A Machine Learning-Based Predictive Maintenance Method for Flight Simulators

Abstract

This invention belongs to the field of flight simulator maintenance technology, specifically relating to a machine learning-based predictive maintenance method for flight simulators. It addresses the problems of existing maintenance methods relying on periodic inspections and passive repairs, delayed fault warnings, and high false alarm / missing rate. This invention acquires historical operational data, sensor time-series data, fault records, and environmental parameters from the flight simulator, and performs preprocessing including cleaning, labeling, and feature fusion. It constructs a composite health feature set containing statistical features, dynamic health status values, and aircraft material reliability parameters. A hybrid prediction model (random forest feature selection + LSTM time-series prediction + adaptive correction reliability assessment) is used to train the model. Combining prediction result fusion analysis and multi-level decision rule post-processing, maintenance work orders and spare parts demand plans are output. This invention improves the accuracy and timeliness of fault prediction, realizing a shift from passive response to proactive prediction in maintenance, significantly reducing maintenance costs and the risk of unplanned downtime.

Description

Technical Field

[0001] This invention belongs to the field of flight simulator equipment maintenance technology, and specifically relates to a method for predictive maintenance of flight simulators based on machine learning. Background Technology

[0002] As the core equipment for civil aviation flight training, the operational stability of flight simulators directly determines the continuity and reliability of training missions. Current flight simulator maintenance and management methods mainly rely on periodic inspections and reactive repairs. The core approach employs a threshold / rule-based status monitoring and alarm scheme. This scheme consists of a sensor subsystem, a data acquisition / transmission unit, a threshold determination model, and an alarm / work order system. The workflow involves sensors collecting real-time data such as vibration, temperature, and current, which is then reported by the acquisition unit to the threshold determination model. The model compares each measurement item against preset thresholds; if the data exceeds the limit, an alarm is triggered and a maintenance notification is generated.

[0003] However, this traditional maintenance method has significant technical defects: on the one hand, it adopts a threshold comparison method and lacks cross-variable joint modeling capabilities, which makes it unable to effectively identify complex and multivariate fault modes, often resulting in false positives (false alarms) or false negatives (missed alarms); on the other hand, this method has no historical fault mode learning capability and can only alarm when a fault has occurred or the indicators deviate significantly, which cannot achieve early warning. As a result, maintenance is still mainly based on passive response, making it difficult to move the maintenance window forward and avoid unplanned downtime. This leads to problems such as low maintenance efficiency, high cost, and unreasonable maintenance cycle, which seriously affects the normal conduct of flight training missions.

[0004] To address the shortcomings of existing technologies, this invention proposes a predictive maintenance method for flight simulators based on machine learning. By integrating historical operational data, sensor time-series data, and fault records, an accurate predictive model is established to proactively identify potential faults and accurately predict the remaining lifespan of components. This optimizes maintenance strategies and improves the operational reliability and maintenance economy of flight simulators. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, namely the passive response, delayed fault prediction, and low maintenance efficiency of current flight simulator maintenance methods, this invention, in its first aspect, proposes a predictive maintenance method for flight simulators based on machine learning. This method includes the following steps:

[0006] Historical operational data of the flight simulator is acquired and preprocessed to obtain a structured training dataset. The preprocessing includes data cleaning and data labeling.

[0007] Based on the structured training dataset, statistical features are extracted and a reliability model is used to calculate the dynamic health status value of the component. The structured training dataset and the dynamic health status value are fused to obtain a composite health feature set for training.

[0008] Using the composite health feature set used for training, a hybrid prediction model is trained, which includes a data-driven prediction model and an adaptively corrected reliability assessment model.

[0009] The real-time operating data of the flight simulator is acquired and preprocessed to obtain the composite health feature set to be predicted;

[0010] The composite health feature set to be predicted is input into the hybrid prediction model; the data-driven prediction model outputs a first prediction result, and the reliability assessment model outputs a second prediction result; wherein, the first prediction result includes the probability of future failure of the component and a first remaining lifetime estimate, and the second prediction result includes the maintenance requirement time based on the task reliability determination and a second remaining lifetime estimate;

[0011] The first prediction result and the second prediction result are fused and analyzed, and processed according to the preset decision rules to obtain the maintenance work order of the flight simulator components.

[0012] The beneficial effects of this invention are:

[0013] 1) By constructing a composite health feature set that includes "statistical features, dynamic health status values, and task parameters", it is possible to capture the complex correlation between the operating status and deterioration trend of key components of the simulator (such as motion system and visual system), which significantly improves the accuracy of identifying fault precursors compared to traditional single threshold indicators.

[0014] 2) A hybrid prediction model combining "random forest feature selection + LSTM time series prediction + adaptive correction reliability assessment" is adopted, balancing the generalization ability of data-driven methods with the interpretability of the reliability physical model, effectively reducing false alarms and false negatives caused by changes in operating conditions (such as high-intensity training subjects). 3) The model has good scalability; only the Weibull distribution parameters and network hyperparameters need to be adjusted to adapt to the predictive maintenance needs of different simulator models or different key components, without the need to reconstruct the core model architecture. 4) Through a closed-loop iterative mechanism of "prediction-maintenance-relearning," the maintenance process is made intelligent and automated. Precise maintenance plans and spare parts requirements can be formulated without much manual intervention, significantly improving maintenance efficiency, reducing the risk of unplanned downtime, and ensuring the continuity of flight training plans. Attached Figure Description

[0015] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0016] Figure 1 This is a flowchart illustrating the steps of a predictive maintenance method for a flight simulator based on machine learning, as described in this invention. Detailed Implementation

[0017] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] To more clearly illustrate the predictive maintenance method for flight simulators based on machine learning according to the present invention, the following will be combined with... Figure 1 The steps in the embodiments of the present invention will be described in detail below.

[0020] This invention proposes a predictive maintenance method for flight simulators based on machine learning. (See also...) Figure 1 The method includes the following steps:

[0021] Step S1: Obtain historical operating data of the flight simulator and preprocess it to obtain a structured training dataset. The preprocessing includes data cleaning and data labeling.

[0022] In this embodiment, the historical operating data of the flight simulator includes sensor data (e.g., real-time monitoring data of key components such as hydraulic system pressure, motor vibration acceleration, transmission system temperature, and visual system current), training task logs (e.g., task duration of each training session, load level: light / medium / heavy, training subjects, number of takeoffs and landings, etc.), maintenance records (time of failure, failure type: minor / complex / major, maintenance measures, replacement part model, maintenance duration, etc.), and environmental parameters (e.g., flight simulator room temperature 5℃-35℃, humidity 30%-70%, power supply voltage stability 220V±5V, etc.).

[0023] Data cleaning includes missing value handling, outlier handling, and data alignment, as detailed below:

[0024] Missing value handling: Short-term missing data (missing duration ≤ 5 min) in sensor data are filled using linear interpolation. Data segments with long-term missing data (missing duration > 5 min) are marked as invalid data and removed.

[0025] Outlier handling: Outliers (data exceeding the mean ± 3 standard deviations) in sensor data are identified based on the 3σ criterion, and replaced with a moving average of adjacent data.

[0026] Data alignment: Based on timestamps, sensor data, task logs, maintenance records, and environmental parameters are aligned at the millisecond level to ensure consistency in the time dimension of the data;

[0027] The data annotation includes:

[0028] Data samples occurring within a time window spanning from a first duration to a second duration prior to a preset fault time point are marked as positive samples indicating fault precursors, wherein the first duration is longer than the second duration; specifically:

[0029] Preset fault time point: Based on the time of fault occurrence clearly stated in the maintenance record, mark it as the preset fault time point t_f;

[0030] Positive sample labeling: Data samples that occur within a time window from the first time period (72h) to the second time period (1h) before the preset fault time point, i.e., t∈[t_f-72h,t_f-1h], are labeled as positive samples indicating the precursor to the fault. A total of 8640 positive samples were labeled (divided into 1h / sample).

[0031] Data samples whose time intervals from any of the preset fault time points are greater than the first duration, and whose data have no fault records within the subsequent preset minimum guaranteed health duration, are marked as negative samples indicating a healthy state; specifically:

[0032] Data samples whose time intervals from any preset fault time point are greater than the first duration (t_f-72h, t_f-1h) (e.g., 144h), that is, data whose absolute value of the difference between t and all t_f is >144h and whose data have no fault records in the next 144h, are marked as negative samples indicating a healthy state. A total of 34,560 negative samples were marked.

[0033] Based on the time difference between the positive sample and its corresponding preset failure time point, a remaining lifespan label for the positive sample is generated; the remaining lifespan label for the negative sample is set to a value not less than the minimum guaranteed health duration; specifically, for all labeled samples, the regression task label RUL(t) is calculated, with the formula RUL(t) = t_f - t (t is the timestamp corresponding to the sample), where RUL(t) for positive samples ∈ [1h, 72h], and RUL(t) for negative samples ≥ 144h;

[0034] The remaining useful life labels of all samples are normalized to obtain the final regression labels; specifically, RUL(t) is normalized and mapped to the [0,1] interval using the min-max normalization method.

[0035] Dataset partitioning: The structured data labeled with regression task tags is divided into training set, validation set and test set in a ratio of 7:2:1 to avoid information leakage caused by random partitioning. Time series partitioning is adopted (the first 70% of the time is the training set, the middle 20% is the validation set and the last 10% is the test set).

[0036] Step S2: Based on the structured training dataset, extract statistical features and use a reliability model to calculate the dynamic health status value of the component. Then, fuse the structured training dataset and the dynamic health status value to obtain a composite health feature set for training.

[0037] In this embodiment, the extraction of statistical features specifically involves:

[0038] Time window setting: Set a sliding window according to the training mission cycle of the flight simulator (Δt=1h), with a window step size of 30min to ensure the continuity of features;

[0039] Feature calculation: Statistical analysis is performed on the continuous sensor data in each window to extract the mean (reflecting the central tendency of the data), variance (reflecting the dispersion of the data), kurtosis (reflecting the steepness of the data distribution), skewness (reflecting the symmetry of the data distribution), peak factor (ratio of peak value to effective value), and kurtosis (reflecting the impact characteristics) (6) types of statistical features, resulting in a total of (72) basic features (12 types of sensors × 6 types of statistical features);

[0040] The key to calculating the dynamic health status values ​​of flight simulator components using a reliability model lies in determining the time-varying failure rate λ(t) of the components. The specific determination process is as follows:

[0041] Model selection and preliminary analysis: For historical failure time data of specific components (such as motors), a non-parametric empirical distribution estimation method is first used for preliminary analysis to assess the data distribution pattern and provide a reference for parametric modeling. In particular, when the historical failure sample size is small, the fitting error of directly using a parametric model is large due to the limited statistical data. In this case, an empirical distribution function can be used first for estimation and correction. Commonly used empirical failure distribution functions include:

[0042] ;

[0043] ;

[0044] ;

[0045] in, Let N be the cumulative number of faults observed up to time t, and N be the total number of test pieces or the number of observed samples. Calculate the empirical fault distribution function using one of the above formulas. Therefore, the empirical reliability function can be obtained. The revised empirical reliable lifespan It can be solved Obtain, where y is the specified reliability level (e.g., 0.9);

[0046] Parametric Model Fitting and Validation: When sufficient failure data is available and preliminary analysis indicates that the data may follow a specific distribution, parametric modeling can be further employed. For example, based on engineering experience, it can be initially assumed that the component lifespan follows a Weibull distribution. The maximum likelihood estimation method can be used to fit the shape parameter β and scale parameter α of the Weibull distribution to obtain the theoretical failure rate function. and reliability function ;

[0047] To ensure the applicability of the model, a goodness-of-fit test (such as the KS test) is used to evaluate the degree of agreement between the Weibull model and historical data. If the test passes (p-value > 0.05), the model is accepted as the benchmark reliability model.

[0048] Baseline Model Determination: Based on the data and modeling objectives, select and determine the final baseline reliability model to be used for calculations.

[0049] If the parametric model path is adopted and the test is passed, the fitted Weibull model is used as the baseline model, and its functions λ(t) and R(t) are as described above;

[0050] If the sample size of the fault data is small (e.g., less than 30), or the parametric model test fails (p-value less than 0.05), then the calculated empirical reliability function will be used. As a benchmark, the time-varying failure rate λ(t) can be obtained from the numerical differentiation method. Export, or directly use empirical distributions for reliability assessment;

[0051] Calculate the dynamic health status value H(t) based on the determined time-varying failure rate λ(t) (whether from a parametric model or derived from an empirical distribution), and set a preset observation period (such as the component design life t_design). Calculate the maximum value λmax of λ(t) within this observation period, and then calculate the dynamic health status value H(t) according to the following formula:

[0052] ;

[0053] Where H(t)∈[0,1], the closer H(t) is to 1, the better the health of the component, and the closer it is to 0, the closer the component is to a faulty state;

[0054] The fusion processing of the structured training dataset and the dynamic health status values ​​is as follows:

[0055] Feature standardization: Perform Z-score standardization on (e.g., 72-dimensional) multidimensional statistical features (mean 0, variance 1) to eliminate the impact of dimensional differences on model training;

[0056] Feature fusion: The standardized statistical features (72 dimensions) are concatenated with the dynamic health status value H(t) (1 dimension), and task parameters (task duration, load level encoding, training times) and environmental parameters (temperature, humidity, voltage stability) are added to form a total of (8) dimensions of features, and finally a (81) dimension composite health feature set for training is formed.

[0057] Short window feature construction: In order to capture state changes on a shorter time scale, the statistical mean and standard deviation of the basic features based on the three most recent consecutive short time windows (such as 10-minute windows) are calculated and compressed by principal component analysis to generate a set of short window related features (46 dimensions) for real-time correction. This set of features is generated and stored synchronously with the 81-dimensional basic features during the training phase and is used for the calculation of dynamic correction coefficients in the subsequent reliability assessment model.

[0058] Step S3: Using the composite health feature set used for training, train a hybrid prediction model, which includes a data-driven prediction model and an adaptively corrected reliability assessment model.

[0059] In this embodiment, a hybrid prediction model is trained using a composite health feature set for training. The specific implementation process is as follows:

[0060] Training of the data-driven prediction model: The data-driven prediction model adopts an architecture of cascading random forest and long short-term memory network (LSTM), in which random forest is used for feature selection and long short-term memory network is used for time series prediction (including classification branch and regression branch).

[0061] The specific feature selection for random forest is as follows: model hyperparameter settings: n_estimators=200 (number of decision trees), max_depth=20 (maximum depth of a single decision tree), min_samples_leaf=2 (minimum number of samples in a leaf node), criterion='squared_error' (mean squared error is used as the splitting criterion);

[0062] Feature selection process: Input the 81-dimensional composite health feature set into the random forest model, calculate the importance score of each feature, and select the top 30 features with the highest importance scores (such as vibration acceleration variance, temperature kurtosis, H(t), task load, environmental humidity, etc.) as input features of the LSTM model to reduce the computational load of the model and improve the generalization ability.

[0063] The Long Short-Term Memory (LSTM) network used for time series prediction is as follows: the input sequence length is 200 (corresponding to the feature data of the most recent 200 1-hour windows), there are 2 hidden layers with 128 neurons in each layer, the dropout rate is 0.3 (to prevent overfitting), the activation function of the input layer and the hidden layer is tanh; the activation function of the output layer of the classification branch is sigmoid, and the activation function of the output layer of the regression branch is linear.

[0064] The optimizer is Adam, the initial learning rate is lr=1e-3, the batch size is 32, and the number of training epochs is 100.

[0065] Loss function selection: For classification tasks (faulty / non-faulty), the Binary Cross Entropy loss function is used, combined with FocalLoss to handle sample imbalance (positive samples: negative samples ≈ 1:4). The FocalLoss formula is:

[0066] ;

[0067] Where αt=0.75 (positive sample weight), γ=2 (difficulty coefficient), is the model's prediction probability of a positive sample (fault) for the t-th sample; the regression task (remaining lifetime prediction) uses the Huber loss function, with the objective value being the normalized RUL_norm(t) to avoid the impact of outliers on the loss.

[0068] Validation strategy: Time series cross-validation (rolling window) is used with a window size of 30 days. After each validation, the learning rate is adjusted (learning rate decay factor = 0.9). An early stopping strategy (patience = 10) is set, that is, training is stopped when the validation set loss (classification loss + regression loss, weight ratio 1:1) does not decrease for 10 consecutive rounds, and the optimal model parameters are saved.

[0069] The adaptively modified reliability assessment model is implemented in the following way:

[0070] Based on the aforementioned statistical characteristics, a neural network is used to generate dynamic correction coefficients;

[0071] The baseline reliability function is modified based on the dynamic correction coefficient to obtain the modified reliability function;

[0072] Specifically, the baseline reliability function is based on the baseline reliability function R(t) or the time-varying failure rate λ(t) determined in step S2. When determining maintenance requirements, the core evaluation index is the task reliability R(t0,Δt), which is defined as the conditional probability that a component will successfully execute for a future duration Δt (such as a single training task cycle) after it has been working for time t0.

[0073] ;

[0074] To improve the adaptability of reliability assessment to complex operating conditions, a dynamic correction coefficient kt is incorporated into the baseline reliability function, resulting in a corrected reliability function R′(t) = [R(t)]kt. Alternatively, the corrected failure rate function is λ′(t) = kt⋅λ(t), and the corresponding corrected mission reliability is:

[0075] ;

[0076] Wherein: when kt=1, the component is under baseline conditions, and the failure risk is consistent with the theoretical model; when kt>1, it indicates that the current environment is harsh (such as high temperature and high humidity) or the task load is too heavy, which leads to an increased failure risk; when kt<1, it indicates that the current operating conditions are relatively good (such as suitable temperature and humidity and stable load), and the failure risk is lower than the baseline level.

[0077] The dynamic correction coefficient kt is generated based on real-time running data through a lightweight fully connected neural network g(⋅). ​​The design and training of this network are as follows:

[0078] Network structure: A single fully connected neural network is used. The input layer receives the processed real-time feature vector, the hidden layer uses ReLU activation function, and the output layer uses Softplus activation function to ensure that the output value is positive.

[0079] Input features Xt include statistical features (or PCA compressed features) of the three most recent short time windows (such as 10-minute windows), current environmental parameters (temperature, humidity), time since the last preventive maintenance, and current task load indicators. These features together characterize the instantaneous stress and cumulative damage state of the component at time tt.

[0080] The network directly outputs a non-negative original correction value. To control the correction magnitude and enhance robustness, the final dynamic correction coefficient kt is obtained through a sensitivity adjustment and cutoff function:

[0081] ;

[0082] in: The original output of the neural network is α; α is the correction sensitivity coefficient (α=0.5 in this embodiment), used to adjust the influence of the network output on the final correction magnitude; kmin and kmmax are the lower and upper limits of the correction coefficient (set to 0.8 and 1.2 respectively in this embodiment), used to avoid over-correction due to abnormal input and ensure the stability of the evaluation; clip(⋅) is the truncation function, which restricts the input value to the interval [kmin, kmmax];

[0083] Neural Network Training Data Construction and Training: Label Construction: To train the network g(·), it is necessary to construct labels for the "ideal correction coefficients" corresponding to the historical data. The method is as follows:

[0084] a. Divide historical operational data into data segments based on fixed durations (e.g., 1 hour);

[0085] b. For each data segment, calculate its corresponding input feature vector Xt;

[0086] c. For the component to which the data segment belongs, use the dynamic health status value H(t) calculated in S2 to calculate the actual observed rate of decrease v_actual of H(t) during the data segment;

[0087] d. Based on the baseline reliability model, calculate the theoretical predicted rate of decrease of H(t) v_theory within the same time period;

[0088] e. Use the ratio of the actual descent rate to the theoretical descent rate (after smoothing) as the label for this data segment, i.e. = v_actual / v_theory, this value reflects the accelerating or mitigating effect of actual operating conditions on the health degradation of components;

[0089] Network training: using a large number of (Xt, The data pairs are composed of Adam optimizer (learning rate 1e-4) and MSE loss function to train the network g(·) for multiple training rounds (preferably 50).

[0090] Step S4: Obtain the real-time operating data of the flight simulator and perform preprocessing to obtain the composite health feature set to be predicted;

[0091] In this embodiment, real-time operational data from the flight simulator is acquired and preprocessed to obtain a composite health feature set to be predicted. The specific implementation process is as follows:

[0092] The flight simulator's real-time data acquisition system (TCP / IP transmission protocol) acquires sensor data, mission parameters, and environmental parameters in real time. The acquisition frequency is consistent with historical data (sensor data 1Hz-50Hz, mission parameters and environmental parameters once every 5 minutes), and the data transmission delay is ≤1s.

[0093] Real-time data preprocessing includes data cleaning, feature extraction, and real-time calculation of dynamic health status values. Specifically, data cleaning uses the same 3σ criterion as historical data to identify outliers, and replaces them in real time using the moving average method. For missing data (missing duration ≤ 1 min), the forward imputation method is used to ensure real-time data availability.

[0094] Feature extraction extracts statistical features (mean, variance, kurtosis, etc.) in real time according to a 1-hour time window, calculates the mean and standard deviation of the statistical features of the three most recent short windows (36 dimensions), and compresses the basic features of the three short windows (10 dimensions) through PCA (PCA model trained with historical data) to form 46-dimensional short window related features;

[0095] Real-time calculation of dynamic health status value: Based on the real-time running time t, λ(t) is calculated by substituting into the Weibull distribution failure rate formula, and then H(t) is obtained by combining λmax.

[0096] Construction of the feature set to be predicted: The statistical features (72 dimensions), short window related features (46 dimensions), real-time H(t), real-time task parameters and environmental parameters extracted in real time are concatenated and then standardized by Z-score (using the mean and variance of historical data) to form a 127-dimensional composite health feature set to be predicted, ensuring that it is consistent with the feature format of the training set.

[0097] Step S5: Input the composite health feature set to be predicted into the hybrid prediction model; the data-driven prediction model outputs a first prediction result, and the reliability assessment model outputs a second prediction result; wherein, the first prediction result includes the probability of future failure of the component and a first remaining lifetime estimate, and the second prediction result includes the maintenance requirement time based on the task reliability determination and a second remaining lifetime estimate;

[0098] In this embodiment, the composite health feature set to be predicted is input into the hybrid prediction model to obtain a first prediction result and a second prediction result. The specific implementation process is as follows:

[0099] Data-driven predictive model predictions:

[0100] Input processing: The 127-dimensional composite health feature set to be predicted is arranged in time series order to form an input sequence of length 200, which is then input into the trained LSTM model.

[0101] First prediction result output:

[0102] Future failure probability: The classification branch outputs the failure probability P_fail(t,72h) within the next 72 hours through the Softmax classifier. For example, if a certain motor component outputs P_fail=0.32, it indicates that the probability of failure within the next 72 hours is 32%.

[0103] First remaining lifetime estimation: The regression branch outputs the normalized remaining lifetime prediction value RUL_norm_pred, and the actual remaining lifetime E[RUL] is calculated by inverse normalization. The inverse normalization formula is E[RUL]=RUL_norm_pred×(RUL_max-RUL_min)+RUL_min. At the same time, the confidence level σ is calculated (σ=±20h in this embodiment). For example, if the regression branch outputs RUL_norm_pred=0.061, after inverse normalization E[RUL]=0.061×(360-1)+1≈23.9h, which is rounded to 24h. The final output E[RUL]=240h (after correction based on the model's time series prediction trend) and the confidence level σ=±20h indicates that the remaining lifetime of this component is approximately 220h-260h.

[0104] Reliability assessment model predictions:

[0105] Dynamic correction coefficient calculation: Extract 50-dimensional input features (46-dimensional short window related features + 2-dimensional environmental parameters + 1-dimensional time since the last maintenance + 1-dimensional current task load index) from the set of composite health features to be predicted, input them into the trained fully connected neural network, and after ReLU activation and softplus output, obtain the current dynamic correction coefficient k_t=1.05.

[0106] Based on the modified reliability function, the reliability of a component continuing to execute a task for a future duration of Δt (Δt = 24 hours, a single training task cycle) after operating for a time t0 (e.g., currently running for 3500 hours) is calculated. In actual calculations, since k_t can be considered approximately constant within a short time window [t0, t0 + Δt] (or represented by the value k_{t0} at time t0), the modified reliability function can be simplified to:

[0107] ;

[0108] Where R(t) is the baseline reliability function determined in step S2;

[0109] When the reliability of the task When the reliability threshold γ is lower than the preset threshold, the determination time t0 is the time when maintenance is required; for example, the preset reliability threshold γ = 0.9 (set according to the importance of the training task, core task γ = 0.95, ordinary task γ = 0.9), because The current time t0=3500h is determined to be the time when maintenance is required;

[0110] Second remaining lifetime estimation: Based on the modified reliability function R'(t), solve for the time t_0.5 (median lifetime) corresponding to R'(t)=0.5, that is, substitute k_t=1.05, and calculate t_0.5=3730h. The second remaining lifetime estimation t_0.5−T=230h;

[0111] Output format: The first and second prediction results are output in JSON format, including prediction timestamp, component number, P_fail, E[RUL], confidence level σ, k_t, maintenance requirement time, and key information of the second remaining lifetime estimate;

[0112] Step S6: Perform a fusion analysis on the first prediction result and the second prediction result, and process them according to the preset decision rules to obtain the maintenance work order for the flight simulator components;

[0113] In this embodiment, the fusion analysis includes consistency verification and arbitration of the probability of future failure and the timing of maintenance needs; specifically: consistency verification: compare the first remaining lifetime estimate (240h) and the second remaining lifetime estimate (230h), the difference between the two is 10h, which is less than the threshold of 20h (the preset difference threshold is 10% of E[RUL]), and the prediction results are determined to be consistent;

[0114] Arbitration mechanism: If the difference between the two is greater than or equal to the threshold, the final remaining lifetime is determined by weighted fusion. The weights are set based on the model training accuracy (0.6 for data-driven prediction model and 0.4 for reliability assessment model). For example, if the first remaining lifetime is 200h and the second remaining lifetime is 150h, the final remaining lifetime = 200 × 0.6 + 150 × 0.4 = 180h.

[0115] The decision-making rules include a hysteresis triggering mechanism and a multi-level probability threshold rule; specifically: multi-level probability threshold determination: the current P_fail=0.32, which satisfies the condition 0.2≤P_fail<0.5, corresponding to the planned maintenance work order level;

[0116] Hysteresis triggering mechanism: The component is predicted twice consecutively (1 hour apart). The second prediction has P_fail=0.35 and k_t=1.07, which still meets the conditions for planned maintenance work order, thus triggering the work order generation (to avoid false triggering caused by single noisy data).

[0117] Maintenance work order generation includes work order content and work order issuance; specifically:

[0118] Work order content: The automatically generated work order includes the part number (motor M-003), risk category (P_fail=0.35, RUL=235h), dynamic correction factor k_t=1.07, maintenance requirement time (3500h), 3 main contributing characteristics (abnormal vibration acceleration variance, high ambient temperature, excessive cumulative running time), recommended maintenance measures (shutdown to check motor bearings, replace lubricating oil, calibrate vibration sensor), and recommended maintenance window (within the next 72 hours).

[0119] Work order issuance: Through the maintenance management system of the Civil Aviation Training Center, work orders are pushed to the maintenance personnel's terminals (mobile APP + computer client) and SMS reminders are sent to ensure that maintenance personnel receive them in a timely manner;

[0120] Step S7, after obtaining the maintenance work order for the flight simulator components, also includes the prediction of spare parts demand, which is implemented as follows:

[0121] Based on historical failure data, a failure interval is divided using an early failure period end time model to determine the current failure stage of the target aircraft material. The failure stage includes the early failure period and the wear-out failure period. Specifically, the early failure period is characterized by a high failure rate that decreases rapidly over time. Using the average residual number method, the failure time is divided into k intervals from smallest to largest, with the time for each interval being... The corresponding failure rate is ,when At that time, the early failure period ends, and the end time is the [number]th [time]. The time point in the interval, the end time of the early fault period is The sum of the time lengths of the intervals is expressed as: ;

[0122] Based on the end time of the early failure period, the failure interval is divided by historical failure data to determine the failure period of the aircraft material.

[0123] If the system is determined to be in a wear-out failure period, the wear-out failure period task demand prediction model is used to calculate the predicted spare parts demand. The wear-out failure period task demand prediction model is as follows:

[0124] ;

[0125] in, For the first Type of aircraft materials The equipment is in the prediction cycle Forecasted demand for spare parts within the country. To predict the start time, This refers to the total number of installed units currently in the wear-out / failure period. For the first Type of aircraft materials The failure rate function of the equipment during the wear and tear period is obtained by fitting historical data (the method of obtaining it can be the time-varying failure rate λ(t), which will not be described in detail here).

[0126] If the system is determined to be in the early failure phase, the early failure phase consumption prediction model is used to calculate the predicted spare parts demand. The early failure phase consumption prediction model is as follows:

[0127] ;

[0128] in, For the first Type of aircraft materials Predicted spare parts demand for the equipment during the early failure period. The planned flight mission time for the target aircraft material within the forecast period. For the first The baseline statistics / standard mission time of the equipment during the early failure period. For historical sample index, indicating the first... Historical equipment samples, For the first The actual flight mission time performed by the historical equipment during the early failure period. For the first The cumulative working time of each piece of historical equipment from the time it was put into use to the current statistical time. This represents the fault decay time constant during the early failure phase. For the first Type of aircraft materials Total number of available historical samples of the equipment;

[0129] Step S8, after obtaining the maintenance work order for the flight simulator components, also includes: updating and training the hybrid prediction model using the new data collected after maintenance;

[0130] In this embodiment, new data acquisition: After maintenance is completed, the post-maintenance operation data (sensor data, health status value, maintenance effect) of the component is collected and marked as new training samples (if the fault is eliminated after maintenance, it is marked as a negative sample, and RUL(t) is calculated and normalized according to the actual fault-free interval; if there is still an anomaly, it is marked as a positive sample, and RUL(t) is calculated and normalized according to the time from the fault time). At the same time, the 50-dimensional dynamic correction coefficient input features and actual correction effect corresponding to the new sample are recorded.

[0131] Model Iterative Training: Add new samples to the training set, incrementally train the hybrid prediction model according to the original training parameters, update the weights of the data-driven prediction model and the dynamically corrected coefficient fully connected neural network, and the iteration cycle is 1 week (update the model once a week) to ensure that the model adapts to the state changes after component aging and maintenance and maintains high prediction accuracy.

[0132] In summary, by systematically fusing and preprocessing historical operational data, sensor time-series data, fault records, and environmental parameters of flight simulators, a structured composite health feature set is constructed. Key influencing factors are screened using random forests, and a data-driven prediction model is established by combining LSTM networks and time-series cross-validation. Then, a reliability assessment model, optimized through a Weibull distribution reliability model and adaptive dynamic correction coefficients, is constructed and calibrated in real time. This ultimately enables comprehensive prediction of component failure probability, remaining lifespan, and maintenance timing, accurately outputting tiered maintenance work orders and spare parts demand plans. This approach not only overcomes the limitations of traditional threshold methods, which rely on single indicators, suffer from delayed warnings, and have high false alarm / false negative rates, but also significantly improves the accuracy and timeliness of predictive maintenance under complex operating conditions through hybrid model fusion and closed-loop iterative learning. Furthermore, it achieves a leap from passive response maintenance to proactive predictive maintenance, meeting the "full-cycle, high-precision, low-cost" maintenance management requirements of flight simulators. This helps civil aviation training centers accurately control equipment health status, optimize maintenance resources and spare parts inventory allocation, reduce the risk of unplanned downtime, ensure the continuous and stable progress of flight training missions, and achieve a dual improvement in maintenance efficiency and operational reliability.

[0133] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related explanations of the methods described above can be found in the corresponding processes in the foregoing system embodiments, and will not be repeated here.

[0134] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The programs corresponding to the software modules and method steps can be placed in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the invention.

[0135] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0136] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0137] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for predictive maintenance of flight simulators based on machine learning, characterized in that, The method includes the following steps: Historical operational data of the flight simulator is acquired and preprocessed to obtain a structured training dataset. The preprocessing includes data cleaning and data labeling. Based on the structured training dataset, statistical features are extracted and a reliability model is used to calculate the dynamic health status value of the component. The structured training dataset and the dynamic health status value are fused to obtain a composite health feature set for training. Using the composite health feature set used for training, a hybrid prediction model is trained, which includes a data-driven prediction model and an adaptively corrected reliability assessment model. The real-time operating data of the flight simulator is acquired and preprocessed to obtain the composite health feature set to be predicted; The composite health feature set to be predicted is input into the hybrid prediction model; the data-driven prediction model outputs a first prediction result, and the reliability assessment model outputs a second prediction result; wherein, the first prediction result includes the probability of future failure of the component and a first remaining lifetime estimate, and the second prediction result includes the maintenance requirement time based on the task reliability determination and a second remaining lifetime estimate; The first prediction result and the second prediction result are fused and analyzed, and processed according to the preset decision rules to obtain the maintenance work order of the flight simulator components; The adaptively modified reliability assessment model is implemented in the following way: Based on the aforementioned statistical characteristics, a neural network is used to generate dynamic correction coefficients; The baseline reliability function is modified based on the dynamic correction coefficient to obtain the modified reliability function; The maintenance requirement time based on task reliability is determined by the following method: based on the modified reliability function, the task reliability of the component continuing to execute a task for a future duration of Δt after working time T is calculated. When the reliability of the task is lower than the preset reliability threshold γ, the determination time T is the time of the maintenance requirement; The data-driven prediction model adopts a cascaded architecture of random forest and long short-term memory network, where random forest is used for feature selection and long short-term memory network is used for time series prediction.

2. The method for predictive maintenance of flight simulators based on machine learning according to claim 1, characterized in that, The data annotation includes: Data samples that occur within a time window from a first duration to a second duration prior to a preset fault time point are marked as positive samples indicating fault precursors, wherein the first duration is longer than the second duration. Data samples that are all within a time interval greater than the first duration from any of the preset fault time points, and that have no fault records within the subsequent preset minimum guaranteed health duration, are marked as negative samples indicating health status. Based on the time difference between the positive sample and its corresponding preset failure time point, a remaining service life label for the positive sample is generated; the remaining service life label for the negative sample is set to a value not less than the minimum guaranteed health duration. The remaining useful life labels of all samples are normalized to obtain the final regression labels.

3. The method for predictive maintenance of a flight simulator based on machine learning according to claim 1, characterized in that, The calculation of the dynamic health status value of the flight simulator components using a reliability model specifically involves: calculating the time-varying failure rate λ(t) based on the Weibull distribution model, and calculating the dynamic health status value H(t) according to the following formula: H(t)=1 λ(t) / λmax; Wherein, λmax is the maximum failure rate within the preset observation period.

4. The method for predictive maintenance of flight simulators based on machine learning according to claim 1, characterized in that, The fusion analysis includes consistency verification and arbitration of the probability of future failures and the timing of maintenance requirements.

5. The method for predictive maintenance of a flight simulator based on machine learning according to claim 1, characterized in that, The decision rules include a hysteresis triggering mechanism and a multi-level probability threshold rule.

6. The method for predictive maintenance of a flight simulator based on machine learning according to claim 1, characterized in that, After obtaining the maintenance work orders for the flight simulator components, the process also includes forecasting the required quantity of spare parts. The implementation method is as follows: Based on historical fault data, the fault interval is divided using an early fault period end time model to determine the current fault stage of the target aircraft material; the fault stage includes the early fault period and the wear-out fault period. If the system is determined to be in a wear-out failure period, the wear-out failure period task demand prediction model is used to calculate the predicted spare parts demand. The wear-out failure period task demand prediction model is as follows: ； in, For the first Type of aircraft materials The equipment is in the prediction cycle Forecasted demand for spare parts within the country. To predict the start time, This represents the total number of installed units currently in the wear and tear / failure period. For the first Type of aircraft materials The failure rate function of the equipment during the wear and tear period is obtained by fitting historical data. If the system is determined to be in the early failure phase, the early failure phase consumption prediction model is used to calculate the predicted spare parts demand. The early failure phase consumption prediction model is as follows: ； in, For the first Type of aircraft materials Predicted spare parts demand for the equipment during the early failure period. The planned flight mission time for the target aircraft material within the forecast period. For the first The baseline statistics / standard mission time of the equipment during the early failure period. For historical sample index, indicating the first... Historical equipment samples, For the first The actual flight mission time performed by the historical equipment during the early failure period. For the first The cumulative working time of each piece of historical equipment from the time it was put into use to the current statistical time. This represents the fault decay time constant during the early failure phase. For the first Type of aircraft materials Total number of available historical samples of the equipment.

7. The method for predictive maintenance of a flight simulator based on machine learning according to claim 1, characterized in that, After obtaining the maintenance work order for the flight simulator components, the process also includes: updating and training the hybrid prediction model using the new data collected after maintenance.

Images