Intelligent detection method and system for aircraft accelerator control state

Abstract

The invention belongs to the technical field of aviation safety and flight data analysis, and discloses an intelligent detection method and system for the control state of an aircraft accelerator, and the method comprises the steps: obtaining the operation parameter time sequence data of an engine of an aircraft; the engine rotating speed and the fuel oil flow are screened out to serve as key monitoring parameters by comparing significant differences of correlation coefficients of all the operating parameters and the air inlet pressure in the normal operating state and the coarse and high operating state; and inputting the key parameters into a machine learning classification model formed by principal component analysis and logistic regression, and carrying out three classifications of'normal manipulation ', 'fast push manipulation' and'coarse and fierce manipulation '. The method further comprises the improvement of dynamically adjusting classification boundaries according to external environment parameters such as atmospheric temperature and the like, and performing real-time early warning when dangerous manipulation is detected. According to the method, hardware transformation is not needed, accurate, robust and automatic classification evaluation and active safety early warning of accelerator control are achieved, and the flight training quality and the safety level are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of aviation safety and flight data analysis technology, specifically relating to an intelligent detection method and system for aircraft throttle control status. Background Technology

[0002] In the field of general aviation, especially for primary trainer aircraft such as the Cessna 172R, the pilot's control of the throttle and air-fuel mixture directly affects engine performance, fuel economy, and even flight safety. "Rough handling," which involves rapidly and drastically changing the throttle position within a short period, can cause engine lag due to a momentary imbalance in the air-fuel mixture, or even engine stall in mid-air, posing a serious safety hazard.

[0003] However, many primary trainer aircraft's onboard flight data recording systems (such as those recording data via SD cards) do not directly record the specific physical angles of the throttle lever and mixture lever due to cost and structural limitations. Existing monitoring methods typically rely on manual observation by flight instructors or post-flight data playback analysis, subjectively evaluating handling quality by indirectly observing changes in parameters such as engine speed (RPM), intake pressure (MAP), and fuel flow (FFLOW). This approach has two major problems: first, the evaluation standards are inconsistent, relying on expert experience and being highly subjective; second, without direct control stick angle data, it is difficult to establish an accurate and automated model to distinguish between "normal handling," "rapid-push handling" (between normal and aggressive handling), and "aggressive handling."

[0004] Therefore, there is an urgent need for an intelligent detection method and system that can accurately, objectively, and automatically classify pilots' throttle control behavior using only conventional engine operating parameters without relying on joystick angle sensors. Summary of the Invention

[0005] The present invention aims to at least partially solve the aforementioned technical problems. Therefore, the present invention aims to provide an intelligent detection method and system for aircraft throttle control status. This method does not rely on direct joystick angle data, can achieve accurate three-dimensional classification of throttle control behavior, and has real-time early warning capabilities, thereby improving flight training quality and flight safety.

[0006] To achieve the above objectives, the present invention provides an intelligent detection method for aircraft throttle control status, comprising the following steps:

[0007] First, acquire time-series data of multiple engine operating parameters, including intake pressure (MAP), engine speed (RPM), and fuel flow rate (FFLOW), output by the aircraft's flight data recording system.

[0008] Then, based on pre-set normal operation state datasets and coarse operation state datasets, the key monitoring parameters most relevant to changes in operation state are selected by comparing the differences in correlation coefficients between various engine operating parameters and intake pressure parameters under the two states. Under coarse operation, the engine response becomes distorted, causing a significant decrease in the correlation between certain parameters (such as RPM and FFLOW) and MAP, which represents the operation intention. Utilizing this "disconnection" characteristic, the most sensitive indicators can be effectively screened. This invention identifies engine speed (RPM) and fuel flow rate (FFLOW) as key monitoring parameters.

[0009] Next, the time series data of the selected key monitoring parameters are preprocessed, such as standardized, to eliminate the influence of units and facilitate model processing.

[0010] Finally, the preprocessed key monitoring parameters are input into a pre-trained machine learning classification model, which directly outputs a three-class classification result of the current throttle control state as "normal control", "fast push control" or "aggressive control".

[0011] Furthermore, to improve the accuracy of the screening, the steps for screening the key monitoring parameters most relevant to changes in operating state can be as follows: calculate the Pearson correlation coefficient between each candidate engine operating parameter and intake pressure (MAP) parameter in the normal operating state dataset and the rough operating state dataset, respectively, to obtain the first correlation coefficient value and the second correlation coefficient value; then calculate the difference between the two correlation coefficient values, and select the parameter whose difference exceeds a predetermined threshold as the key monitoring parameter.

[0012] Furthermore, the data preprocessing step preferably includes standardizing the key monitoring parameter data, and the calculation formula is as follows: This helps accelerate model convergence and improve performance.

[0013] Furthermore, the machine learning classification model preferably employs a combined model, including a principal component analysis (PCA) module and a logistic regression classification module. The PCA module is responsible for feature extraction and dimensionality reduction of key monitoring parameters in two or higher dimensions, while the logistic regression classification module uses the extracted principal component features to calculate the probability of belonging to three types of manipulation through three trained classifiers, and outputs the most probable result.

[0014] As a significant improvement of this invention, the method may further include a dynamic adaptive step. This involves acquiring the aircraft's current atmospheric temperature (OAT) or density altitude calculated from air pressure as external environmental parameters, and dynamically adjusting the classification boundary of the classification model based on these parameters. For example, a preset compensation function f(E) is used to adjust the bias term b in the logistic regression model, resulting in badjusted.adj =b+f(E). This improvement enables the model to adapt to changes in engine response under different flight environments, significantly improving the robustness of detection.

[0015] As another important improvement of this invention, this method can be deployed on airborne equipment for real-time execution. When a "aggressive" throttle operation is detected, or when "rapid thrust" is detected multiple times in a short period of time, a real-time warning signal can be generated for the pilot through in-cockpit warning devices (such as screen flashing, audible prompts, etc.). This elevates this invention from a post-event analysis tool to a proactive safety defense system.

[0016] Accordingly, the present invention also provides an intelligent detection system for aircraft throttle control status. This system includes a data acquisition module, a processing module, and a classification module, the functions of which correspond to the steps of the method described above. Preferably, the system further includes an alarm module for triggering real-time warnings when dangerous control actions are detected.

[0017] The beneficial effects of this invention are as follows:

[0018] This invention proposes a feature-filtering method that analyzes the differences in the correlation between parameters under different control states, enabling precise monitoring without expensive hardware modifications. It can identify not only dangerous "rough maneuvers" but also potentially risky "rapid thrust maneuvers," providing richer and more refined data for flight training and quality assessment.

[0019] The dynamic adaptive classification boundary introduced in this invention enables the model to adapt to different flight environments, avoiding misjudgments or omissions caused by environmental changes. By adding a real-time warning function, the pilot can be alerted the moment a dangerous maneuver occurs, preventing accidents from happening. Attached Figure Description

[0020] Figure 1 This is a diagram illustrating the working principle of a fuel regulator.

[0021] Figure 2 This is a comparison chart of the correlation coefficients between the characteristic variables and the target variable under normal and aggressive manipulation.

[0022] Figure 3 This is a graph showing the results of Bootstrap confidence interval calculation.

[0023] Figure 4 This is a diagram illustrating the working principle of the identification model based on principal component analysis.

[0024] Figure 5 This is a flowchart of the process for building a robust manipulation recognition model.

[0025] Figure 6 This is a structural block diagram of the intelligent detection system in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] It should be understood that, and also noted, in the embodiments, the functions / actions may appear in a different order than those shown in the figures. For example, depending on the functions / actions involved, they may actually be performed substantially concurrently, or sometimes the two figures shown consecutively may be performed in reverse order.

[0028] This embodiment uses a Cessna 172R aircraft (equipped with a Lycoming 10-360-L2A engine) as an example to illustrate in detail the intelligent detection method for aircraft throttle control status of the present invention.

[0029] Step 1: Data Acquisition and Dataset Preparation

[0030] The flight parameters recorded at a frequency of 1Hz are obtained from the aircraft's flight data recording system (such as SD card data exported from the Garmin G1000 system). These parameters include at least: intake pressure (MAP), engine speed (RPM), fuel flow rate (FFLOW), engine exhaust temperature (EGT), and lubricating oil pressure (OilP).

[0031] By analyzing known engine shutdown event data and manual annotation of a large amount of flight training data by expert instructors, two basic datasets were established:

[0032] Normal handling state dataset: contains smooth, gentle throttle control segments that conform to standard operating procedures (SOP).

[0033] Rough and aggressive handling state dataset: contains segments of rapid, large-amplitude throttle manipulation that cause engine lag or stall.

[0034] Step 2: Screening of key monitoring parameters based on correlation differences

[0035] This step aims to find the parameter that is most sensitive to changes in the manipulation state.

[0036] Determine the target variable: Select intake pressure (MAP) as the target variable Y, because it can most directly reflect the pilot's intention and magnitude when pushing or pulling the throttle lever.

[0037] Determine candidate feature variables: RPM, FFLOW, EGT, OilP, etc. are selected as candidate feature variables X (as shown in Table 1).

[0038] Table 1. Definitions of Candidate Feature Variables and Target Variables

[0039] Variable symbol Variable name Full English name / unit Variable Roles / Descriptions MAP Intake pressure Manifold AbsolutePressure(inHg) The target parameters are used for feature selection, and their changes directly represent the pilot's control intentions. RPM Engine speed Revolutions Per Minute The key characteristic parameters that were ultimately selected showed significant differences in their correlation with MAP under different manipulation conditions. FFLOW Fuel flow Fuel Flow(GPH-Gallons PerHour) The key characteristic parameters that were ultimately selected showed significant differences in their correlation with MAP under different manipulation conditions. EGT Exhaust temperature Exhaust Gas Temperature (°C) Parameters that are candidates but not selected are excluded in the feature selection step because their correlation differences are not significant. OilP Lubricating oil pressure Oil Pressure (PSI) Parameters that are candidates but not selected are excluded in the feature selection step because their correlation differences are not significant. OilT Lubricating oil temperature Oil Temperature (°C) Parameters that are candidates but not selected are excluded in the feature selection step because their correlation differences are not significant. Y Throttle control status Throttle ManipulationState The target variable is the classification result that the detection model of this invention ultimately needs to output. Its values ​​include: • coarse manipulation • fast pushing manipulation • normal manipulation

[0040] Using the normal manipulation state dataset, calculate the Pearson correlation coefficient r between each candidate feature variable X and the target variable Y (MAP). 正常 The calculation formula is:

[0041]

[0042] Where: x i y is the i-th sample value of feature variable X. i It is the i-th sample value of the target variable Y. These are the means of variables X and Y, respectively. n is the total number of samples.

[0043] Using the coarse-grained manipulation state dataset, the same correlation coefficient r was calculated. 粗猛 .

[0044] Analyze the differences and filter: Compare the results (e.g.) Figure 2 As shown), the findings are as follows:

[0045] FFLOW:r 正常 =0.82, r 粗猛 =0.61, the difference is significant.

[0046] RPM: r 正常 =0.61, r 粗猛 =0.19, a huge difference.

[0047] EGT: r 正常 =0.54, r 粗猛 =0.16, there is a difference, but the response of this parameter is lagging.

[0048] OilP: r 正常 =0.89, r 粗猛 =0.23, a huge difference, but it is the result of manipulation, not the direct cause.

[0049] Confidence interval analysis using Bootstrap (e.g.) Figure 3 As shown in the figure, the difference between FFLOW and RPM is statistically significant. This is further validated by the working principle of the fuel regulator (e.g., Figure 1Rapidly accelerating directly causes a surge in FFLOW, which RPM cannot keep up with, and is the root cause of fuel-rich engine stalling. Therefore, fuel flow (FFLOW) and engine speed (RPM) were ultimately chosen as the key monitoring parameters.

[0050] Step 3: Data Preprocessing

[0051] The selected FFLOW and RPM time series data are standardized to eliminate differences in their units and numerical ranges. The calculation formula is as follows:

[0052]

[0053] Among them, X ij For the original data points, Z ij For standardized data points, Let be the mean of the features in column j. Let be the standard deviation of the j-th column feature.

[0054] Step 4: Model Building and Training

[0055] A combined model using PCA and logistic regression (e.g.) Figure 4 , Figure 5 (As shown).

[0056] Dimensionality reduction (PCA):

[0057] The standardized FFLOW and RPM data are used to construct an n×2 matrix X'.

[0058] Calculate its covariance matrix C:

[0059] in: It is a 2×2 matrix, where 2 is the number of feature data and n is the total number of data samples.

[0060] Calculate the eigenvalues ​​and eigenvectors of the covariance matrix C.

[0061] Projecting the data X' onto the principal component direction determined by the eigenvectors yields the dimensionality-reduced one-dimensional or two-dimensional feature data P.

[0062] Classifier training (logistic regression):

[0063] Prepare three types of labeled data: normal operation (label 2), fast push operation (label 1), and rough operation (label 0).

[0064] Train three one-to-many (One-vs-Rest) logistic regression classifiers. Each classifier is used to predict the probability that a sample belongs to its corresponding class.

[0065] The probability calculation formula for the k-th classifier is:

[0066]

[0067] Using gradient descent, the optimal parameters for each classifier are iteratively solved with the goal of minimizing the loss function. .

[0068] The formula for calculating the loss function is:

[0069]

[0070] When the model predicts value Compared with the actual label value The greater the difference, the higher the value output by the loss function. The larger.

[0071] Minimize the loss function using gradient descent. Gradient descent involves repeated adjustments through numerous iterative calculations. , , The value of causes the value of the loss function to gradually decrease. The formula for gradient descent is:

[0072]

[0073] in The learning rate determines the step size for parameter updates; here, the learning rate is set to 0.05. It is a loss function pair The partial derivatives are calculated as follows:

[0074] , ,

[0075] After training, three sets of model coefficients were obtained.

[0076] Step 5: Real-time detection, adaptive adjustment, and early warning

[0077] Real-time detection process:

[0078] The trained model is then deployed to an airborne computer or a dedicated monitoring device.

[0079] The system acquires FFLOW and RPM data in real time at a frequency of 1Hz.

[0080] Standardization and PCA dimensionality reduction are performed sequentially.

[0081] The dimensionality-reduced features are input into three logistic regression classifiers, and the probabilities P0, P1, and P2 of data x1 belonging to coarse-force manipulation data are calculated respectively:

[0082]

[0083]

[0084]

[0085] Select the category with the highest probability as the manipulation state at the current moment and output it.

[0086] Dynamic adaptive classification boundary:

[0087] The system also acquires atmospheric temperature (OAT) parameters.

[0088] Set a reference temperature T base (e.g., 20°C).

[0089] Define a compensation function, such as a linear function: , where k is a preset sensitivity coefficient.

[0090] Before calculating P0, P1, and P2, the bias terms of each classifier are adjusted:

[0091]

[0092] For example, the higher the temperature, the worse the engine response may be, so the boundary between "normal" and "fast push" can be appropriately relaxed.

[0093] like Figure 6 As shown, the intelligent detection system 100 includes a data acquisition module 101, a processing module 102, a classification module 103, and an alarm module 104. The processing module 102 and the classification module 103 can be integrated into a central processing unit (CPU).

[0094] When the classification module 103 outputs "rough and violent manipulation" (label 0), the alarm module 104 is immediately triggered.

[0095] The alarm module 104 sends a command to the multifunction display (MFD) in the cockpit to display a red visual warning “ROUGH THROTTLE” or to broadcast a voice prompt “Warning, rough throttle operation” through the cockpit audio system.

[0096] In addition, a counter can be set up so that if three or more "fast thrust maneuvers" (label 1) are detected within 10 seconds, a yellow alert level warning "CAUTION: FAST THROTTLE" is triggered to remind the pilot to pay attention to the smoothness of the maneuvers.

[0097] Through the above embodiments, the present invention not only achieves accurate classification of throttle control states with F1 scores all above 0.9 (as shown in Table 2), but also significantly improves the system's practicality and flight safety assurance capabilities by introducing dynamic adaptation and real-time warning functions.

[0098] Table 2. Model Classification Performance Validation Table

[0099] Category Data volume (test set) Precision Recall F1 score Rough manipulation 500 0.96 0.94 0.95 Quick push operation 1,500 0.97 0.98 0.98 Normal operation 8,000 0.99 0.99 0.99

[0100] Where F1 = 2 * (precision * recall) / (precision + recall)

[0101] Those skilled in the art should understand that the above embodiments are merely examples, and various modifications and variations can be made without departing from the spirit and scope of the invention. For example, the machine learning classification model can be replaced with SVM, neural networks, etc.; other statistical indicators can also be used to screen key monitoring parameters. All such modifications and variations should fall within the protection scope of this invention.

Claims

1. An intelligent detection method for aircraft throttle control status, characterized in that, Includes the following steps: Acquire time-series data of multiple engine operating parameters, including intake pressure, engine speed, and fuel flow, from the aircraft's flight data recording system. Based on a preset normal operating state dataset and a rough operating state dataset, by comparing the differences in the correlation coefficients between each engine operating parameter and the intake pressure parameter under the two states, the key monitoring parameters most relevant to the changes in operating state are selected, wherein the key monitoring parameters include engine speed and fuel flow. Data preprocessing is performed on the time series data of the key monitoring parameters; The preprocessed key monitoring parameters are input into a pre-trained machine learning classification model, which outputs a three-classification result of the current throttle control state.

2. The method according to claim 1, characterized in that, The specific steps for selecting the key monitoring parameters most relevant to changes in control status are as follows: Calculate the Pearson correlation coefficients between each engine operating parameter and the intake pressure parameter in the normal operating state dataset and the coarse and rough operating state dataset respectively, and obtain the first correlation coefficient value and the second correlation coefficient value. Calculate the difference between the first correlation coefficient value and the second correlation coefficient value; Engine operating parameters whose differences exceed a predetermined threshold are selected as the key monitoring parameters.

3. The method according to claim 1, characterized in that, The data preprocessing step includes standardizing the time-series data of the key monitoring parameters, and the calculation formula is as follows: Among them, X ij For the original data points, Z ij For standardized data points, Let be the mean of the features in column j. Let be the standard deviation of the j-th column feature.

4. The method according to claim 1, characterized in that, The machine learning classification model includes a principal component analysis module and a logistic regression classification module. The principal component analysis module is used to extract features and reduce the dimensionality of the preprocessed key monitoring parameters to obtain principal component features. The logistic regression classification module receives the principal component features and calculates the probability that the input data belongs to normal manipulation, fast push manipulation, and coarse manipulation, respectively, and uses the one with the highest probability as the final classification result.

5. The method according to claim 1, characterized in that, Also includes: Obtain the aircraft's current atmospheric temperature or density altitude as an external environmental parameter; Based on the external environment parameters, the classification boundary used to distinguish different manipulation states in the machine learning classification model is dynamically adjusted.

6. The method according to claim 5, characterized in that, The specific step of dynamically adjusting the classification boundary is as follows: adjusting the bias term b in the machine learning classification model to b adj =b+f(E), where E is the external environment parameter and f(E) is the preset compensation function.

7. The method according to claim 1, characterized in that, The method is executed in real time during aircraft flight and also includes: When the throttle is detected to be operated roughly or repeatedly as being operated rapidly, a real-time warning signal is generated to the pilot via the warning device in the cockpit.

8. The method according to claim 7, characterized in that, The real-time warning signal is a visual alarm or an auditory alarm.

9. An intelligent detection system for aircraft throttle control status, characterized in that, include: The data acquisition module is used to acquire time-series data of multiple engine operating parameters, including intake pressure, engine speed, and fuel flow, from the aircraft's flight data recording system. The processing module is configured as follows: Based on a preset normal operating state dataset and a rough operating state dataset, key monitoring parameters are selected by comparing the differences in the correlation coefficients between each engine operating parameter and the intake pressure parameter under the two states. The key monitoring parameters include engine speed and fuel flow. Data preprocessing is performed on the time series data of the key monitoring parameters; The classification module receives preprocessed key monitoring parameters and outputs a three-class classification result of the current throttle control state through a pre-trained machine learning classification model.

10. The system according to claim 9, characterized in that, Also includes: The alarm module, connected to the processing module and the classification module, is used to drive the alarm device in the cockpit to generate a real-time warning signal when the output result of the classification module is coarse control during flight.

Images