Helicopter flight state identification method and device based on PINN

Abstract

The invention relates to a PINN-based helicopter flight state identification method and device, and the method comprises the steps: obtaining a plurality of flight parameters of a helicopter at the same moment; carrying out preprocessing and physical rationality verification on the flight parameters to obtain a flight parameter sequence; inputting the flight parameter sequence into a helicopter flight state recognition network fused with PINN to obtain a helicopter flight state recognition result; wherein the PINN-fused helicopter flight state recognition network introduces a physical information constraint module, and the physical information constraint module is used for converting a helicopter flight physical rule into network constraints and generating physical consistent features through the network constraints based on key features of flight parameters. The method can reduce the data dependence, and improves the recognition precision and generalization in a complex scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aircraft state recognition, in particular to a helicopter flight state recognition method and device based on PINN. BACKGROUND

[0002] Helicopters have strong maneuverability and can realize 35 typical flight states such as ground effect hovering, vertical ascent, and diving pull-up. The damage degree of moving parts (such as rotors and transmission shafts) and serviceable parts is significantly different under different states. Precise recognition of flight states is crucial for key component life prediction and improving aviation safety.

[0003] In the prior art, helicopter flight state recognition relies on feature extraction and classification of multiple flight parameters (such as indicated airspeed, overload, engine speed, etc.), but has the following shortcomings:

[0004] Strong data dependence: Traditional machine learning methods (such as SVM and random forest) and single deep learning networks require a large amount of high-quality historical flight data for training. However, data is scarce during the test flight phase of new helicopters and under special working conditions (such as extreme weather), which significantly reduces the model's generalization ability.

[0005] Insufficient physical coupling processing: Helicopter flight parameters are influenced by multiple physical fields such as dynamics (overload and acceleration), aerodynamics (airspeed and altitude), and power systems (speed and torque). Existing methods rely solely on data-driven fitting and do not incorporate physical law constraints. When there is noise or abnormality in the parameters, it is easy to make "data fitting but physically unreasonable" errors in recognition.

[0006] Poor generalization of small samples: When the amount of training data is insufficient (such as when a new helicopter model only accumulates a few hundred flight data), existing feature extraction networks have difficulty fully learning the mapping relationship between parameters and states, resulting in a significant decrease in recognition accuracy. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a helicopter flight state recognition method and device based on PINN, which can reduce data dependence and improve recognition accuracy and generalization in complex scenarios.

[0008] The technical solution adopted by the present application to solve its technical problem is to provide a helicopter flight state recognition method based on PINN, comprising the following steps:

[0009] Obtain multiple flight parameters of the helicopter at the same time;

[0010] Preprocess and physically verify the flight parameters to obtain a flight parameter sequence;

[0011] input the flight parameter sequence into a helicopter flight state recognition network of the fusion PINN to obtain a recognition result of the helicopter flight state;

[0012] The helicopter flight state recognition network of the fusion PINN comprises:

[0013] An encoder module is configured to capture long-distance data features in the flight parameter sequence.

[0014] A multi-scale feature extraction module is configured to extract deep data features of different scales from the long-distance data features.

[0015] A residual module is configured to find key features of the flight parameters from the deep data features of different scales.

[0016] A physical information constraint module is configured to convert helicopter flight physical laws into network constraints and generate physically consistent features based on the key features of the flight parameters through the network constraints.

[0017] A classification module is configured to recognize the helicopter flight state according to the spliced key features and physically consistent features of the flight parameters to obtain the recognition result of the helicopter flight state.

[0018] The physical reasonableness verification refers to effectively screening the collected flight parameters based on helicopter design principles and flight physical laws, and eliminating abnormal data that obviously violate physical constraints.

[0019] The physical reasonableness verification at least comprises one of the following:

[0020] Single-parameter absolute value verification, that is, judging whether each flight parameter exceeds the limit range;

[0021] Multi-parameter correlation verification, that is, verifying the reasonableness of the combination of flight parameters based on the physical coupling relationship between the flight parameters.

[0022] The physical information constraint module comprises:

[0023] A physical law modeling unit is configured to construct a core physical constraint equation based on helicopter flight mechanics, aerodynamics and power system principles.

[0024] A constraint loss calculation unit is configured to substitute the key features of the flight parameters into the core physical constraint equation, calculate a physical residual, and take the L2 norm of the physical residual as a physical constraint loss.

[0025] A loss fusion unit is configured to fuse the physical constraint loss and the classification loss of the residual module according to weights to construct a hybrid loss function.

[0026] The core physical constraint equation constructed by the physical law modeling unit includes a dynamic constraint, which is expressed as: Wherein, N x is a longitudinal overload, N z is a normal overload, m is a fuselage mass of the helicopter, g is a gravity acceleration, a x is a longitudinal acceleration of the helicopter, a z is a normal acceleration of the helicopter.

[0027] The core physical constraint equation constructed by the physical law modeling unit includes an aerodynamic constraint, which is expressed as: Wherein, V i is an air speed of the helicopter, p is an atmospheric density, P total is total pressure sensor data, P static is static pressure sensor data.

[0028] The core physical constraint equation constructed by the physical law modeling unit includes a power system constraint, which is expressed as: Wherein, M fa is an engine torque of the helicopter, N r is a rotor speed of the helicopter, P HP is a motor power (unit: horsepower).

[0029] The technical solution adopted by the present application to solve its technical problems is to provide a helicopter flight state recognition device based on PINN, comprising:

[0030] An acquisition module is configured to acquire multiple flight parameters of a helicopter at the same time;

[0031] A preprocessing module is configured to preprocess and physically check the flight parameters to obtain a flight parameter sequence;

[0032] A recognition module is configured to input the flight parameter sequence into a helicopter flight state recognition network of a fusion PINN to obtain a recognition result of the helicopter flight state;

[0033] Wherein, the helicopter flight state recognition network of the fusion PINN comprises:

[0034] An encoder module is configured to capture long-distance data features in the flight parameter sequence;

[0035] A multi-scale feature extraction module is configured to extract deep data features of different scales from the long-distance data features;

[0036] A residual module is configured to find key features of the flight parameters from the deep data features of different scales;

[0037] A physical information constraint module is configured to convert helicopter flight physical laws into network constraints, and generate physical consistent features based on key features of the flight parameters through the network constraints.

[0038] A classification module is configured to identify the helicopter flight state according to the key features and the physical consistent features of the spliced flight parameters, and obtain an identification result of the helicopter flight state.

[0039] The technical solution adopted by the present application to solve its technical problems is to provide an electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the helicopter flight state identification method based on PINN.

[0040] The technical solution adopted by the present application to solve its technical problems is to provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the helicopter flight state identification method based on PINN.

[0041] Advantages

[0042] Compared with the prior art, the present application has the following advantages and positive effects: the encoder module of the PINN can accurately capture long-distance data features in the sequence; at the same time, the physical law constraint is introduced through the PINN, and in the small sample scene, the recognition accuracy can still reach more than 98.5%, which alleviates the problem of lack of helicopter test flight data; the physical consistency is improved: the physical constraint module can correct the noise parameters (such as "abnormal speed" caused by sensor errors), avoid "data fitting but physically unreasonable" false identification, and the introduction of the PINN physical constraint mechanism can improve the interpretability, generalization and accuracy of the network; after the fusion of the physical law, the network can be migrated to the same series of different models of helicopters (only the physical parameters such as the body mass m need to be fine-tuned), without the need to collect a large amount of data again, reducing the model adaptation cost. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a flowchart of the helicopter flight state identification method of the first embodiment of the present application;

[0044] Figure 2 is a flowchart of the flight parameter preprocessing in the first embodiment of the present application;

[0045] Figure 3 is a structural schematic diagram of the helicopter flight state identification network in the first embodiment of the present application;

[0046] Figure 4 is a structural schematic diagram of the encoder unit in the first embodiment of the present application;

[0047] Figure 5 is a structural schematic diagram of a multi-scale feature extraction module in the first embodiment of the application;

[0048] Figure 6 is a structural schematic diagram of a residual module in the first embodiment of the application;

[0049] Figure 7 is a structural schematic diagram of a channel attention layer in the first embodiment of the application. DETAILED DESCRIPTION

[0050] The application will be further described below in connection with specific embodiments. It should be understood that these embodiments are only used to illustrate the application and not used to limit the scope of the application. Furthermore, it should be understood that after reading the content taught by the application, those skilled in the art can make various modifications or changes to the application, and these equivalent forms also fall within the scope defined by the appended claims.

[0051] The first embodiment of the application relates to a PINN-based helicopter flight state recognition method, as shown in Figure 1 , comprising the following steps:

[0052] Step 1: Obtain multiple flight parameters of the helicopter at the same time. There are 23 flight parameters obtained in this step, as shown in Table 1.

[0053] Table 1: Helicopter flight parameter table

[0054]

[0055]

[0056] Step 2: Preprocess and physically check the flight parameters to obtain a flight parameter sequence.

[0057] Due to the complex and changeable environment during helicopter flight, the flight parameter recorder will introduce external noise and cross-linking coupling effects during its collection and recording, which is easy to produce abnormal phenomena such as data loss and data distortion. And these abnormal data will directly affect the recognition rate of the flight state, so it is necessary to preprocess the flight parameter data before flight state recognition, to repair distorted data, fill in missing data, and reduce unnecessary noise and unpredictable data in the state recognition process.

[0058] As shown in Figure 2 , when preprocessing the flight parameters, the following steps are performed in sequence: outlier removal (3σ criterion), repair (linear interpolation), amplitude limiting (constrained according to the design parameter range of the helicopter), smoothing (Gaussian filtering), and fitting (polynomial fitting completion).

[0059] The core role of the physical rationality check in this step is to filter out abnormal data that obviously violates physical constraints based on the helicopter design principle and flight physical law, and to provide high-quality input for the subsequent network model. The physical rationality check in this step includes single-parameter absolute value check and multi-parameter correlation check.

[0060] The single-parameter absolute value check refers to judging whether each flight parameter exceeds the limit range, for example, the engine speed N g : 98% ~ 102% (design manual constraint), exceeding this range is determined as abnormal (such as flameout or over-speed failure); the normal overload N z : N z ≥ 0.3 (flight mechanics constraint, the normal overload of a helicopter will not be less than 0.3 when flying normally); the total distance lever displacement H c : 0° ~ 30° (mechanical structure constraint, the maximum deflection angle of the total distance lever is limited by mechanical limiting); the barometric altitude H p : -500m ~ 8000m (task scenario constraint, civil helicopters usually do not enter the airspace outside this range).

[0061] The multi-parameter correlation check refers to verifying the rationality of the combination of flight parameters based on the physical coupling relationship between each flight parameter. Typical correlation rules include dynamics correlation, aerodynamics correlation, and attitude correlation. Among them, the dynamics correlation: the longitudinal overload N x and the longitudinal acceleration a x need to satisfy N x ≈ a x / g, the error should be ≤15%, otherwise it is determined that the acceleration sensor or overload sensor is abnormal; the aerodynamics correlation: the indicated airspeed V i and the true airspeed V t need to meet V t = V i T / T0 (T is the atmospheric temperature and T0 is the standard temperature), a deviation of more than 20% is determined as an airspeed system failure; the power system correlation: the engine torque M fa and the rotor speed N r need to satisfy a negative correlation (i.e. when the rotor speed N r increases, the engine torque M fa decreases), if a positive correlation occurs, it is determined that the power transmission is abnormal; the attitude correlation: the pitch angle θ and the elevator deflection angle δ e need to change in the same direction (when the pitch angle θ increases, the elevator deflection angle δ e should be positively deflected), and if they change in opposite directions, it is determined that the flight control system is abnormal.

[0062] Step 3, inputting the flight parameter sequence into the helicopter flight state recognition network of the fusion PINN to obtain a recognition result of the helicopter flight state. The recognition result includes 35 flight states, as shown in Table 2.

[0063] Table 2: Table of flight states to be identified

[0064] Number Flight condition Number Flight condition 1 Hover in-ground effect 19 Cruise speed, level, left 10° 2 Hover out of ground effect 20 Cruise speed, level, right 10° 3 Back flight 21 Maximum cruise speed, level, 0° 4 Right side flight 22 Maximum cruise speed, level, left 5° 5 Left side flight 23 Maximum cruise speed, level, right 5° 6 Vertical climb 24 Horizontal descent 7 Oblique climb 25 Horizontal acceleration 8 Maximum continuous power, oblique climb 26 Descent 9 Climb turn or hover 27 Descent pull-up 10 Take-off acceleration 28 Endurance speed, level turn or hover 11 Transition speed, level 29 180 km / h speed, level turn 12 Endurance speed, level, 0° 30 Maximum cruise speed, turn 13 Endurance speed, level, left 10° 31 Autorotation 14 Endurance speed, level, right 10° 32 Approach, glide, descent 15 180 km / h speed, level, 0° 33 Flare landing 16 180 km / h speed, level, left 10° 34 Vertical descent 17 180 km / h speed, level, right 10° 35 Ground idle 18 Cruise speed, level, 0°

[0065] The helicopter flight state recognition network of the fusion PINN (PINN-CTFM-Net) in the embodiment is constructed based on the CTFM-Net network and the fusion PINN, as shown in Figure 3 includes an encoder module, a multi-scale feature extraction module, a residual module, a physical information constraint module, and a classification module. The encoder module is used to capture long-distance data features in the flight parameter sequence. The multi-scale feature extraction module is used to extract deep data features of different scales from the long-distance data features. The residual module is used to find key features of the flight parameters from the deep data features of different scales. The physical information constraint module is used to convert the physical laws of helicopter flight into network constraints and generate physically consistent features based on the key features of the flight parameters through the network constraints. The classification module is used to identify the helicopter flight state according to the spliced key features and physically consistent features of the flight parameters to obtain the recognition result of the helicopter flight state.

[0066] The encoder module includes three consecutive encoding units, as shown in Figure 4 Each encoding unit includes a first layer normalization layer, a multi-head attention layer, a second layer normalization layer, and a feedforward neural network layer arranged in sequence. The input of the second layer normalization layer is the fusion feature of the output of the multi-head attention layer and the input of the first layer normalization layer. The fusion feature of the output of the feedforward neural network layer and the input of the second layer normalization layer is taken as the output of the encoding unit.

[0067] As can be seen, the encoding unit mainly consists of a multi-head self-attention mechanism and a feedforward neural network. In addition to the above two network layers, the structure of each encoding unit also includes residual connection and layer normalization for accelerating and stabilizing the training of the model. The residual connection adds the input and output of the encoder unit, and the layer normalization standardizes the output of each sub-layer, making the model more easily converge.

[0068] The multi-head self-attention layer obtains a weighted representation of each position of the input sequence by performing self-attention calculation on all positions of the input sequence, where the weight is calculated by a query vector, a key vector and a value vector. The query vector represents the features of the current time step, and the key vector and the value vector represent the features of other time steps in the sequence. The multi-head attention mechanism maps the input query vector and the key vector to different vector spaces and calculates the similarity between them to obtain an attention distribution. Then, the attention distribution is multiplied by the value vector to obtain a weighted value vector, and finally, the weighted vectors of multiple heads are spliced as the output. The calculation formula of each self-attention is as follows:

[0069]

[0070] In the formula, d k represents the channel dimension, Q, K and V represent the query vector, the key vector and the value vector respectively, and Softmax is an activation function. The multiple self-attention weighted vectors mapped to different vector spaces are spliced to obtain the final multi-head attention vector. The final multi-head attention output is obtained by mapping back to the original vector dimension through matrix transformation.

[0071] In the encoding unit, the multi-head self-attention mechanism layer performs a weighted calculation on each vector of the input, thereby capturing the mutual relationship between all vectors in the sequence.

[0072] The feedforward neural network layer is arranged after the multi-head self-attention mechanism layer and performs a separate fully connected feedforward neural network transformation on each vector representation. It is composed of two consecutive linear layers and a nonlinear activation function ReLU is added in between, so that the network learns more complex nonlinear relationships and enhances the expression ability of the model.

[0073] Different scales of convolution extract features with different levels of detail. The smaller the scale of convolution, the easier it is to extract local sensitive features. If only the same scale of convolution is used in a single layer, other precision features may be ignored, resulting in the inability to fully extract the feature information of each scale in the input data. To improve the feature extraction capability of the network and improve the flight state recognition accuracy, the multi-scale feature extraction module in the embodiment includes, as shown in Figure 5

[0074] ​The multi-scale feature extraction module superimposes two feature splicing units to perform feature extraction of different fineness on the input flight parameters, improve the richness and diversity of the features, and enhance the perception ability of the network to different scale targets in the flight parameters. When processing, first, the input features pass through a four-branch multi-scale convolution layer, and on each branch, a series of convolution and pooling operations of different scales are performed on the input features to extract feature information of each dimension; then, the feature information of different dimensions is spliced to complete the first stage of multi-scale feature extraction; then, the combined data features pass through a three-branch multi-scale down-sampling layer to further extract deeper data features while reducing the data dimension; finally, the deep features of three different scales are spliced.

[0075] The residual module enables the PINN-CTFM-Net to obtain deeper data features while avoiding the problems of gradient disappearance and gradient explosion. Since the residual module introduces a skip connection, i.e., directly adding the input to the output, the neural network can propagate the gradient from the back to the deeper network without losing information, effectively alleviating the degradation problem of the neural network. At the same time, the channel attention mechanism is used at the input and output ends of the residual module to enable the network to better utilize the features of different channels, thereby improving the expression ability and generalization of the network. The residual module structure in the embodiment is shown in Figure 6 The residual module structure in the embodiment is shown in

[0076] The channel attention layer can focus the network on the key features of the flight parameters through the channel attention mechanism, making the model more suitable for flight state classification tasks. As shown in Figure 7As shown, the channel attention layer comprises: an extraction unit for obtaining a vector of the number of channels and the length of features of deep data features; two fully connected layers for obtaining a weight vector of the channel direction; and an output layer for multiplying each element of each channel of the deep data features with the corresponding weight vector. When processing the input features, first, the vector of the number of channels T and the length L of the deep data features is obtained; then, the weight vector of the channel direction is obtained through two fully connected layers, wherein the first fully connected layer has T nodes, uses a Swish activation function, the second fully connected layer also has T nodes, and uses a Sigmoid activation function; finally, the elements of each channel of the input features are multiplied with the corresponding weight vector, and each channel has different weights, so that different attention degrees can be obtained for the features of each channel.

[0077] The physical information constraint module in the embodiment comprises:

[0078] A physical law modeling unit is configured to construct a core physical constraint equation based on helicopter flight mechanics, aerodynamics and power system principles, wherein the constructed core physical constraint equation comprises a dynamics constraint, an aerodynamic constraint and a power system constraint.

[0079] The dynamics constraint is expressed as:

[0080]

[0081] wherein, N x is the longitudinal overload, N z is the normal overload, m is the mass of the helicopter, g is the acceleration of gravity, a x is the longitudinal acceleration of the helicopter, and a z is the normal acceleration of the helicopter.

[0082] The aerodynamic constraint is expressed as:

[0083]

[0084] wherein, V i is the airspeed of the helicopter, p is the atmospheric density, P total is the total pressure sensor data, and P static is the static pressure sensor data.

[0085] The power system constraint is expressed as:

[0086]

[0087] wherein, M fa is the engine torque of the helicopter, N r is the rotor speed of the helicopter, and P HP is the motor power (unit: horsepower).

[0088] a constraint loss calculation unit for substituting key features of flight parameters into the core physical constraint equation, calculating a physical residual (i.e. "the difference between the model predicted parameters and the values calculated by the physical equation"), and taking the L2 norm of the physical residual as the physical constraint loss;

[0089] a loss fusion unit for fusing the physical constraint loss and the classification loss of the residual module by weight to build a hybrid loss function. The hybrid loss function is represented as:

[0090] L total = aL cls + bL phy ;

[0091] wherein a is the weight of the classification loss, taking a value of 0.6-0.8, and b is the weight of the physical constraint loss, taking a value of 0.2-0.4. The network is ensured to simultaneously fit the data distribution and the physical law by tuning through the validation set.

[0092] The physical information constraint module in the embodiment embeds the flight physical law of the helicopter, builds a constraint equation, calculates a physical residual to obtain a constraint loss, and fuses and optimizes the network with the classification loss. That is, the residual module and the physical information constraint module in the embodiment are not independently operated, but are deeply coupled through "feature interaction-double path fusion-joint optimization". Specifically, the "feature interaction-double path fusion-joint optimization" includes:

[0093] Feature interaction: the residual module provides "key parameter features" for the physical constraint. The core function of the residual module is to extract the most critical parameter features for flight state recognition from multi-scale features. These features are synchronously input into the physical information constraint module as the "data basis" for the calculation of the physical equation. The "key feature vector" (dimension 256, containing the filtered parameter weights and coupling relationships) output by the residual module is directly transmitted to the classification module in part, and is decomposed into parameter values required by the physical equation (such as extracting the quantized values of key parameters such as the longitudinal acceleration a x of the helicopter, the rotor speed N r of the helicopter, and the total pressure sensor data P total , etc.) to input into the physical law modeling unit, so as to ensure that the parameters used for the calculation of the physical constraint have been filtered by the feature of the residual module, focus on the most critical physical quantities for state recognition, and reduce the interference of redundant parameters.

[0094] Dual-path fusion: Based on the key parameters input by the residual module, the physical information constraint module generates "physically consistent features" (dimension 128, including the physical coupling strength between parameters, residual correction coefficients, etc.) through physical equation verification, and inputs "key features" output by the residual module into the classification module, forming dual-path feature fusion, realizing the complementary enhancement of physically consistent features and key features, and in subsequent operations, the "key feature vector" (256 dimensions) and the "physically consistent feature vector" (128 dimensions) are spliced into 384-dimensional fusion features, which are input into the fully connected layer of the classification module.

[0095] Joint optimization: the physical loss L phy of the physical information constraint module is combined with the classification loss L cls of the residual module to form the total loss L total Binding, synchronously updating the parameters of the two modules during the backpropagation process, realizing collaborative optimization. When the network is trained, the gradient of the total loss L total is backpropagated to the residual module and the physical information constraint module at the same time. The residual module adjusts the convolution kernel weight and the channel attention coefficient, so that the extracted key features are more consistent with the physical law (such as reducing the weight of parameters with large physical residuals); the physical information constraint module optimizes the adaptive coefficients in the physical equation, so that the physical constraint is more consistent with the actual flight data.

[0096] The classification module of the embodiment splices the "physically consistent features" output by the physical information constraint module and the "key features" output by the residual module on the basis of the fully connected layer, and then inputs them into the Softmax layer to output the probability distribution of 35 flight states. The state with the highest probability is the recognition result. The hidden layer of the fully connected layer is used to connect the input data with the weight and apply an activation function for calculation, and the Softmax layer converts the output into a probability distribution.

[0097] The training method of the helicopter flight state recognition network in the embodiment is as follows:

[0098] First, the historical flight parameters of the helicopter are corresponded with the flight states to obtain a flight shape original database. Each flight state corresponds to a period of continuously collected flight parameter signals.

[0099] Next, the historical flight parameters in the flight shape original database are preprocessed. The preprocessing method is the same as that in step 2 above.

[0100] Then, the Pearson correlation coefficient is used for sensitive flight parameter selection, that is, the Pearson correlation coefficient between each historical flight parameter after preprocessing is calculated, and the historical flight parameter whose Pearson correlation coefficient does not meet the requirement is removed.

[0101] To determine the correlation between data items, the present embodiment uses the Pearson correlation coefficient, which is a linear correlation coefficient. If two sets of sample data {(x i ,y i ), i = 1, 2,... n} (n is the number of samples) are bivariate normal distribution , then the correlation coefficient is:

[0102]

[0103] where and are the sample means of X and Y, respectively. The absolute value of the correlation coefficient is closer to 1, the stronger the correlation between the two, and closer to 0, the worse the correlation between the two.

[0104] After that, the historical flight parameters in the processed flight mode original database are divided into training set, validation set and test set. The total sample size of the flight parameter data set used in the present embodiment is 782400, and the data is randomly extracted in each flight state sample in the ratio of 7:2:1 for training, validation and testing.

[0105] To verify the experimental environment and parameter settings built by the present embodiment are as follows:

[0106] Hardware: CPU Intel Xeon Gold 6248, GPU NVIDIA A100;

[0107] Software: PyTorch 2.0, Python 3.9;

[0108] Physical parameters: the fuselage mass of the helicopter m = 4500 kg, the gravitational acceleration g = 9.8 m / s 2 , power transmission coefficient k = 0.002, atmospheric density p calculated by air pressure height Hp according to the international standard atmospheric model; hybrid loss weight: a = 0.7, b = 0.3.

[0109] The structure parameters of the helicopter flight state recognition network are shown in Table 3. In order to increase the nonlinearity of the model, the output of each layer uses the ReLu nonlinear activation function to enhance the feature expression ability of the network.

[0110] Table 3 PINN-CTFM-Net network model parameter table

[0111]

[0112] The network training adopts an Adam optimizer, sets an initial learning rate as 0.01, a batch sample data size as 1000, a training iteration number as 100, and selects a cross-entropy loss function. The input of the network is a column vector of 21x1, the network is trained using training samples, after completing an iteration, test samples are used for testing to obtain a test accuracy rate. After the iteration of the neural network model is completed, the network is tested using a test set. The recognition accuracy comparison in different scenes is shown in Table 4.

[0113] Table 4: Recognition accuracy comparison in different scenes

[0114]

[0115] The flight state recognition results of the test set are shown in Table 5.

[0116] Table 5: Flight state recognition results table

[0117]

[0118] As shown in Table 5, among the recognition results of 35 flight states of CTFM-Net, the recognition accuracy of 18 flight states reaches 100%, only the recognition rate of flight state No. 30 is 94.5%, which is relatively low, and the recognition rate of most states is more than 99%, among the recognition results of 35 flight states of PINN-CTFM-Net, the recognition accuracy of 22 states reaches 100%, the lowest accuracy is 96.2% of state No. 30 (maximum cruise speed turning) (the original method is 94.5%), the average accuracy of 35 states reaches 99.57% (CTFM-Net: 99.33%), which proves that the method of the embodiment has improved the recognition stability of various states, and proves the superiority of the network model PINN-CTFM-Net in the flight state recognition task.

[0119] It is not difficult to find that the helicopter flight state recognition network of the application fusing PINN can accurately capture long-distance data features in the sequence; at the same time, by introducing the physical law constraint through PINN, the recognition accuracy can still reach more than 98.5% in the small sample scene, which alleviates the problem of lack of helicopter test flight data; improve the physical consistency: the physical constraint module can correct the noise parameters (such as "abnormal speed" caused by sensor errors), avoid "data fitting but physically unreasonable" false recognition; after fusing the physical law, the network can be migrated to the same series of different models of helicopters (only the physical parameters such as the body mass m need to be fine-tuned), without the need to collect a large amount of data again, reducing the model adaptation cost. And introduce the PINN physical constraint mechanism to improve the explainability, generalization and accuracy of the network.

[0120] The second embodiment of the application relates to a PINN-based helicopter flight state recognition device, comprising:

[0121] An acquisition module is configured to acquire multiple flight parameters of the helicopter at the same time.

[0122] A preprocessing module is configured to preprocess and physically check the flight parameters to obtain a flight parameter sequence.

[0123] A recognition module is configured to input the flight parameter sequence into a helicopter flight state recognition network of a fusion PINN to obtain a recognition result of the helicopter flight state.

[0124] The helicopter flight state recognition network of the fusion PINN comprises:

[0125] An encoder module is configured to capture long-distance data features in the flight parameter sequence.

[0126] A multi-scale feature extraction module is configured to extract deep data features of different scales from the long-distance data features.

[0127] A residual module is configured to find key features of the flight parameters from the deep data features of different scales.

[0128] A physical information constraint module is configured to convert helicopter flight physical laws into network constraints and generate physically consistent features based on the key features of the flight parameters through the network constraints.

[0129] A classification module is configured to recognize the helicopter flight state according to the key features and the physically consistent features of the spliced flight parameters to obtain the recognition result of the helicopter flight state.

[0130] The physical reasonableness check refers to effectively screening the collected flight parameters based on the helicopter design principle and flight physical laws, and eliminating abnormal data that obviously violate physical constraints.

[0131] The physical reasonableness check at least comprises one of the following:

[0132] Single-parameter absolute value check, that is, judging whether each flight parameter exceeds the limit range;

[0133] Multi-parameter correlation check, that is, verifying the rationality of the flight parameter combination based on the physical coupling relationship between the flight parameters.

[0134] The physical information constraint module comprises:

[0135] A physical law modeling unit is configured to construct a core physical constraint equation based on helicopter flight mechanics, aerodynamics and power system principles.

[0136] A constraint loss calculation unit is configured to substitute key features of the flight parameters into the core physical constraint equation, calculate a physical residual, and take the L2 norm of the physical residual as a physical constraint loss.

[0137] A loss fusion unit is configured to fuse the physical constraint loss and the classification loss of the residual module by weights to construct a hybrid loss function.

[0138] The core physical constraint equation constructed by the physical law modeling unit includes a dynamics constraint, which is expressed as: wherein N x is a longitudinal overload, N z is a normal overload, m is a body mass of the helicopter, g is a gravity acceleration, a x is a longitudinal acceleration of the helicopter, a z is a normal acceleration of the helicopter.

[0139] The core physical constraint equation constructed by the physical law modeling unit includes an aerodynamic constraint, which is expressed as: wherein V i is an air speed of the helicopter, p is an atmospheric density, P total is total pressure sensor data, P static is static pressure sensor data.

[0140] The core physical constraint equation constructed by the physical law modeling unit includes a power system constraint, which is expressed as: wherein M fa is an engine torque of the helicopter, k is a power transmission coefficient, N r is a rotor speed of the helicopter, P HP is motor power.

[0141] The third embodiment of the present application relates to an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the PINN-based helicopter flight state recognition method of the first embodiment when executing the computer program.

[0142] The fourth embodiment of the present application relates to a computer readable storage medium having stored thereon a computer program, the computer program implementing the steps of the PINN-based helicopter flight state recognition method of the first embodiment when executed by a processor.

[0143] Those skilled in the art will appreciate that embodiments of the application can be devised for a method, a system, or a computer program product. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage etc.) embodying computer readable program code.

[0144] The present application is described in reference to the flowchart and / or block diagram of the method, apparatus (system) and computer program product according to embodiments of the application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and a combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, a special purpose computer, an embedded processor or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the function specified in the flowchart and / or block diagram block or blocks.

[0145] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the function specified in the flowchart and / or block diagram block or blocks.

[0146] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the function specified in the flowchart and / or block diagram block or blocks.

[0147] The above description is only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A helicopter flight state recognition method based on PINN, characterized in that, Includes the following steps: Obtain multiple flight parameters of the helicopter at the same time; The flight parameters are preprocessed and their physical rationality is verified to obtain a flight parameter sequence. The flight parameter sequence is input into the helicopter flight status recognition network fused with PINN to obtain the helicopter flight status recognition result; The helicopter flight status recognition network fused with PINN includes: The encoder module is used to capture long-range data features in the flight parameter sequence; A multi-scale feature extraction module is used to extract deep data features of different scales from the long-distance data features; The residual module is used to find key features of flight parameters from the deep data features at different scales; The physical information constraint module is used to transform the physical laws of helicopter flight into network constraints, and generate physically consistent features through network constraints based on the key features of flight parameters. The classification module is used to identify the helicopter's flight status based on the key features and physical consistency features of the spliced ​​flight parameters, and to obtain the identification result of the helicopter's flight status.

2. The helicopter flight status recognition method based on PINN according to claim 1, characterized in that, The physical rationality verification refers to the process of screening the collected flight parameters based on helicopter design principles and flight physics laws, and eliminating abnormal data that obviously violate physical constraints.

3. The helicopter flight status recognition method based on PINN according to claim 2, characterized in that, The physical rationality verification includes at least one of the following: Single parameter absolute value verification, that is, determining whether each flight parameter exceeds the limit range; Multi-parameter correlation verification, that is, verifying the rationality of the combination of flight parameters based on the physical coupling relationship between various flight parameters.

4. The helicopter flight status recognition method based on PINN according to claim 1, characterized in that, The physical information constraint module includes: The physical law modeling unit is used to construct core physical constraint equations based on the principles of helicopter flight mechanics, aerodynamics, and propulsion systems. The constraint loss calculation unit is used to substitute the key features of the flight parameters into the core physical constraint equation, calculate the physical residual, and use the L2 norm of the physical residual as the physical constraint loss. The loss fusion unit is used to fuse the physical constraint loss and the classification loss of the residual module according to weights to construct a hybrid loss function.

5. The helicopter flight status recognition method based on PINN according to claim 1, characterized in that, The core physical constraint equations constructed by the physical law modeling unit include dynamic constraints, which are expressed as follows: Where, N x For longitudinal overload, N z Indicates normal overload, m is the helicopter's fuselage mass, g is the acceleration due to gravity, and a x For the longitudinal acceleration of the helicopter, a z This represents the normal acceleration of the helicopter.

6. The helicopter flight status recognition method based on PINN according to claim 1, characterized in that, The core physical constraint equations constructed by the physical law modeling unit include aerodynamic constraints, which are expressed as follows: Among them, V i Let P be the airspeed of the helicopter, ρ be the atmospheric density, and P be the airspeed of the helicopter. total For total pressure sensor data, P static This is data from a hydrostatic sensor.

7. The helicopter flight status recognition method based on PINN according to claim 1, characterized in that, The core physical constraint equations constructed by the physical law modeling unit include dynamic system constraints, which are expressed as follows: Among them, M fa N is the engine torque of the helicopter. r P is the rotor speed of the helicopter. HP This refers to the motor power.

8. A helicopter flight status identification device based on PINN, characterized in that, include: The acquisition module is used to acquire multiple flight parameters of the helicopter at the same time. The preprocessing module is used to preprocess and verify the physical rationality of flight parameters to obtain a flight parameter sequence; The identification module is used to input the flight parameter sequence into the helicopter flight status identification network fused with PINN to obtain the identification result of the helicopter flight status. The helicopter flight status recognition network fused with PINN includes: The encoder module is used to capture long-range data features in the flight parameter sequence; A multi-scale feature extraction module is used to extract deep data features of different scales from the long-distance data features; The residual module is used to find key features of flight parameters from the deep data features at different scales; The physical information constraint module is used to transform the physical laws of helicopter flight into network constraints, and generate physically consistent features through network constraints based on the key features of flight parameters. The classification module is used to identify the helicopter's flight status based on the key features and physical consistency features of the spliced ​​flight parameters, and to obtain the identification result of the helicopter's flight status.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the PINN-based helicopter flight status recognition method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the PINN-based helicopter flight status recognition method as described in any one of claims 1-7.

Images