Aero-engine remaining useful life prediction method and system based on stable anchoring soft working condition residual memory

Abstract

A kind of aero-engine remaining useful life prediction method and system based on stable anchoring soft working condition residual memory, only utilize sensor sequence and operating condition visible up to current time in each prediction window, based on the soft belonging probability of training side clustering center calculation current working condition;Respectively with the first time scale, build faster soft working condition memory, with slower second time scale, build stable working condition anchor point, and form short-term working condition residual with the difference between the two;Then according to the novelty of working condition, memory entropy, distribution change, volatility, state switching rate and residence time form reliability and stability auxiliary quantity.Sensor sequence obtains degradation representation by time series encoder, basic soft working condition branch, stable working condition anchor point branch, short-term working condition residual branch and auxiliary quantity branch are fused after being projected independently, and the degradation representation is fused, and the remaining useful life is output;The application provides double time scale working condition representation for the adaptability of life prediction under dynamic working condition and tail fluctuation control.

Description

Technical Field

[0001] This invention relates to the fields of aero-engine prediction and health management, complex operating condition time series modeling and intelligent life prediction technology, specifically to a method and system for predicting the remaining service life of aero-engines based on stable anchored soft operating condition residual memory. Background Technology

[0002] Aero-engines operate under varying flight missions, environmental conditions, and thrust requirements, and degradation information in sensor sequences is highly coupled with operational conditions. The same health state may exhibit different sensor responses under different operational conditions, and similar sensor changes may correspond to different degradation meanings in different historical operational contexts. Therefore, temporal encoding only a fixed-length sensor window can easily lead to misinterpreting operational condition changes as degradation; completely ignoring historical operational conditions will result in the loss of information related to long-term load exposure.

[0003] In multi-condition lifetime prediction, if only historical condition statistics and time-series coding results are fused at the readout end, and only a single recursive time scale is used, it is difficult to simultaneously characterize the long-term stable condition structure and recent condition shifts. When the context memory is too sensitive to short-term changes, the prediction is prone to tail spikes; when the memory is too smooth, it will mask recent condition changes that are valuable for lifetime prediction. If the condition vector is directly used for strong gating or deep modulation, unreliable condition signals may be amplified when the condition reliability changes over time, leading to a decrease in prediction accuracy and stability.

[0004] Furthermore, if the operating condition clustering, standardization, or reliability features utilize future windows or complete trajectory statistics of the test engine, data leakage will occur; if future data is used for smoothing or curve repair after the model output, it will also fail to meet the requirements of online prediction. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention aims to provide a method and system for predicting the remaining service life of aero-engines based on the residual memory of stable anchored soft operating conditions. Under strict prefix visibility conditions, the long-term stable operating conditions, short-term operating condition offsets, and contextual reliability are explicitly separated and can be used independently by the model. The structure is clear, the prefix is ​​safe, and it can be modularly integrated into the time series prediction model.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting the remaining service life of an aero-engine based on the residual memory of stable anchored soft operating conditions includes the following steps: S1, acquire the sensor sequence and operating condition sequence of the aero-engine in the current prediction window and its historical windows, wherein neither the sensor sequence nor the operating condition sequence contains future data after the current prediction window; S2, calculate the soft attribution probability vector of the current operating condition to the K operating condition cluster centers based on the K operating condition cluster centers obtained by fitting only the operating conditions of the training engine; S3, perform a first causal recursion on the soft attribution probability vector to obtain a first time-scale soft operating condition memory; perform a second causal recursion on the soft attribution probability vector with a time scale slower than the first causal recursion to obtain a slow stable operating condition anchor point; calculate the difference between the first time-scale soft operating condition memory and the slow stable operating condition anchor point to obtain the short-term operating condition residual. S4. Based on the current and historical operating conditions, calculate the operating condition contribution reliability (characterizing the reliability of the operating conditions), the operating condition stability reliability (characterizing the smoothness of the changes in the operating conditions), and the amplitude of the short-term operating condition residuals to form reliability and dynamic auxiliary quantities. S5, combine the first time-scale soft working condition memory and its statistics, slow stable working condition anchor point, short-term working condition residual and reliability with dynamic auxiliary quantity to form a composite working condition feature. S6, input the sensor sequence into the time encoder to obtain the degradation time series characterization; extract the basic soft condition component, slow stable condition anchor point component, short-term condition residual component, and reliability and dynamic auxiliary component from the composite operating condition features, and set the basic soft condition branch, slow stable condition anchor point branch, short-term condition residual branch, and reliability and dynamic auxiliary component branch. The basic soft condition branch includes the first time scale soft condition memory and its statistics; project the four branches into the embedding space, and perform late fusion of each branch with the degradation time series characterization; S7, input the fused representation into the remaining lifetime prediction head, output the non-negative remaining lifetime corresponding to the current prediction window, and save the first time scale soft condition memory and slow stable condition anchor point for the next prediction window recursion; the reliability and dynamic auxiliary quantity are independently projected as parallel feature inputs, not used to perform multiplication gating on the four branches, nor used to modulate the write rate of the first causal recursion or the second causal recursion.

[0007] Step S2 includes: calculating the current operating condition vector. Clustering center with the kth working condition Distance between And based on temperature parameters The soft attribution probability is obtained using the following formula: ; Where t represents the time index of the current prediction window. This represents the standardized operating condition vector of the current window. This represents the k-th cluster center obtained by fitting the operating conditions of the training engine. express and The distance between them This represents the soft-attribution temperature parameter determined by the training engine. This represents the soft-attribution probability that the current operating condition belongs to the k-th operating condition, where j represents the summation index and K represents the total number of operating condition cluster centers; The operating condition cluster center, standardized operating condition parameters, and temperature parameters are all determined from the training engine data. The standardized operating condition parameters include the training mean and training standard deviation of each operating condition variable, and are frozen before model deployment. They are used to perform the same standardization process on the current and historical operating conditions of the engine to be predicted.

[0008] The first causal recursion and the second causal recursion in step S3 are respectively: , , , Where t represents the time index of the current prediction window. This represents the current soft attribution probability vector. This represents the soft operating condition memory at the first time scale. Indicates anchor points for slow, stable operating conditions. Indicates short-term residuals. This represents the smoothing coefficient of the first causal recursion. This represents the smoothing coefficient of the second causal recursion, and 0 ≤ < <1.

[0009] The reliability of the operating condition contribution in step S4 is obtained by weighting the operating condition novelty and the normalized entropy of the soft operating condition memory at the first time scale; the operating condition novelty is obtained by normalizing the distance between the current operating condition and the nearest operating condition cluster center relative to the distance quantile on the training side.

[0010] The operating condition stability reliability in step S4 is obtained by weighting at least two of the following: the Jensen-Shannon divergence between the current soft-attribute probability and the recent soft-attribute probability statistics, the recent operating condition distribution volatility, the operating condition state switching rate, and the current operating condition residence time; the operating condition contribution reliability, operating condition stability reliability, and short-term operating condition residual norm are all limited to a preset value range.

[0011] In step S5, the composite operating condition features sequentially include: K-dimensional first-time-scale soft operating condition memory; a five-dimensional statistic composed of soft-attribution probability entropy, first-time-scale soft operating condition memory entropy, maximum soft-attribution probability, the difference between the maximum and second-largest soft-attribution probabilities, and operating condition novelty; K-dimensional slow-stable operating condition anchor points; K-dimensional short-term operating condition residuals; and a three-dimensional auxiliary quantity composed of operating condition contribution reliability, operating condition stability reliability, and the normalized norm of short-term operating condition residuals; these are used by the four projection branches according to their corresponding semantic components, making the dimension of the composite operating condition features 3K+8. K is set to 32, and the dimension of the composite operating condition features is 104; the smoothing coefficient α of the first causal recursion is set to 0.9998, and the smoothing coefficient of the second causal recursion... Take 0.99995.

[0012] The temporal encoder in step S6 is a Transformer encoder, which maps the sensor sequence to a token representation. The degraded temporal representation is obtained after passing through at least one self-attention coding block and temporal pooling. The four operating condition branches are connected to the degraded temporal representation after passing through a linear layer, a nonlinear activation layer and a random deactivation layer, and then input into the multilayer perceptron prediction head.

[0013] Training, validation, and testing are divided according to engine units; the future window, complete trajectory statistics, and remaining service life labels of the test engine are not involved in cluster center fitting, standardized parameter calculation, reliability calibration, model selection, and online prediction, and the output results are not smoothed, monotonicized, or replaced by posterior curves using future windows.

[0014] A system for predicting the remaining service life of an aero-engine based on stable anchored soft condition residual memory includes: a data acquisition module, a soft condition coding module, a dual-timescale condition memory module, a reliability calculation module, a temporal degradation coding module, a branch projection fusion module, and a remaining service life output module; each module is configured to execute a method for predicting the remaining service life of an aero-engine based on stable anchored soft condition residual memory.

[0015] An electronic device includes a processor, a memory, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement a method for predicting the remaining service life of an aero-engine based on stable anchored soft condition residual memory.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for predicting the remaining service life of an aero-engine based on the residual memory of stable anchored soft operating conditions.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: (1) Using soft attribution probability to represent the working condition can express the transition relationship between multiple typical states of the working condition, avoiding information abrupt changes caused by hard clustering; (2) By using causal contextual memory at two different time scales, the long-term stable operating conditions and recent operating condition changes can be explicitly distinguished; (3) Treating short-term residuals as an independent branch allows the model to utilize directional changes relative to stable anchor points without having to rewrite or suppress all historical memories. (4) By using the reliability contribution of operating conditions and the reliability of stable operating conditions as auxiliary quantities, the prediction head can identify novel, unstable or frequently changing operating condition ranges; (5) Each context component is projected independently to preserve different temporal semantics and reduce instability caused by premature mixing and strong gating; (6) All context and reliability quantities are calculated from current and historical prefixes, without the need for future windows, complete test trajectory statistics and post-prediction smoothing, making it suitable for online lifetime prediction; (7) This method can be combined with Transformer, temporal convolutional network, recurrent neural network and state space model, and can be extended to gas turbine, wind power equipment and other mechanical systems with significant changes in operating conditions. Attached Figure Description

[0018] Figure 1 This is a general flowchart of the method in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of dual-timescale soft condition memory and short-term residual generation in an embodiment of the present invention.

[0020] Figure 3 A schematic diagram illustrating the calculation of reliability contribution under operating conditions and reliability stability under operating conditions in embodiments of the present invention.

[0021] Figure 4 This is a structural diagram of the branch projection fusion and remaining service life prediction system according to an embodiment of the present invention.

[0022] Figure 5 This is a grouped bar chart comparing the overall root mean square error and mean absolute error of the method in this embodiment of the invention with the single-timescale soft working condition control method under three groups of fixed random seeds.

[0023] Figure 6 This is a comparison chart of the actual remaining service life and the predicted remaining service life trajectory on two predefined reserve test units, comparing the method of this invention with the single time-scale soft operating condition comparison method. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to embodiments and accompanying drawings. It should be understood that the following parameters are used to describe an executable embodiment, and those skilled in the art can adjust them according to the sampling frequency, engine type, and operating conditions.

[0025] like Figure 1 As shown, a method for predicting the remaining service life of an aero-engine based on the residual memory of stable anchored soft operating conditions includes the following steps: S1, acquire the sensor sequence and operating condition sequence of the aero-engine in the current prediction window and its historical windows, wherein neither the sensor sequence nor the operating condition sequence contains future data after the current prediction window; This embodiment uses full-life or truncated-life sequences of multiple aero-engine units, with each sample containing a sensor window of length L. and the operating condition data visible up to the end of the window. Sensor dimension D can include variables related to temperature, pressure, rotational speed, flow rate, and efficiency; operating conditions can include variables such as altitude, Mach number, thrust command, ambient temperature, and flight status. S2, calculate the soft attribution probability vector of the current operating condition to the K operating condition cluster centers based on the K operating condition cluster centers obtained by fitting only the operating conditions of the training engine; Calculate the current operating condition vector Clustering center with the kth working condition Distance between And based on temperature parameters The soft attribution probability is obtained using the following formula: ; Where t represents the time index of the current prediction window. This represents the standardized operating condition vector of the current window. This represents the k-th cluster center obtained by fitting the operating conditions of the training engine. express and The distance between them This represents the soft-attribution temperature parameter determined by the training engine. This represents the soft-attribution probability that the current operating condition belongs to the k-th operating condition, where j represents the summation index and K represents the total number of operating condition cluster centers; The operating condition cluster center, standardized operating condition parameters, and temperature parameters are all determined from the training engine data. The standardized operating condition parameters include the training mean and training standard deviation of each operating condition variable, and are frozen before model deployment. They are used to perform the same standardization process on the current and historical operating conditions of the engine to be predicted.

[0026] This embodiment uses only the operating condition samples of the training engine to fit K operating condition cluster centers. The clustering method can be K-means, Gaussian mixture model, prototype network or other unsupervised methods that can output centers or prototypes. The standardized mean, standard deviation, distance quantile and temperature parameters are also calculated only from the training data. For the current working condition vector First, calculate the squared distance from each cluster center: The soft-attribution probability is then obtained from the squared distance: .

[0027] Soft attribution probability vector Each element of the is non-negative and sums to 1, which can describe the degree of similarity between the current operating condition and multiple typical operating conditions.

[0028] S3, as Figure 2 As shown, a first causal recursion is performed on the soft attribution probability vector to obtain a first time-scale soft operating condition memory; a second causal recursion is performed on the soft attribution probability vector with a time scale slower than the first causal recursion to obtain a slow stable operating condition anchor point; the difference between the first time-scale soft operating condition memory and the slow stable operating condition anchor point is calculated to obtain the short-term operating condition residual. In this embodiment, the first time-scale soft operating condition memory uses a relatively fast first exponential moving average: ; The anchor point for slow, stable operating conditions uses a slower second exponential moving average: ; The short-term residual is: ; Where t represents the time index of the current prediction window. This represents the current soft attribution probability vector. This represents the soft operating condition memory at the first time scale. Indicates anchor points for slow, stable operating conditions. Indicates short-term residuals. This represents the smoothing coefficient of the first causal recursion. This represents the smoothing coefficient of the second causal recursion, and 0 ≤ < <1.

[0029] When the operating conditions are stable over a long period of time and Towards consensus Approaching zero; when the recent operating condition distribution shifts relative to the long-term background. Prior to change, Record the direction and magnitude of the changes. The slow, stable operating condition anchor does not replace the adaptive memory, and the short-term operating condition residual does not multiply the adaptive memory. The three are provided to the prediction model as parallel information with different temporal semantics.

[0030] In this embodiment, , , The above coefficients correspond to the current data sampling and window step size; when the sampling frequency is lower or the operating conditions change faster, the corresponding coefficients can be reduced; the recursive initial value can be set as the first soft-attribution probability, uniform distribution, or initialized by the average operating condition distribution on the training side.

[0031] S4, as Figure 3 As shown, based on the current and historical operating conditions, the reliability of the operating condition contribution, which characterizes the credibility of the operating conditions, the reliability of the operating condition stability, which characterizes the smoothness of the operating condition changes, and the amplitude of the short-term operating condition residuals are calculated to form reliability and dynamic auxiliary quantities. The reliability of the operating condition contribution is obtained by weighting the operating condition novelty and the normalized entropy of the soft operating condition memory at the first time scale; the operating condition novelty is obtained by normalizing the distance between the current operating condition and the nearest operating condition cluster center relative to the distance quantile on the training side; The operating condition stability reliability is obtained by weighting at least two of the following: the Jensen-Shannon divergence between the current soft-attribute probability and the recent soft-attribute probability statistics, the recent operating condition distribution volatility, the operating condition state switching rate, and the current operating condition residence time; the operating condition contribution reliability, operating condition stability reliability, and short-term operating condition residual norm are all limited to a preset value range.

[0032] Novelty of the operating conditions in this embodiment Minimum cluster distance relative to the training side median and high quantile Normalize: ; in, Indicates the novelty of the current operating condition. This represents the minimum distance from the current operating condition to the cluster centers of each operating condition. This represents the median of the minimum clustering distance for the training engines. This represents the preset high quantile of the minimum clustering distance for training engines. Indicates the upper limit of novelty, (x,a,b) represents a function that truncates the independent variable x to the interval [a,b].

[0033] Memory entropy Reflecting the dispersion of soft operating condition memory at the first time scale, the operating condition contributes to reliability. It can be composed of 1- With 1- The weighted combination is obtained; when the novelty is higher or the memory is more dispersed, the reliability contribution of the operating condition decreases accordingly. Operating condition stability and reliability The parameters are derived from at least two of the following: the Jensen-Shannon divergence between the current soft-attribution probability and the average probability over the most recent M windows; the probability volatility over the most recent N windows; the hard-condition label switching rate; and the dwell time of the current dominant condition. Higher divergence, volatility, and switching rate generally indicate lower condition stability reliability; longer dwell time generally indicates higher condition stability reliability. Each parameter uses the training-side quantiles. and Perform calibration; for signals where lower values ​​indicate higher reliability, the following method is used: ; in, This indicates the normalized reliability of the index ξ, where lower values ​​indicate higher reliability. ξ represents the value of the index to be calibrated. and These represent the 20th and 80th percentiles of the metric on the training engine, respectively. (x,0,1) represents a function that truncates the independent variable x to the interval [0,1].

[0034] For higher dwell times, indicating greater reliability, the following approach is adopted: ; in, Indicators that indicate higher values ​​are more reliable The corresponding normalized reliability; where , , and The meaning of (x,0,1) is the same as described above; when calibrating the dwell time... Dwell time under the current dominant operating condition .

[0035] In the embodiment: , , in, Indicates the reliability contribution under operating conditions. Indicates stable and reliable operating conditions. and These represent the normalized reliability of the low-value reliability index and the high-value reliability index, respectively. This represents the normalized entropy of the soft condition memory at the first timescale. , , and These represent recent distribution deviation, probability volatility, operating condition switching rate, and current dominant operating condition dwell time, respectively; each numerical coefficient represents a fixed combination weight determined by the training engine in this embodiment.

[0036] The recent windows can be 5 and 20 prefix windows respectively; Further calculation of the short-term residual norm: ; in, The normalized norm of the short-term residuals is represented by the normalized norm. Represents short-term residuals The 2-norm, (x,0,1) represents a function that truncates the independent variable x to the interval [0,1].

[0037] In the embodiment, the reliability auxiliary quantity is [ , , It can also increase the product of the two, memory entropy, operating condition switching confidence, or other quantities obtained solely from the prefix.

[0038] S5, combine the first time-scale soft working condition memory and its statistics, slow stable working condition anchor point, short-term working condition residual and reliability with dynamic auxiliary quantity to form a composite working condition feature. The composite operating condition features sequentially include: K-dimensional first-timescale soft operating condition memory; a five-dimensional statistic composed of soft-attribution probability entropy, first-timescale soft operating condition memory entropy, maximum soft-attribution probability, the difference between the maximum and second-largest soft-attribution probabilities, and operating condition novelty; K-dimensional slow-stable operating condition anchor points; K-dimensional short-term operating condition residuals; and a three-dimensional auxiliary quantity composed of operating condition contribution reliability, operating condition stability reliability, and the normalized norm of short-term operating condition residuals; and are used by the four projection branches according to their corresponding semantic components, making the dimension of the composite operating condition features 3K+8; when K is 32, the dimension of the composite operating condition features is 104; the smoothing coefficient α of the first causal recursion is 0.9998, and the smoothing coefficient of the second causal recursion... Take 0.99995; The basic soft condition features in this embodiment consist of K-dimensional first-time-scale soft condition memory and five statistics: the current soft assignment probability entropy, the first-time-scale soft condition memory entropy, the maximum soft assignment probability, the difference between the maximum and second-largest soft assignment probabilities, and the condition novelty. The composite condition features are: , in, This indicates the composite operating condition characteristics of the current prediction window. This represents a five-dimensional statistic; therefore The dimension is hour, It has 104 dimensions.

[0039] S6, input the sensor sequence into the time encoder to obtain the degradation time series characterization; extract the basic soft condition component, slow stable condition anchor point component, short-term condition residual component, and reliability and dynamic auxiliary component from the composite operating condition features, and set the basic soft condition branch, slow stable condition anchor point branch, short-term condition residual branch, and reliability and dynamic auxiliary component branch. The basic soft condition branch includes the first time scale soft condition memory and its statistics; project the four branches into the embedding space, and perform late fusion of each branch with the degradation time series characterization; The timing encoder is a Transformer encoder that maps the sensor sequence into a token representation. The degraded timing representation is obtained after passing through at least one self-attention coding block and timing pooling. The four operating condition branches are connected to the degraded timing representation after passing through a linear layer, a nonlinear activation layer and a random deactivation layer, respectively, and then input into the multilayer perceptron prediction head.

[0040] To avoid premature mixing of features with different semantics, this embodiment extracts basic soft operating conditions, slow stable operating condition anchor points, short-term operating condition residuals, and reliability and dynamic auxiliary quantities from the composite operating condition features according to their semantic components, and sets up four independent projection branches: ; Each φ can consist of a linear layer, a GELU activation layer, and a random deactivation layer; the four embedding dimensions can be the same or different; this structure allows the prediction head to learn the effects of basic soft conditions, long-term stable background, recent offset, and reliability status on lifetime prediction; the four information channels are only spliced ​​together late at the read end, and the reliability auxiliary quantity is not multiplied by any condition branch, nor is it adjusted for the write coefficients of the two exponential moving averages.

[0041] S7, input the fused representation into the remaining lifetime prediction head, output the non-negative remaining lifetime corresponding to the current prediction window, and save the first time scale soft condition memory and slow stable condition anchor point for the next prediction window recursion; the reliability and dynamic auxiliary quantity are independently projected as parallel feature inputs, not used to perform multiplication gating on the four branches, nor used to modulate the write rate of the first causal recursion or the second causal recursion.

[0042] Sensor window in this embodiment First, the token is transposed into a token sequence of length L. Then, a d-dimensional token embedding is obtained through a linear mapping of the input, followed by several Transformer encoding blocks: ; in, Indicates the current sensor window, This represents the mapping of a timing encoder with parameter θ. θ represents the token representation of the sensor window output by the timing encoder, and θ represents the encoder parameters learned from the training engine data and frozen before deployment.

[0043] right Perform adaptive average pooling or attention pooling along the time dimension to obtain a degenerate temporal representation. .Will Connected to four context embeddings: ; in, Indicates fusion representation, This represents the degenerate temporal representation obtained through temporal pooling. , , and These represent the projection embeddings corresponding to the basic soft condition component, the slow stable condition anchor point component, the short-term condition residual component, and the reliability and dynamic auxiliary component, respectively.

[0044] Fusion characterization Inputting the multilayer perceptron and the non-negative output layer yields the current predicted remaining lifetime: in, This represents the non-negative remaining useful life prediction value for the current prediction window. This represents the prediction head of a multilayer perceptron. This represents the fusion representation, and max(0,·) represents a function that restricts the predicted values ​​to non-negative values.

[0045] Sensor dimensions in this embodiment Window length The temporal encoder contains 4 Transformer coding blocks, a hidden dimension of 128, 8 attention heads, and a 256-dimensional feedforward layer; the embedding dimension of the four context branches is 16, and the hidden layer dimensions of the prediction head are 128 and 64.

[0046] Training loss can be root mean square error, mean square error, Huber loss, or a combination of losses weighted by late errors. Training and validation units are used for model training, early stopping, and checkpoint selection; test engine labels are not used.

[0047] In this embodiment, training, validation, and test data are divided according to engine units. Window formation, soft attribution of operating conditions, two context memories, and reliability quantities are all recursively calculated in chronological order. Input at any given time only includes data up to the end of the current window. The mean of the complete trajectory of the test unit, the distribution of future operating conditions, the actual termination time, and the remaining future service life are not included in the current prediction. The model output is not smoothed, monotonicized, or curve-patched using future windows.

[0048] In an offline validation embodiment using N-CMAPSS with fixed data partitioning, a fixed training protocol, and a fixed random seed, nine data subsets are used, with training, validation, and testing comprising 51, 9, and 39 engine unit paths, respectively; the window shape is 20×50. The overall root mean square error (RMSE) of this embodiment is 7.6876, and the mean absolute error (MAO) is 5.5542. Compared with the single-timescale soft-condition baseline, the RMSE is 8.6076 and the MAO is 6.0315. Among the 39 test units, the RMSE of 26 units decreased, the maximum tail jump of 24 units decreased, and the total number of peaks decreased by 625. In the case of additional random seeds, the overall RMSE is still lower than the baseline with the same seed, but the MAO and the total number of peaks do not decrease under all random seeds. Therefore, the above results are only used to illustrate the feasibility and operational boundaries of the method, and do not indicate that all units, random seeds, or tail stability indicators will necessarily improve, nor do they constitute a numerical limitation on the scope of protection.

[0049] like Figure 5 As shown, under the same N-CMAPSS fixed data partitioning and training protocol, compared with the single-timescale soft condition control method, the overall root mean square error of random seeds 20260616, 20260617, and 20260618 decreased from 8.6076, 8.4552, and 8.5952 to 7.6876, 8.1830, and 7.8606, respectively, with improvements observed in all three random seed groups. The mean absolute error changed from 6.0315, 5.9400, and 6.1200 to 5.5542, 6.0242, and 5.8863, with improvements observed in 20260616 and 20260618, but not in 20260617. The above results are only used to illustrate the change in overall error under a fixed protocol and do not necessarily indicate that all random seeds and all evaluation indicators will improve.

[0050] like Figure 6As shown, under random seed 20260616, the predefined DS08a-009 Unit12 tail stability case and DS03-012 Unit12 overall error case were selected, and the complete predicted trajectories of the actual remaining useful life, the single-timescale soft condition control method, and the structure of this embodiment were plotted respectively. In the former, the root mean square error change of this embodiment compared to the control method was -3.7640 h, and in the latter, it was -9.4268 h. Each curve in the figure is connected in the time sequence of all original prediction windows without smoothing, interpolation, downsampling, curve repair, or posterior processing using future windows; the shaded area only represents the evaluation tail segment defined after the prediction is completed and is not used as model input. The units shown are determined according to predefined selection rules and are used to visually illustrate the fit between the predicted trajectory and the actual trajectory, and do not indicate that all test units have obtained the same improvement.

[0051] like Figure 4 As shown, an aero-engine remaining service life prediction system based on stable anchored soft condition residual memory includes a data acquisition module, a soft condition encoding module, a dual-timescale context memory module, a reliability calculation module, a temporal degradation encoding module, a branch projection fusion module, and a remaining service life output module. The dual-timescale context memory only needs to store the data from the previous time step. and It also features a finite-length recent statistics cache, suitable for online recursion. The system can be deployed on ground servers, test benches, or edge devices. When resources are limited, the K value can be reduced or a lightweight timing encoder can be used. When migrating to other mechanical equipment, the operating condition and monitoring data are replaced, while retaining soft operating condition attribution, dual time-scale anchoring, and residual fusion mechanisms.

Claims

1. A method for predicting the remaining service life of an aero-engine based on the residual memory of stable anchored soft operating conditions, characterized in that, Includes the following steps: S1, acquire the sensor sequence and operating condition sequence of the aero-engine in the current prediction window and its historical windows, wherein neither the sensor sequence nor the operating condition sequence contains future data after the current prediction window; S2, calculate the soft attribution probability vector of the current operating condition to the K operating condition cluster centers based on the K operating condition cluster centers obtained by fitting only the operating conditions of the training engine; S3, perform a first causal recursion on the soft attribution probability vector to obtain the first time-scale soft operating condition memory; A second causal recursion with a time scale slower than the first causal recursion is performed on the soft attribution probability vector to obtain the slow stable operating condition anchor point. The difference between the soft operating condition memory at the first time scale and the slow stable operating condition anchor point is calculated to obtain the short-term operating condition residual. S4. Based on the current and historical operating conditions, calculate the operating condition contribution reliability (characterizing the reliability of the operating conditions), the operating condition stability reliability (characterizing the smoothness of the changes in the operating conditions), and the amplitude of the short-term operating condition residuals to form reliability and dynamic auxiliary quantities. S5, combine the first time-scale soft working condition memory and its statistics, slow stable working condition anchor point, short-term working condition residual and reliability with dynamic auxiliary quantity to form a composite working condition feature. S6, input the sensor sequence into the time encoder to obtain the degradation time series characterization; extract the basic soft condition component, slow stable condition anchor point component, short-term condition residual component, and reliability and dynamic auxiliary component from the composite operating condition features, and set the basic soft condition branch, slow stable condition anchor point branch, short-term condition residual branch, and reliability and dynamic auxiliary component branch. The basic soft condition branch includes the first time scale soft condition memory and its statistics; project the four branches into the embedding space, and perform late fusion of each branch with the degradation time series characterization; S7, input the fused representation into the remaining lifetime prediction head, output the non-negative remaining lifetime corresponding to the current prediction window, and save the first time scale soft condition memory and slow stable condition anchor point for the next prediction window recursion; the reliability and dynamic auxiliary quantity are independently projected as parallel feature inputs, not used to perform multiplication gating on the four branches, nor used to modulate the write rate of the first causal recursion or the second causal recursion.

2. The method according to claim 1, characterized in that, Step S2 includes: calculating the current operating condition vector. Clustering center with the kth working condition Distance between And based on temperature parameters The soft attribution probability is obtained using the following formula: ； Where t represents the time index of the current prediction window. This represents the soft-attribution temperature parameter determined by the training engine. This represents the soft-attribution probability that the current operating condition belongs to the k-th operating condition, where j represents the summation index and K represents the total number of operating condition cluster centers; The operating condition cluster center, standardized operating condition parameters, and temperature parameters are all determined from the training engine data. The standardized operating condition parameters include the training mean and training standard deviation of each operating condition variable, and are frozen before model deployment. They are used to perform the same standardization process on the current and historical operating conditions of the engine to be predicted.

3. The method according to claim 1, characterized in that, The first and second causal recursions in step S3 are as follows: ， ， ， Where t represents the time index of the current prediction window. This represents the current soft attribution probability vector. This represents the soft operating condition memory at the first time scale. Indicates anchor points for slow, stable operating conditions. Indicates short-term residuals. This represents the smoothing coefficient of the first causal recursion. This represents the smoothing coefficient of the second causal recursion, and 0 ≤ < <1.

4. The method according to claim 1, characterized in that, The reliability of the operating condition contribution in step S4 is obtained by weighting the operating condition novelty and the normalized entropy of the soft operating condition memory at the first time scale; the operating condition novelty is obtained by normalizing the distance between the current operating condition and the nearest operating condition cluster center relative to the distance quantile on the training side.

5. The method according to claim 1, characterized in that, The operating condition stability reliability in step S4 is obtained by weighting at least two of the following: the Jensen-Shannon divergence between the current soft-attribution probability and the recent soft-attribution probability statistics, the recent operating condition distribution volatility, the operating condition state switching rate, and the current operating condition residence time; the operating condition contribution reliability, operating condition stability reliability, and short-term operating condition residual norm are all limited to a preset value range.

6. The method according to claim 1, characterized in that, The composite operating condition features described in step S5 sequentially include: K-dimensional first-time-scale soft operating condition memory; a five-dimensional statistic composed of soft-attribution probability entropy, first-time-scale soft operating condition memory entropy, maximum soft-attribution probability, the difference between the maximum and second-largest soft-attribution probabilities, and operating condition novelty; K-dimensional slow-stable operating condition anchor points; K-dimensional short-term operating condition residuals; and a three-dimensional auxiliary quantity composed of operating condition contribution reliability, operating condition stability reliability, and the normalized norm of short-term operating condition residuals; which are then used by the four projection branches according to their corresponding semantic components, making the dimension of the composite operating condition features 3K+8; K is set to 32, and the dimension of the composite operating condition features is 104; the smoothing coefficient α of the first causal recursion is set to 0.9998, and the smoothing coefficient of the second causal recursion is set to... Take 0.99995.

7. The method according to claim 1, characterized in that, The temporal encoder in step S6 is a Transformer encoder, which maps the sensor sequence to a token representation. The degraded temporal representation is obtained after passing through at least one self-attention coding block and temporal pooling. The four operating condition branches are connected to the degraded temporal representation after passing through a linear layer, a nonlinear activation layer and a random deactivation layer, and then input into the multilayer perceptron prediction head.

8. A prediction system for the remaining service life of an aero-engine based on the residual memory of stable anchored soft operating conditions, characterized in that, include: The module includes a data acquisition module, a soft condition coding module, a dual time-scale condition memory module, a reliability calculation module, a time-series degradation coding module, a branch projection fusion module, and a remaining useful life output module. Each module is configured to perform the method of any one of claims 1 to 7.

9. An electronic device comprising a processor, a memory, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images