All-electric drive aircraft tractor control system and method

Abstract

The invention relates to the field of electric digital data processing, in particular to a control system and method for an all-electric drive aircraft tractor. The method mainly comprises nine steps from S1 to S9. The system comprises a data storage module, a data acquisition module, a data expansion module, a data processing module, a tractor control module and an execution module. An existing full-electric drive aircraft tractor control system has the problems that the data processing capacity is poor and the decision-making accuracy is low. The control system and method for the all-electric drive aircraft tractor are high in data processing capacity and high in decision-making accuracy.

Description

Technical Field

[0001] The invention relates to the field of electric digital data processing, and in particular to a control system and method for a fully electric driven aircraft tractor. Background Art

[0002] Aircraft tractors are a type of support equipment used to tow aircraft on the ground at airports. For energy and other reasons, they are usually fully electric. Due to the large mass and size of aircraft and their high price, fully electric aircraft tractors are usually operated by experienced pilots to ensure the safety of aircraft. However, with the increasing number and types of aircraft, fully electric aircraft tractors are used more frequently. Relying solely on experienced pilots can no longer meet current driving needs. Therefore, people have developed a fully electric aircraft tractor control system to achieve functions such as automatic steering to reduce the workload of the driver and reduce the requirements for the driver's ability.

[0003] The existing all-electric drive aircraft tractor control system usually obtains driving data for the system to analyze and judge and make operating instructions through sensors, GPS, etc., maps the driving data to a certain one of the given operating instruction categories to obtain specific operating instructions, and then executes the operating instructions to assist driving. However, due to the further increase in airport area and other reasons, the complexity of the airport operating environment has increased, and more data needs to be collected as driving data, so the complexity of driving data has greatly increased. The data processing of the existing all-electric drive aircraft tractor control system mainly includes feature extraction and feature dimensionality reduction. Feature extraction is mostly based on classical neural networks or linear methods, which are difficult to process data with highly complex relationships and high dimensions, resulting in poor feature extraction effects and inability to effectively capture important information in the data; feature dimensionality reduction usually cannot dynamically optimize the feature space, and is prone to losing important information or retaining irrelevant information, resulting in the inability of the reduced dimensionality data to reduce the dimension while ensuring data integrity, making the data processing capability of the existing all-electric drive aircraft tractor control system relatively weak. At the same time, the classifier algorithm of the existing all-electric aircraft tractor control system lacks an adaptive adjustment mechanism when faced with complex data, and is unable to optimize the classification decision boundary according to real-time data distribution changes, resulting in a decrease in classification accuracy and insufficient generalization ability when dealing with complex or extreme situations, resulting in low decision-making accuracy.

[0004] Therefore, the existing all-electric drive aircraft tractor control system has the problems of weak data processing capability and low decision-making accuracy. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a full-electric drive aircraft tractor control system and method with strong data processing capability and high decision-making accuracy.

[0006] In order to solve the above technical problems, the present invention provides a control method for an all-electrically driven aircraft tractor, comprising: S1. Collecting real driving data collected in the past by the data collection module, where the driving data includes environmental monitoring data and sensor data; S2, marking the operation instructions to be performed corresponding to each real driving data, and after the marking is completed, sending the marked real driving data to the data storage module for storage; The marking method is manual marking, and the operation instructions include driving instructions, steering instructions, braking instructions and emergency stop instructions; S3, sending the annotated real driving data to the data expansion module, the data expansion module uses a generative adversarial network algorithm based on adversarial perturbations to generate expanded driving data, and sends the expanded driving data to the storage module for storage; S4, the annotated real driving data and the expanded driving data together constitute training samples, and the training samples are input into a data processing module, the data processing module includes a feature extraction model based on a neural network algorithm of quantum dynamic evolution and a feature dimensionality reduction model based on an autoencoder neural network of feature refinement; First, the training samples are used to train the feature extraction model of the neural network algorithm based on quantum dynamic evolution to obtain the training samples after feature extraction; then, the training samples after feature extraction are used to train the feature dimension reduction model based on the autoencoder neural network based on feature refinement to obtain the training samples after feature dimension reduction; S5, inputting the training samples after feature dimension reduction into the tractor control module, the tractor control module includes a classifier of the integrated gradient boosting tree algorithm based on the dynamic decision boundary, and using the training samples after feature dimension reduction to train the classifier of the integrated gradient boosting tree algorithm based on the dynamic decision boundary; S6. Collect new real driving data through the data collection module, and send the new real driving data to the data storage module for storage; S7, inputting the new real driving data into the data processing module, first performing feature extraction on the new real driving data using the feature extraction model to obtain new real driving data after feature extraction, and performing feature dimension reduction on the new real driving data after feature extraction using the feature dimension reduction model to obtain new real driving data after feature dimension reduction; S8, inputting the new real driving data after feature dimension reduction into the tractor control module, using a classifier to analyze the new real driving data after feature dimension reduction to obtain specific operation instructions and sending the operation instructions to the execution module; S9. The execution module receives and executes the operation instruction.

[0007] As a further improvement of the present invention: the step of generating the expanded driving data by using the generative adversarial network algorithm based on adversarial perturbation in the data expansion module in S3 includes: S301, initialize the network structures of the generator and the discriminator respectively; S302: Using the topological structure constraint strategy in the generator, by encoding the geometric spatial relationship of each driving data point, establish its position association in the real driving data, and the topological structure constraint The calculation formula is: , In the formula, is the topological structure constraint weight factor, For real driving data dimensions, and Generators The output feature vectors of the two extended driving data are: is the feature vector of the i-th extended driving data, is the feature vector of the j-th extended driving data, for norm, For the generator The standard deviation of the output augmented driving data, is a small random number in the topological structure constraint, and is the feature vector of the corresponding real driving data, is the feature vector of the i-th real driving data, is the feature vector of the jth real driving data; is the standard deviation of real driving data; S303: According to the adversarial perturbation defense strategy, a perturbation term is added during the generator training process to apply random perturbations to the driving data generated by the generator. The calculation formula is: , In the formula, is the disturbance intensity represents the gradient of the input vector, represents the hyperbolic tangent function, represents the parameter that controls the gradient scaling, is the disturbance intensity, represents the loss function of the discriminator, represents the discriminator function, The feature vector representing the real driving data, represents the true label, represents the quantum bit parameters; S304: The generator generates a batch of extended driving data similar to the real driving data through the initial noise input. The extended driving data output by the generator after adding disturbance It is expressed as: , In the formula, is the generator function, is the input noise of the generator, is an adversarial perturbation; S305: The discriminator performs binary classification training on the real driving data and the expanded driving data through the cross entropy loss function, and updates the weight parameters of the discriminator according to the correctness of the classification. The update method is expressed as: , In the formula, is the learning rate of the discriminator, is the loss function relative to the discriminator weights The Hessian matrix of express The inverse matrix of represents the gradient of the discriminator weight, represents the loss function of the discriminator, is the classification result of the discriminator on the expanded driving data, represents the true label; S306. Optimize the generator by combining the quantization loss function. The update method is expressed as: , In the formula, is the learning rate of the generator, represents the gradient of the generator weight, is the loss function of the generator, represents the discriminator function, is the generator function, is the input noise of the generator, is the false sample label, is the regularization parameter that controls the strength of the topological constraint, is the topological constraint, To control the parameters of the regularization effect, is the regularization term of the generator parameters; S307, repeating iterations S302 to S306 until a preset stop iteration condition is met.

[0008] As a further improvement of the present invention: the step of training the feature extraction model of the neural network algorithm based on quantum dynamic evolution in S4 includes: S411. Initialize the structure of the neural network based on quantum dynamic evolution. Assume that each layer of the neural network contains a number of quantum neurons, each quantum neuron has the ability to perform quantum gate operations, each quantum neuron parameter is represented by a quantum state, and the weight parameter of the first layer of the neural network is Defined as: , In the formula, is the number of quantum neurons in the neural network; are the complex probability amplitudes of the kth quantum neuron in the neural network, indicating the probability amplitudes of the quantum being in the ground state and the excited state, ensuring ; is the first ground state of the quantum bit, is the second ground state of the quantum bit; S412. Define a quantum gate operation for each neuron. The quantum gate operation is represented by a matrix of the following form: , In the formula, is the quantum gate operation of the neural network layer l, is an imaginary unit, is the parameter that controls the rotation of the quantum state, is the Pauli-Y matrix; S413. When the input driving data is forward propagated through the network, the quantum neurons of each layer process the driving data according to their quantum gate operations. The process of forward propagation of the driving data through the quantum neural network is represented as follows: , In the formula, is the state of the l+1th layer after quantum gate and weight operation, is the quantum gate operation of the neural network layer l, is the weight parameter of the lth layer of the neural network, is the quantum neuron state of the lth layer; S414. Define a quantum-inspired loss function to evaluate the effect of feature extraction, which not only considers the traditional error index, but also includes the fidelity index of the quantum state. Use this loss function for backpropagation to update the parameters of each quantum gate in the network and the parameters of the traditional neural network. The loss function of the neural network The calculation formula is: , In the formula, is the number of samples input to the neural network in the current batch, is the target value of the nth sample, is the expected value of the quantum state output by the network, is the regularization parameter of the neural network, is the quantum fidelity function, is the quantum neuron state of the Lth layer, is an ideal quantum state; S415, repeating iterations S412 to S414 until a preset stop iteration condition is met.

[0009] As a further improvement of the present invention: the feature dimension reduction model of the autoencoder neural network based on feature refinement is trained in S4, including: S421, assuming that the data input to the autoencoder neural network is , the encoder uses a multi-layer nonlinear mapping structure to map high-dimensional data to the initial low-dimensional feature space, which can be expressed as: , In the formula, represents the initial low-dimensional features, is the multi-layer Sigmoid activation function of the encoder, is the weight matrix of the encoder, is the data input to the autoencoder neural network, is the bias vector of the encoder; S422, after the low-dimensional features are generated, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space. When the module is initialized, all features are given the same initial weight so that they can be gradually adjusted according to the feature contribution in subsequent steps. The initial weight matrix The calculation formula is: , In the formula, is a function for extracting the diagonal elements of a matrix, is the feature weight vector, Each element in Initialized to the same value, indicating that all features have the same importance in the initial stage; S423, the feature adjustment module recursively optimizes the initially generated low-dimensional features. In each round of iteration, the module adjusts the weights of each feature according to the performance of the features in the previous round, gradually enhances those features with important influences, and gradually weakens redundant or noise features. Assume that in the tth round of iteration, the weight update rule is as follows: , In the formula, is the feature weight of the t+1th iteration, is the feature weight of the t-th iteration, is the learning rate of the autoencoder neural network, is the gradient of the loss function with respect to the feature weight, is the loss function of the autoencoder neural network, is the feature representation after feature weight adjustment, is the label data; S424, in order to ensure the effectiveness of the dimensionality reduction process, the decoder remaps the low-dimensional features back to the high-dimensional space to ensure that the dimensionality reduction process does not lose important information; the reconstruction process of the decoder is expressed as: , In the formula, is the reconstructed high-dimensional data, is the multi-layer Sigmoid activation function of the decoder, is the weight matrix of the decoder, is the feature representation after feature weight adjustment, is the bias vector of the decoder; S425, repeat the iterations from S422 to S424 until a preset stop iteration condition is met.

[0010] As a further improvement of the present invention: the classifier of the integrated gradient boosting tree algorithm based on the dynamic decision boundary training in S5 includes: S501, initializing the integrated gradient boosting tree classifier and setting the initial prediction value; S502: Calculate the residual between the initial prediction value and the actual category label for all training samples. The residual reflects the current prediction error of the model and serves as the target variable for subsequent training of weak classifiers. Specifically, the calculation formula for the residual of the sample is: , In the formula, is the residual of the i-th sample in round 0, is the symbol of partial derivative, is the objective function of the integrated gradient boosting tree, is the predicted value of the sample, is the predicted value of the initial model for the i-th sample, is the true label of the i-th training sample; S503, iteratively training weak classifiers; S504: After each round of iteration, evaluate whether the model meets the stopping condition. The stopping condition may be that a preset maximum number of iterations is reached, or the error of the model drops to an expected threshold. When the stopping condition is met, the iteration process is terminated.

[0011] The present invention also provides a full-electric drive aircraft tractor control system, which uses the above-mentioned full-electric drive aircraft tractor control method, including: A data storage module, used for storing data; A data acquisition module, used to collect real driving data and send it to the data storage module; A data expansion module, used for generating expanded driving data based on real driving data, and sending the expanded driving data to the data storage module; A data processing module is used to perform feature extraction on the real driving data and the expanded driving data or on the real driving data alone to obtain feature-extracted data and perform feature dimension reduction on the feature-extracted data to obtain feature-reduced data; The tractor control module is used to analyze the data after feature dimension reduction to obtain specific operation instructions and send the operation instructions to the execution module; The execution module is used to receive and execute operation instructions.

[0012] The beneficial effects of the present invention are as follows: the all-electric drive aircraft tractor control system and method provided by the present invention have strong data processing capabilities and high decision-making accuracy.

[0013] Before training the feature extraction model of the neural network algorithm based on quantum dynamic evolution and the feature dimension reduction model of the autoencoder neural network based on feature refinement, the present invention first generates the extended driving data by using the generative adversarial network algorithm based on adversarial perturbation in the data expansion module with the annotated real driving data, expands the data source, and effectively prevents the generator from overfitting certain specific driving data points by introducing adversarial perturbation defense strategies, thereby improving the diversity and robustness of driving data expansion; the topological structure constraint of the generator encodes the topological relationship of the driving data stream, thereby retaining the spatial structure characteristics of the extended driving data, ensuring that the extended driving data has higher similarity and rationality. Therefore, the present invention can use a large number of training samples to train the feature extraction model and the feature dimension reduction model. In addition, when extracting features, the feature extraction model uses the concept of quantum computing to enhance the driving data processing capability of the traditional neural network, especially when processing driving data with high dimensions and complex relationships, the dynamic evolution of the quantum state is used to simulate and optimize the feature extraction process, so that each neural network node not only processes classical information, but also can use the superposition and entanglement state of quantum bits to enhance its information processing capability. Each quantum neuron can not only process classical information, but also use the characteristics of quantum computing to optimize the feature extraction process, making the extracted features more effective. During feature dimensionality reduction, the feature dimensionality reduction model dynamically adjusts the feature space after dimensionality reduction through recursive feature adaptive optimization, so that important features can be strengthened and secondary features are gradually weakened, so that the feature representation after dimensionality reduction has better simplicity while retaining data information. The decoder ensures the reversibility of the dimensionality reduction process, prevents the loss of important information, and enhances the accuracy and effectiveness of the dimensionality reduction process. Therefore, the data processing capability of the all-electric drive aircraft tractor control system and method provided by the present invention is relatively strong.

[0014] The tractor control module of the present invention includes a classifier of an integrated gradient boosting tree algorithm based on a dynamic decision boundary, and uses an integrated gradient boosting tree as a classification algorithm. Through a dynamic decision boundary adjustment mechanism, the split point and leaf node output are adaptively adjusted during the decision tree training process to adapt to the real-time changes in data distribution, and the classification decision boundary is adaptively adjusted to improve the generalization ability of the model and the ability to distinguish complex samples. The adjustment mechanism combines information gain with sample distribution differences to improve the generalization ability and classification accuracy of the classifier under complex data. Therefore, the decision accuracy of the all-electric drive aircraft tractor control system and method provided by the present invention is relatively high. DETAILED DESCRIPTION

[0015] The specific implementation modes of the present invention are further described in detail below.

[0016] A control method for a fully electric driven aircraft tractor provided by the present invention comprises: 1. Training on the control system of all-electric aircraft tractor (S1~S5) S1. Collect the real driving data collected in the past by the data acquisition module. The driving data includes environmental monitoring data and sensor data. In actual application, it may also include operation data, all-electric drive vehicle performance data, etc.

[0017] Environmental data is collected by the weather station, including temperature, humidity, wind speed, etc. Sensor data is measured by multiple sensors, including attitude sensors, inertial measurement units, radars, etc. equipped on the tractor, which can collect information such as the angle, speed, acceleration of the tractor and the distance to surrounding obstacles in real time.

[0018] In actual applications, the attributes of driving data are usually more than 10 attributes, and the number of attributes of driving data may reach dozens or even hundreds. Example 1 is given to illustrate the format and types of driving data: In Example 1, the data in the data storage module is saved in a structured format, specifically JSON or XML format, and the attributes of driving data include: Angle Ra, indicating the current direction angle of the tractor; Speed ​​Va, records the instantaneous speed of the tractor; Acceleration Aa, which indicates how fast the tractor accelerates or decelerates; Obstacle distance Da, indicating the distance to the nearest obstacle; Ambient temperature Ea, affects the operating efficiency of the tractor system; Ambient humidity Fa, affecting the accuracy of the sensor; Wind speed Ga, which affects the stability and energy consumption of the tractor; The attitude angle Ha provides the inclination of the tractor relative to the ground; Tire pressure Ia, ensuring driving safety; Battery charge Ja, monitors the energy usage of the tractor.

[0019] S2, marking the operation instructions to be performed corresponding to each real driving data, and after the marking is completed, sending the marked real driving data to the data storage module for storage; The labeling method is manual labeling, and the operation instructions include driving instructions, steering instructions, braking instructions, and emergency stop instructions. The situation of manual labeling of the real driving data in Example 1 is shown in Table 1: Table 1

[0020] S3, sending the annotated real driving data to the data expansion module, the data expansion module uses a generative adversarial network algorithm based on adversarial perturbations to generate expanded driving data, and sends the expanded driving data to the storage module for storage; The generative adversarial network consists of a generator and a discriminator. The generator generates new driving data samples by learning the distribution of real driving data. The task of the discriminator is to judge the difference between the generated samples and the real driving data. In this system, the input of the generator is random noise, and the output is the expanded driving data with a distribution similar to that of the real driving data; the discriminator performs binary classification on the samples output by the generator and the real driving data. The adversarial perturbation defense strategy is a strategy to improve the robustness of the generator, so that the generator can not only generate diversified data, but also resist possible malicious perturbations. This strategy ensures that the generator can still maintain output quality and consistency when facing various data and condition changes. On the basis of the traditional generative adversarial network, the present invention uses the adversarial perturbation defense strategy to add perturbation terms in the generator training process, and applies random perturbations to the expanded driving data generated by the generator to prevent the generator from relying on a single mode or a certain type of feature. At the same time, combined with the generator constraints of the topological structure, the expressiveness of the generator and the diversity of the expanded driving data are enhanced. In S3, the steps of generating expanded driving data by the data expansion module using the generative adversarial network algorithm based on adversarial perturbation include: S301, initialize the network structures of the generator and the discriminator respectively; the initialization method is expressed as: , , In the formula, represents the weight matrix of the generator, Indicates a parameter update operation. The function represents the generation of random weights for network initialization. The function represents the number of neurons in the network layer that adapts the generator or discriminator as the dimension of the initial parameter matrix. represents a generator, represents the weight matrix of the discriminator, represents the discriminator; S302: Using the topological structure constraint strategy in the generator, by encoding the geometric spatial relationship of each driving data point, establish its position association in the real driving data. The topological structure constraint supervises the generator to retain the geometric characteristics of the real driving data during the driving data expansion process, thereby improving the physical rationality of the expanded driving data. The calculation formula is: , In the formula, is the topological structure constraint weight factor, Set to 0.3; For real driving data dimensions, and Generators The output feature vectors of the two extended driving data are: is the feature vector of the i-th extended driving data, is the feature vector of the j-th extended driving data, for norm, For the generator The standard deviation of the output augmented driving data; is a small random number in the topological structure constraint, Set to 0.001; and is the feature vector of the corresponding real driving data, is the feature vector of the i-th real driving data, is the feature vector of the jth real driving data; is the standard deviation of real driving data; S303. According to the adversarial perturbation defense strategy, by adding perturbation terms during the generator training process, random perturbations are applied to the driving data generated by the generator to prevent the generator from relying on a single mode or a certain type of feature, thereby improving the diversity of the expanded driving data, so that the generator can still generate valid samples when responding to perturbations. Nonlinear transformations are performed based on the hyperbolic tangent function, and the gradient values ​​are scaled and smoothed to ensure the effectiveness and stability of the perturbations. The calculation formula is: , In the formula, is the disturbance intensity represents the gradient of the input vector, Set to 0.015; represents the hyperbolic tangent function; represents the parameter that controls the gradient scaling, Set to 0.5; is the disturbance intensity, represents the loss function of the discriminator, represents the discriminator function, The feature vector representing the real driving data, represents the true label, represents the quantum bit parameters; The quantum bit parameters are based on quantum heuristic optimization strategies, which improve the traditional generator influencing factors by imitating quantum behavior. The calculation formula is: , In the formula, is the complex probability amplitude of the corresponding state of the first ground state of the qubit, is the first ground state of the quantum bit; is the complex probability amplitude of the corresponding state of the second ground state of the quantum bit, and satisfies ; It is the second ground state of the quantum bit.

[0021] In each iteration, the superposition characteristics of the quantum state are used to search for parameters, the rotation of the quantum state is simulated to explore the new parameter space, and the quantum bit parameters are updated, which can be expressed as: , , In the formula, is the updated quantum bit parameter; It is a quantum gate operation, used to update the quantum bit state; represents the quantum bit parameters, is the adjustment of the quantum bit parameters based on the gradient of the loss function, represents the imaginary unit, is the Pauli-Y matrix; S304: The generator generates a batch of augmented driving data similar to the real driving data through the initial noise input. The discriminator receives the real driving data and the augmented driving data and determines whether they are real. The goal of the generator is to confuse the discriminator so that it cannot distinguish between the augmented driving data and the real driving data. The augmented driving data output by the generator after adding disturbances is similar to the real driving data. It is expressed as: , In the formula, is the generator function, is the adversarial perturbation, is the input noise of the generator; The discriminator's judgment method is expressed as: , In the formula, is the classification result of the discriminator on the expanded driving data, represents the discriminator function, The augmented driving data output by the generator after adding disturbances.

[0022] S305: The discriminator performs binary classification training on the real driving data and the expanded driving data through the cross entropy loss function, and updates the weight parameters of the discriminator according to the correctness of the classification. The update method is expressed as: , In the formula, is the learning rate of the discriminator, is the loss function relative to the discriminator weights The Hessian matrix of express The inverse matrix of Hessian enables the optimization of the discriminator weights to more accurately reflect the local curvature of the loss surface; represents the gradient of the discriminator weight, represents the loss function of the discriminator, is the classification result of the discriminator on the expanded driving data, represents the true label; S306. Optimize the generator by combining the quantization loss function. The goal of the generator is to minimize the distribution difference between the augmented driving data and the real driving data. By calculating the quantization loss function, the generator continuously optimizes its output in each iteration to make the augmented driving data more realistic and diverse. The regularization term is used to prevent overfitting and improve the generalization ability of the model. The weight of the generator The update method is expressed as: , In the formula, is the learning rate of the generator, Set to 0.01; represents the gradient of the generator weight, is the loss function of the generator, represents the discriminator function, is the generator function, is the input noise of the generator, is a false sample label; is a regularization parameter that controls the strength of the topological constraint, Set to 0.3; is the topological structure constraint; is a parameter that controls the effect of regularization, Set to 0.2; is the regularization term of the generator parameters. The regularization term of the generator parameters adopts Regularization; S307, repeating iterations S302 to S306 until a preset stop iteration condition is met, which means that the model training is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0023] S4. The annotated real driving data and the expanded driving data together constitute training samples. If the real driving data is 800 and the expanded driving data is 200, the training samples are 1000. The training samples are input into the data processing module, which includes a feature extraction model based on a quantum dynamic evolution neural network algorithm and a feature dimensionality reduction model based on a feature refinement autoencoder neural network; In order to effectively capture and extract key information from driving data, the feature extraction model of the neural network algorithm based on quantum dynamic evolution is first trained with training samples to obtain training samples after feature extraction. Quantum dynamic evolution refers to integrating quantum gate operations in each neuron of the neural network, allowing neurons to use the superposition and entanglement states of quantum bits for data processing, dynamically adjusting the parameters of the quantum gate to adapt to the characteristics of the input data, and using a composite loss function that includes traditional errors and quantum state fidelity for network training. The neural network parameter update combines gradient descent and swarm intelligence strategies, as well as a dynamically adjusted annealing algorithm to achieve the global optimal parameter configuration, thereby greatly improving the neural network's ability to process complex tractor data. The steps for training the feature extraction model of the neural network algorithm based on quantum dynamic evolution in S4 include: S411. Initialize the structure of the neural network based on quantum dynamic evolution. Assume that each layer of the neural network contains a number of quantum neurons, each quantum neuron has the ability to perform quantum gate operations, each quantum neuron parameter is represented by a quantum state, and the weight parameter of the first layer of the neural network is Defined as: , In the formula, is the number of quantum neurons in the neural network; are the complex probability amplitudes of the kth quantum neuron in the neural network, indicating the probability amplitudes of the quantum being in the ground state and the excited state, ensuring ; is the first ground state of the quantum bit, is the second ground state of the quantum bit; S412. Define a quantum gate operation for each neuron. The quantum gate operation is represented by a matrix of the following form: , In the formula, is the quantum gate operation of the neural network layer l, is an imaginary unit; It is the parameter that controls the rotation of the quantum state, specifically the angle parameter of the gate operation, which affects the rotation of the quantum state; is the Pauli-Y matrix of the lth layer of the neural network, which is used to implement the rotation of the quantum bit; The quantum gate operation uses a combination of Pauli matrix and Hadamard gate, and the calculation formula is: , In the formula, is an imaginary unit, is the parameter that controls the rotation of the quantum state; is the Pauli-Y matrix of the lth layer of the neural network, which is used to implement the rotation of the quantum bit; is the identity matrix; is the Pauli-Y matrix, which is in the form of ; In order to make quantum gate operations more suitable for complex driving data structures, conditional rotation is used to make the rotation angle It becomes a function of the input driving data, and the calculation formula of the parameters controlling the rotation of the quantum state is: , In the formula, is the scaling parameter of the parameter that controls the rotation of the quantum state, which is used to control the angle size. Set to 0.2; are the weights of the parameters that control the rotation of the quantum state, specifically the weights associated with the inputs of the neural network; is the jth input of the neural network, is the bias term of the parameters that control the rotation of the quantum state; S413. When the input driving data is forward propagated through the network, the quantum neurons of each layer process the driving data according to their quantum gate operations. The process of forward propagation of the driving data through the quantum neural network is represented as follows: , In the formula, is the state of the l+1th layer after quantum gate and weight operations, is the quantum gate operation of the neural network layer l, is the weight parameter of the lth layer of the neural network, is the quantum neuron state of layer l; S414. Define a quantum-inspired loss function to evaluate the effect of feature extraction, which not only considers the traditional error index, but also includes the fidelity index of the quantum state. Use this loss function for backpropagation to update the parameters of each quantum gate in the network and the parameters of the traditional neural network. The loss function of the neural network The calculation formula is: , In the formula, is the number of samples input to the neural network in the current batch, is the target value of the nth sample, is the expected value of the quantum state output by the network; is the regularization parameter of the neural network, Set to 0.3; is the quantum fidelity function, is the quantum neuron state of layer L, is an ideal quantum state; The gradient descent method is used to update the parameters that control the rotation of the quantum state. In each iterative optimization process, the swarm intelligence strategy is used to adjust the parameters of the quantum gate, imitating the exploration behavior of quantum particles to find the global optimal network parameter configuration. The update method is expressed as: , In the formula, is the parameter that controls the rotation of the quantum state, Indicates parameter update operation; is the learning rate of the neural network, Set to 0.01; is the annealing temperature coefficient of the tth iteration, which is adjusted and calculated according to the simulated annealing algorithm; is the symbol of partial derivative, is the loss function of the neural network; is a coefficient that controls the strength of the local search, Set to 0.3; is the optimal solution found in the history of the current particle; is a coefficient that controls the global search intensity, Set to 0.5; is the optimal solution among all particles, where some parameters are explained in the same way as a class of parameters in the simulated annealing algorithm; The annealing temperature coefficient is dynamically adjusted with the number of iterations to control the step size of the gradient update, thereby achieving more refined global optimization. The calculation formula is: , In the formula, is the initial temperature coefficient, Set to 0.2; Represents the current iteration step, where some parameters are explained in the same way as a class of parameters in the simulated annealing algorithm.

[0024] The quantum fidelity function measures the fidelity between the output quantum state and the ideal quantum state, making it possible to accurately measure the similarity between quantum states. The calculation formula is: , In the formula, is the quantum fidelity function, is the quantum neuron state of layer L, is an ideal quantum state, represents the inner product of two quantum states, and the square of its modulus gives the fidelity of the two states; Represents the inner product of the quantum state output by the network and the ideal quantum state; S415, repeating iterations S412 to S414 until a preset stop iteration condition is met, which means that the model training is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0025] Then, the feature-extracted training samples are used to train the feature dimension reduction model based on the feature-refined autoencoder neural network to obtain the training samples after feature dimension reduction; The autoencoder neural network based on feature refinement is used as a feature dimensionality reduction model to recursively optimize the initially generated low-dimensional features through the feature refinement strategy. In each round of iteration, the module will adjust the weights of each feature according to the performance of the previous round of features, gradually enhance those features with important influences, and gradually weaken redundant or noise features to achieve adaptive refinement of features. The autoencoder neural network based on feature refinement consists of three parts: encoder, decoder, and feature adjustment module. The encoder is responsible for mapping high-dimensional input data to low-dimensional feature space, and the decoder is used to reconstruct the reduced features back to the original space to ensure the reversibility of the dimensionality reduction process. The feature dimensionality reduction model for training the autoencoder neural network based on feature refinement in S4 includes: S421, assuming that the data input to the autoencoder neural network is , the encoder uses a multi-layer nonlinear mapping structure to map high-dimensional data to the initial low-dimensional feature space, which can be expressed as: , In the formula, represents the initial low-dimensional features, is the multi-layer Sigmoid activation function of the encoder, is the weight matrix of the encoder, is the data input to the autoencoder neural network, is the bias vector of the encoder; S422, after the low-dimensional features are generated, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space. When the module is initialized, all features are given the same initial weight so that they can be gradually adjusted according to the feature contribution in subsequent steps. The initial weight matrix The calculation formula is: , In the formula, is a function for extracting the diagonal elements of a matrix, is the feature weight vector, Each element in Initialized to the same value, indicating that all features have the same importance in the initial stage; Feature representation after feature weight adjustment for: , In the formula, is the initial weight matrix.

[0026] S423, the feature adjustment module recursively optimizes the initially generated low-dimensional features. In each round of iteration, the module adjusts the weights of each feature according to the performance of the features in the previous round, gradually enhances those features with important influences, and gradually weakens redundant or noise features. Assume that in the tth round of iteration, the weight update rule is as follows: , In the formula, represents the feature weight of the t+1th iteration, represents the feature weight of the tth iteration; is the learning rate of the autoencoder neural network, Set to 0.01; represents the gradient of the loss function with respect to the feature weight, is the loss function of the autoencoder neural network, is the feature representation after feature weight adjustment, is the label data; the loss function of the autoencoder neural network is the reconstruction error loss function; S424, in order to ensure the effectiveness of the dimensionality reduction process, the decoder remaps the low-dimensional features back to the high-dimensional space to ensure that the dimensionality reduction process does not lose important information. The reconstruction process of the decoder is expressed as: , In the formula, is the reconstructed high-dimensional data, is the multi-layer Sigmoid activation function of the decoder, is the weight matrix of the decoder, is the feature representation after feature weight adjustment, is the bias vector of the decoder; S425, repeating iterations S422 to S424 until a preset stop iteration condition is met, which means that the model training is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0027] S5, inputting the training samples after feature dimension reduction into the tractor control module, the tractor control module includes a classifier of the integrated gradient boosting tree algorithm based on the dynamic decision boundary, and using the training samples after feature dimension reduction to train the classifier of the integrated gradient boosting tree algorithm based on the dynamic decision boundary; Dynamic decision boundary refers to the process of training weak classifiers, that is, when building decision trees, dynamically adjusting the splitting points and leaf node outputs of decision trees according to the distribution of samples in feature space and error gradient, so that the classification decision boundary can better adapt to the complexity of data. The integrated gradient boosting tree algorithm based on dynamic decision boundary is used as the classification algorithm. By constructing multiple weak classifiers (i.e., the decision trees in the corresponding integrated gradient boosting tree), the overall performance of the model is gradually improved, and the data after feature dimensionality reduction is accurately classified. The classifiers trained in S5 based on the integrated gradient boosting tree algorithm based on dynamic decision boundary include: S501, initialize the integrated gradient boosting tree classifier and set the initial prediction value; set the initial prediction value to the mean of the target values ​​of all training samples, the prediction value of the initial model It is expressed as: , In the formula, The input sample of the integrated gradient boosting tree classifier is the driving data after feature extraction; The number of training samples input to the ensemble gradient boosting tree for the current batch; is the true label of the i-th training sample, specifically the control instruction category of the all-electric aircraft tractor.

[0028] S502: Calculate the residual between the initial prediction value and the actual category label for all training samples. The residual reflects the current prediction error of the model and serves as the target variable for subsequent training of weak classifiers. Specifically, the calculation formula for the residual of the sample is: , In the formula, is the residual of the i-th sample in round 0, is the symbol of partial derivative, is the objective function of the integrated gradient boosting tree, and the objective function of the integrated gradient boosting tree adopts the log-likelihood loss function; is the predicted value of the sample, is the predicted value of the initial model for the i-th sample, is the true label of the i-th training sample; S503, iteratively train weak classifiers, for m=1 to ( is the maximum number of iterations), perform the following steps: S5031. Training weak classifiers: Using the current residual as the new target variable, a decision tree weak classifier is trained. The weak classifier is designed to fit the residual of the current model, thereby reducing the overall prediction error, which is expressed as: , In the formula, For input The mth weak classifier at time is the mth weak classifier, The number of training samples input to the ensemble gradient boosting tree for the current batch; is the weight of the i-th sample in the m-1th round, initially ; is the residual of the current m-1th round, which is used as the target value; is the mth weak classifier when the i-th sample is input; To integrate the regularization parameters of the gradient boosting tree and control the impact of model complexity on the loss function; is the complexity penalty term of the mth weak classifier; The complexity penalty term of the weak classifier controls the complexity of the tree to prevent overfitting and improve the generalization ability of the model. The complexity penalty term of the mth weak classifier The calculation formula is: , In the formula, is the number of leaf nodes of the mth tree, is the penalty coefficient for each leaf node, is the leaf node weight Regularization coefficient, is the output value of the jth leaf node of the mth tree.

[0029] S5032, Dynamic Decision Boundary Adjustment: In the training process of the weak classifier, that is, when building the decision tree, the split point and leaf node output of the decision tree are dynamically adjusted according to the distribution of samples in the feature space and the error gradient through the dynamic decision boundary adjustment mechanism, so that the classification decision boundary can better adapt to the complexity of the data. Specifically, for each split node, the split point is adjusted to minimize the following objective function: , In the formula, is the objective function of the integrated gradient boosting tree, To split the node, and are the sample sets of the left and right child nodes respectively; is the weight of the i-th sample in the m-1th round, initially ; is the residual of the current m-1th round; and are the penalty coefficients for the left and right child nodes respectively; To integrate the tuning parameters of the gradient boosting tree, control the impact of information gain and sample distribution on splitting; is the information gain at the split point, is the difference in sample distribution density at the split point; Information gain at the split point The calculation formula is: , In the formula, is the entropy of the parent node, is the sample set of the parent node, and are the sample sets of the left and right child nodes respectively; is the entropy of the left and right child nodes.

[0030] Entropy of the parent node The calculation formula is: , In the formula, is the number of sample categories, For the parent node The probability of the k-th class sample in ; Difference in sample distribution density at the split point The calculation formula is: , In the formula, and are the sample sets of the left and right child nodes respectively; is the weight of the i-th sample in the m-1th round. ; By comprehensively considering the information gain and the difference in sample distribution density, the split point is dynamically adjusted so that each split can not only effectively distinguish the residuals, but also consider the weight distribution of the samples, thereby improving the adaptability of the model to complex data.

[0031] The calculation formula for the penalty coefficients of the left and right child nodes is: , , In the formula, and are the sample sets of the left and right child nodes respectively; is the weight of the i-th sample in the m-1th round. ; is the residual of the current m-1th round; dynamic decision boundary adjustment is achieved by adjusting the split point in real time , so that each split can better distinguish the residuals and optimize the model performance.

[0032] S5033, Update the integrated model: Add the newly trained weak classifier to the integrated model, update the prediction value of the model, and improve the overall performance of the model by accumulating the output of each weak classifier. The method of updating the prediction value of the integrated model is expressed as: , In the formula, is the prediction value of the integrated model in round m; is the prediction value of the integrated model in round m-1; is the learning rate of the ensemble gradient boosting tree, and ; For input The mth weak classifier at time ; S5034, automatic optimization of dynamic integration: According to the performance index of the current model and the error distribution of the sample, the weight of each weak classifier is dynamically adjusted. The present invention adopts the sample difficulty coefficient to give higher weights to samples with larger errors, so that subsequent weak classifiers pay more attention to these difficult-to-predict samples. The calculation formula of the sample difficulty coefficient is: , In the formula, is the difficulty coefficient of the i-th sample in the m-th round; It is the difficulty adjustment parameter, which controls the sensitivity of the difficulty coefficient to the residual; is the residual of the current m-1th round; The difficulty adjustment parameter is obtained through adaptive calculation based on sample gradient and high-order statistics. The calculation formula is: , In the formula, is the difficulty scaling factor, The number of training samples input to the integrated gradient boosting tree for the current batch, is the gradient of the i-th sample in the m-1th round, is the mean of the gradient, A tiny constant to prevent division by zero; The gradient of the sample in the m-1th round The calculation formula is: , In the formula, is the objective function of the integrated gradient boosting tree, is the predicted value of the sample, is the prediction value of the integrated model in the m-1th round when the i-th sample is input; Mean of the gradient The calculation formula is: , In the formula, The number of training samples input to the integrated gradient boosting tree for the current batch, is the gradient of the i-th sample in the m-1th round; the standard deviation and mean of the gradient are used to make the sample difficulty coefficient more sensitive to reflect the learning difficulty of the sample under the current model, and then adaptively adjust the sample weight.

[0033] Update the sample weights. The update of sample weights makes the model pay more attention to samples with large errors and improves the classifier's learning ability for difficult samples. The update method is expressed as: , In the formula, is the weight of the i-th sample in the m-th round; is the weight of the i-th sample in the m-1th round, is the difficulty coefficient of the i-th sample in the m-th round.

[0034] Adjust the weak classifier weight according to the updated sample weight, and calculate the error of the weak classifier, which is expressed as: , In the formula, is the weighted error rate of the mth weak classifier, The number of training samples input to the integrated gradient boosting tree for the current batch, is the weight of the i-th sample in the m-th round; It is an indicator function, which takes 1 if the condition is met, and 0 otherwise; is the true label of the i-th training sample, is the mth weak classifier when the i-th sample is input; The updated residual is used as the target value of the next iteration for the next round of weak classifier training, gradually reducing the model error. The update method is expressed as: , In the formula, is the residual of the mth round, is the true label of the i-th training sample, is the prediction value of the integrated model in the mth round when the i-th sample is input, is the weighted error rate of the mth weak classifier; S504: After each round of iteration, evaluate whether the model meets the stopping condition. The stopping condition may be that the preset maximum number of iterations is reached, or the model error drops to the expected threshold. When the stopping condition is met, the iteration process is terminated. The preset maximum number of iterations is , the residual sum of squares threshold is , if the number of iterations is greater than , or the residual sum of squares Less than Then stop the iteration and output the integrated gradient boosting tree classifier model after the final training is completed.

[0035] 2. Processing new real driving data (S6~S9) S6. Collect new real driving data through the data collection module, and send the new real driving data to the data storage module for storage; S7, inputting the new real driving data into the data processing module, first performing feature extraction on the new real driving data using the feature extraction model to obtain new real driving data after feature extraction, and performing feature dimension reduction on the new real driving data after feature extraction using the feature dimension reduction model to obtain new real driving data after feature dimension reduction; S8, inputting the new real driving data after feature dimension reduction into the tractor control module, using a classifier to analyze the new real driving data after feature dimension reduction to obtain specific operation instructions and sending the operation instructions to the execution module; S9. The execution module receives and executes the operation instruction.

[0036] The present invention also provides a control system for an all-electrically driven aircraft tractor using the above-mentioned all-electrically driven aircraft tractor control method, comprising: A data storage module, used for storing data and transmitting the data to the central processing unit through a wireless network through a wireless communication module for backup in a server; A wireless communication module is used to transmit data to a central processing unit through a wireless network and back it up in a server; A data acquisition module, used to collect real driving data and send it to the data storage module; A data expansion module, used for generating expanded driving data based on real driving data, and sending the expanded driving data to the data storage module; A data processing module is used to perform feature extraction on the real driving data and the expanded driving data or on the real driving data alone to obtain feature-extracted data and perform feature dimension reduction on the feature-extracted data to obtain feature-reduced data; The tractor control module is used to analyze the data after feature dimension reduction to obtain specific operation instructions and send the operation instructions to the execution module; The execution module is used to receive and execute operation instructions.

[0037] When applied, the system also includes multiple control methods. The system supports operation through wireless remote control and body control panel. The wireless remote control can be used for remote operation, and the body panel is used as an auxiliary method for on-site operation. The command issued by the wireless remote control or the body panel is transmitted to the PLC controller through the communication bus, and then the corresponding traction action is executed by the PLC controller. The system efficiently transmits all instructions to each execution module of the tractor through the communication bus to ensure the accuracy and timeliness of the instruction execution. The central IMU attitude sensor can continuously monitor the attitude changes of the tractor and feed back this data to the control system. By comparing the difference between the actual and expected values, the system can adjust the driving strategy in real time to ensure that the tractor remains stable in complex terrain or unstable conditions. The radar monitoring system ensures the safety of the tractor in operation. When the tractor approaches an obstacle during operation, the system can automatically issue a braking or emergency stop command to avoid a collision accident. The emergency braking function of the system combines multi-sensor data to detect obstacles within milliseconds and respond to ensure the safe operation of the tractor.

Claims

1. A method for controlling a fully electric driven aircraft tractor, characterized in that: include: S1. Collecting real driving data collected in the past by a data collection module, wherein the driving data includes environmental monitoring data and sensor data; S2, marking the operation instructions to be performed corresponding to each of the real driving data, and after the marking is completed, sending the marked real driving data to the data storage module for storage; The marking method is manual marking, and the operation instructions include driving instructions, steering instructions, braking instructions and emergency stop instructions; S3, sending the annotated real driving data to a data expansion module, the data expansion module uses a generative adversarial network algorithm based on adversarial perturbations to generate expanded driving data, and sending the expanded driving data to a storage module for storage; S4, the annotated real driving data and the expanded driving data together constitute a training sample, and the training sample is input into a data processing module, wherein the data processing module includes a feature extraction model based on a neural network algorithm of quantum dynamic evolution and a feature dimensionality reduction model based on an autoencoder neural network of feature refinement; First, the feature extraction model of the neural network algorithm based on quantum dynamic evolution is trained with the training samples to obtain the training samples after feature extraction; then, the feature dimension reduction model of the autoencoder neural network based on feature refinement is trained with the training samples after feature extraction to obtain the training samples after feature dimension reduction; S5, inputting the training samples after the feature dimension reduction into a tractor control module, the tractor control module includes a classifier of an integrated gradient boosting tree algorithm based on a dynamic decision boundary, and using the training samples after the feature dimension reduction to train the classifier of the integrated gradient boosting tree algorithm based on a dynamic decision boundary; S6, collecting new real driving data through the data collection module, and sending the new real driving data to the data storage module for storage; S7, inputting the new real driving data into the data processing module, first performing feature extraction on the new real driving data using the feature extraction model to obtain new real driving data after feature extraction, and performing feature dimension reduction on the new real driving data after feature extraction using the feature dimension reduction model to obtain new real driving data after feature dimension reduction; S8, inputting the new real driving data after the feature dimension reduction into the tractor control module, using the classifier to analyze the new real driving data after the feature dimension reduction to obtain specific operation instructions and sending the operation instructions to the execution module; S9. The execution module receives and executes the operation instruction.

2. A method for controlling a fully electric aircraft tractor according to claim 1, characterized in that: The step of the data expansion module in S3 using a generative adversarial network algorithm based on adversarial perturbations to generate expanded driving data includes: S301, initialize the network structures of the generator and the discriminator respectively; S302: Using a topological structure constraint strategy in the generator, by encoding the geometric spatial relationship of each driving data point, establish its position association in the real driving data. The calculation formula is: ， In the formula, is the topological structure constraint weight factor, Data dimension for real driving and Generators The output feature vectors of the two extended driving data are: is the feature vector of the i-th extended driving data, is the feature vector of the j-th extended driving data, for norm, For the generator The standard deviation of the output augmented driving data, is a small random number in the topological structure constraint, and is the feature vector of the corresponding real driving data, is the feature vector of the i-th real driving data, is the feature vector of the jth real driving data; is the standard deviation of real driving data; S303: According to the adversarial perturbation defense strategy, a perturbation term is added during the generator training process to apply random perturbations to the driving data generated by the generator. The calculation formula is: ， In the formula, is the disturbance intensity represents the gradient of the input vector, represents the hyperbolic tangent function, represents the parameter that controls the gradient scaling, is the disturbance intensity, represents the loss function of the discriminator, represents the discriminator function, The feature vector representing the real driving data, represents the true label, represents the quantum bit parameters; S304: The generator generates a batch of extended driving data similar to the real driving data through the initial noise input. The extended driving data output by the generator after adding disturbance It is expressed as: ， In the formula, is the generator function, is the input noise of the generator, is an adversarial perturbation; S305: The discriminator performs binary classification training on the real driving data and the expanded driving data through the cross entropy loss function, and updates the weight parameters of the discriminator according to the correctness of the classification. The update method is expressed as: ， In the formula, is the learning rate of the discriminator, is the loss function relative to the discriminator weights The Hessian matrix of express The inverse matrix of represents the gradient of the discriminator weight, represents the loss function of the discriminator, is the classification result of the discriminator on the expanded driving data, represents the true label; S306. Optimize the generator by combining the quantization loss function. The update method is expressed as: ， In the formula, is the learning rate of the generator, represents the gradient of the generator weight, is the loss function of the generator, represents the discriminator function, is the generator function, is the input noise of the generator, is the false sample label, is the regularization parameter that controls the strength of the topological constraint, is the topological constraint, To control the parameters of the regularization effect, is the regularization term of the generator parameters; S307, repeating iterations S302 to S306 until a preset stop iteration condition is met.

3. The method for controlling a fully electric aircraft tractor according to claim 1, characterized in that: The step of training the feature extraction model of the neural network algorithm based on quantum dynamic evolution in S4 includes: S411. Initialize the structure of the neural network based on quantum dynamic evolution. Assume that each layer of the neural network contains a number of quantum neurons, each quantum neuron has the ability to perform quantum gate operations, each quantum neuron parameter is represented by a quantum state, and the weight parameter of the first layer of the neural network is Defined as: ， In the formula, is the number of quantum neurons in the neural network; are the complex probability amplitudes of the kth quantum neuron in the neural network, indicating the probability amplitudes of the quantum being in the ground state and the excited state, ensuring ; is the first ground state of the quantum bit, is the second ground state of the quantum bit; S412, defining a quantum gate operation for each neuron, wherein the quantum gate operation is represented by a matrix of the following form: ， In the formula, is the quantum gate operation of the neural network layer l, is an imaginary unit, is the parameter that controls the rotation of the quantum state, is the Pauli-Y matrix; S413. When the input driving data is forward propagated through the network, the quantum neurons of each layer process the driving data according to their quantum gate operations. The process of forward propagation of the driving data through the quantum neural network is represented as follows: ， In the formula, is the state of the l+1th layer after quantum gate and weight operation, is the quantum gate operation of the neural network layer l, is the weight parameter of the lth layer of the neural network, is the quantum neuron state of the lth layer; S414. Define a quantum-inspired loss function to evaluate the effect of feature extraction, which not only considers the traditional error index, but also includes the fidelity index of the quantum state. Use this loss function for backpropagation to update the parameters of each quantum gate in the network and the parameters of the traditional neural network. The loss function of the neural network The calculation formula is: ， In the formula, is the number of samples input to the neural network in the current batch, is the target value of the nth sample, is the expected value of the quantum state output by the network, is the regularization parameter of the neural network, is the quantum fidelity function, is the quantum neuron state of the Lth layer, is an ideal quantum state; S415, repeating the iterations of S412 to S414 until a preset condition for stopping the iteration is met.

4. The method for controlling a fully electric aircraft tractor according to claim 1, characterized in that: The feature dimension reduction model of the autoencoder neural network based on feature refinement is trained in S4, including: S421, assuming that the data input to the autoencoder neural network is , the encoder uses a multi-layer nonlinear mapping structure to map high-dimensional data to the initial low-dimensional feature space, which can be expressed as: ， In the formula, represents the initial low-dimensional features, is the multi-layer Sigmoid activation function of the encoder, is the weight matrix of the encoder, is the data input to the autoencoder neural network, is the bias vector of the encoder; S422, after the low-dimensional features are generated, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space. When the module is initialized, all features are given the same initial weight so that they can be gradually adjusted according to the feature contribution in subsequent steps. The initial weight matrix The calculation formula is: ， In the formula, is a function for extracting the diagonal elements of a matrix, is the feature weight vector, Each element in Initialized to the same value, indicating that all features have the same importance in the initial stage; S423, the feature adjustment module recursively optimizes the initially generated low-dimensional features. In each round of iteration, the module adjusts the weights of each feature according to the performance of the features in the previous round, gradually enhances those features with important influences, and gradually weakens redundant or noise features. Assume that in the tth round of iteration, the weight update rule is as follows: ， In the formula, is the feature weight of the t+1th iteration, is the feature weight of the t-th iteration, is the learning rate of the autoencoder neural network, is the gradient of the loss function with respect to the feature weight, is the loss function of the autoencoder neural network, is the feature representation after feature weight adjustment, is the label data; S424, in order to ensure the effectiveness of the dimensionality reduction process, the decoder remaps the low-dimensional features back to the high-dimensional space to ensure that the dimensionality reduction process does not lose important information; the reconstruction process of the decoder is expressed as: ， In the formula, is the reconstructed high-dimensional data, is the multi-layer Sigmoid activation function of the decoder, is the weight matrix of the decoder, is the feature representation after feature weight adjustment, is the bias vector of the decoder; S425, repeatedly iterate the steps S422 to S424 until a preset stop iteration condition is met.

5. The method for controlling a fully electric aircraft tractor according to claim 1, characterized in that: The classifier of training the integrated gradient boosting tree algorithm based on dynamic decision boundary in S5 includes: S501, initializing the integrated gradient boosting tree classifier and setting the initial prediction value; S502: Calculate the residual between the initial prediction value and the actual category label for all training samples. The residual reflects the current prediction error of the model and serves as the target variable for subsequent training of weak classifiers. Specifically, the calculation formula for the residual of the sample is: ， In the formula, is the residual of the i-th sample in round 0, is the symbol of partial derivative, is the objective function of the integrated gradient boosting tree, is the predicted value of the sample, is the predicted value of the initial model for the i-th sample, is the true label of the i-th training sample; S503, iteratively training weak classifiers; S504: After each round of iteration, evaluate whether the model meets the stopping condition. The stopping condition may be that a preset maximum number of iterations is reached, or the error of the model drops to an expected threshold. When the stopping condition is met, the iteration process is terminated.

6. A fully electric drive aircraft tractor control system, characterized in that: A method for controlling a fully electric aircraft tractor according to any one of claims 1 to 5, comprising: A data storage module, used for storing data; A data collection module, used to collect real driving data and send it to the data storage module; a data expansion module, configured to generate expanded driving data based on the real driving data, and send the expanded driving data to the data storage module; a data processing module, configured to perform feature extraction on the real driving data and the expanded driving data or on the real driving data alone to obtain feature-extracted data and perform feature dimension reduction on the feature-extracted data to obtain feature-reduced data; A tractor control module, used for analyzing the feature dimension-reduced data to obtain specific operation instructions and sending the operation instructions to an execution module; The execution module is used to receive and execute the operation instruction.

Images