A predictive health management system for aircraft engines
By constructing an integrated teacher model and a lightweight diagnostic system, the problems of early prediction and accurate diagnosis in the traditional aircraft engine maintenance mode are solved, and predictive health management of aircraft engines with real-time fault identification and resource optimization is realized.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN ZHUOHUI ELECTRONIC TECH CO LTD
- Filing Date
- 2026-04-21
- Publication Date
- 2026-05-26
AI Technical Summary
Traditional aircraft engine maintenance methods cannot achieve early prediction and accurate diagnosis, resulting in unplanned downtime and waste of resources. Existing airborne system algorithms are rigid and cannot identify complex failure modes.
A teacher integration model is constructed, and a lightweight diagnostic model is generated through knowledge distillation and pruning quantization. This model is then combined with airborne sensor data for real-time diagnosis and continuously optimized through a ground update module.
It enables real-time monitoring of multi-dimensional parameters during aircraft flight, identifies complex fault modes and outputs early prediction probabilities, reduces onboard computing requirements, and forms a closed-loop system of data acquisition, model optimization and operational feedback.
Smart Images
Figure CN122087589A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of aircraft engine failure prediction technology, and specifically relates to an aircraft engine predictive health management system. Background Technology
[0002] Traditional aircraft engine maintenance methods are mainly divided into two types: reactive maintenance and planned maintenance. Reactive maintenance refers to repairs performed only after a failure occurs. This method often leads to unplanned downtime, causing flight delays or cancellations, resulting in direct economic losses and, in extreme cases, threats to flight safety. Planned maintenance, on the other hand, involves preventative replacement or maintenance of engine components based on fixed flight hours or takeoff and landing cycles. While this fixed-cycle maintenance method improves reliability to some extent, it is essentially conservative over-maintenance, with many components being replaced far before the end of their service life, resulting in significant cost waste and resource idleness.
[0003] With the development of avionics technology, aircraft engine maintenance has begun to be based on flight condition maintenance. Data analysis is used to capture abnormal trends, enabling condition-based maintenance and maximizing component lifespan while ensuring safety. Existing aircraft condition monitoring systems are typically installed on the aircraft to monitor key engine physical parameters in real time, including temperature, pressure, and vibration. When a parameter exceeds a preset threshold, the system sends a warning signal to the pilot in the cockpit. After landing, detailed data recorded during flight is downloaded to a ground data center for in-depth analysis and fault diagnosis by professional engineers or ground analysis software. However, due to the limited computing power of avionics hardware, it cannot support complex machine learning models that require massive computing resources. Furthermore, the algorithmic logic of traditional airborne systems is rigid, employing simple judgment rules based on fixed thresholds. This pre-programmed model cannot identify complex fault modes formed by subtle changes in multi-dimensional parameters, and can only provide passive alarms, failing to achieve early prediction and accurate diagnosis. Summary of the Invention
[0004] The purpose of this invention is to overcome the deficiencies in the prior art and provide a predictive health management system for aircraft engines.
[0005] This invention provides a predictive health management system for aircraft engines, comprising: The model building module is used to receive and process flight data and maintenance records from the fleet, and to iteratively train the flight data and maintenance records as input to obtain a teacher ensemble model for outputting fault diagnosis results, wherein the diagnosis results include the classification probability for each preset fault mode. The lightweight model module takes unlabeled flight data as input and uses the fault diagnosis results generated by the teacher ensemble model as the supervised target for iterative training to obtain a student model, so that the output of the student model approximates the fault diagnosis results; the lightweight model module is also used to prune and quantize the trained student model to generate a lightweight diagnostic model; An airborne diagnostic module, installed on the aircraft, includes a memory and a processor. The memory stores the lightweight diagnostic model, and the processor collects multi-dimensional sensor data of the engine in real time. The lightweight diagnostic model outputs the real-time matching probability between the current engine status and each preset fault mode based on the multi-dimensional sensor data. When the real-time matching probability exceeds a preset threshold, an alarm message is generated. The cockpit display module, connected to the airborne diagnostic module via an avionics bus, is used to receive the alarm information and display it in a designated display area of the engine indication and crew warning system or electronic flight instrument system.
[0006] A further embodiment is that the model building module includes a data acquisition unit and a model training unit; The data acquisition unit is used to acquire historical flight data and corresponding maintenance records of the same type of aircraft. Based on the engine's unique identifier as an index, the historical flight data is matched with the maintenance records to construct a real engine fault sample set. The model training unit trains at least two neural network units based on the fault sample set, and combines the trained neural network units to form a teacher ensemble model.
[0007] A further embodiment is that the neural network unit includes at least one convolutional neural network and at least one recurrent neural network. The convolutional neural network slides along the time dimension to capture the spatial distribution patterns of sensor parameters formed over a short period of time, including the cooperative changes between temperature, pressure, and vibration parameters. The recurrent neural network adopts a long short-term memory network or a gated recurrent unit structure to handle long-term dependencies and learn the trend of sensor parameters slowly drifting from a normal state to an abnormal state. The convolutional neural network and the recurrent neural network are combined in series or in parallel, and the output feature map of the convolutional neural network is flattened and used as the input of the recurrent neural network.
[0008] A further solution is that the pruning process includes: based on a preset sparsity threshold, evaluating the contribution of each convolutional kernel or neuron channel in the student model, deleting the weight connections of convolutional kernels or corresponding channels whose contribution is lower than the sparsity threshold, and fine-tuning the remaining network structure to restore model accuracy. The quantization process includes: converting the floating-point weight parameters and activation values of the pruned student model into low-bit-width integer fixed-point numbers, generating a parameter quantization table, and converting the numerical operations in the lightweight diagnostic model calculation process from floating-point form to fixed-point form based on the quantization table.
[0009] A further approach is that the fine-tuning includes: iteratively training the lightweight diagnostic model on a fault sample set, using cross-entropy loss as the supervision target, updating the retained weight parameters, so that the model restores its classification accuracy for preset fault modes.
[0010] A further embodiment is that the model lightweighting module also includes a weight parameter update unit, which is configured to perform the following operations: When constructing the student model, the lightweight model module simultaneously extracts a batch of data from the fault sample set as hard-label supervision. The unlabeled flight data and the data from the fault sample set are merged or alternately input into the student model for iterative training, and the KL divergence loss corresponding to the soft label and the cross-entropy loss corresponding to the hard label are calculated respectively. We perform a weighted summation of the KL divergence loss and cross-entropy loss, and then use the backpropagation algorithm to update the weight parameters of the student model.
[0011] A further approach involves using KL divergence loss to measure the difference between the soft labels output by the student model and the teacher ensemble model, and cross-entropy loss to measure the difference between the student model and the true labels. The weights of KL divergence loss and cross-entropy loss are balanced using hyperparameters.
[0012] A further embodiment includes a data preprocessing unit in the model building module, which is configured to perform the following operations: Data points that exceed the preset mean standard deviation range are marked as outliers and removed. Normalize and map sensor parameters of different dimensions to the same numerical range; The continuous time series data is sliced into multiple data sample units according to a preset fixed length, and each data sample unit covers the time window before the fault occurs.
[0013] A further embodiment is that the processor also includes a cache unit, wherein the processor continuously collects multi-dimensional sensor data of the engine at a fixed sampling frequency during flight, and uses a sliding window mechanism to temporarily store the latest continuous data in the cache unit; The length of the sliding window is the same as the length of the time window for the training samples. The processor reads the data sequence within the current window from the cache unit and inputs it into the lightweight diagnostic model. The lightweight diagnostic model calculates the matching probability of the current data sequence with all preset fault modes in parallel and outputs the probability value. When the probability value of the fault mode continuously exceeds the preset threshold, an alarm message is generated.
[0014] A further embodiment is that the system also includes a ground model update module, which is used to periodically collect new flight data and maintenance records of the fleet, re-execute fault labeling and teacher integrated model training, generate an updated lightweight diagnostic model, and push the model update package to the airborne diagnostic module of each aircraft through a dedicated aviation communication network. When the aircraft is in ground maintenance mode, the airborne diagnostic module automatically detects model updates. Ground staff connect to the airborne maintenance interface via a portable data loader to complete the verification and loading of the model update package. After the update is completed, the airborne diagnostic module records the new model version number and enables the new model for real-time diagnostics during subsequent flights.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention, through a technical architecture of ground training and airborne deployment, embeds data-driven complex diagnostic capabilities into the airborne hardware platform using lightweight methods such as knowledge distillation and pruning quantization. This enables the aircraft to capture subtle changes in multidimensional parameters in real time, identify complex fault modes, and output early prediction probabilities during flight, fundamentally solving the technical defects of traditional airborne systems, such as fixed algorithms, passive alarm capabilities, and inability to make early predictions.
[0016] The model building module of this invention is based on operational data and maintenance records accumulated from a fleet of aircraft of the same model. Using the engine's unique code as an index, it establishes a correspondence between sensor parameters recorded during flight and maintenance records after landing, forming a fault sample set with real fault labels. Based on this fault sample set, various neural network models with different structures are trained, and the outputs of these heterogeneous models are fused for decision-making to form a teacher ensemble model. When new flight data is input, the teacher ensemble model outputs the classification probability for each preset fault mode, providing a high-quality source of supervisory information for subsequent lightweight deployment.
[0017] The lightweight model module of this invention uses unlabeled flight data to guide the lightweight diagnostic model to mimic the output of the teacher ensemble model, enabling the lightweight diagnostic model to learn the ensemble model's judgment criteria and boundary information for various fault modes. It also simplifies the execution structure of the lightweight diagnostic model, significantly reducing the computing power requirements and memory usage of the onboard computer, allowing the deep diagnostic capabilities that could only operate on the ground to be invoked in real time during flight.
[0018] The airborne diagnostic module of this invention has a built-in data buffer that continuously collects sensor data from various engine monitoring points. It employs a sliding time window approach to always retain the latest continuous data sequence. The processor periodically reads the data sequence within the current window from the buffer and feeds it into a lightweight diagnostic model for parallel computation. The model simultaneously outputs a score indicating the degree of matching between the current data segment and each known fault mode. When the matching score for a certain fault mode consistently exceeds a preset warning threshold, the system automatically generates a warning signal. This enables the system to continuously monitor the engine status throughout flight, identifying potential risks at an early stage of fault characteristics through subtle changes in multidimensional parameters.
[0019] This invention also includes a ground update module that periodically collects newly added operational data and maintenance records from the fleet, re-labels samples, and retrains the integrated model. The updated diagnostic model is then distributed to each aircraft via a dedicated aviation wireless data link. During ground maintenance, aircraft automatically detect available updates. Maintenance personnel use handheld data terminals to connect to the airborne maintenance interface to verify and load the update package. After the update is complete, the system records the new model version identifier, and the updated model is used for real-time diagnostics during subsequent flights. This allows the diagnostic capabilities of the entire system to continuously evolve with the accumulation of fleet operational data, adapting to performance changes and the emergence of new faults throughout the engine's entire lifecycle, forming a complete closed loop of data acquisition, model optimization, airborne deployment, and operational feedback. Attached Figure Description
[0020] The following figures are for illustrative purposes only and are not intended to limit the scope of the invention, wherein: Figure 1 : System connection block diagram of the present invention; In the diagram: 1. Model building module; 2. Model lightweighting module; 3. Airborne diagnostic module; 4. Cockpit display module; 5. Data acquisition unit; 6. Model training unit; 7. Data preprocessing unit; 8. Weight parameter update unit; 9. Neural network unit; 10. Convolutional neural network; 11. Recurrent neural network; 12. Memory; 13. Processor; 14. Cache unit; 15. Ground model update module. Detailed Implementation
[0021] To make the objectives, technical solutions, design methods, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.
[0022] like Figure 1As shown, this invention provides a predictive health management system for aircraft engines, including a model building module 1, a lightweight model module 2, an onboard diagnostic module 3, and a cockpit display module 4. The model building module 1 is deployed in a ground data center and connected to the airline's operations control center via a data interface. It receives and processes flight data and maintenance records from the entire fleet. The model building module 1 uses the flight data and maintenance records as input for iterative training to obtain a teacher ensemble model that outputs fault diagnosis results. The diagnostic results output by the teacher ensemble model include the classification probability for each preset fault mode. The lightweight model module 2 is connected to the model building module 1. The lightweight model module 2 uses unlabeled flight data as input to a student model and iteratively trains the student model using the fault diagnosis results generated by the teacher ensemble model as the supervised target. This allows the student model's output to continuously approximate the fault diagnosis results of the teacher ensemble model. The lightweight model module 2 also prunes and quantizes the trained student model, ultimately generating a lightweight diagnostic model. The airborne diagnostic module 3 is installed on the aircraft and includes a memory 12 and a processor 13. The memory 12 stores the lightweight diagnostic model generated by the lightweight model module 2. The processor 13 is connected to the engine sensor system via the airborne data bus to collect multi-dimensional sensor data of the engine in real time. The processor 13 calls the lightweight diagnostic model in the memory 12 and outputs the real-time matching probability of the current engine status with each preset fault mode based on the real-time collected multi-dimensional sensor data. When the real-time matching probability of a certain fault mode exceeds a preset threshold, the processor 13 generates an alarm message. The cockpit display module 4 is connected to the airborne diagnostic module 3 via the avionics bus. It is used to receive the alarm messages sent by the airborne diagnostic module 3 and display the alarm messages in a designated display area of the engine indication and crew warning system or electronic flight instrument system, so that the pilot can intuitively obtain the engine health status and fault warning information.
[0023] In the above, the model building module 1 further includes a data acquisition unit 5 and a model training unit 6. The data acquisition unit 5 is connected to the airline's operation control database and maintenance record database to acquire historical flight data and corresponding maintenance records of the same type of aircraft. Using the engine's unique identifier as an index, the data acquisition unit 5 precisely matches the historical flight data of the same engine with its past maintenance records to construct a real engine fault sample set. Each sample in this fault sample set contains a sequence of sensor parameters for a period of time prior to the fault occurrence and a corresponding fault type label. The model training unit 6 is connected to the data acquisition unit 5, receives the constructed real engine fault sample set, and trains at least two neural network units 9 based on this fault sample set. The model training unit 6 then combines the trained neural network units 9 using an ensemble learning approach to form a teacher ensemble model. Through this data matching and ensemble training method, the teacher ensemble model can learn the characteristic patterns of various fault modes from real fault cases, providing high-quality supervision information for subsequent model lightweighting. In this embodiment, the neural network unit 9 specifically includes a convolutional neural network 10 and a recurrent neural network 11. The convolutional neural network 10 and the recurrent neural network 11 are combined in parallel, that is, the convolutional neural network 10 and the recurrent neural network 11 process the input data separately, and the output features of the two are concatenated or weighted summed in the fusion layer. The convolutional neural network 10 adopts a one-dimensional convolutional structure, sliding along the time dimension to capture the spatial distribution pattern of sensor parameters formed in a short time. Specifically, the convolutional neural network 10 extracts the cooperative change relationship of multiple sensor parameters such as temperature, pressure, and vibration within the same moment or a short time window by sliding the convolutional kernel along the time axis, forming a feature map reflecting local features. The recurrent neural network 11 adopts a long short-term memory network or a gated recurrent unit structure to process the long-term dependency relationship of sensor parameters over time. It learns the trend features of sensor parameters slowly drifting from a normal state to an abnormal state through memory units and gating mechanisms.
[0024] In this embodiment, the pruning process includes: based on a preset sparsity threshold, evaluating the contribution of each convolutional kernel or neuron channel in the student model to the diagnostic results. Contribution evaluation can be conducted using various methods such as the absolute value of weights, gradient information, or activation value distribution. Convolutional kernels or weighted connections of corresponding channels with contributions below the sparsity threshold are removed from the network. This sparsity threshold is set as the median or the top 10% quantile of the contribution distribution. The specific value is determined through cross-validation based on the balance between model compression ratio and accuracy loss, ranging from 0.01 to 0.1, to ensure effective model size reduction without significantly reducing diagnostic accuracy. After the deletion operation, the remaining network structure is fine-tuned, and the model recovers diagnostic accuracy through a small number of iterative training iterations. Quantization processing includes converting the floating-point weights and activation values of the pruned student model into low-bit-width fixed-point integers, such as converting 32-bit floating-point numbers to 8-bit integers. During this conversion, a parameter quantization table is generated, recording the mapping relationship between floating-point and integer values. Based on this table, the lightweight diagnostic model converts numerical calculations that originally required floating-point operations into fixed-point operations, significantly reducing the complexity of each multiplication-addition operation. Through these two-stage pruning and quantization processes, the student model's size is significantly reduced, and computational efficiency is greatly improved, making it suitable for the limited resources of onboard computers. The fine-tuning operation after pruning has a specific implementation method. The fine-tuning process includes iteratively training the lightweight diagnostic model on a fault sample set. During training, the cross-entropy loss function is used as the supervision target. The cross-entropy loss measures the difference between the model's output probability distribution and the true fault labels. The retained weight parameters are updated using the backpropagation algorithm, allowing the model to gradually restore its classification accuracy for preset fault modes. During fine-tuning, the learning rate is typically set lower than the initial training value to avoid excessive perturbation of the retained weights. The number of iterations is also reduced accordingly, requiring only a few training rounds to bring the model's accuracy back to near its pre-pruning level. This targeted fine-tuning eliminates the accuracy loss caused by pruning while maintaining the compact structure of the pruned model.
[0025] In the above, the lightweight model module 2 also includes a weight parameter update unit 8. When constructing the student model, the weight parameter update unit 8 simultaneously randomly selects a batch of labeled data from the fault sample set as a hard-labeled supervision source. During training, unlabeled flight data and labeled data from the fault sample set are input into the student model either together or alternately for iterative training. For unlabeled data, the KL divergence loss between the student model output and the soft labels output by the teacher ensemble model is calculated; KL divergence measures the difference between two probability distributions. For labeled data, the cross-entropy loss between the student model output and the true fault labels is calculated. The weight parameter update unit 8 performs a weighted summation of the KL divergence loss and the cross-entropy loss to obtain the total loss value. The weights of the KL divergence loss and the cross-entropy loss are balanced by hyperparameters α and β, where α corresponds to the KL divergence loss and β corresponds to the cross-entropy loss, with values ranging from [0,1] and satisfying α+β=1. In the initial training phase, β is set to 0.7–0.9 to accelerate convergence, and α is increased to 0.3–0.5 in the later stages to enhance the learning ability of the teacher model distribution. The backpropagation algorithm is used to calculate the gradient based on the total loss value and update the weight parameters of the student model. Through this joint training method, the student model learns the similarity relationships between failure modes from the teacher ensemble model and the accurate classification boundaries from real failure samples. In this embodiment, KL divergence loss is used to measure the difference between the soft labels output by the student model and the teacher ensemble model. The soft labels contain the probability judgments of the teacher ensemble model for various failure modes, reflecting the similarity relationships between different failure categories. Cross-entropy loss is used to measure the difference between the student model and the real failure labels. The real labels provide accurate classification supervision information. The weights of both are balanced through hyperparameters, with values between 0 and 1, adjusted based on performance on the validation set. When the student model has not yet converged in the early stages of training, the weight of the cross-entropy loss can be appropriately increased to accelerate convergence; when fine-tuning is needed in the later stages of training, the weight of the KL divergence loss can be appropriately increased to better learn the output distribution of the teacher model. By using weighted balancing, the student model can comprehensively utilize the prior knowledge of the teacher model and the accurate information from the real labels to achieve better diagnostic performance.
[0026] In the above, the model building module 1 also includes a data preprocessing unit 7, used for outlier removal, normalization, and time series slicing of the data. Specifically, outlier removal calculates the mean and standard deviation of each sensor parameter in historical data for the input flight data, marks data points exceeding the preset mean and standard deviation range as outliers and removes them from the data sequence, avoiding interference from sensor momentary failures or noise on model training. Normalization is used for sensor parameters with different units, such as temperature, pressure, and vibration, which have different numerical ranges. Max-min normalization is used to map all parameters to between 0 and 1, eliminating the impact of unit differences on model training. Time series slicing involves slicing continuous time series data into slices of a preset fixed length. The slice length is usually set to cover a period of time before the failure occurs, such as 30 minutes or the first 100 sampling points before the failure. Each slice forms an independent data sample unit, and each sample unit covers the complete evolution process from the normal state to the time before the failure.
[0027] In the above description, the processor 13 of the airborne diagnostic module 3 also includes a cache unit 14. During flight, the processor 13 continuously acquires multi-dimensional sensor data from the engine at a fixed sampling frequency, consistent with the data sampling rate used during model training. The cache unit 14 always temporarily stores the latest acquired continuous data segment, and its size is equal to the length of the sliding window. The length of the sliding window is set to be consistent with the time window length of the training samples, for example, 100 consecutive sampling points, ensuring that the data format input to the model in real time completely matches the training samples. This fixed length is determined based on the fault evolution cycle and sampling frequency, covering the period from 5 to 30 minutes before the fault occurs. The specific value is optimized by analyzing the average warning lead of historical fault data, with a range of 50 to 200 sampling points. At each sampling moment, the processor 13 reads the latest data sequence within the current window from the cache unit 14 and inputs it into the lightweight diagnostic model. The lightweight diagnostic model uses parallel computing to simultaneously calculate the matching probability of the current data sequence with all preset fault modes, outputting the probability value for each fault mode. The processor 13 continuously monitors the probability output of each fault mode. When the probability value of a certain fault mode exceeds a preset threshold multiple times consecutively, it determines that the fault mode may occur and generates corresponding alarm information. Through this sliding window caching and continuous probability output mechanism, the system can continuously monitor the engine status during flight and promptly detect early signs of faults.
[0028] To adapt to performance changes and the emergence of new faults throughout the engine's lifecycle, the system also includes a ground model update module 15. This module works in conjunction with the fleet's data acquisition unit 5 and the onboard diagnostic modules 3 of each aircraft to ensure continuous updates to diagnostic capabilities. Specifically, the ground model update module 15 periodically collects new flight data and maintenance records from the entire fleet. When the new data accumulates to a certain scale, it re-executes the fault labeling process, labeling new fault modes in the new data, and retrains the teacher ensemble model using the updated dataset to generate an updated lightweight diagnostic model. The ground model update module 15 generates the updated lightweight diagnostic model and manually transmits the model update package to the aircraft via storage medium, where ground crew loads it while the aircraft is in ground maintenance mode. When the aircraft is in ground maintenance mode and the engine is off, the onboard diagnostic module 3 automatically detects whether there is an available model update and, upon detection, notifies maintenance personnel on the ground maintenance terminal. Ground crew connect to the airborne maintenance interface via a portable data loader to verify and load the model update package. The verification process ensures the integrity and security of the update package, while the loading process writes the updated lightweight diagnostic model into the memory 12 of the airborne diagnostic module 3. After the update is complete, the airborne diagnostic module 3 records the new model version number in the configuration file and automatically activates the new model for real-time diagnostics on the next flight.
[0029] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.
Claims
1. An aircraft engine prognostic health management system, characterized by, The method comprises the following steps: a model construction module is configured to receive and process flight data and maintenance records from a fleet of aircraft, and iteratively train the flight data and maintenance records as input to obtain a teacher ensemble model for outputting fault diagnosis results, the diagnosis results including classification probabilities for each pre-set fault mode; a model lightening module is configured to iteratively train a student model using no-labeled flight data as input and the fault diagnosis results generated by the teacher ensemble model as a supervision target, so that the output of the student model approximates the fault diagnosis results; the model lightening module is further configured to prune and quantize the trained student model to generate a lightened diagnosis model; an on-board diagnosis module is provided on an aircraft and comprises a memory and a processor, the memory stores the lightened diagnosis model, and the processor is configured to collect multi-dimensional sensor data of an engine in real time; the lightened diagnosis model outputs real-time matching probabilities of a current engine state and each pre-set fault mode based on the multi-dimensional sensor data; when the real-time matching probability exceeds a pre-set threshold, an alarm information is generated; a cockpit display module is connected to the on-board diagnosis module through an avionics bus, configured to receive the alarm information and display the alarm information on a designated display area of an engine indication and crew alerting system or an electronic flight instrument system.
2. A prognostic health management system for an aircraft engine as defined in claim 1, wherein, The model construction module comprises a data acquisition unit and a model training unit; the data acquisition unit is configured to acquire historical flight data and corresponding maintenance records of the same type of aircraft, match the historical flight data and the maintenance records based on an engine unique identification number as an index, and construct a real engine fault sample set; the model training unit is configured to train at least two neural network units based on the fault sample set, and combine the trained multiple neural network units to form a teacher ensemble model.
3. A prognostic health management system for an aircraft engine as defined in claim 2, wherein, The neural network unit comprises at least one convolutional neural network and at least one recurrent neural network, the convolutional neural network slides in the time dimension to capture the spatial distribution pattern formed by sensor parameters in a short time, including the cooperative change relationship among temperature, pressure and vibration parameters; the recurrent neural network adopts a long short-term memory network or a gated recurrent unit structure to process long-time dependence and learn the trend of sensor parameters slowly drifting from a normal state to an abnormal state; the convolutional neural network and the recurrent neural network are combined in series or in parallel, and the output feature map of the convolutional neural network is flattened and used as the input of the recurrent neural network.
4. The aircraft engine prognostic health management system of claim 3, wherein, The pruning process comprises: based on a pre-set sparsification threshold, evaluating the contribution degree of each convolution kernel or neuron channel in the student model, deleting the weight connection of the convolution kernel or the corresponding channel with a contribution degree lower than the sparsification threshold, and fine-tuning the remaining network structure to restore the model accuracy; the quantization process comprises: converting the floating-point weight parameters and activation values of the pruned student model into low-bit-width integer fixed-point numbers, generating a parameter quantization table, and converting the numerical operations in the calculation process of the lightened diagnosis model from floating-point numbers to fixed-point numbers based on the quantization table.
5. A prognostic health management system for an aircraft engine as defined in claim 4, wherein, The fine-tuning includes: iteratively training the lightweight diagnostic model on the fault sample set, taking cross-entropy loss as a supervision target, updating the reserved weight parameters, and restoring the classification accuracy of the model to the preset fault mode.
6. A prognostic health management system for an aircraft engine as defined in claim 5, wherein, The model lightweight module further includes a weight parameter updating unit configured to perform the following operations: When constructing the student model, the weight parameter updating unit simultaneously randomly extracts a batch of data from the fault sample set as hard label supervision; The unlabelled flight data and the data in the fault sample set are merged or alternately input into the student model for iterative training, and the KL divergence loss corresponding to the soft label and the cross-entropy loss corresponding to the hard label are calculated respectively; The KL divergence loss and the cross-entropy loss are weighted and summed, and the weight parameters of the student model are updated using the back propagation algorithm.
7. A prognostic health management system for an aircraft engine as defined in claim 6, wherein, The KL divergence loss measures the difference between the soft label output by the student model and the teacher integrated model, and the cross-entropy loss measures the difference between the student model and the real label. The weights of the KL divergence loss and the cross-entropy loss are balanced by hyperparameters.
8. A prognostic health management system for an aircraft engine as defined in claim 7, wherein, The model construction module further includes a data preprocessing unit configured to perform the following operations: Data points exceeding the preset mean standard deviation range are marked as abnormal and excluded; Different dimensional sensor parameters are normalized and mapped to the same value interval; Continuous time series data is sliced according to a preset fixed length to form a plurality of data sample units, each of which covers a time window before the occurrence of a fault.
9. A prognostic health management system for an aircraft engine as defined in claim 8, wherein, The processor further includes a cache unit, and the processor continuously collects multi-dimensional engine sensor data at a fixed sampling frequency during flight and temporarily stores the latest continuous data in the cache unit using a sliding window mechanism; The length of the sliding window is consistent with the length of the training sample time window. The processor reads the data sequence in the current window from the cache unit and inputs it into the lightweight diagnostic model. The lightweight diagnostic model calculates the matching probability of the current data sequence and all preset fault modes in parallel and outputs the probability value. When the probability value of the fault mode continuously exceeds a preset threshold, an alarm information is generated.
10. The aircraft engine prognostic health management system of claim 9, wherein, The system further includes a ground model updating module for periodically collecting newly added flight data and maintenance records of the fleet, re-executing fault labeling and teacher integrated model training, generating an updated lightweight diagnostic model, and pushing the model update package to the on-board diagnostic module of each aircraft through an aviation special communication network; When the aircraft is in a ground maintenance state, the on-board diagnostic module automatically detects model updates, ground personnel connect the on-board maintenance interface through a portable data loader to complete the verification and loading of the model update package; after the update is completed, the on-board diagnostic module records the new model version number and enables the new model for real-time diagnosis in subsequent flights.