performance and deployment is ensured by introducing gradient approximation mechanisms (such as pass-through estimators) during the model training phase, enabling end-to-end training of non-differentiable quantization operations. This enables the parameters of the feature quantization encoder, quantization-to-time mapper, and downstream prediction network to be optimized collaboratively, avoiding the significant performance degradation that may be caused by traditional "post-training quantization" methods, and ensuring that the model can maintain high prediction accuracy after deployment.

[0141] Obviously, those skilled in the art should understand that the modules or steps of the above-described embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

[0142] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping, characterized in that, include: The sensor data generated during the operation of the aero-engine is acquired, and the sensor data is preprocessed and labeled with the remaining life tag to generate preprocessed long sequence multi-channel sensor data, wherein the sensor data is multi-channel and continuous time series data; Temporal sample construction is performed on the preprocessed long sequence multi-channel sensor data to generate a sample tensor of continuous real-value sensor data. A temporal deep model including a feature quantization encoder, a quantization-to-temporal mapper, and a prediction network is constructed, and the temporal deep model is trained to generate a trained temporal deep model. Configure the learnable scaling factor of the feature quantization encoder a i And configure the bit width of the low-bit quantization operator. b The learnable scaling factor is used to adaptively adjust the dynamic range of the input features. i For sensor channel index, b The sample tensor of the continuous real-valued sensor data is a positive integer less than 8. x As input, the sample tensor x The first in i Each input of each channel x i Computational quantization operator ,in, i For sensor channel index, a i For the first i Each channel corresponds to a learnable scaling factor, `round()` is the rounding function, and `clip()` is the truncation function; based on the quantization operator... Q i Generated low-bit discrete amplitude sequence Q ( x This is the output of the feature quantization encoder; Configure a pulse coding strategy for the quantization-to-time mapper, the pulse coding strategy being used to convert discrete amplitude information into pulse events in the time domain, wherein the pulse coding strategy includes level-frequency coding, level-delay coding, and threshold crossover coding; configure a dynamic threshold unit for the quantization-to-time mapper, through which a time-varying threshold θ is generated. t Wherein, the threshold that changes over time θ t The adjustment rules are one or more combinations of the following: changes in firing time, calculation based on statistics of the input discrete amplitude sequence within a local time window, and pulse coding strategy adaptive adjustment based on the state of neuron membrane potential or historical firing rate during threshold crossing coding; through the low-bit discrete amplitude sequence Q ( x Based on the dynamic threshold that changes over time θ t And the pulse coding strategy, to generate a pulse sequence. S ( t ); The sample tensor of the continuous real-value sensor data is input into the trained feature quantization encoder for feature quantization encoding to generate a low-bit discrete amplitude sequence. The low-bit discrete amplitude sequence is then converted into a pulse sequence in the time domain by the trained quantization-to-time mapper. The pulse sequence in the time domain is input into the trained prediction network to output the predicted value of the remaining life of the aero-engine.

2. The method for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping as described in claim 1, characterized in that, Acquire sensor data generated during the operation of the aero-engine, preprocess the sensor data and label it with remaining life tags, and generate preprocessed long-sequence multi-channel sensor data, including: For the sensor data of each sensor channel, wavelet transform algorithm and adaptive median filtering algorithm are used to smooth and denoise the data, and generate denoised sensor data. Calculate the mean and standard deviation of each sensor channel in the training dataset, and use the mean and standard deviation to perform Z-score normalization on the denoised sensor data in the training dataset to generate normalized sensor data. The end-of-life time of the aircraft engine is determined based on the equipment failure definition and / or maintenance threshold. For any moment during the operation of an aircraft engine t A piecewise linear decay model is used to evaluate the remaining lifetime of the tag. RUL ( t ) is used for calculation, where, , R max This is a preset maximum remaining lifespan limit based on the stated end-of-life time. t th This is the moment when degradation begins; The normalized sensor data is arranged in chronological order, and the remaining lifetime tag corresponding to each timestamp is... RUL ( t )and t The sensor data at different times are correlated to generate a long sequence of preprocessed multi-channel sensor data.

3. The method for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping as described in claim 1, characterized in that, The preprocessed long-sequence multi-channel sensor data is used to construct time-series samples to generate a sample tensor of continuous real-value sensor data, including: Set the parameters of the sliding window, including the window length L and the sliding step size S; Using the sliding window with the window length L and the sliding step size S, starting from the beginning time of the preprocessed long sequence multi-channel sensor data, sliding and truncating are performed in the time dimension. Each time, a continuous data segment of length L containing all M sensor channels is slid and truncated to form an independent window sample. The sliding and truncating operation is repeated until the entire preprocessed long sequence multi-channel sensor data is traversed, generating N window samples arranged in chronological order. The N window samples are stacked in the batch dimension to generate a sample tensor of three-dimensional continuous real-valued sensor data. ; Set the remaining lifetime label y corresponding to the i-th window sample. i = RUL ( ),in, RUL ( t )for t The remaining lifespan label at any given moment. The end time of the window; Collect the labels of all window samples to generate the remaining lifetime label vector Y corresponding to the sample tensor X; The sample data pair consisting of the sample tensor X and the remaining lifetime label vector Y is output as a continuous real-valued sensor data sample tensor.

4. The method for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping as described in claim 1, characterized in that, Construct a temporal deep model including a feature quantization encoder, a quantization-to-temporal mapper, and a prediction network, including: A feature quantization encoder is constructed by a learnable scaling factor and a low-bit quantization operator, and the feature quantization encoder outputs a low-bit discrete amplitude sequence, wherein the feature quantization encoder is used to map continuous inputs to discrete amplitudes; Construct the quantization-to-time mapper, input the low-bit discrete amplitude sequence into the quantization-to-time mapper, convert the low-bit discrete amplitude sequence into a sparse pulse sequence in the time domain and output it; A prediction network is constructed, and the sparse pulse sequence in the time domain is input into the prediction network to output the predicted value of the remaining life of the aero-engine.

5. The method for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping as described in claim 1, characterized in that, Through the low-bit discrete amplitude sequence Q ( x Based on the dynamic threshold that changes over time θ t And the pulse coding strategy, to generate a pulse sequence. S ( t ), include: pulse sequence S ( t Set the sequence to empty, initialize the time step counter, and initialize the neuron membrane potential state V. Traversing the low-bit discrete amplitude sequence Q ( x All time steps t and their corresponding discrete amplitudes in ) q t This process continues until all input data has been processed, generating the final pulse sequence. S ( t ): If the pulse coding strategy is the level-frequency coding, calculate the current time step. t The number of pulses n generated by the internal response t =F( q t ), at time step t Within, generate n according to the preset pattern. t A pulse is generated, and the timestamp t of the pulse is recorded in the pulse sequence. S ( t In ), where F( q t F is a predefined amplitude-frequency mapping function, and F() is a non-decreasing function; If the pulse coding strategy is the level-delay coding, calculate the first pulse relative to the time step. t Delay in issuing the starting point τ t = D( q t ), at the point in time t + τ t A pulse is generated at the specified time point. t + τ t Recorded as a timestamp in the pulse sequence S ( t In ), where D( q t D is a predefined magnitude-delay mapping function, and D() is a non-increasing function; If the pulse coding strategy is the threshold crossing coding, the discrete amplitude will be... q t As an input current, it is accumulated to the neuron membrane potential V( t )=V( t- 1)+ q t Through the neuronal membrane potential V( t Determine whether a pulse is generated, and record the timestamp t of the pulse into the pulse sequence. S ( t In the process, after the pulse is fired, the neuron membrane potential V( t Reset to the preset reset potential, wherein the pulse sequence , For dynamic threshold θ t The upper threshold, For dynamic threshold θ t The lower threshold.

6. The method for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping as described in claim 1, characterized in that, Training the aforementioned time-series deep model to generate a trained time-series deep model includes: The temporal deep model is trained and iterated as follows until a preset convergence condition is met: The prediction loss for the current batch is calculated based on the loss function L, and the quantization operator of the time-series deep model is calculated using the prediction loss. Q i Operator gradient and learnable scaling factor a i The scaling factor gradient; According to the optimization algorithm, the learnable scaling factor is updated through the scaling factor gradient. a i ; For the quantization operator Q i Configure a gradient approximation mechanism and use the gradient approximation mechanism as the training strategy during end-to-end training of the temporal deep model; The parameters of the quantization-to-time mapper are optimized, including the coefficients of the dynamic threshold adjustment function, the mapping relationship between pulse firing rate and amplitude, and the membrane potential reset value.

7. The method for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping as described in claim 1, characterized in that, Constructing a prediction network includes: A feature extraction layer is constructed for the prediction network. The feature extraction layer is used to extract temporal features related to engine degradation from the input pulse sequence. The feature extraction layer includes a spiking neural network layer, a recurrent neural network layer, and an attention network layer. The spiking neural network layer is used to perform asynchronous event-driven processing on the pulse events. The recurrent neural network layer is used to convert the pulse sequence into a dense vector and then input it into the recurrent neural network layer of the long short-term memory network. The attention network layer is used to convert the pulse sequence into a binary or rate-coded form. A regression output layer of the prediction network is constructed, which is used to map the extracted high-dimensional features into a continuous scalar value as a prediction of the remaining life of the aero-engine.

8. A device for predicting the remaining life of an aero-engine based on feature quantization encoding and quantization pulse mapping, characterized in that, include: The preprocessing module is used to acquire sensor data generated by the aero-engine during operation, preprocess the sensor data and construct a remaining life tag, and generate preprocessed long-sequence multi-channel sensor data, wherein the sensor data is multi-channel and continuous time series data. A sample tensor construction module is used to construct time-series samples from the preprocessed long-sequence multi-channel sensor data to generate a sample tensor of continuous real-value sensor data. The model training module is used to construct a temporal deep model including a feature quantization encoder, a quantization-to-temporal mapper, and a prediction network, and to train the temporal deep model to generate a trained temporal deep model. The model training module is also used to configure the learnable scaling factor of the feature quantization encoder. a i And configure the bit width of the low-bit quantization operator. b The learnable scaling factor is used to adaptively adjust the dynamic range of the input features. i For sensor channel index, b The sample tensor of the continuous real-valued sensor data is a positive integer less than 8. x As input, the sample tensor x The first in i Each input of each channel x i Computational quantization operator ,in, i For sensor channel index, a i For the first i Each channel corresponds to a learnable scaling factor, `round()` is the rounding function, and `clip()` is the truncation function; based on the quantization operator... Q i Generated low-bit discrete amplitude sequence Q ( x This is the output of the feature quantization encoder; The model training module is further configured to configure a pulse coding strategy for the quantization-to-time mapper, wherein the pulse coding strategy is used to convert discrete amplitude information into pulse events in the time domain, and wherein the pulse coding strategy includes level-frequency coding, level-delay coding, and threshold crossing coding; and to configure a dynamic threshold unit for the quantization-to-time mapper, through which a threshold θ that varies with time is generated. t Wherein, the threshold that changes over time θ t The adjustment rules are one or more combinations of the following: changes in firing time, calculation based on statistics of the input discrete amplitude sequence within a local time window, and pulse coding strategy adaptive adjustment based on the state of neuron membrane potential or historical firing rate during threshold crossing coding; through the low-bit discrete amplitude sequence Q ( x Based on the dynamic threshold that changes over time θ t And the pulse coding strategy, to generate a pulse sequence. S ( t ); The prediction output module is used to input the sample tensor of the continuous real-value sensor data into the trained feature quantization encoder for feature quantization encoding to generate a low-bit discrete amplitude sequence, and convert the low-bit discrete amplitude sequence into a pulse sequence in the time domain through the trained quantization-to-time mapper. The pulse sequence in the time domain is then input into the trained prediction network to output the predicted value of the remaining life of the aero-engine.

Images