A helicopter maneuvering flight state recognition method
Patent Information
- Application Number
- CN202510418121.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-03
- Publication Date
- 2026-10-16
AI Technical Summary
[0004]本发明技术方案提供一种针对直升机机动飞行状态的识别方法，该方法结合直升机机动飞行状态的典型特征和神经网络模糊识别的优势，能够明显提高机动飞行状态识别的准确率，并降低神经网络的识别难度，以解决直升机机动飞行状态动作复杂，飞参中对应参数变化多样，目前的飞行状态识别方法在机动状态识别准确率不高的问题
[0041] This invention provides a method for identifying the maneuvering flight state of a helicopter. This method combines the typical features of the helicopter's maneuvering flight state with the advantages of neural network fuzzy recognition, which can significantly improve the accuracy of maneuvering flight state identification and reduce the recognition difficulty of neural networks.
Smart Images
Figure CN122925923A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of helicopter flight status recognition technology, and particularly relates to a method for recognizing helicopter maneuvering flight status. Background Technology
[0002] With the continuous updating of helicopter flight status identification methods, the main methods currently available include logic tree analysis, image analysis, and neural network methods. The logic tree analysis method decomposes the helicopter status layer by layer using a logic tree based on flight parameters; the image analysis method mainly analyzes flight video images of the helicopter to obtain individual flight statuses; and the neural network method uses backpropagation (BP) neural networks or deep learning neural networks to identify and analyze flight parameter data to obtain the flight status.
[0003] However, current methods primarily target steady-state flight and cannot accurately identify maneuvering flight states. Helicopter flight states reflect their operational status and are crucial for guiding design, operation, and maintenance, especially maneuvering flight states, which significantly impact a helicopter's lifespan. Helicopter maneuvers often involve significant loads, causing varying degrees of wear and tear on different components. Users cannot monitor these loads in real-time and must rely on flight states to assess potential damage and lifespan; therefore, accurate identification of maneuvering flight states is essential. Furthermore, helicopter maneuvers are complex, with diverse parameters in flight data, and current flight state identification methods still require improvement in accuracy. Summary of the Invention
[0004] The present invention provides a method for identifying the maneuvering flight state of a helicopter. This method combines the typical characteristics of the helicopter's maneuvering flight state with the advantages of neural network fuzzy recognition, which can significantly improve the accuracy of maneuvering flight state identification and reduce the recognition difficulty of the neural network. This solves the problem that current flight state identification methods have low accuracy in maneuvering flight state identification due to the complexity of helicopter maneuvering flight state actions and the diverse changes in corresponding parameters in flight parameters.
[0005] To achieve the above objectives, the present invention employs the following technical solution.
[0006] A method for identifying the maneuvering flight state of a helicopter, the method comprising:
[0007] S1, Determine the flight state F during helicopter maneuvering flight. i Typical parameters f0, f1, f2, ..., f n The characteristics of change;
[0008] S2, based on the flight state F iThe inherent parameter changes determine the first typical parameter f0. This first typical parameter is then scanned within the flight data to find the parameters that enable flight state F. i The position of the variation characteristics of the first typical parameter f0 ensures the maneuvering flight state F i It must occur in these locations;
[0009] S3, according to flight status F i Duration feature l i Determine the data truncation window length L i L i >l i Based on the position in S2, the flight state F i All typical parameters f0, f1, f2, ..., f n Data is extracted;
[0010] S4, based on typical parameter characteristics, gradually selects the true maneuvering flight state F i First, exclude 100% non-F. i After determining the state, a fuzzy recognition neural network is then used to analyze the maneuvering flight state F. i Perform identification and screening;
[0011] S5, for the selected maneuvering flight state F i The boundaries are corrected to obtain a more accurate flight status.
[0012] Furthermore,
[0013] When flight state F i Typical parameters for a dive / dive pull-up maneuver include barometric altitude, heading angle, pitch angle, longitudinal control, and normal overload.
[0014] The dive / pull-up maneuver is inevitably accompanied by a rapid decrease in altitude. Using barometric altitude as the first typical parameter f0, and based on the duration characteristics of the dive / pull-up maneuver, the window length is set to 60s.
[0015] Normal descent conditions are eliminated based on a consistently positive pitch angle, and sharp descent and turning conditions are eliminated based on a stable heading angle.
[0016] Furthermore,
[0017] When flight state F i Typical parameters for a jump-inverted maneuver include barometric altitude, indicated airspeed, heading angle, pitch angle, roll angle, longitudinal control, lateral control, and normal overload.
[0018] The leap-inverted maneuver is inevitably accompanied by a rapid increase in altitude. Using air pressure altitude as the first typical parameter f0, and based on the duration characteristics of the leap-inverted maneuver, the window length is set to 100s.
[0019] Based on changes in heading angle, we can rule out rapid climbs and sharp ascents during turns; based on changes in indicated airspeed, we can rule out circular turns.
[0020] Furthermore,
[0021] In S4, a fuzzy recognition neural network is used to analyze the maneuvering flight state F. i Before identification and screening;
[0022] The corresponding neural network is trained using a single-parameter feature.
[0023] Flight status F i Typical parameters f0, f1, f2, ..., f n For parameters that cannot be directly identified, a corresponding neural network is used to identify them.
[0024] If the parameters that cannot be directly identified all show characteristics of a certain flight state when identified by the neural network corresponding to the parameter, then the flight state is determined to be a true maneuvering state.
[0025] Furthermore,
[0026] When training the corresponding neural network using a single-parameter feature, the sample data under maneuvering conditions is expanded using the following method:
[0027] Obtain the variation curve of a single parameter during maneuvering flight;
[0028] By adjusting the slope and intercept of the change curve, more sample data can be generated without changing the overall trend of a single parameter.
[0029] Furthermore,
[0030] S5, the boundaries of the selected maneuvering flight states Fi are corrected, specifically as follows:
[0031] The boundary of the selected maneuvering flight state Fi is segmented using a boundary point segmentation algorithm to obtain multiple initial state boundary points;
[0032] False boundary points caused by environmental interference or pilot jitter are removed from multiple initial state boundary points; and a convolutional neural network is used to segment the state of typical data parameters.
[0033] Furthermore,
[0034] A convolutional neural network is used to segment the state of typical data parameters, specifically as follows:
[0035] The flight data with false boundary points removed is used to create a training dataset. Each sample data in the training dataset has a preset data length, and the label of the sample data is the state of the middle data point of the preset data length.
[0036] A convolutional neural network was trained using the training dataset, and the convolutional neural network was used to perform state segmentation on flight data.
[0037] Furthermore,
[0038] The boundary point segmentation algorithm is as follows:
[0039] First, let x = x0 be the dividing point, and then perform linear function fitting on both sides of the data dividing point;
[0040] Scan the data points (xi, yi) of a typical parameter within a preset window, let x0 = xi, and calculate the mean square error E of the fit. The point with the smallest E is taken as the initial state boundary point.
[0041] This invention provides a method for identifying the maneuvering flight state of a helicopter. This method combines the typical features of the helicopter's maneuvering flight state with the advantages of neural network fuzzy recognition, which can significantly improve the accuracy of maneuvering flight state identification and reduce the recognition difficulty of neural networks. Attached Figure Description
[0042] Figure 1 A schematic diagram illustrating the evolution and expansion of sample data for barometric altitude parameters provided in an embodiment of the present invention;
[0043] Figure 2 This is a schematic diagram illustrating the principle of the boundary point segmentation algorithm provided in an embodiment of the present invention;
[0044] Figure 3 This is a schematic diagram of parameter segmentation samples provided in an embodiment of the present invention;
[0045] Figure 4 This is a schematic diagram of the sample data labeling method provided in an embodiment of the present invention;
[0046] Figure 5 This is a schematic diagram of the flight parameter status identification results provided in an embodiment of the present invention. Detailed Implementation
[0047] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.
[0048] This invention provides a method for identifying the maneuvering flight state of a helicopter. This method combines the typical features of the helicopter's maneuvering flight state with the advantages of neural network fuzzy recognition, which can significantly improve the accuracy of maneuvering flight state identification and reduce the recognition difficulty of neural networks.
[0049] Identification steps:
[0050] S1, determine the variation characteristics of typical parameters (f0, f1, f2, f3...) under each flight state (e.g., dive / dive pull-up, serpentine maneuver, climb and reverse, control reciprocation, dove jump) in helicopter maneuvering flight states;
[0051] S2, determine the maneuvering flight state to be searched for F i Based on the parameter change characteristics that inevitably accompany this flight state, the first typical parameter f0 (such as altitude parameter) is determined. This typical parameter is then scanned in the flight data to find the parameters that indicate the presence of the flight state F. i By analyzing the changing characteristics of the corresponding parameter f0 (such as altitude change characteristics), position information is obtained to ensure the maneuvering flight state F. i It must have occurred near these locations;
[0052] S3, according to flight status F i Duration feature l i Determine the length L of the data truncation window. i (L i >l i ), based on the position information in S2, the flight state F i Extract all typical parameter (f0, f1, f2, f3...) data;
[0053] S4, based on typical parameter characteristics, gradually selects the true maneuvering flight state F i The selection of typical features follows a sequence from simple to complex, for example, flight state F. i The parameter f1 has the characteristics of being simple to change and easy to search (for example, a dive and pull-up must involve a nose-down, and the pitch angle is negative). It can be filtered first using simple mathematical methods (such as filtering out data segments where the pitch angle is always positive), and only excluding 100% non-F1 parameters. i The state, along with the parameters f0 f1f2 f3 f4…, also exhibit complex feature changes. These feature changes are highly ambiguous in their recognition, so a neural network, which is adept at fuzzy recognition, is used for identification and filtering.
[0054] Each parameter f0 f1 f2 f3 f4… is identified using a neural network corresponding to that parameter, and each neural network is trained with a single parameter feature; if the identification result of each parameter f0 f1 f2 f3 f4… is a feature of a certain flight state, the flight state is determined to be a real maneuver state.
[0055] S5, for the final filtered state F i The boundaries are corrected to obtain a more accurate flight status.
[0056] Key technical points:
[0057] 1) Analysis of typical parameters of maneuvering flight state
[0058] Take a dive / dive pull-up and a climb inversion as examples.
[0059] Dive / Pull-up: Typical parameters include barometric altitude, heading angle, pitch angle, longitudinal control input, and normal load. A dive / pull-up is inevitably accompanied by a rapid decrease in altitude; barometric altitude can be used as the primary typical parameter f0. Based on the duration characteristics of the dive / pull-up, a window length of approximately 60 seconds is selected. Normal descent states are filtered out based on a consistently positive pitch angle, and states such as sharp descents and turns are filtered out based on the stability of the heading angle. The changes in pitch angle and control input during dive / pull-up are complex and lack clear ranges, requiring the use of fuzzy recognition characteristics of neural networks for differentiation. Finally, the load intensity of the dive / pull-up is further subdivided based on normal load.
[0060] Climb-invert: Typical parameters include barometric altitude, indicated airspeed, heading angle, pitch angle, roll angle, longitudinal control input, lateral control input, and normal load. Climb-invert inevitably involves a rapid increase in altitude; barometric altitude can be used as the primary typical parameter f0. Based on the characteristics of climb-invert duration, a window length of approximately 100 seconds is selected. Rapid climbs and sharp climb-turns are excluded based on changes in heading angle, and circular turns are excluded based on changes in indicated airspeed. Parameters such as barometric altitude, pitch angle, roll angle, longitudinal control input, and lateral control input also exhibit characteristic variations that are difficult to define precisely; these are identified and filtered using neural networks.
[0061] 2) Neural Network Recognition Method
[0062] First, an LSTM (Long Short Term Mermory network) neural network is introduced for recognition and screening. LSTM is a special type of RNN (Recurrent Network) that can solve the long-term dependency problem of RNN neural networks and improve the recognition accuracy of long-term time series data.
[0063] Secondly, using neural networks to identify individual parameter features, such as f0, f1, f2, f3, f4… all exhibit variable characteristics that are difficult to define precisely. For each parameter, a corresponding neural network can be trained separately. This has two advantages: first, it reduces the complexity requirements of the neural network. If a single neural network were to identify multiple parameter features simultaneously, its structure would need to be more complex, its network depth greater, and the number of training samples required more. Second, it solves the problem of limited sample size for maneuvering states. Maneuvering states account for a small proportion of the helicopter's entire lifespan, making samples difficult to obtain and resulting in a low sample size. By using neural networks to identify individual parameter features, the training samples are for a single parameter, and the system can evolve and expand based on actual flight test results to obtain more sample data. Figure 1 The barometric altitude parameters show that the data changes are mainly divided into rising, stable, and falling. By fine-tuning the slope and intercept of the curve, more sample data can be generated without changing the overall trend.
[0064] 3) Boundary Correction Algorithm
[0065] There are two main types of boundary correction algorithms: one is the boundary point segmentation algorithm, and the other is the convolutional neural network segmentation algorithm.
[0066] Because state boundaries are often accompanied by significant inflection points in certain parameter curves, boundary point segmentation algorithms can accurately locate these inflection points. The boundary point segmentation algorithm is as follows.
[0067] First, let x = x0 be the dividing point, and then perform linear function fitting on both sides of the data dividing point.
[0068] Scan the entire data segment (xi, yi), let x0 = xi, and calculate the mean square error E. The point with the smallest E is the true state boundary point, such as... Figure 2 As shown.
[0069] Boundary point segmentation algorithms can identify the boundary points of most parameter changes, but for flight data with large fluctuations, sometimes local fluctuations cannot truly represent the trend of parameter changes. Boundary point segmentation will have certain recognition errors when encountering large local fluctuations. Therefore, a convolutional neural network segmentation algorithm is introduced on the basis of boundary point segmentation algorithm.
[0070] First, the parameters are segmented using a boundary point segmentation algorithm. Errors are corrected at locations with significant local fluctuations to obtain, as shown below. Figure 3 The parameters shown are used to segment the samples.
[0071] Data labels are created based on the segmented samples. Data segments are truncated according to a suitable flight parameter reference length, such as 5 seconds of data. The reference length of the sample data is 5 seconds. The data labels use the status markers of the intermediate data points within the reference length (e.g., ...). Figure 4 That is, the state of a point is determined based on data from 2.5 seconds before and after it;
[0072] The convolutional neural network (CNN) segmentation algorithm fully considers the changes in flight parameters and the sample labeling reflects the labeler's understanding of the state boundaries. Therefore, it can accurately segment the state based on the labeler's understanding of the state boundaries. Using 10 minutes of flight data as samples for labeling, a CNN was trained. The trained network was then used to identify another segment of data. The identification results are as follows: Figure 5 As shown.
[0073] Convolutional neural network segmentation algorithms can accurately segment data based on the segmentation habits of samples as the amount of sample data increases, overcoming the influence of local fluctuations.
[0074] This method can more accurately identify the flight maneuver status of helicopters. Through identification and comparison verification of scientific research flight test data, 17 out of 18 test flight records of dove-jump status were accurately identified by automatic identification; 18 out of 19 test flight records of serpentine maneuver status were accurately identified by automatic identification; all 83 test flight records of dive / dive-pull-up status were accurately identified, and more low-G dive / dive-pull-up statuses that were not previously recorded in the test flights were also identified; all 10 test flight records of maneuvering reciprocating test status were accurately identified; and 34 out of 35 test flight records of climb and inversion were accurately identified.
[0075] Based on the identification tests of the flight test data, the accuracy rate of identifying dolphin jumps, serpentine maneuvers, dives / pull-ups, maneuvering back and forth, and climbs and inversions can all reach about 95%. The states that were not accurately identified all had some issues with non-standard state actions. This problem can be gradually resolved by continuously increasing the number of samples.
[0076] This invention provides a method for identifying the maneuvering flight state of a helicopter. This method combines the typical features of the helicopter's maneuvering flight state with the advantages of neural network fuzzy recognition, which can significantly improve the accuracy of maneuvering flight state identification and reduce the recognition difficulty of neural networks.
Claims
1. A method for recognizing the maneuvering flight state of a helicopter, characterized in that, The method includes: S1, Determine the flight state F during helicopter maneuvering flight. i Typical parameters f0, f1, f2, ..., f n The characteristics of change; S2, based on the flight state F i The inherent parameter changes determine the first typical parameter f0. This first typical parameter is then scanned in the flight data to find the parameters that enable flight state F. i The position of the variation characteristics of the first typical parameter f0 ensures the maneuvering flight state F i It must occur in these locations; S3, according to flight status F i Duration feature i Determine the data truncation window length L i L i >l i Based on the position in S2, the flight state F i All typical parameters f0, f1, f2, ..., f n Data is extracted; S4, based on typical parameter characteristics, gradually selects the true maneuvering flight state F i First, exclude 100% non-F. i After determining the state, a fuzzy recognition neural network is then used to analyze the maneuvering flight state F. i Perform identification and screening; S5, for the selected maneuvering flight state F i The boundaries are corrected to obtain a more accurate flight status.
2. The helicopter maneuvering flight state identification method according to claim 1, characterized in that, When flight state F i Typical parameters for a dive / dive pull-up maneuver include barometric altitude, heading angle, pitch angle, longitudinal control, and normal overload. The dive / pull-up maneuver is inevitably accompanied by a rapid decrease in altitude. Using barometric altitude as the first typical parameter f0, and based on the duration characteristics of the dive / pull-up maneuver, the window length is set to 60s. Normal descent conditions are eliminated based on a consistently positive pitch angle, and sharp descent and turning conditions are eliminated based on a stable heading angle.
3. The helicopter maneuvering flight state identification method according to claim 1, characterized in that, When flight state F i Typical parameters for a jump-inverted maneuver include barometric altitude, indicated airspeed, heading angle, pitch angle, roll angle, longitudinal control, lateral control, and normal overload. The leap-inverted maneuver is inevitably accompanied by a rapid increase in altitude. Using air pressure altitude as the first typical parameter f0, and based on the duration characteristics of the leap-inverted maneuver, the window length is set to 100s. Based on changes in heading angle, we can rule out rapid climbs and sharp ascents during turns; based on changes in indicated airspeed, we can rule out circular turns.
4. The helicopter maneuvering flight state identification method according to claim 1, characterized in that, In S4, a fuzzy recognition neural network is used to analyze the maneuvering flight state F. i Before identification and screening; The corresponding neural network is trained using a single-parameter feature. Flight status F i Typical parameters f0, f1, f2, ..., f n For parameters that cannot be directly identified, a corresponding neural network is used to identify them. If the parameters that cannot be directly identified all show characteristics of a certain flight state when identified by the neural network corresponding to the parameter, then the flight state is determined to be a true maneuvering state.
5. The helicopter maneuvering flight state identification method according to claim 4, characterized in that, When training the corresponding neural network using a single-parameter feature, the sample data under maneuvering conditions is expanded using the following method: Obtain the variation curve of a single parameter during maneuvering flight; By adjusting the slope and intercept of the change curve, more sample data can be generated without changing the overall trend of a single parameter.
6. The helicopter maneuvering flight state identification method according to claim 4, characterized in that, S5, the boundaries of the selected maneuvering flight states Fi are corrected, specifically as follows: The boundary of the selected maneuvering flight state Fi is segmented using a boundary point segmentation algorithm to obtain multiple initial state boundary points; False boundary points caused by environmental interference or pilot jitter are removed from multiple initial state boundary points; and a convolutional neural network is used to segment the state of typical data parameters.
7. The helicopter maneuvering flight state recognition method according to claim 6, characterized in that, A convolutional neural network is used to segment the state of typical data parameters, specifically as follows: The flight data with false boundary points removed is used to create a training dataset. Each sample data in the training dataset has a preset data length, and the label of the sample data is the state of the middle data point of the preset data length. A convolutional neural network was trained using the training dataset, and the convolutional neural network was used to perform state segmentation on flight data.
8. The helicopter maneuvering flight state identification method according to claim 6, characterized in that, The boundary point segmentation algorithm is as follows: First, let x = x0 be the dividing point, and then perform linear function fitting on both sides of the data dividing point; Scan the data points (xi, yi) of a typical parameter within a preset window, let x0 = xi, and calculate the mean square error E of the fit. The point with the smallest E is taken as the initial state boundary point.