System and method for predicting residual life of aero-engine

Abstract

The invention discloses a residual life prediction system and method for an aero-engine, and belongs to the technical field of life prediction and the technical field of digital circuit design, and the system comprises a multi-modal data collection module, a data preprocessing module, a working mode arbitration module, a low-power-consumption health state evaluation module and a self-adaptive neural network processing module. The multi-modal data acquisition module is connected with the data preprocessing module, the data preprocessing module is connected with the working mode arbitration module, the working mode arbitration module is connected with the low-power-consumption health state evaluation module and the adaptive neural network processing module, and the low-power-consumption health state evaluation module is connected with the adaptive neural network processing module. According to the method, the calculation precision is adaptively adjusted according to the health condition of the aero-engine, so that the limited airborne calculation capability, memory and power consumption are dealt with.

Description

Technical Field

[0001] This application relates to the fields of life prediction technology and digital circuit design, and in particular to a system and method for predicting the remaining life of an aero-engine. Background Technology

[0002] As one of the most critical components of an aircraft, the health of an aero-engine directly impacts flight safety, operating costs, and mission reliability. Therefore, accurate prediction of the engine's remaining service life is crucial for efficient maintenance and preventing catastrophic accidents, and is also the core of modern aviation fault prediction and health management systems. Traditional remaining service life prediction methods often rely on physical models or single vibration signal analysis, which have limitations such as model complexity, weak generalization ability, and insensitivity to multiple failure modes. With the deep application of IoT and big data technologies in the aviation field, engines generate massive amounts of multimodal data during operation, including high and low pressure turbine outlet temperatures, high and low pressure turbine outlet pressures, fuel flow rates, and fan speeds. This data reflects the engine's health degradation process from different dimensions, providing a foundation for data-driven intelligent prediction.

[0003] However, while multimodal data can be used for remaining lifetime prediction, the emergence of massive amounts of sensor data and its inherent complexity present two core challenges to achieving accurate airborne prediction. The primary challenge lies in the effective fusion and value extraction of the data. The data collected by sensors throughout the engine exhibits significant heterogeneity in physical meaning and data format, forming a complex system that is high-dimensional, nonlinear, and with unclear coupling relationships. Extracting unified and robust health status indicators from this data is the primary challenge. Furthermore, intelligent remaining lifetime prediction systems must consider the resource constraints of airborne edge computing scenarios. The prediction system must be strictly constrained within limited airborne computing power, memory, and power consumption budgets, making it difficult to directly deploy many complex models that perform well in the cloud. Therefore, finding the optimal solution between accuracy and efficiency is key to achieving a high-precision, practical prediction system. Summary of the Invention

[0004] This application aims to provide a system and method for predicting the remaining life of an aero-engine. The system constructs a collaborative processing architecture that integrates data perception, processing, arbitration, and intelligent computing. After performing physical model threshold judgment on the preprocessed multimodal sensor data, it can adaptively adjust the calculation accuracy according to the health status of the aero-engine, thereby coping with limited onboard computing power, memory, and power consumption. Ultimately, it realizes closed-loop intelligent remaining life prediction of the aero-engine based on adaptive decision-making from multimodal sensor data.

[0005] To achieve the above objectives, the technical solution of this application is as follows: An aero-engine remaining life prediction system includes a multimodal data acquisition module, a data preprocessing module, a working mode arbitration module, a low-power health status assessment module, and an adaptive neural network processing module. The multimodal data acquisition module is connected to the data preprocessing module, the data preprocessing module is connected to the working mode arbitration module, the working mode arbitration module is connected to the low-power health status assessment module and the adaptive neural network processing module, and the low-power health status assessment module is connected to the adaptive neural network processing module. The working mode arbitration module is used to identify the operand features of the operands corresponding to the preprocessed regularized data, perform working mode matching based on the operands after feature identification, and output the working mode matching result signal. The low-power health status assessment module is used to set a health status threshold according to the health status assessment requirements during the system initialization phase, compare the preprocessed regularized data with the health status threshold, and output a health status indication signal. The adaptive neural network processing module is used to select neural network models of different precision based on health status indication signals, select corresponding working mode configuration parameters based on working mode matching result signals, perform neural network adaptive prediction calculations on preprocessed regularized data, and output prediction results of corresponding precision and working mode.

[0006] Optionally, the data preprocessing module includes: a data analysis and selection unit, a normalization unit, a sliding window unit, and a segmented degradation processing unit. The data analysis and selection unit, the normalization unit, the sliding window unit, and the segmented degradation processing unit are connected in sequence. The data analysis and selection unit is connected to the multimodal data acquisition module, and the segmented degradation processing unit is connected to the data preprocessing module. The segmented degradation processing unit is used to perform segmented degradation processing on the continuous time-series data segments input by the sliding window unit and output preprocessed regular data. The segmented degradation processing includes: calculating the mean and variance of the input continuous time-series data segments as degradation identification features characterizing the degradation of the aero-engine, and obtaining preprocessed regular data with degradation identification features.

[0007] Optionally, the working mode arbitration module includes: an operand decision module and a working mode matching module, with the operand decision module connected to the working mode matching module; the operand decision module is connected to the data preprocessing module, and the working mode matching module is connected to the low-power health status assessment module and the adaptive neural network processing module.

[0008] Optionally, the low-power health status assessment module includes: a physical model threshold comparison module and a health status threshold setting module, which are connected together; the physical model threshold comparison module is connected to the working mode arbitration module and the adaptive neural network processing module.

[0009] Optionally, the adaptive neural network processing module includes: a working mode configuration module, a neural network module, a memory module, and a two-to-one selector. The working mode configuration module is connected to the neural network module and the memory module, and the neural network module is connected to the memory module and the two-to-one selector. The working mode configuration module and the neural network module are connected to the working mode arbitration module, and the neural network module and the two-to-one selector are connected to the low-power health status assessment module. A two-to-one selector is used to gating and switching the intermediate results of neural network predictions output by the neural network module. The gating and switching includes: delivering data by outputting the prediction results with the corresponding accuracy and working mode according to the activated prediction mode.

[0010] Optionally, the neural network module includes a low-power quantization model and a high-precision computing model, which are respectively connected to the working mode configuration module, the memory module, and the two-to-one selector.

[0011] Optionally, the low-power quantization model is obtained by training an artificial intelligence neural network model with real flight datasets to obtain a floating-point precision model with multiple parameters, and then using fixed-point quantization technology to reduce the number of bits of the model parameters.

[0012] Optionally, the high-precision computing model is a floating-point precision model with multiple parameters obtained by training an artificial intelligence neural network model on a real flight dataset.

[0013] A method for predicting the remaining life of an aero-engine, applied to the aero-engine remaining life prediction system described above, includes, Step S1: Collect raw data from the multimodal sensor; Step S2: Data preprocessing; Step S3: Operand decision; Step S4: Determine if the working mode matches. If yes, proceed to step S5; otherwise, proceed to step S6. Step S5: Keep the configuration parameters and proceed to step S7; Step S6: Reload configuration parameters, proceed to step S7; Step S7: Assess whether the health status is healthy. If yes, proceed to step S8; otherwise, proceed to step S9. Step S8: Activate low power prediction mode and proceed to step S10; Step S9: Activate high-precision prediction mode and proceed to step S10; Step S10: Output the prediction results.

[0014] Optionally, determining whether the working mode matches includes: determining whether it is the same as the previous working mode based on the working mode matching decision signal; Assessing whether a health status is healthy includes: comparing the preprocessed regularized data with a health status threshold, determining whether the preprocessed regularized data is within the threshold range, and outputting a health status indication signal.

[0015] The aero-engine remaining life prediction system and method provided in this application can dynamically adjust the prediction strategy based on the preprocessing results of the collected multimodal sensor raw data, the aero-engine health status assessment conclusions, and the matching of operating modes, flexibly selecting either a low-power prediction mode or a high-precision prediction mode. When the aero-engine is in a healthy state and the system operating mode is matched, the system activates the low-power prediction mode, simplifying the prediction process through a low-power quantization model and reducing the overall system power consumption. Conversely, when the aero-engine is in a sub-healthy state, the system switches to the high-precision prediction mode, refining the prediction process using a high-precision calculation model with hybrid precision quantization, ensuring the accuracy of the remaining life prediction results, and providing reliable support for aero-engine fault early warning and maintenance decisions. This dynamic adaptation mechanism can achieve the optimal balance between prediction accuracy, system power consumption, and operational reliability under different aero-engine operating scenarios, meeting the remaining life prediction needs throughout the entire aero-engine lifecycle, optimizing energy consumption while ensuring the effectiveness of the prediction results, and contributing to the safe and stable operation of aero-engines.

[0016] To make the above-mentioned features and advantages of the application more apparent and understandable, specific embodiments are provided below, and detailed descriptions are given in conjunction with the accompanying drawings. Attached Figure Description

[0017] Figure 1 The structural diagram of the aircraft engine remaining life prediction system provided in this application.

[0018] Figure 2 This is a diagram of the multi-modal sensor data preprocessing module for aero-engines in this application.

[0019] Figure 3 This is a diagram of the arbitration module for the intelligent predictive operating mode of an aero-engine in this application.

[0020] Figure 4 This is a diagram of the low-power health status assessment module for aero-engines in this application.

[0021] Figure 5 This is a block diagram of the aero-engine adaptive neural network processor of this application.

[0022] Figure 6 This is a flowchart of the method for predicting the remaining life of an aero-engine according to this application. Detailed Implementation

[0023] To make the objectives and technical solutions of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0024] In one embodiment of this application, please refer to Figure 1 , Figure 1 The present application provides a block diagram of the remaining life prediction system for an aero-engine, which includes: a multimodal data acquisition module 1, a data preprocessing module 2, a working mode arbitration module 3, a low-power health status assessment module 4, and an adaptive neural network processing module 5.

[0025] The multimodal data acquisition module 1 is connected to the data preprocessing module 2. The data preprocessing module 2 is connected to the working mode arbitration module 3. The working mode arbitration module 3 is connected to the low-power health status assessment module 4 and the adaptive neural network processing module 5. The low-power health status assessment module 4 is connected to the adaptive neural network processing module 5.

[0026] As an example, the multimodal data acquisition module 1 is used to acquire and output raw multimodal sensor data from the multimodal sensors deployed on various parts of the aero-engine during stable operation, and to acquire and output operand data according to the external operating conditions of the aero-engine.

[0027] Specifically, the components of an aero-engine include fan blades, low-pressure compressor, high-pressure compressor, combustion chamber, low-pressure rotor, high-pressure rotor, high-pressure turbine, low-pressure turbine, and exhaust nozzle. Each component is equipped with multi-modal sensors such as temperature sensors, pressure sensors, fuel flow measurement sensors, and fan speed sensors to collect and output raw data. Simultaneously, units for external operating conditions are preset, and operands are collected and output based on these conditions. For example, if the collected external operating conditions include: flight altitude 8km, Mach number 0.8Ma, and throttle lever angle 100°, operands 8, 0.8, and 100 are output based on these conditions, providing a basis for subsequent operand decisions.

[0028] Data preprocessing module 2 is used to preprocess the acquired raw data from the multimodal sensors to obtain preprocessed regular data, making the acquired raw data from the multimodal sensors with different dimensions and amplitudes more regular, which is convenient for subsequent feature extraction of neural networks.

[0029] The working mode arbitration module 3 is used to identify the operand features of the operands corresponding to the preprocessed regularized data, perform working mode matching based on the operands after feature identification, and output the working mode matching result signal to provide a basis for subsequent parameter adjustment of the neural network model.

[0030] The low-power health status assessment module 4 is used to set the health status threshold according to the health status assessment requirements in the system initialization phase, compare the preprocessed regularized data with the health status threshold, output the health status indication signal, and provide the basis for selecting the neural network model in the subsequent neural network model prediction stage.

[0031] The adaptive neural network processing module 5 is used to select neural network models of different precision based on health status indication signals, select corresponding working mode configuration parameters based on working mode matching result signals, perform neural network adaptive prediction calculations on preprocessed regular data, and output prediction results with corresponding precision and working mode; the prediction mode of the system is dynamically switched through the physical model threshold method, thereby achieving a balance between prediction accuracy and computational efficiency.

[0032] The aero-engine remaining life prediction system provided in this application can dynamically adjust its prediction strategy based on the preprocessing results of raw data from multimodal sensors, the aero-engine health status assessment conclusions, and the matching of operating modes. It can flexibly select between a low-power prediction mode and a high-precision prediction mode. When the aero-engine is in a healthy state and the system's operating mode is matched, the system activates the low-power prediction mode, simplifying the prediction process through a low-power quantization model and reducing overall system power consumption. When the aero-engine is in a sub-healthy state, the system switches to the high-precision prediction mode, refining the prediction process using a high-precision calculation model with hybrid precision quantization, ensuring the accuracy of the remaining life prediction results, and providing reliable support for aero-engine fault early warning and maintenance decisions. This dynamic adaptation mechanism can achieve the optimal balance between prediction accuracy, system power consumption, and operational reliability under different aero-engine operating scenarios, meeting the remaining life prediction needs throughout the entire aero-engine lifecycle. It optimizes energy consumption while ensuring the effectiveness of prediction results, contributing to the safe and stable operation of aero-engines.

[0033] For example, please refer to Figure 2 , Figure 2This is a schematic diagram of the structure of the data preprocessing module 2 of this application. The data preprocessing module 2 includes: a data analysis and selection unit 21, a normalization unit 22, a sliding window unit 23, and a segmented degradation processing unit 24. The data analysis and selection unit 21, the normalization unit 22, the sliding window unit 23, and the segmented degradation processing unit 24 are connected in sequence. The data analysis and selection unit 21 is connected to the multimodal data acquisition module 1, and the segmented degradation processing unit 24 is connected to the data preprocessing module 2.

[0034] The data analysis and selection unit 21 is used to perform feature analysis and selection on the collected multimodal sensor raw data, and output different feature selection signals. The feature analysis and selection includes selecting a non-constant number of multimodal sensor raw data as feature selection signals. The non-constant number of multimodal sensor raw data reflects the difference in correlation between the multimodal sensor raw data and the engine degradation process. During the engine degradation process, the multimodal sensor raw data of the degraded parts will show a positive or negative correlation over time, while the sensor data that is not affected by degradation remains constant. Based on the difference in correlation between the multimodal sensor raw data and the engine degradation process, the subsequent normalization unit 22 is driven to perform differentiated focusing and processing on key health indicators.

[0035] Normalization unit 22 is used to normalize the input feature selection signal and output regularized data. The normalization process includes: eliminating dimensional deviations and unifying the value ranges based on the differences in the dimensional type and value range distribution of each feature selection signal, to obtain regularized data of each feature selection signal with eliminated dimensional deviations, which facilitates the unification of data ranges and the timing processing adaptation for subsequent sliding window unit 23; The sliding window unit 23 is used to perform sliding window processing on the input regularized data and output a continuous time-series data segment of fixed length. The sliding window processing includes: aligning the regularized data in time according to the temporal correlation and dynamic change frequency differences; extracting 30 or 50 timestamp data in the same time period during each prediction to obtain a continuous time-series data segment of fixed length, which serves as the input data for a prediction and supports the subsequent segmented degradation processing unit 24 for degradation feature extraction; wherein, due to the temporal correlation and dynamic change frequency differences of the data, the sampling frequency of each sensor data in the same time period is not necessarily completely synchronized and uniform, so the sensor with the lowest sampling frequency is used as a single timestamp reference; since the time period is very short, the number of operands corresponding to the regularized data in this time period will not change, that is, the number of operands will not change in a prediction; the length of the continuous time-series data segment is determined at initialization and remains fixed in subsequent predictions.

[0036] The segmented degradation processing unit 24 is used to perform segmented degradation processing on the input continuous time-series data segments and output preprocessed regularized data. The segmented degradation processing includes: calculating the mean and variance of the input continuous time-series data segments as degradation identifier features characterizing aero-engine degradation, obtaining preprocessed regularized data with degradation identifier features, and providing feature recognition decision basis for the subsequent working mode arbitration module 3. The degradation identifier features, as manually extracted features, reflect the trend change patterns and abrupt change characteristics of the continuous time-series data segments, and can be used to characterize the degradation stage of the aero-engine. A larger variance indicates a deeper degree of aero-engine degradation. Using the preprocessed regularized data with degradation identifier features as input to the neural network model makes the prediction more accurate.

[0037] For example, please refer to Figure 3 , Figure 3 This is a schematic diagram of the working mode arbitration module 3 of this application. The working mode arbitration module 3 includes: an operand decision module 31 and a working mode matching module 32, with the operand decision module 31 connected to the working mode matching module 32; the operand decision module 31 is connected to the data preprocessing module 2, and the working mode matching module 32 is connected to the low-power health status assessment module 4 and the adaptive neural network processing module 5.

[0038] The operand decision module 31 is used to perform operand feature recognition on the operands corresponding to the preprocessed regularized data. The operand feature recognition includes: performing working mode matching based on the operands corresponding to the preprocessed regularized data to obtain a working mode matching decision signal, and then performing mode adaptation on the subsequent working mode matching module 32. Since the operands corresponding to the external operating conditions may change each time raw sensor data is acquired compared to the previous acquisition, different operating modes are required for prediction under different external operating conditions. Therefore, different working modes are defined according to the range of operands, and the operands corresponding to each working mode are pre-set. For example, if the flight altitude is set to (6,8] km, Mach number to (0.6,1.0] Ma, and throttle angle to (80,100]° as operating mode 1, and the flight altitude is set to (8,10] m, Mach number to (0.6,1.0] Ma, and throttle angle to (80,100]° as operating mode 2, then when the operands corresponding to the preprocessed normalized data are 8, 0.8, and 100, the operating mode matching decision signal for operating mode 1 will be output; if the operands corresponding to the preprocessed normalized data are 9, 0.8, and 100, then the operating mode matching decision signal for operating mode 2 will be output.

[0039] The working mode matching module 32 is used to perform working mode matching based on the input working mode matching decision signal and the system's previous working mode. The working mode matching includes: determining whether the working mode is the same as the previous working mode based on the working mode matching decision signal, and outputting a working mode matching result signal. This provides a basis for the adaptive neural network processing module 5 to adjust the neural network model parameters and clarifies the direction of parameter adjustment for the neural network model. For example, if the working mode matching decision signal is working mode 1 and the previous working mode was also working mode 1, then the output working mode matching result signal is "matched"; if the working mode matching decision signal is working mode 1 and the previous working mode was working mode 2, then the output working mode matching result signal is "not matched".

[0040] For example, please refer to Figure 4 , Figure 4 This is a schematic diagram of the low-power health status assessment module 4 of this application. The low-power health status assessment module 4 includes: a physical model threshold comparison module 41 and a health status threshold setting module 42, which are connected; the physical model threshold comparison module 41 is connected to the working mode arbitration module 3 and the adaptive neural network processing module 5.

[0041] The physical model threshold comparison module 41 is used to compare the input preprocessed regularized data with physical model thresholds and output a health status indication signal. The physical model threshold comparison includes: comparing the preprocessed regularized data with the health status thresholds to determine whether the preprocessed regularized data is within the threshold range, and outputting a health status indication signal to provide a basis for neural network model selection in the subsequent neural network model prediction stage of the adaptive neural network processing module 5. Specifically, if none of the preprocessed regularized data exceeds the health status threshold, the current aero-engine is determined to be in a healthy state, and a healthy status indication signal is output; if any preprocessed regularized data exceeds the health status threshold, the current aero-engine is determined to be in a sub-healthy state, and a sub-healthy status indication signal is output.

[0042] The health status threshold setting module 42 is used to set health status thresholds according to the health status assessment requirements during the system initialization phase. Setting the health status thresholds includes: outputting an appropriate health status threshold based on the preset operational safety standards and life assessment requirements of the aero-engine, providing a comparison benchmark for the subsequent physical model threshold comparison module 41. Since normalization was performed in the data preprocessing module 2, the value of each preprocessed normalized data point is between 0 and 1. During the aero-engine degradation process, the device's health status is output before the preprocessed normalized data points drop below the health status threshold; after the preprocessed normalized data points drop below the health status threshold, the device is output as being in a sub-healthy state.

[0043] In one embodiment of this application, the health status threshold is 0.7.

[0044] For example, please refer to Figure 5 , Figure 5 This is a schematic diagram of the adaptive neural network processing module 5 of this application. The adaptive neural network processing module 5 includes: a working mode configuration module 51, a neural network module 52, a memory module 53, and a two-to-one selector 54. The working mode configuration module 51 is connected to the neural network module 52 and the memory module 53, respectively. The neural network module 52 is connected to the memory module 53 and the two-to-one selector 54, respectively. The working mode configuration module 51 and the neural network module 52 are connected to the working mode arbitration module 3, respectively. The neural network module 52 and the two-to-one selector 54 are connected to the low-power health status assessment module 4, respectively.

[0045] The working mode configuration module 51 is used to configure the working mode based on the input working mode matching result signal. The working mode configuration includes selecting the corresponding working mode configuration parameters according to the working mode matching result signal. Specifically, if the working mode matching result signal is a match, it indicates that the working mode of the current prediction is the same as the previous working mode, and a hold signal is output to the neural network module 52 and the memory module 53; if the working mode matching result signal is a mismatch, it indicates that the working mode of the current prediction is different from the previous working mode, and a reload signal is output to the neural network module 52 and the memory module 53, calling the working mode configuration parameters of the current prediction stored in the memory module 53.

[0046] Specifically, each neural network model is pre-trained for different working modes to obtain corresponding working mode configuration parameters. Each neural network model has the same topology, but the corresponding working mode configuration parameters need to be overloaded in different working modes. The configuration parameters include parameters such as the weights and biases of the neural network model.

[0047] As an example, please continue reading Figure 5 The neural network module 52 includes a low-power quantization model 521 and a high-precision calculation model 522, which are respectively connected to the working mode configuration module 51, the memory module 53, and the two-to-one selector 54. The neural network module 52 is used to perform adaptive neural network prediction calculations on the input preprocessed regularized data, corresponding working mode configuration parameters, and health status indication signals, and outputs intermediate neural network prediction results for the corresponding working mode.

[0048] The low-power quantization model 521 is used to activate the low-power prediction mode when the received health status indication signal is sub-healthy. It performs rapid low-power neural network prediction on the preprocessed regularized input data and corresponding working mode configuration parameters, outputting intermediate neural network prediction results in the corresponding working mode to reduce prediction computation power consumption. Specifically, if the working mode configuration module 51 outputs a hold signal, the working mode configuration parameters remain unchanged; if the working mode configuration module 51 outputs a model reload signal, the working mode configuration parameters are reloaded.

[0049] Specifically, the low-power quantization model 521 is a lightweight model obtained by training an artificial intelligence neural network model on a real flight dataset to obtain a 32-bit floating-point precision model with a large number of parameters, and then using fixed-point quantization technology to reduce the number of bits of the model parameters. This results in a sacrifice of prediction accuracy, reduced resource consumption, and faster computation speed. The artificial intelligence neural network models used include: Long Short-Term Memory Network (LSTM), Convolutional Neural Network (CNN), and Fully Connected Neural Network (FCN).

[0050] In one embodiment of this application, the number of bits used to quantize the model parameters is 8 bits or 4 bits using fixed-point quantization technology.

[0051] The high-precision calculation model 522 is used to activate the high-precision prediction mode when a health status indication signal is received. It performs high-precision neural network prediction on the preprocessed regularized input data and corresponding working mode configuration parameters, outputting intermediate neural network prediction results under the corresponding working mode to improve prediction accuracy. Specifically, if the working mode configuration module 51 outputs a hold signal, the working mode configuration parameters remain unchanged; if the working mode configuration module 51 outputs a model reload signal, the working mode configuration parameters are reloaded.

[0052] Specifically, the high-precision computing model 522 is a 32-bit floating-point precision model with a large number of parameters, trained using an artificial intelligence neural network model on a real flight dataset. Due to the high precision of the data type, it has high prediction accuracy, but also consumes a lot of resources. The artificial intelligence neural network models used include: Long Short-Term Memory Network (LSTM), Convolutional Neural Network (CNN), and Fully Connected Neural Network (FCN).

[0053] The memory module 53 is used to store and read the working mode configuration parameters in response to the parameter calling requirements of the input working mode configuration module 51. The storage and reading management of the working mode configuration parameters includes: storing the corresponding working mode configuration parameters obtained by pre-training each working mode, and providing the corresponding working mode configuration parameters to the adaptive neural network module 52 according to the current predicted working mode. The parameter calling requirements include: hold signal and reload signal.

[0054] A 2-to-1 selector 54 is used to gating and switching the intermediate prediction results output by the neural network module 52. This gating and switching includes: delivering data by outputting prediction results corresponding to the activated prediction mode and the working mode. Specifically, if the health status indicator signal is healthy, a low-power quantized prediction result is output; if the health status indicator signal is sub-healthy, a high-precision calculation prediction result is output. If the low-power health status assessment module 4 provides a healthy signal, it outputs a low-power quantized model prediction result; if the low-power health status assessment module 4 provides a sub-healthy signal, it outputs a high-precision calculation model prediction result.

[0055] Specifically, the low-power quantization model 521 and the high-precision calculation model 522 are only activated in their respective prediction modes, and the output results are only valid in their respective prediction modes. Ultimately, the corresponding prediction result is selected for output, rather than both models working simultaneously.

[0056] In another embodiment of this application, a method for predicting the remaining life of an aero-engine is also provided, applied to the aforementioned aero-engine remaining life prediction system. Please refer to... Figure 6 , Figure 6 The flowchart of the aircraft engine remaining life prediction method proposed in this application includes steps S1 to S10.

[0057] Step S1: Collect raw data from the multimodal sensor; Step S2: Data preprocessing; Step S3: Operand decision; Step S4: Determine if the working mode matches. If yes, proceed to step S5; otherwise, proceed to step S6. Step S5: Keep the configuration parameters and proceed to step S7; Step S6: Reload configuration parameters, proceed to step S7; Step S7: Assess whether the health status is healthy. If yes, proceed to step S8; otherwise, proceed to step S9. Step S8: Activate low power prediction mode and proceed to step S10; Step S9: Activate high-precision prediction mode and proceed to step S10; Step S10: Output the prediction results.

[0058] In step S3, the operand decision includes: operand feature identification for the operands corresponding to the preprocessed regularized data; operand feature identification includes: obtaining a working mode matching decision signal by performing working mode matching based on the operands corresponding to the preprocessed regularized data.

[0059] In step S4, determining whether the working mode matches includes: judging whether it is the same as the previous working mode based on the working mode matching decision signal.

[0060] In steps S5 and S6, the configuration parameters include: each neural network model is pre-trained for different working modes, and the corresponding parameters such as weights and biases of the neural network model under the corresponding working mode are obtained.

[0061] In step S7, assessing whether the health status is healthy includes: comparing the preprocessed regularized data with the health status threshold, determining whether the preprocessed regularized data is within the threshold range, and outputting a health status indication signal.

[0062] The aero-engine remaining life prediction system and method provided in this application can dynamically adjust the prediction strategy based on the preprocessing results of the collected multimodal sensor raw data, the aero-engine health status assessment conclusions, and the matching of operating modes, flexibly selecting either a low-power prediction mode or a high-precision prediction mode. When the aero-engine is in a healthy state and the system operating mode is matched, the system activates the low-power prediction mode, simplifying the prediction process through a low-power quantization model and reducing the overall system power consumption. Conversely, when the aero-engine is in a sub-healthy state, the system switches to the high-precision prediction mode, refining the prediction process using a high-precision calculation model with hybrid precision quantization, ensuring the accuracy of the remaining life prediction results, and providing reliable support for aero-engine fault early warning and maintenance decisions. This dynamic adaptation mechanism can achieve the optimal balance between prediction accuracy, system power consumption, and operational reliability under different aero-engine operating scenarios, meeting the remaining life prediction needs throughout the entire aero-engine lifecycle, optimizing energy consumption while ensuring the effectiveness of the prediction results, and contributing to the safe and stable operation of aero-engines.

[0063] Although this application has been disclosed above with reference to embodiments, it is not intended to limit this application. Anyone skilled in the art may make some modifications and refinements without departing from the spirit and scope of this application. Therefore, the scope of protection of this application shall be determined by the appended claims.

Claims

1. An aeroengine remaining life prediction system, characterized in that, The system comprises a multi-modal data acquisition module, a data preprocessing module, a working mode arbitration module, a low-power health state evaluation module, and an adaptive neural network processing module. The multi-modal data acquisition module is connected with the data preprocessing module, the data preprocessing module is connected with the working mode arbitration module, the working mode arbitration module is connected with the low-power health state evaluation module and the adaptive neural network processing module, and the low-power health state evaluation module is connected with the adaptive neural network processing module. The working mode arbitration module is configured to perform operation number feature recognition on the preprocessed regular data, perform working mode matching according to the recognized operation number, and output a working mode matching result signal. The low-power health state evaluation module is configured to set a health state threshold according to the health state evaluation requirement in the system initialization stage, compare the preprocessed regular data with the health state threshold, and output a health state indication signal. The adaptive neural network processing module is configured to select different precision neural network models according to the health state indication signal, select corresponding working mode configuration parameters according to the working mode matching result signal, perform neural network adaptive prediction calculation on the preprocessed regular data, and output a prediction result under the corresponding precision and working mode.

2. The aeroengine remaining useful life prediction system in accordance with claim 1, wherein, The data preprocessing module comprises a data analysis and selection unit, a normalization unit, a sliding window unit, and a segmented degradation processing unit, which are connected in sequence. The segmented degradation processing unit is configured to perform segmented degradation processing on the continuous time series data segment input by the sliding window unit, and output preprocessed regular data.

3. The aeroengine remaining useful life prediction system in accordance with claim 1, wherein, The working mode arbitration module comprises an operation number judgment module and a working mode matching module, which are connected.

4. The aeroengine remaining useful life prediction system in accordance with claim 1, wherein, The low-power health state evaluation module comprises a physical model threshold comparison module and a health state threshold setting module, which are connected.

5. The aeroengine remaining useful life prediction system in accordance with claim 1, wherein, The adaptive neural network processing module comprises a working mode configuration module, a neural network module, a memory module, and a two-alternative selector, which are connected. The two-alternative selector is used for gating switching of the neural network prediction intermediate result output by the neural network module, and the gating switching comprises: according to the activated prediction mode, data delivery of the prediction result in the corresponding precision and working mode is output.

6. The aeroengine remaining useful life prediction system in accordance with claim 5, wherein, The neural network module comprises a low-power quantization model and a high-precision calculation model, and the low-power quantization model and the high-precision calculation model are connected with the working mode configuration module, the memory module and the two-alternative selector.

7. The aeroengine remaining useful life prediction system in accordance with claim 6, wherein, The low-power quantization model is a model obtained by using a fixed-point quantization technology to reduce the bit number of model parameters after a multi-parameter floating-point precision model is obtained by training a real flight data set using an artificial intelligence neural network model.

8. The aeroengine remaining useful life prediction system in accordance with claim 6, wherein, The high-precision calculation model is a multi-parameter floating-point precision model obtained by training a real flight data set using an artificial intelligence neural network model.

9. The method of claim 1-8, applied to the aero-engine remaining useful life prediction system of any one of claims 1-8, wherein, Comprise, Step S1: collect multi-modal sensor original data; Step S2: data preprocessing; Step S3: operand judgment; Step S4: judge whether the working mode is matched, if yes, go to step S5; if no, go to step S6; Step S5: configuration parameter remains, go to step S7; Step S6: configuration parameter reload, go to step S7; Step S7: evaluate whether the health state is healthy, if yes, go to step S8; if no, go to step S9; Step S8: activate the low-power prediction mode, go to step S10; Step S9: activate the high-precision prediction mode, go to step S10; Step S10: output the prediction result.

10. The aero-engine remaining life prediction method of claim 9, wherein, Judging whether the working mode is matched comprises: judging whether the working mode is the same as the last working mode according to a working mode matching judgment signal; Evaluating whether the health state is healthy comprises: comparing the preprocessed regular data with a health state threshold value to judge whether the preprocessed regular data is within the threshold value range, and outputting a health state indication signal.

Images