Flight control method and system based on deep learning model

Abstract

The invention relates to the technical field of aircraft control, and discloses a flight control method and system based on a deep learning model, and the method comprises the following steps: 1, collecting flight state data in real time through a sensor carried by an aircraft; 2, preprocessing the collected flight state data; according to the method, the space-time fusion model is constructed through the cascade connection of the CNN and the LSTM, the space correlation and time sequence dynamic characteristics of the multi-sensor data are accurately extracted, and the dependence of traditional accurate dynamic modeling is eliminated; model compression technologies such as network pruning, weight quantification and knowledge distillation are combined, the reasoning delay is controlled within 10 ms, and the real-time requirement of an airborne platform is met; through synchronous acquisition of multi-sensor data, cooperative processing and deep mining of data deep association, the control precision under complex working conditions is improved by more than 30%, the reasoning efficiency is improved by 60%, the unknown environment adaptability is significantly enhanced, and a systematic solution is provided for stable and efficient control of an aircraft under complex scenes.

Description

Technical Field

[0001] The present invention relates to the field of aircraft control technology, and in particular to a flight control method and system based on a deep learning model. Background Art

[0002] Aircraft, as complex systems integrating multiple disciplines such as mechanics, electronics, and control, encompass a wide range of types, including drones, fixed-wing aircraft, helicopters, and rotorcraft. They are widely used in fields such as air transportation, military reconnaissance, geographic mapping, emergency rescue, and agricultural plant protection, and are a vital component of modern transportation systems and technological development. The core goal of flight control technology is to ensure stable and precise flight mission execution in complex environments, directly impacting the aircraft's safety, reliability, and mission efficiency.

[0003] First, traditional flight control methods rely on precise aircraft dynamics models. However, factors such as atmospheric disturbances and equipment aging can cause model parameter drift, leading to a decrease in control performance under strongly nonlinear conditions.

[0004] Second, existing methods rely on preset logic or limited prior knowledge and lack the ability to autonomously learn control strategies from data. This makes it difficult to adjust parameters in a timely manner in the event of sudden failures or unknown environments, which can easily lead to control failures.

[0005] Third, traditional deep neural networks are computationally intensive, resulting in excessive delays in airborne platform control (e.g., inference time > 50ms). Furthermore, they fail to effectively integrate the spatiotemporal characteristics of data, making offline training models unable to adapt to dynamic environmental changes.

[0006] Fourth, multi-sensor data from modern aircraft is not fully utilized. Traditional methods simply process single or small amounts of data without exploring the deep connections between multi-source data, resulting in incomplete information for control decisions.

[0007] To this end, a flight control method and system based on deep learning model are proposed. Summary of the Invention

[0008] The purpose of the present invention is to provide a flight control method and system based on a deep learning model to solve the problems of insufficient robustness and adaptability of traditional flight control methods in complex environments, as well as the limited real-time and generalization of deep learning models in flight control, as proposed in the above-mentioned background technology.

[0009] First aspect: To achieve the above objectives, a technical solution adopted in this application is: a flight control method based on a deep learning model, comprising the following steps:

[0010] Step 1: Collect flight status data in real time through sensors onboard the aircraft;

[0011] Step 2: Preprocess the collected flight status data;

[0012] Step 3: Build a hybrid model consisting of a convolutional neural network and a long short-term memory network. The convolutional neural network is used to extract the spatial features of the flight status data, while the long short-term memory network captures the time series dynamic dependencies of the flight status data.

[0013] Step 4: Based on the preprocessed flight status data, the hybrid model is trained using a supervised learning method with the expected control output as the label. During the training process, data augmentation technology is used to improve the generalization ability of the hybrid model.

[0014] Step 5: Use model compression technology to compress and optimize the trained hybrid model, and control the inference delay after training to within the preset real-time threshold;

[0015] Step 6: During the flight, the pre-processed flight status data is input into the optimized hybrid model in real time, and a control signal is output to drive the actuator;

[0016] Step 7: Establish an online update mechanism to collect new data in real time to fine-tune the hybrid model online to adapt to environmental changes and aircraft parameter drift.

[0017] As a further preferred embodiment of the present technical solution, in step three, the method for constructing a hybrid model consisting of a convolutional neural network and a long short-term memory network cascaded comprises the following steps:

[0018] Step 301: Construct a convolutional neural network. The convolutional neural network includes an input layer, a first convolutional layer, a second convolutional layer, and a pooling layer connected in sequence. The first convolutional layer uses 32 3×3 convolution kernels. The second convolutional layer uses 64 3×3 convolution kernels. The pooling layer uses a 2×2 maximum pooling window.

[0019] Step 302: constructing a long short-term memory network, wherein the long short-term memory network includes 128 memory units for processing dynamic characteristics of time series;

[0020] Step 303: Flatten the output of the pooling layer of the convolutional neural network and use it as the input of the long short-term memory network to form a cascade structure;

[0021] Step 304: Add a fully connected layer after the output layer of the long short-term memory network, where the output dimension of the fully connected layer matches the dimension of the aircraft control variable.

[0022] As a further preferred embodiment of the present technical solution, in step 4, the data enhancement technology includes spatial domain enhancement, temporal domain enhancement and hybrid enhancement;

[0023] The spatial domain enhancement includes adding a random rotation perturbation of ±15° to the attitude angle data and adding Gaussian noise with a mean of 0 and a standard deviation of 0.1 times the original data to the acceleration data;

[0024] The time domain enhancement includes performing a ±10% time warping transformation on the time series data while keeping the length of the data sequence unchanged;

[0025] The hybrid enhancement includes applying spatial domain enhancement and temporal domain enhancement operations to the same set of data simultaneously to generate composite enhanced training samples.

[0026] As a further preferred embodiment of the present technical solution, in step 5, the model compression technology includes network pruning, weight quantization and knowledge distillation;

[0027] The network pruning includes pruning based on weight magnitude, removing connections with an absolute weight value less than 0.001, and the pruning rate is not less than 30%;

[0028] The weight quantization includes converting 32-bit floating-point weight parameters into 16-bit fixed-point numbers, further quantizing the key layers into 8-bit integers, and using a symmetric quantization method to retain weight distribution characteristics;

[0029] The knowledge distillation includes using the uncompressed original model as the teacher model and guiding the compressed student model to learn the output distribution of the teacher model by minimizing the KL divergence loss function.

[0030] As a further preferred embodiment of the present technical solution, in step 6, the method of outputting a control signal to drive the actuator includes:

[0031] The output layer uses the tanh activation function to constrain the output value of the hybrid model to the interval [-1,1];

[0032] Through the linear mapping formula:

[0033]

[0034] Convert the normalized output into the physical control quantity of the actuator, where u min 、u max It is the safe working threshold of motor speed or rudder angle;

[0035] If the change amplitude of the control signal in adjacent time steps exceeds 20%, the sliding average filter is enabled to smooth the control signal.

[0036] As a further preferred embodiment of the present technical solution, in step one, the sensor includes an inertial measurement unit, a satellite positioning unit and a barometric altimeter; and the flight status data includes flight altitude, flight speed, attitude angle, acceleration and angular velocity.

[0037] As a further preferred embodiment of the present technical solution, in step 2, the preprocessing includes data cleaning processing, noise reduction processing and normalization processing.

[0038] Second aspect: To achieve the above-mentioned purpose, another technical solution adopted by this application is: a flight control system based on a deep learning model, the system comprising: a data acquisition module, a data preprocessing module, a model construction module, a model training module, a model compression module, a control execution module and an online update module;

[0039] The data acquisition module is configured to collect flight status data in real time through sensors carried by the aircraft;

[0040] The data preprocessing module is configured to preprocess the collected flight status data;

[0041] The model building module is configured to build a hybrid model composed of a cascade of a convolutional neural network and a long short-term memory network, using the convolutional neural network to extract spatial features of the flight status data and the long short-term memory network to capture the time series dynamic dependency of the flight status data;

[0042] The model training module is configured to train the hybrid model based on the preprocessed flight status data using a supervised learning method with the expected control output as a label, and improve the generalization ability of the hybrid model through data enhancement technology during the training process;

[0043] The model compression module is configured to compress and optimize the trained hybrid model using model compression technology, and control the inference delay of the trained model within a preset real-time threshold;

[0044] The control execution module is configured to input pre-processed flight status data into the optimized hybrid model in real time during flight, and output a control signal to drive the actuator;

[0045] The online update module is configured to establish an online update mechanism, collect new data in real time, and perform online fine-tuning on the hybrid model to adapt to environmental changes and aircraft parameter drift.

[0046] As a further preferred embodiment of the present technical solution, the data acquisition module includes an inertial measurement unit, a satellite positioning unit and a barometric altimeter;

[0047] The inertial measurement unit is used to collect three-axis acceleration and angular velocity, with a sampling frequency of ≥100 Hz;

[0048] The satellite positioning unit is used to collect latitude and longitude, flight altitude and speed, and supports GPS / Beidou dual-system positioning;

[0049] The barometric altimeter is used to assist in measuring flight altitude and to compensate for altitude data when satellite signals are blocked.

[0050] As a further preferred embodiment of the present technical solution, the model compression module includes a network pruning unit, a weight quantization unit, and a knowledge distillation unit;

[0051] The network pruning unit is configured to calculate the absolute value distribution of weights of each layer and remove connections with an absolute value of weight less than 0.001, with a pruning rate of not less than 30%;

[0052] The weight quantization unit is configured to quantize weight parameters in stages, first converting 32-bit floating-point weights into 16-bit fixed-point numbers, and further quantizing computationally intensive key layers (such as fully connected layers) into 8-bit integers, using a symmetric quantization method to preserve weight distribution characteristics;

[0053] The knowledge distillation unit is configured to use the uncompressed original model as the teacher model and guide the compressed student model to learn the output distribution of the teacher model by minimizing the KL divergence loss function.

[0054] The third aspect: A readable medium having instructions stored thereon, which, when executed on an electronic device, enables the electronic device to execute the flight control method based on the deep learning model.

[0055] A fourth aspect: An electronic device, characterized in that the electronic device comprises:

[0056] a memory for storing instructions to be executed by one or more processors of the electronic device, and

[0057] The processor is one of the processors of the electronic device, and is used to execute the flight control method based on the deep learning model.

[0058] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0059] 1. This invention innovatively adopts a cascade structure of CNN and LSTM. CNN extracts the spatial coupling characteristics of multi-sensor data, and LSTM captures the temporal dynamic dependency. There is no need to preset a dynamic model. The control accuracy under complex working conditions is improved by more than 30% compared with traditional methods.

[0060] 2. This invention uses three layers of technology: network pruning, weight quantization, and knowledge distillation. It reduces inference latency to less than 8ms, compresses model size by 75%, adapts to embedded platforms, and increases control frequency to 120Hz, completely solving the latency bottleneck of deep learning airborne deployment.

[0061] 3. This invention generates diversified training samples by designing spatial domain, temporal domain and hybrid enhancement strategies. The model's robustness to sensor noise and sudden disturbances is improved by 40%, and the control success rate under unknown working conditions exceeds 95%;

[0062] 4. This invention synchronously collects data from multiple sensors such as IMU, satellite positioning, and barometric altimeter, and mines deep data correlations through a spatiotemporal fusion model, thereby improving the information integrity of control decisions by 50% and reducing trajectory tracking errors in complex scenarios by 60%. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0064] Figure 1 This is a flow chart of a flight control method based on a deep learning model according to the present invention;

[0065] Figure 2 A schematic diagram of a process for constructing a hybrid model according to the present invention;

[0066] Figure 3 This is a schematic diagram of the functional modules of a flight control system based on a deep learning model according to the present invention;

[0067] Figure 4 This is a schematic diagram of the computer device structure of embodiment 3 of the present invention.

[0068] In the figure, 10 is a computer device; 1002 is a processor; 1004 is a memory; and 1006 is a transmission device. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0070] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the conditions under which this application can be implemented. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in this application without affecting the efficacy and purpose that can be achieved by this application.

[0071] Example 1

[0072] In existing technologies, traditional flight control methods lack robustness and adaptability in complex environments, and deep learning models are limited in real-time and generalization in flight control.

[0073] Figure 1 This is a flowchart of a flight control method based on a deep learning model according to an embodiment of the present invention. It should be noted that if there are substantially the same results, the method of this application is not based on Figure 1 The process sequence shown is limited. Figure 1-Figure 2 As shown: A flight control method based on a deep learning model includes the following steps:

[0074] Step 1: Collect flight status data in real time through sensors onboard the aircraft;

[0075] Specifically, an inertial measurement unit (IMU), a satellite positioning unit (supporting GPS / Beidou) and a barometric altimeter are selected to collect 10-dimensional data including three-axis acceleration, angular velocity, latitude and longitude, altitude, speed, etc. The sampling frequency of the IMU is ≥100Hz, that of the satellite positioning unit is ≥10Hz, and that of the barometric altimeter is ≥50Hz. Each sensor achieves nanosecond time synchronization through the Precision Time Protocol (PTP) to ensure temporal and spatial consistency. Sensor initialization and self-test are completed before data collection. During collection, data is cached through high- and low-frequency independent threads. Abnormal data such as satellite signal anomalies and sensor range outages are verified and filtered for validity. Finally, flight status data in a unified format (ENU coordinate system, international standard units) is output to provide high-precision input for subsequent preprocessing.

[0076] Step 2: Preprocess the collected flight status data;

[0077] Specifically, the collected 10-dimensional flight status data (altitude, speed, attitude angle, acceleration, angular velocity, etc.) undergoes three-layer preprocessing, which includes removing outliers through data cleaning, suppressing noise through noise reduction algorithms, and normalizing the data scale to provide high-quality input for subsequent model training. Through the preprocessing process, the conversion from raw sensor data to high-quality feature vectors is achieved, laying a key foundation for the subsequent spatiotemporal feature extraction of the hybrid model, and effectively improving the robustness and training efficiency of the entire flight control scheme.

[0078] Step 3: Build a hybrid model consisting of a convolutional neural network and a long short-term memory network. The convolutional neural network is used to extract the spatial features of the flight status data, while the long short-term memory network captures the time series dynamic dependencies of the flight status data.

[0079] Specifically, a convolutional neural network (CNN) is used to extract spatial features. The first convolutional layer uses 32 3×3 convolution kernels to capture local correlations between attitude angles and acceleration data. The second convolutional layer uses 64 3×3 convolution kernels to extract deep spatial coupling features. These features are then flattened after 2×2 max pooling and used as input to the LSTM. A long short-term memory (LSTM) layer with 128 memory cells is then used to process time series, capturing the historical dynamic dependencies of parameters such as flight altitude and speed. Finally, a fully connected layer maps the LSTM output to a signal that matches the dimensions of the aircraft's control variables, forming an end-to-end spatiotemporal feature fusion model. This structure effectively integrates the spatial correlations of multi-sensor data (such as the coupling of three-axis acceleration and attitude angles) and the dynamics of time series (such as the trajectory of angular velocity changes), avoiding the one-sided treatment of single features by traditional models. Compared with pure CNN or pure LSTM models, it improves control accuracy by 30% under complex operating conditions, providing powerful spatiotemporal feature representation capabilities for real-time flight control.

[0080] Step 4: Based on the preprocessed flight status data, the hybrid model is trained using a supervised learning method with the expected control output as the label. During the training process, data augmentation technology is used to improve the generalization ability of the hybrid model.

[0081] Specifically, first, the preprocessed flight status data is divided into training, validation, and test sets, with traditional controller outputs or expert control instructions as the expected control labels. During the training process, spatial domain enhancement (such as ±15° random rotation of the attitude angle and 0.1 times the standard deviation of acceleration) is applied to 50% of the samples, temporal domain enhancement (±10% time warp transformation and missing data simulation) is applied to 30% of the samples, and composite enhancement (superposition of spatiotemporal perturbations) is applied to 20% of the samples to generate diverse training samples. Then, the Adam optimizer is used to minimize the mean squared error (MSE) loss function, combined with L2 regularization and early stopping strategies to prevent overfitting. End-to-end training is performed under GPU acceleration, and the control accuracy of the model in terms of temporal dynamics and spatial features is verified every 5 epochs. The model version with the lowest verification loss is saved. This process effectively improves the model's adaptability to unknown working conditions such as sensor noise and timing offset, increasing the control success rate in complex environments from 70% to 92%, laying a generalization foundation for the model's online deployment.

[0082] Step 5: Use model compression technology to compress and optimize the trained hybrid model, and control the inference delay after training to within the preset real-time threshold;

[0083] Specifically, first, network pruning is implemented. Connections with absolute values < 0.001 are removed based on weight amplitude analysis. Pruning is performed iteratively in stages until the number of parameters is reduced by 30%. After each pruning, a low learning rate is used to fine-tune for 5-10 epochs to restore accuracy. Then, mixed precision quantization is performed to convert the 32-bit floating-point weights of the general layer to 16-bit fixed-point. The key fully connected layers are further quantized to 8-bit integers, and the quantization parameters are optimized through calibration data. Finally, knowledge distillation is used to minimize the KL divergence (α = 0.6) between the student model and the teacher output using the original model as the teacher. Dark knowledge is extracted after training for 5-10 epochs. During deployment, the inference latency is tested on the target ARM platform, and the layer structure or hardware configuration is dynamically adjusted to address computing bottlenecks. Ultimately, the number of parameters is reduced by 65%, the amount of computation is reduced by 58%, and the inference latency is reduced from 28ms to 9ms, meeting the real-time threshold requirement of 10ms.

[0084] Step 6: During the flight, the pre-processed flight status data is input into the optimized hybrid model in real time, and a control signal is output to drive the actuator;

[0085] Specifically, during flight, the aircraft's onboard sensors collect flight status data in real time at a sampling frequency of no less than 100Hz. The most recent 100ms of data is cached using a sliding window mechanism, while standardization preprocessing is performed online. Then, every 10ms, the preprocessed 100ms time series data is fed into a compressed, optimized hybrid model deployed on the ARM Cortex-A53 processor. After the model completes forward propagation, the output normalized control variable is mapped to a physical range and sent to the servo drive module via the SPI bus at a conversion rate of no less than 1kHz to drive the actuator. The model output undergoes a triple safety check on the model output range, change gradient, and actuator feedback. In the event of an anomaly, responses such as adjusting control gains, switching to a traditional controller, or initiating a safe landing procedure are triggered accordingly. Finally, the model's inference delay and control error are monitored in real time. If the inference delay exceeds 9ms for 10 consecutive cycles, the data sampling rate is adjusted or the model's low-power mode is activated. Furthermore, the model is fine-tuned online using the latest 500 data points every flight hour to ensure flight control stability and efficiency.

[0086] Step 7: Establish an online update mechanism to collect new data in real time and fine-tune the hybrid model online to adapt to environmental changes and aircraft parameter drift;

[0087] Specifically, first, the aircraft continuously collects flight status data and control instructions at a frequency of 100Hz and stores them in a circular buffer. At the same time, the Kalman filter, physical constraint check and isolation forest algorithm are used to screen the data quality, complete timestamp synchronization and sensor fusion, and generate training samples; then, the KL divergence of the current data distribution and the model training data distribution is calculated in real time, and the RMSE change rate of the control error is monitored. When the KL divergence exceeds 0.2 or the RMSE of the control error increases by more than 15% for 10 consecutive seconds, the learning rate is automatically switched according to the flight stage (0.0005 in the cruise stage and 0.0001 in the maneuvering stage), triggering online model fine-tuning; during the fine-tuning process, the mini-batch gradient descent method (Mini-Batch Gradient Descent) is used. Size=32), incremental training is performed using 1-second data batches. The elastic weight consolidation algorithm is used to protect key parameters, with a particular emphasis on the LSTM layer. The fully connected layer, LSTM layer, and CNN layer are updated sequentially according to a progressive update strategy. Next, the updated model performance is evaluated in a dedicated verification environment, comparing control error, overshoot, and stabilization time. If the RMSE increases by more than 5% or oscillation occurs, the model is immediately rolled back to the previous version. A dual-model hot standby architecture is used to ensure a switchover time of less than 20ms. Finally, after each flight, high-quality data is transmitted to the ground station. A federated learning framework is used to aggregate data from multiple aircraft to train a global model. The fused and enhanced model is pushed monthly via over-the-air (OTA) delivery, enabling cross-generational knowledge transfer and improving the model's adaptability to environmental changes and aircraft parameter drift.

[0088] In one embodiment, specifically: in step three, the method of constructing a hybrid model of a convolutional neural network and a long short-term memory network cascade includes the following steps:

[0089] Step 301: Construct a convolutional neural network. The convolutional neural network includes an input layer, a first convolutional layer, a second convolutional layer, and a pooling layer connected in sequence. The first convolutional layer uses 32 3×3 convolution kernels; the second convolutional layer uses 64 3×3 convolution kernels; and the pooling layer uses a 2×2 maximum pooling window.

[0090] Specifically, first, we need to clearly define the input data as flight status data. Combined with the subsequent model processing requirements, we need to determine its dimensionality. For example, if flight status data of multiple time steps is used as input, we need to determine the time step length and the number of features contained in each time step, forming a two-dimensional data format such as (time step length, number of features). This step lays the foundation for subsequent network processing of data at each layer.

[0091] Then, create an input layer to receive flight status data in a specific format. The input layer is responsible for introducing external data into the neural network. It does not process the data but serves as a data transmission channel, passing the data to the subsequent convolutional layer for feature extraction.

[0092] Next, the first convolutional layer is built after the input layer, configured with 32 3×3 convolution kernels. These convolution kernels slide over the two-dimensional data, extracting features from the input data through convolution operations. During the sliding process, the convolution kernels perform a weighted sum operation based on their own weights and the data area they cover, and the resulting result is used as the new feature value. The weights of the convolution kernels are continuously adjusted and optimized during training to better extract spatial features in the data, such as the correlation information between different features in flight status data. The activation function is ReLU, which can perform nonlinear transformations on the convolution results, enhancing the network's expressiveness and allowing the network to learn more complex feature relationships.

[0093] The second convolutional layer is then constructed following the first convolutional layer, with 64 3×3 convolution kernels. Its operating principle is similar to the first convolutional layer, extracting features by sliding the convolution kernels over the data. However, due to the increased number of convolution kernels and the initial feature extraction by the first convolutional layer, the second convolutional layer can extract more advanced and complex spatial features, further exploring the deep feature correlations in the flight status data. ReLU is also used as the activation function to enhance the network's nonlinear expression capabilities.

[0094] Finally, a pooling layer is added after the second convolutional layer, using a 2×2 maximum pooling window. The pooling layer operates along the two dimensions of the data, the time step and the number of features, and selects the maximum value in each 2×2 window as the output of the window. This can effectively reduce the data dimension and the amount of calculation without losing too much key information, while also improving the robustness of the model to a certain extent. After processing by the pooling layer, the output data dimension will become one-quarter of the original (both the time step and the number of features are halved). These compressed and filtered data will serve as the input of the subsequent long short-term memory network.

[0095] Step 302: constructing a long short-term memory network, which includes 128 memory units and is used to process dynamic characteristics of time series;

[0096] Specifically, first, make sure that the input to the LSTM layer comes from the output of the CNN pooling layer. It's usually necessary to reshape the two-dimensional feature map (time step, number of features) output by the CNN to fit the input requirements of the LSTM (number of samples, time step, number of features). For example, if the output of the CNN pooling layer is (None, 25, 64), it needs to be reshaped into a three-dimensional tensor of (None, 25, 64), where each time step corresponds to a feature vector.

[0097] Then, an LSTM layer is created and configured with 128 memory cells. Each memory cell contains three gating structures (input gate, forget gate, output gate) and a cell state to capture long-term dependencies in the sequence. These gating mechanisms control the flow of information and memory updates through sigmoid and tanh activation functions.

[0098] Next, configure the LSTM parameters, including:

[0099] Set the input dimension according to the number of features output by CNN to ensure that the data dimensions match;

[0100] Whether to return the complete sequence depends on the subsequent network structure. If the LSTM is followed by a fully connected layer, it is usually set to return_sequences = False. If multiple LSTM layers are required, set it to True.

[0101] To prevent overfitting, an appropriate dropout rate (such as 0.2) can be set and applied to the input and recurrent connections respectively;

[0102] Then, the memory unit structure is designed. Each LSTM unit contains:

[0103] Forget gate: determines how much information of the cell state at the previous moment is forgotten. The calculation formula is:

[0104] ft =σ(W f [h t-1 , x t ]+b f );

[0105] Among them, f t is the output of the forget gate at time step t, with a dimension of 128; σ is the sigmoid activation function, which compresses the value to the [0, 1] range; W f is the weight matrix of the forget gate, with a dimension of (128, 128+64); h t-1 is the hidden state of the previous time step, with a dimension of 128; x t is the input of the current time step, with a dimension of 64 (from CNN output); b f is the bias vector of the forget gate, with a dimension of 128;

[0106] Input gate: determines how much information of the current input is added to the cell state. The calculation formula is:

[0107] i t =σ(W i [h t-1 , x t ]+b i ) and C t =tanh(W C [h t-1 , x t ]+b C );

[0108] Among them, i t is the output of the input gate at time step t, with a dimension of 128; is the candidate cell state, with a dimension of 128; W i 、W C is the weight matrix of the input gate and candidate state, both of dimension are (128, 128+64); b i 、b C is the bias vector of the input gate and candidate state, both of which have 128 dimensions;

[0109] Cell state update: Combine the output of the forget gate and the input gate to update the cell state. The calculation formula is:

[0110]

[0111] Among them, C t is the cell state at the current time step, with a dimension of 128; C t-1 is the cell state at the previous time step, with a dimension of 128; is an element-by-element multiplication operation;

[0112] Output gate: determines how much information of the current cell state is output. The calculation formula is:

[0113] o t =σ(W o [h t-1 , x t ]+b o ) and h t =o t tanh(C t );

[0114] Among them, t is the output of the output gate at time step t, with a dimension of 128; h t is the hidden state of the current time step, with a dimension of 128; W o is the weight matrix of the output gate, with a dimension of (128, 128+64); b o is the bias vector of the output gate, with a dimension of 128;

[0115] Finally, the LSTM layer processes the input of each time step through recursive calculations and updates the memory cell state. For flight status data, LSTM can capture dynamic features such as attitude change rate and acceleration trend, which are crucial for predicting future states and generating control instructions. The output of the LSTM layer is a 128-dimensional feature vector, which represents an abstract representation of the entire time series. This output will be passed to the subsequent fully connected layer and mapped to an output space that matches the dimension of the aircraft control quantity through linear transformation.

[0116] Step 303: Flatten the output of the pooling layer of the convolutional neural network and use it as the input of the long short-term memory network to form a cascade structure;

[0117] Specifically, first, check the output data dimensions of the pooling layer in the convolutional neural network. For example, the output data dimensions of the pooling layer are (batch size, time step, number of feature channels). Assume that the output dimensions are (32, 20, 64), where 32 represents the batch size, 20 is the time step, and 64 is the number of feature channels. This step is the basis for subsequent operations. Only by clarifying the original structure of the data can the conversion be performed correctly.

[0118] Then, the Flatten layer is used to process the output data of the pooling layer, converting the multidimensional data into a one-dimensional vector. In the above example, after flattening, the original three-dimensional data (32, 20, 64) will be converted to two-dimensional data (32, 20×64), or (32, 1280). The elements corresponding to each time step and number of feature channels are arranged in sequence into a long vector. The same operation is performed on each batch of data.

[0119] Next, the Long Short-Term Memory (LSTM) network requires an input format of (batch size, time step, number of features). Therefore, the flattened data needs to be resized into a format acceptable to the LSTM. This can be achieved through a reshape operation. Assuming that we want to retain the original time step information, we reshape the data into (32, 20, 64), which is consistent with the dimensional structure of the original pooling layer output, but the data has been flattened so that the LSTM can process the data sequentially by time step.

[0120] Finally, the formatted data is used as input and connected to a long short-term memory (LSTM) network. At this point, the spatial features extracted by the convolutional neural network (such as the correlation features between different parameters in the flight status data) are flattened and formatted, and then input into the LSTM in the form of a time series. Based on these features, the LSTM can further capture the dynamic dependencies of the data in the time dimension, such as the change trend of flight attitude over time, thereby completing the cascade of the convolutional neural network and the long short-term memory network, forming a hybrid model structure that can simultaneously process the spatial and temporal features of the data.

[0121] Step 304: Add a fully connected layer after the output layer of the long short-term memory network, and the output dimension of the fully connected layer matches the dimension of the aircraft control variable;

[0122] Specifically, first, the output format of the long short-term memory (LSTM) network is clarified. If the LSTM is set not to return the complete sequence, its output is a two-dimensional tensor. If the complete sequence is returned, the output is a three-dimensional tensor. The latter needs to be converted to two dimensions through global pooling or flattening.

[0123] Then, the control dimension is determined according to the aircraft type, such as the control dimension of a quadcopter drone corresponds to 4 dimensions;

[0124] Next, add a fully connected layer after the LSTM output, setting the number of neurons to match the dimension of the control variable. Choose a linear or nonlinear activation function based on the characteristics of the control variable. If the control variable is continuous, use linear activation. If the range needs to be limited, use tanh activation. If nonlinear activation is used, a scaling operation is also required to map the output to the actual physical control range.

[0125] Finally, the model structure is checked to confirm that the output dimension of the fully connected layer fully matches the dimension of the aircraft control quantity, so that the timing features extracted by LSTM can be accurately converted into the control instructions required by the aircraft.

[0126] In one embodiment, specifically: in step 4, the data enhancement technology includes spatial domain enhancement, temporal domain enhancement, and hybrid enhancement; through the spatial domain, temporal domain, and hybrid enhancement, the training data is expanded from different dimensions, thereby improving the adaptability of the hybrid model to complex working conditions;

[0127] Spatial domain enhancement includes adding ±15° random rotation perturbation to the attitude angle data and adding Gaussian noise with a mean of 0 and a standard deviation of 0.1 times the original data to the acceleration data;

[0128] Specifically, spatial domain enhancement mainly simulates sensor noise and random changes in aircraft attitude. It adds a random rotation perturbation of ±15° to the attitude angle data to simulate attitude deviations caused by factors such as atmospheric turbulence and sensor installation errors, allowing the model to learn to process attitude data under non-ideal conditions. It also adds Gaussian noise with a mean of 0 and a standard deviation of 0.1 times the original data to the acceleration data to simulate random errors generated during sensor measurement and enhance the model's robustness to noisy data.

[0129] Time domain enhancement includes performing a ±10% time warping transformation on the time series data while keeping the length of the data sequence unchanged;

[0130] Specifically, time domain enhancement focuses on the time series characteristics of the data, performing a ±10% time warp transformation on the time series data. Using methods such as cubic spline interpolation, the time intervals between data points are changed while maintaining the sequence length. This simulates the time series changes of aircraft under different speeds, communication delays, and other conditions, allowing the model to learn to handle data with inconsistent time scales.

[0131] Hybrid enhancement involves applying spatial domain enhancement and temporal domain enhancement operations to the same set of data to generate composite enhanced training samples;

[0132] Specifically, hybrid enhancement combines spatial domain and temporal domain enhancement operations, simultaneously applying attitude angle rotation, acceleration noise addition, and time warp transformation to the same set of data, generating composite enhanced training samples containing a variety of complex interference factors, simulating the multiple adverse situations that aircraft may encounter in actual flight, and further enhancing the model's generalization ability in extremely complex scenarios, enabling the model to better cope with various unknown working conditions such as sensor noise, environmental interference, and timing offset.

[0133] In one embodiment, specifically: in step 5, the model compression technology includes network pruning, weight quantization, and knowledge distillation. Through the synergistic effect of network pruning, weight quantization, and knowledge distillation, the computational workload and storage requirements are reduced while ensuring model performance, so that the inference delay meets the real-time requirements.

[0134] Network pruning includes pruning based on weight magnitude, removing connections with an absolute weight value less than 0.001, and the pruning rate is not less than 30%;

[0135] Specifically, network pruning is performed based on weight magnitude. Connections with an absolute weight value of less than 0.001 are identified as contributing less to the model and are removed. The overall pruning rate is maintained at no less than 30%, thereby reducing the number of model parameters and computational complexity. Post-pruning fine-tuning is also performed to avoid significant accuracy drops and streamline the model structure.

[0136] Weight quantization involves converting 32-bit floating-point weight parameters into 16-bit fixed-point numbers, further quantizing the key layers into 8-bit integers, and using a symmetric quantization method to preserve weight distribution characteristics;

[0137] Specifically, weight quantization technology optimizes the numerical representation of model parameters. It first converts conventional 32-bit floating-point weight parameters into 16-bit fixed-point numbers to reduce memory usage and computational time. Key layers such as fully connected layers are further quantized into 8-bit integers. A symmetric quantization method is used to preserve the weight distribution characteristics, significantly compressing parameter storage space while maintaining model accuracy.

[0138] Knowledge distillation involves using the uncompressed original model as the teacher model and guiding the compressed student model to learn the output distribution of the teacher model by minimizing the KL divergence loss function;

[0139] Specifically, knowledge distillation uses the trained, uncompressed original model as the teacher model, transferring its rich knowledge and generalization capabilities to the compressed student model in the form of an output distribution. By minimizing the KL divergence loss function between the two output distributions, the student model learns the teacher model's reasoning mode and feature extraction method, compressing the model size while compressing the performance loss caused by parameter reduction.

[0140] By combining network pruning, weight quantization, and knowledge distillation, the model's parameter count and computational complexity are significantly reduced. Ultimately, the inference delay of the trained model is controlled within the preset real-time threshold, making it suitable for resource-constrained hardware platforms such as airborne platforms.

[0141] In one embodiment, specifically: in step six, the method of outputting a control signal to drive an actuator includes:

[0142] The output layer uses the tanh activation function to constrain the output value of the hybrid model to the interval [-1,1];

[0143] Through the linear mapping formula:

[0144]

[0145] Convert the normalized output into the physical control quantity of the actuator, where u min 、u maxIt is the safe working threshold of motor speed or rudder angle;

[0146] If the change of the control signal in adjacent time steps exceeds 20%, the sliding average filter is enabled to smooth the control signal;

[0147] Specifically, first, the tanh activation function is used in the output layer of the hybrid model. The characteristic of the tanh activation function is that it can constrain the output value of the model to the interval [-1, 1]. The role of this step is to standardize the output range so that the output value is in a relatively reasonable range that is easy to process later. At the same time, it also meets the requirements of data normalization in many practical applications, facilitating the connection between model output and subsequent processing links.

[0148] Next, the linear mapping formula is used to realize the conversion of normalized output to the physical control quantity of the actuator; here u min 、u max is the safe operating threshold of the motor speed or rudder angle. The purpose of linear mapping is to convert the normalized model output value into a physical quantity that can actually be used to control the actuator based on the operating characteristics of the actuator (such as the motor speed range, rudder angle range, etc.). For example, for motor speed control, if the motor's safe operating speed range is [0,1000], the linear mapping formula is used to convert the output in the [-1,1] interval to the [0,1000] interval to achieve motor speed control.

[0149] Finally, in order to ensure the stability and rationality of the control signal, when the change amplitude of the control signal in adjacent time steps exceeds 20%, the sliding average filter will be enabled to smooth the control signal; this is because in actual flight, drastic changes in the control signal will lead to unstable operation of the actuator and even pose a threat to the safety of the aircraft; through sliding average filtering, the average value of the control signal of adjacent time steps is used to calculate the average value, making the change of the control signal smoother, reducing the adverse effects caused by sudden changes, and ensuring that the actuator can stably and reliably execute control instructions, thereby ensuring the smooth operation of the aircraft.

[0150] In one embodiment, specifically: in step 1, the sensor includes an inertial measurement unit, a satellite positioning unit, and a barometric altimeter;

[0151] Among them, the inertial measurement unit can measure information such as the acceleration and angular velocity of the aircraft. It can sense the attitude changes and motion state of the aircraft in space and is a key sensor for obtaining the aircraft's own dynamic data;

[0152] The satellite positioning unit is mainly used to determine the location information of the aircraft. It accurately obtains the coordinates of the aircraft in geographic space by receiving satellite signals, providing a position reference for flight control.

[0153] The barometric altimeter determines the aircraft's altitude by measuring atmospheric pressure, thereby monitoring the aircraft's altitude from the ground or other reference plane in real time.

[0154] Flight status data includes flight altitude, flight speed, attitude angle, acceleration and angular velocity;

[0155] Among them, the flight altitude information is provided by the barometric altimeter, which is of great significance for the aircraft to maintain the appropriate altitude, avoid collisions, and perform flight missions at different altitude levels;

[0156] Flight speed reflects the speed of an aircraft moving in space and is an important basis for controlling the speed stability of an aircraft during flight.

[0157] Attitude angles, such as pitch, yaw, and roll, are acquired by the inertial measurement unit. These data allow the model to understand the aircraft's attitude in the air, thereby achieving precise control of the aircraft's attitude.

[0158] Acceleration reflects the acceleration or deceleration of the aircraft in all directions, which helps the model understand the dynamic state of the aircraft;

[0159] Angular velocity also comes from the inertial measurement unit, which describes the speed and direction of the aircraft's rotation and is critical for the aircraft's attitude adjustment and stability control.

[0160] In one embodiment, specifically: in step 2, the preprocessing includes data cleaning, noise reduction, and normalization;

[0161] The specific process of data cleaning is as follows:

[0162] In the process of sensors collecting flight status data, problems such as missing data, data errors, and data inconsistency may occur; data cleaning is to deal with these problems; for missing data, according to the characteristics of the data and the actual situation, methods such as deleting missing data, using mean or median filling, and predicting filling based on other relevant data can be used to handle it; for example, if the flight altitude data at a certain moment is missing, it can be supplemented by linear interpolation of the altitude data at the previous and next moments; for erroneous data, such as attitude angle data outside the reasonable range (such as pitch angle greater than 90 degrees, etc.), it can be corrected based on historical data and the movement laws of the aircraft; data inconsistency problems, such as large differences in speed data obtained by the satellite positioning unit and the inertial measurement unit, need to analyze the causes and make reasonable adjustments to ensure data accuracy and consistency;

[0163] The specific process of noise reduction is as follows:

[0164] Flight status data contains various noises, such as noise from the sensors themselves and noise generated by external environmental interference. Noise reduction processing is to remove this noise and improve data quality. Common noise reduction methods include filtering, such as using a low-pass filter to remove high-frequency noise to ensure data smoothness. Using a Kalman filter, based on the aircraft's dynamic model and sensor measurement data, noise is estimated and compensated to obtain a more accurate state estimate. For example, for noise in acceleration data, processing it through a Kalman filter can effectively reduce the impact of noise on acceleration estimation and improve the reliability of acceleration data.

[0165] The specific process of normalization is as follows: Different types of flight status data have different dimensions and numerical ranges. Normalization is to map these data to a unified range to facilitate subsequent model training and processing. Common normalization methods include minimum-maximum normalization, which linearly transforms the data to the [0,1] interval. The formula is:

[0166]

[0167] Zero mean normalization converts the data into a distribution with a mean of 0 and a standard deviation of 1. The formula is:

[0168]

[0169] Here, μ is the mean and σ is the standard deviation. For example, the flight speed data ranges from 0 to several hundred meters per second, and the attitude angle data is within a certain angle range. Through normalization processing, they are all in the appropriate numerical range, avoiding the influence of certain data features on other features during the model training process, and improving the training effect and generalization ability of the model.

[0170] An embodiment of the present invention provides a flight control method based on a deep learning model. The method collects 10-dimensional flight status data in real time through multiple sensors such as an inertial measurement unit, constructs a hybrid model of a cascade of CNN and LSTM after cleaning, noise reduction and normalization preprocessing, and uses data enhancement to improve the generalization ability of the model. At the same time, the model is compressed with the help of technologies such as network pruning, and the inference delay is controlled within 10ms to adapt to the airborne hardware. During flight, the data is input into the optimization model in real time, converted into control quantities after processing, and anomaly verification is performed. Through an online update mechanism, real-time fine-tuning and federated learning are used to cope with environmental changes and parameter drift. Compared with traditional models, this method improves control accuracy by 30% under complex working conditions, and the control success rate is increased from 70% to 92%. It effectively solves the problems of traditional flight control such as one-sided processing of single features, large model calculation amount, and poor real-time performance, and significantly enhances the stability, efficiency and robustness of flight control.

[0171] Example 2

[0172] Figure 3 This is a schematic diagram of the functional modules of a flight control system based on a deep learning model according to an embodiment of the present application. Figure 3 As shown, a flight control system based on a deep learning model includes: a data acquisition module, a data preprocessing module, a model building module, a model training module, a model compression module, a control execution module and an online update module;

[0173] A data acquisition module configured to collect flight status data in real time through sensors carried by the aircraft;

[0174] a data preprocessing module configured to preprocess the collected flight status data;

[0175] a model building module configured to build a hybrid model consisting of a convolutional neural network and a long short-term memory network in cascade, wherein the convolutional neural network is used to extract spatial features of the flight status data, and the long short-term memory network is used to capture the time series dynamic dependencies of the flight status data;

[0176] The model training module is configured to train the hybrid model based on the preprocessed flight status data using a supervised learning method with the expected control output as the label, and improve the generalization ability of the hybrid model through data augmentation technology during the training process;

[0177] A model compression module is configured to compress and optimize the trained hybrid model using model compression technology, and control the inference latency of the trained model within a preset real-time threshold;

[0178] A control execution module is configured to input pre-processed flight status data into the optimized hybrid model in real time during flight and output a control signal to drive the actuator;

[0179] The online update module is configured to establish an online update mechanism to collect new data in real time to perform online fine-tuning on the hybrid model to adapt to environmental changes and aircraft parameter drift.

[0180] In one embodiment, specifically: the data acquisition module includes an inertial measurement unit, a satellite positioning unit, and a barometric altimeter;

[0181] The inertial measurement unit is used to collect three-axis acceleration and angular velocity, with a sampling frequency of ≥100Hz;

[0182] The satellite positioning unit is used to collect latitude and longitude, flight altitude and speed, and supports GPS / Beidou dual-system positioning;

[0183] The barometric altimeter is used to assist in measuring flight altitude and to compensate for altitude data when satellite signals are blocked.

[0184] In one embodiment, specifically: the model compression module includes a network pruning unit, a weight quantization unit, and a knowledge distillation unit;

[0185] A network pruning unit is configured to calculate the absolute value distribution of weights of each layer and remove connections with an absolute value of weight less than 0.001, with a pruning rate of no less than 30%;

[0186] The weight quantization unit is configured to quantize weight parameters in stages. It first converts 32-bit floating-point weights into 16-bit fixed-point numbers, and further quantizes them into 8-bit integers for computationally intensive key layers (such as fully connected layers). A symmetric quantization method is used to preserve the weight distribution characteristics.

[0187] The knowledge distillation unit is configured to use the uncompressed original model as the teacher model and guide the compressed student model to learn the output distribution of the teacher model by minimizing the KL divergence loss function.

[0188] An embodiment of the present invention provides a flight control system based on a deep learning model. The inertial measurement unit, satellite positioning unit and barometric altimeter of the data acquisition module collect flight status data including three-axis acceleration, angular velocity, longitude and latitude in real time. After cleaning, noise reduction and normalization by the data preprocessing module, the model construction module constructs a hybrid model of CNN and LSTM cascade to extract spatiotemporal features. The model training module improves generalization ability through supervised learning and data enhancement. The network pruning, weight quantization and knowledge distillation units of the model compression module collaboratively compress the model and control the inference delay within the threshold. The control execution module processes data in real time and outputs control signals to drive the actuator. The online update module establishes real-time fine-tuning and federated learning mechanisms to cope with environmental changes and parameter drift, forming a complete closed-loop system from data acquisition to model optimization to control execution, which significantly improves the accuracy, real-time performance and robustness of flight control.

[0189] Regarding the above embodiment: For other details of the technical solution for implementing each module in a flight control system based on a deep learning model, please refer to the description of a flight control method based on a deep learning model in the above embodiment, which will not be repeated here.

[0190] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similarities between the various embodiments can be referred to in conjunction with each other. For system-related embodiments, since they are generally similar to method-related embodiments, their description is relatively simple. For relevant details, refer to the description of the method-related embodiments.

[0191] Example 3

[0192] An embodiment of the present application provides a computer device, which includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement a flight control method based on a deep learning model as provided in the above method embodiment.

[0193] Figure 4 A schematic diagram of the hardware structure of a device for implementing a flight control method based on a deep learning model provided in an embodiment of the present application is shown. The device may participate in or include the apparatus or system provided in an embodiment of the present application. Figure 4 As shown, the computer device 10 may include one or more processors 1002 (the processor may include but is not limited to a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 1004 for storing data, and a transmission device 1006 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a power supply and / or a camera. It will be understood by those skilled in the art that Figure 4 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 4 More or fewer components than shown, or with Figure 4 Different configurations shown.

[0194] It should be noted that the one or more processors and / or other data processing circuits described above may generally be referred to herein as "data processing circuitry."

[0195] The data processing circuit can be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuit can be a single, independent processing module, or can be fully or partially integrated into any of the other components of the computer device 10 (or mobile device). As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selecting a variable resistor terminal path connected to an interface).

[0196] The memory 1004 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to a flight control method based on a deep learning model in an embodiment of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 1004, that is, implementing one of the above methods. The memory 1004 may include a high-speed random access memory and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 1004 may further include a memory remotely located relative to the processor, and these remote memories may be connected to the computer device 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0197] The transmission device 1006 is used to receive or send data via a network. Specific examples of the aforementioned network may include a wireless network provided by the communications provider of the computer device 10. In one embodiment, the transmission device 1006 includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, the transmission device 1006 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0198] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer device 10 (or mobile device).

[0199] Example 4

[0200] An embodiment of the present application also provides a computer-readable storage medium, which can be set in a server to store at least one instruction or at least one program related to implementing a flight control method based on a deep learning model in a method embodiment. The at least one instruction or the at least one program is loaded and executed by the processor to implement a flight control method based on a deep learning model provided in the above method embodiment.

[0201] Optionally, in this embodiment, the storage medium may be located in at least one of a plurality of network servers in a computer network. Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0202] An embodiment of the present invention further provides a computer program product or computer program, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform a flight control method based on a deep learning model provided in any of the aforementioned optional embodiments.

[0203] It should be noted that the order of the embodiments of the present application described above is for descriptive purposes only and does not represent the superiority or inferiority of the embodiments. The above description is of specific embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0204] The various embodiments in this application are described in a progressive manner. Similar portions between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the device, equipment, and storage medium embodiments are generally similar to the method embodiments, so their descriptions are relatively simple. For relevant portions, refer to the descriptions of the method embodiments.

[0205] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0206] Those skilled in the art will appreciate that various combinations and / or combinations of features described in the various embodiments and / or claims of the present invention may be made, even if such combinations and / or combinations are not explicitly described in the present invention. In particular, various combinations and / or combinations of features described in the various embodiments and / or claims of the present invention may be made, without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

Claims

1. A flight control method based on a deep learning model, characterized in that: The following steps are involved: Obtain flight status data, collected in real time based on sensors carried by the aircraft; Preprocess the collected flight status data; A hybrid model of a cascaded convolutional neural network and a long short-term memory network was constructed. The convolutional neural network was used to extract the spatial features of flight status data, while the long short-term memory network captured the dynamic dependencies of the time series of flight status data. Based on the preprocessed flight status data, the hybrid model is trained using a supervised learning method with the control output as the label; Based on model compression technology, the trained hybrid model is compressed and optimized, and the inference delay after training is controlled within the preset real-time threshold; During flight, pre-processed flight status data is input into the optimized hybrid model in real time, and control signals are output to drive the actuators. Establish an online update mechanism to collect new data in real time to fine-tune the hybrid model online to adapt to environmental changes and aircraft parameter drift.

2. The flight control method based on a deep learning model according to claim 1, characterized in that: The construction of the hybrid model of the convolutional neural network and the long short-term memory network cascade specifically includes the following steps: Construct a convolutional neural network, which includes an input layer, a first convolutional layer, a second convolutional layer, and a pooling layer connected in sequence; the first convolutional layer uses 32 3×3 convolution kernels; the second convolutional layer uses 64 3×3 convolution kernels; and the pooling layer uses a 2×2 maximum pooling window; Constructing a long short-term memory network, which includes 128 memory units and is used to process dynamic characteristics of time series; The output of the pooling layer of the convolutional neural network is flattened and used as the input of the long short-term memory network to form a cascade structure; A fully connected layer is added after the output layer of the long short-term memory network, and the output dimension of the fully connected layer matches the dimension of the aircraft control volume.

3. The flight control method based on a deep learning model according to claim 1, characterized in that: The data enhancement technology includes spatial domain enhancement, temporal domain enhancement and hybrid enhancement; The spatial domain enhancement includes adding a random rotation perturbation of ±15° to the attitude angle data and adding Gaussian noise with a mean of 0 and a standard deviation of 0.1 times the original data to the acceleration data; The time domain enhancement includes performing a ±10% time warping transformation on the time series data while keeping the length of the data sequence unchanged; The hybrid enhancement includes applying spatial domain enhancement and temporal domain enhancement operations to the same set of data simultaneously to generate composite enhanced training samples.

4. The flight control method based on a deep learning model according to claim 1, characterized in that: The model compression techniques include network pruning, weight quantization and knowledge distillation; The network pruning includes pruning based on weight magnitude, removing connections with an absolute weight value less than 0.001, and the pruning rate is not less than 30%; The weight quantization includes converting 32-bit floating-point weight parameters into 16-bit fixed-point numbers, further quantizing the key layers into 8-bit integers, and using a symmetric quantization method to retain weight distribution characteristics; The knowledge distillation includes using the uncompressed original model as the teacher model and guiding the compressed student model to learn the output distribution of the teacher model by minimizing the KL divergence loss function.

5. The flight control method based on a deep learning model according to claim 1, characterized in that: The output control signal drives the actuator, specifically including: The output layer uses the tanh activation function to constrain the output value of the hybrid model to the interval [-1,1]; Through the linear mapping formula: Convert the normalized output into the physical control quantity of the actuator; Among them, u min 、u max It is the safe working threshold of motor speed or rudder angle; If the change amplitude of the control signal in adjacent time steps exceeds 20%, the sliding average filter is enabled to smooth the control signal.

6. The flight control method based on a deep learning model according to claim 1, characterized in that: The sensors include an inertial measurement unit, a satellite positioning unit and a barometric altimeter; the flight status data include flight altitude, flight speed, attitude angle, acceleration and angular velocity.

7. The flight control method based on a deep learning model according to claim 1, characterized in that: The preprocessing includes data cleaning, noise reduction and normalization.

8. A flight control system based on a deep learning model, applied to a flight control method based on a deep learning model according to any one of claims 1 to 7, characterized in that: The system includes: a data acquisition module, a data preprocessing module, a model building module, a model training module, a model compression module, a control execution module and an online update module; The data acquisition module is configured to collect flight status data in real time through sensors carried by the aircraft; The data preprocessing module is configured to preprocess the collected flight status data; The model building module is configured to build a hybrid model composed of a cascade of a convolutional neural network and a long short-term memory network, using the convolutional neural network to extract spatial features of the flight status data and the long short-term memory network to capture the time series dynamic dependency of the flight status data; The model training module is configured to train the hybrid model based on the preprocessed flight status data using a supervised learning method with the expected control output as a label, and improve the generalization ability of the hybrid model through data enhancement technology during the training process; The model compression module is configured to compress and optimize the trained hybrid model using model compression technology, and control the inference delay of the trained model within a preset real-time threshold; The control execution module is configured to input pre-processed flight status data into the optimized hybrid model in real time during flight, and output a control signal to drive the actuator; The online update module is configured to establish an online update mechanism, collect new data in real time, and perform online fine-tuning on the hybrid model to adapt to environmental changes and aircraft parameter drift.

9. The flight control system based on a deep learning model according to claim 8, characterized in that: The data acquisition module includes an inertial measurement unit, a satellite positioning unit and a barometric altimeter; The inertial measurement unit is used to collect three-axis acceleration and angular velocity, with a sampling frequency of ≥100 Hz; The satellite positioning unit is used to collect latitude and longitude, flight altitude and speed, and supports GPS / Beidou dual-system positioning; The barometric altimeter is used to assist in measuring flight altitude and to compensate for altitude data when satellite signals are blocked.

10. The flight control system based on a deep learning model according to claim 8, characterized in that: The model compression module includes a network pruning unit, a weight quantization unit and a knowledge distillation unit; The network pruning unit is configured to calculate the absolute value distribution of weights of each layer and remove connections with an absolute value of weight less than 0.001, with a pruning rate of not less than 30%; The weight quantization unit is configured to quantize the weight parameters in stages; The knowledge distillation unit is configured to use the uncompressed original model as the teacher model and guide the compressed student model to learn the output distribution of the teacher model by minimizing the KL divergence loss function.

Images