Transform model-based aircraft control quality evaluation method

Abstract

The invention discloses an aircraft control quality evaluation method based on a transformer model, and belongs to the field of aviation ergonomics. The method comprises the following steps: firstly, designing a flight mission subject primitive, and collecting a multi-modal physiological signal and flight mission performance data of a pilot; key features are extracted from the physiological signals through multi-level screening, and the key features are fused with a task performance grading result; a transform prediction model combining convolution and a multi-head self-attention mechanism is constructed and used for effectively modeling a local feature and cross-modal global dependency relationship and realizing automatic prediction of a manipulation quality level; a few-sample fine tuning strategy based on dynamic weight freezing is adopted, so that the model can quickly adapt to different pilots, and the generalization ability is improved; and finally, the prediction result of the aggregation group is finally evaluated. The objective and automatic evaluation of multi-source data driving is realized, the limitation of a traditional subjective evaluation method is overcome, and the consistency and efficiency of evaluation are improved.

Description

Technical Field

[0001] This invention belongs to the field of aviation human factors engineering and control quality evaluation technology, and in particular relates to an aircraft control quality evaluation method based on the transformer model. It is an evaluation method that integrates flight mission performance and multimodal physiological signals into a unified model, thereby outputting an aircraft control quality score. Background Technology

[0002] For a long time, the evaluation of aircraft handling qualities has primarily relied on the traditional Cooper-Harper scale. This method requires experienced test pilots to make graded judgments based on mission completion and subjective feelings. While this evaluation approach is widely used in civil aviation and has developed relatively unified standards in practice, it remains essentially an experience-based subjective evaluation method. With the rapid development of urban air traffic and electric vertical take-off and landing (eVTOL) aircraft, aircraft configurations are becoming increasingly complex, and the pilot pool is becoming more diverse. Relying solely on the subjective evaluations of a few expert test pilots is no longer sufficient to meet the needs of rapid iterative design, performance verification, and airworthiness certification for next-generation aircraft.

[0003] The Cooper-Harper method has significant limitations in existing technologies. First, subjective ratings are influenced by pilot experience levels and individual differences, making the results susceptible to bias and difficult to guarantee consistency and repeatability. Second, flight testing relying on expert test pilots is not only costly but also time-consuming, significantly hindering aircraft development efficiency. Finally, with the diversification of flight modes and the increasing complexity of operating environments for new aircraft, traditional scales struggle to comprehensively reflect pilot control loads and compensatory behaviors, limiting the guiding role of evaluation results in design optimization and control law improvement. Furthermore, some studies use flight performance indicators as objective measures of control quality, such as measuring mission completion through flight trajectory, flight path angle, and deviations from performance limits, categorizing performance into ideal, adequate, controllable, and uncontrollable levels. While these methods overcome the limitations of purely subjective evaluation to some extent, they primarily focus on the mission outcome itself, neglecting the human loads and compensatory behaviors experienced by pilots during mission completion.

[0004] In recent years, the development of wearable physiological sensors has provided new technological approaches for monitoring pilot condition. Multimodal physiological signals such as electrocardiogram, skin conductance, respiration, electromyography, and eye movement can objectively reflect a pilot's cognitive load and compensation patterns. However, existing studies often separate flight performance and physiological signals, using them separately for research on mission completion quality or pilot condition, lacking a method to integrate the two through models and form a unified framework for predicting handling quality.

[0005] Meanwhile, transformer models based on self-attention mechanisms have been widely applied in natural language processing, time series prediction, and multimodal learning. Compared with traditional recurrent neural networks (RNN, LSTM, etc.), transformers establish dependencies between arbitrary features within the same layer through self-attention, making them more suitable for modeling complex relationships between multimodal features. Some work has begun to explore the application of transformers in physiological signal analysis such as electrocardiograms and electroencephalograms, as well as human-computer interaction state recognition. However, these approaches are mostly aimed at single-modality or general scenarios, and a transformer network structure and training process for evaluating aircraft handling quality has not yet been established.

[0006] In view of the above problems, this invention proposes an aircraft handling quality assessment method based on the transformer model. This method, based on the Mission Task Element (MTE), comprehensively utilizes flight performance indicators and pilot physiological signals, and combines feature selection and deep learning models to achieve objective and automated prediction of handling quality levels. This method not only overcomes the subjectivity of the traditional Cooper-Harper method but also compensates for the shortcomings of relying solely on flight performance while neglecting human factors, providing a new technical approach for the design verification and airworthiness certification of new aircraft. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for evaluating aircraft handling qualities based on a transformer model. This method uses mission performance indicators and multimodal physiological signals as inputs, and utilizes a transformer encoding structure combining convolutional units and multi-head self-attention to automatically learn the cross-modal coupling patterns of pilot compensatory behavior and physiological load. This method effectively overcomes the drawbacks of traditional Cooper-Harper methods, such as reliance on expert subjective ratings, high cost, and low efficiency. By introducing a pilot-in-the-loop (MTE) design into the experimental environment, the evaluation process is ensured to be repeatable and representative. This method improves the consistency and accuracy of aircraft handling quality evaluation and shortens the iterative cycle of aircraft design and validation.

[0008] This invention provides a method for evaluating the handling qualities of an aircraft based on a transformer model, comprising the following steps: Design flight missions and collect multimodal physiological signals and flight mission performance of pilots during the execution of the missions; classify the flight mission performance. Feature extraction and screening of the multimodal physiological signals are performed to obtain key physiological features, which are then combined with the results of flight mission performance grading to form key quantitative indicators of handling quality. Using key indicators of manipulating quality as input, local and global features are jointly modeled by convolutional units and self-attention mechanisms. The global dependencies of features are modeled by positional encoding and multi-head self-attention mechanisms to construct a transformer-based prediction model. The transformer-based prediction model is trained by using few-shot fine-tuning and dynamic weight freezing. The trained transformer-based prediction model is adapted and generalized across pilots to obtain a control quality prediction model. The control quality prediction model is used to predict data from multiple pilots to obtain prediction results for the data from multiple pilots. The final aircraft control quality assessment result is obtained based on the prediction results of the data from multiple pilots.

[0009] Optionally, the designed flight missions include pilots constructing Mission Subject Elementary Modules (MTEs) in a ring-shaped experimental platform to simulate control mission scenarios; by setting the flight mission completion time, control precision, and trajectory deviation, the pilot's flight mission performance when completing the Mission Subject Elementary Modules (MTEs) is converted into multiple levels.

[0010] Optionally, the step of feature extraction and screening of multimodal physiological signals includes: Temporal, frequency, and nonlinear features are extracted from multimodal physiological signals to construct a candidate feature set. At least two different feature importance assessment methods are used to evaluate the candidate features, and the features that rank higher in the at least two assessment methods are retained as the key physiological features.

[0011] Optionally, the feature importance evaluation method includes a statistical analysis method based on Copula entropy, a machine learning method based on Shapley values, and a dynamic analysis method based on channel attention weights.

[0012] Alternatively, the execution steps of a transformer-based prediction model are as follows: Local feature encoding of key physiological features is performed through one-dimensional convolution operation; the encoded features are added with the positional encoding to form a sequence embedding; the sequence embedding is input into at least one transformer encoding layer, which uses a multi-head self-attention mechanism to calculate the global correlation between features; finally, the output of the transformer encoding layer is concatenated with the result of flight mission performance grading, and the predicted probability distribution of the handling quality level is output through a classification layer.

[0013] Optionally, the specific steps of the few-sample fine-tuning are as follows: Based on a pre-trained transformer-based prediction model, the contribution of model parameters to the output is calculated to determine parameter importance; gradient gating coefficients are generated based on the parameter importance to attenuate or mask the gradients of highly important parameters during backpropagation; subsequently, using new pilot data, only parameters that are not frozen or have low gradient attenuation are updated.

[0014] Optionally, the contribution of the model parameters can be calculated using the integral gradient method, and the importance of the parameters can be converted into freezing probabilities based on the contribution, thereby controlling the gradient gating coefficients.

[0015] Optionally, the multimodal physiological signals include at least three of the following: electrocardiogram, skin conductance, respiration, electromyography, and eye movement signals.

[0016] Optionally, the new pilot data refers to the results of multimodal physiological signal samples collected by the new pilot during the flight mission and their corresponding flight mission performance ratings.

[0017] Optionally, the final aircraft handling quality assessment result is the average handling quality perception score obtained by averaging the predicted scores of all participating pilots under the same flight mission.

[0018] Compared with the prior art, the present invention has at least the following beneficial effects: 1. The aircraft handling quality assessment method of this invention realizes the transformation from evaluation relying solely on subjective experience to evaluation driven by multi-source data. Through standardized MTE design and unified experimental benchmarks, it considers both mission completion results and pilot compensation behaviors, effectively improving the comprehensiveness and consistency of handling quality testing.

[0019] 2. The aircraft handling quality assessment method of the present invention constructs a comprehensive framework of multi-level feature screening and deep learning prediction model, which can capture the nonlinear relationship and long-range dependence between mission performance and physiological signals while ensuring feature stability, thereby significantly improving the scientificity and reliability of aircraft handling quality level prediction.

[0020] 3. The aircraft handling quality evaluation method of the present invention achieves model generalization across pilot groups through small-sample fine-tuning. While reducing experimental and training costs, it ensures the applicability of the method to pilot groups with different experience levels and has broad engineering application value. Attached Figure Description

[0021] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0022] Figure 1This is a flowchart of the aircraft handling quality evaluation method based on the transformer model of the present invention; Figure 2 A schematic diagram of the basic design of typical task subjects; Figure 3 This is a diagram illustrating the determination of task performance levels. Figure 4 This is a schematic diagram of a multi-level feature filtering process; Figure 5 This is a schematic diagram of the structure of a transformer-based manipulation quality prediction model. Figure 6 A schematic diagram illustrating the fine-tuning and adaptation process for pilots with a small sample size; Figure 7 This is a schematic diagram of the improved LeNet-5 model of the present invention; Figure 8 This is a schematic diagram of the attention mechanism embedded in the improved LeNet-5 model of the present invention. Detailed Implementation

[0023] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0024] A specific embodiment of the present invention, such as Figures 1-8 This paper discloses a method for evaluating aircraft handling quality based on a transformer model. The method uses a fusion of flight mission performance and multimodal physiological signals as input, and includes the following steps: Step S1: Design flight missions and collect multimodal physiological signals and flight mission performance of pilots during flight missions; Specifically, the designed flight missions include pilots constructing Mission Elementary Tasks (MTEs) in a test environment to simulate typical maneuvering scenarios. Taking an eVTOL aircraft as an example, these MTEs include hovering, heading correction, transition flight, and altitude change. By setting mission indicators such as mission completion time, control precision, and trajectory deviation, flight mission performance is converted into Satisfactory (SAT), Adequate (ADQ), Controllable (CON), or Uncontrollable (UNC) levels. Flight mission performance refers to the degree to which the pilot completes the mission according to preset standards.

[0025] Specifically, the pilot's multimodal physiological signal acquisition utilizes wearable sensors to obtain the pilot's electrocardiogram, skin conductance, respiration, electromyography, and eye movement data, which are recorded synchronously with the Cooper Harper Score.

[0026] Furthermore, the multimodal physiological signals are preprocessed to obtain preprocessed multimodal physiological signals.

[0027] Step S2: Extract and screen features from the preprocessed multimodal physiological signals to obtain key physiological features. Based on the key physiological features and the flight mission performance obtained in step S1, a key indicator for quantitative control quality is formed.

[0028] Features are extracted from the preprocessed multimodal physiological signals, and a candidate feature set is constructed. The feature importance of the features in the candidate feature set is evaluated in a multi-level manner through statistical analysis, machine learning and dynamic weight analysis, and key physiological features are selected. Based on the key physiological features and the flight mission performance level obtained in step S1, a quantitative key indicator of control quality is formed. Preferably, the feature extraction in step S2 includes extracting time-domain, frequency-domain, and nonlinear features from the preprocessed multimodal physiological signals to construct a candidate feature set.

[0029] Preferably, in step S2, the feature importance of the candidate feature set is evaluated using a multi-level method, including statistical methods (Copula entropy calculation), machine learning methods (Shapley value calculation), and dynamic analysis (channel attention weight). After obtaining the three types of indicators, the features ranking in the top 30% in at least two indicator systems are retained as key indicators.

[0030] Specifically, the steps for obtaining three types of indicators by conducting multi-level evaluation of the feature importance of features in the candidate feature set are as follows: First, time-domain, frequency-domain, and nonlinear features were extracted from preprocessed multimodal physiological signals (including electrocardiogram, skin conductance, respiratory, electromyography, and eye movement features) to form a candidate feature set. To evaluate the correlation between these features and the manipulation quality rating (CHR score), three complementary analytical methods were employed: 1. Statistical Method (Copula Entropy Calculation): Copula entropy analysis is used to analyze the joint distribution characteristics between different features and flight performance levels, measuring the nonlinear correlation between variables. This method can identify complex coupling relationships that are difficult to discover using traditional linear statistics.

[0031] 2. Machine Learning-Based Approach (Shapley Score Calculation): A random forest model is trained using physiological features as training input, manipulation quality ratings as training labels, and predicted manipulation quality ratings as output. The Shapley score is then introduced to quantify the marginal contribution of each candidate feature in the random forest model to the prediction results. The Shapley score reflects the degree of influence of different features on the manipulation quality rating at the model level.

[0032] 3. Attention-based approach: Using the improved LeNet-5 framework as a carrier, the weights of each channel are calculated to evaluate the strength of the role of different modal features in the temporal modeling process and the model's attention, thereby revealing the cross-modal dependency structure between features.

[0033] Further, see Figures 7-8 The improved LeNet-5 model structure for calculating attention weights consists of an attention mechanism layer, two levels of convolutional-pooling layers, and two fully connected mapping layers. The network input is an input feature map H*W*C, where H is the height of the feature map, W is the width of the feature map, and C is the channel dimension of the feature map. In this LeNet-5 model structure, it is a 90-dimensional single-channel feature vector. First, a 128-dimensional first feature map is extracted through the first convolutional layer. In this stage, an attention mechanism is introduced to adaptively weight key features to obtain a weighted feature map. The second feature map is then obtained by downsampling through the first pooling layer. The third feature map is obtained after passing through the second convolutional layer. The fourth feature map is obtained by further downsampling from the second pooling layer. After flattening it into a 32-dimensional vector, it is passed through the first and second connection layers for non-linear mapping. Finally, the output layer outputs a 9-dimensional result, representing the probabilities of the 9 categories.

[0034] Furthermore, the expression for the channel weight is: (1) in, L It is the spatiotemporal dimension of the input features. and Represents the weight matrix. and This represents the bias vector. This represents the Sigmoid function. Indicates the first c The first channel i Feature values ​​of a spatiotemporal location This represents the activation function.

[0035] In this invention, each feature is considered as a channel. Therefore, both global average pooling and global max pooling degenerate into the feature value itself, i.e. L =1.

[0036] After obtaining the importance indices of the three categories of features mentioned above, the results are standardized and compared comprehensively. Only the features ranking in the top 30% across at least two methods are retained as the final set of key physiological indicators. This set of key physiological indicators, along with the flight performance rating obtained based on mission standards, forms the input to the handling quality prediction model.

[0037] Step S3: Using the aforementioned key indicators for maneuver quality as input, local and global features are jointly modeled by convolutional units and self-attention mechanisms. The global dependency between position encoding and multi-head self-attention mechanism modeling features is introduced to construct a transformer-based prediction model to predict the maneuver quality level of the aircraft. Preferably, in step S3, the transformer model first extracts local features of the pilot's physiological signals from the key indicators of maneuver quality quantification through convolutional units, and maps the local features of the physiological signals into sequence embeddings with positional encoding to preserve the temporal dependence of the physiological signals. Then, a multi-head self-attention mechanism is used to model the global correlation between different modal features, as shown in Equation (3), thereby identifying the pilot's cross-modal compensatory adjustment patterns during maneuvering.

[0038] This invention uses positional encoding to explicitly provide the network with information about the temporal or sequential order. Convolutional units extract local physiological features (such as heart rate, electromyography, etc.) within each short time window. Positional encoding labels the spatial location of these features extracted during flight control. Self-attention then captures the interaction relationships between features (e.g., whether electromyography signals affect electrocardiogram signals after changes in control load). This invention maps local features of physiological signals to sequence embeddings with positional encoding, adding positional encoding to the input features so that the model can recognize the relative or absolute positional relationships of features. This overcomes the problem that existing transformer self-attention mechanisms are structurally insensitive to input order, all features are processed in parallel, and the model itself cannot distinguish the order of signals.

[0039] This invention employs multi-head self-attention to measure the correlation between any two features in an input sequence, reflecting the simultaneous coupling of different modalities (such as the synergistic regulation between electromyography (EMG) and heart rate). For example, cross-modal interactions involve multiple physiological modalities, such as the combined balance of neuromuscular, autonomic, and cardiovascular systems. This invention's multi-head self-attention reflects the indirect influence and many-to-many coupling across modalities (rather than one-to-one local correlations). For instance, without multi-head attention, it might only detect an increase in EMG when EMG increases, but with multi-head attention, it might detect an increase in EMG only when both EMG and ECG increase, thus improving the accuracy of the assessment results.

[0040] Furthermore, the Layer Normalization layer inside the transformer encoder enhances feature representation through a feedforward network and normalization operations. In the classification layer, a Softmax activation function is used to obtain the predicted probability distribution of the handling quality level, enabling automated prediction of the aircraft's handling quality level. The corresponding handling quality perception score prediction result is output.

[0041] Furthermore, the normalization operation involves standardizing each sample with zero mean and unit variance across the channel dimension, aligning the dimensions of different modalities / features. This facilitates the attention mechanism and subsequent feedforward networks in modeling complex relationships on the same scale. Specifically, the feedforward network independently performs nonlinear mapping and dimensionality expansion on each feature vector to enhance the nonlinear capacity of the representation and its interaction with features, forming a global representation.

[0042] Furthermore, the self-attention mechanism of the transformer-based prediction model uses the input feature sequence... The linear transformation generates a query (Q), key (K), and value (V) matrix, expressed as: (2) in, x n Indicates the first The input feature vector, that is, the nth feature vector in the input sequence. One element; Represents the real number field; This indicates the length of the input sequence, i.e., the number of elements in the input feature sequence; This represents the dimension of each input vector; , and For learnable linear transformation weight matrix, , and This is a bias term.

[0043] It is understandable that the input feature sequenceX To extract local features of the pilot's physiological signals and flight mission performance.

[0044] Then, scaled dot product attention is used to obtain the correlation between the query (Q), key (K), and value (V) matrices, expressed as: (3) in, This represents the vector dimension in the key matrix.

[0045] Furthermore, to enhance the model's ability to express different feature dimensions, a multi-head attention mechanism is introduced, expressed as: (4) in, Indicates the first j The output of each attention head; This represents the multi-head attention mechanism function, which concatenates the outputs of multiple attention heads (using `Concat`) and then performs a linear transformation to output the projection matrix. To merge; Indicates the first j The weight matrix of each attention head. This indicates the output projection matrix.

[0046] Furthermore, through parallel computation of the multi-head attention mechanism function, the transformer-based prediction model can learn multi-scale feature interactions from different subspaces. Following the attention layer, the encoder structure employs a feedforward network to further enhance the nonlinear modeling capability of the transformer-based prediction model, and combines residual connections and layer normalization operations, such as... Figure 5 As shown in the classification module.

[0047] Furthermore, residual joins and layer normalization operations are used to obtain the residual join output. The expression is: (5) in, Y This is the input vector of the corresponding sub-layer, i.e., the input of multi-head attention; It is the output vector of the corresponding sub-layer, that is, the output of multi-head attention. It adds the input to the sub-layer output to alleviate gradient vanishing, stabilize training, and keep the input feature information from being lost. Presentation layer normalization operation.

[0048] Furthermore, the output of the normalization layer is concatenated with the flight mission performance level, and then input into the feedforward network to obtain the predicted probability of the target category, expressed as: (6) in,Z This represents the input of the feedforward network, i.e., the residual connection output in formula (5); and These represent the linear transformation weights of the first and second layer weight matrices, respectively. and These represent the bias terms of the first and second layers, respectively.

[0049] Furthermore, in step S3, during the training of the transformer-based prediction model, the integral gradient method is used to calculate the integral gradient contribution of trainable parameters such as the parameters of the convolutional feature extraction layer, the parameters of the multi-head self-attention sub-layer in each transformer coding layer, the parameters of the feedforward network sub-layer, the layer normalization parameters, and the parameters of the output classification layer to the model output. The expression for this contribution is as follows: (7) in, For the input sample, As the baseline, For the model output function, In the transformer-based prediction model, the first... m The trainable parameters include the convolution kernel and bias of the convolutional unit, the elements in the query / key / value and output projection matrix of the multi-head self-attention sublayer, the weights and biases of each layer of the feedforward network, the scaling and translation parameters of the layer normalization, and the weights and biases of the output layer, etc. These are the integral path coefficients; Indicates input sample x The first in p Each feature component; Indicates baseline input The first in p One component; Indicates the first m trainable parameters In input sample x With limit input The contribution of the integral gradient on the path between them.

[0050] Furthermore, the contribution of the integral gradient described above is used to obtain the first... m trainable parameters In input sample and baseline The path integral gradient between them. For each trainable parameter Calculate its L2 norm to construct the parameter importance matrix. ,in, Represents the real number field. This represents the total number of trainable parameters in the prediction model. The m-th trainable parameter in the parameter importance matrix. Importance score S m The larger the value, the more trainable parameters there are. The more significant the global impact on the output, the higher the importance score S will be. m This is converted into a freeze probability, controlling the degree of freedom in parameter updates. The top 20% of parameters by importance are assigned a high freeze probability, while the bottom 30% of parameters are allowed to be updated freely.

[0051] Furthermore, the importance scores of the parameters are converted into frozen probabilities. The expression is: (8) (9) in, Represents the Sigmoid function; This represents the importance score of the m-th trainable parameter, i.e. The L2 norm; , These are the sets of importance scores { S m The mean and standard deviation of}; This is the steepness coefficient, used to control the slope and discrimination of the Sigmoid function; Represents the m-th trainable parameter The integral gradient contribution corresponding to the kth scalar parameter element after flattening the vector; Represents the m-th trainable parameter The number of scalar elements after flattening into a vector.

[0052] Further, see Figure 6 During support set fine-tuning, each training batch first performs forward propagation and calculates the loss for that batch. Then, backpropagation is performed to obtain the parameter gradients. Next, gradient gating is used to adjust the importance score S. m Apply a gradient mask to the top 30% of trainable parameters, expressed as: (10) in, The masking gradient for the m-th trainable parameter represents the gradient value that the m-th trainable parameter is actually used to update after applying the frozen weight gate. The original gradient of the m-th trainable parameter is the original gradient of the m-th trainable parameter calculated by the loss function during backpropagation.

[0053] Specifically, based on the aforementioned parameter importance scores Calculate the probability of freezing Then, set the gradient gating coefficients for each parameter. When When the gradient is at a certain value, it is decayed to less than 20% of its original value, resulting in a significant decay of the gradient of that parameter. Conversely, the gradient gating coefficient is set to 1 or a value close to 1, resulting in a significant decay of the gradient of that parameter.

[0054] Finally, gradient descent is applied only to the weights that are not frozen, with the update rule as follows: (11) in, and Let represent the values ​​of the m-th trainable parameter at iterations t+1 and t, respectively; η represents the learning rate, which controls the step size of parameter updates and determines the magnitude of weight adjustments made by the model in each iteration. In this invention, η is set to 10. -5 .

[0055] In each iteration, forward propagation is first performed using the current batch of support set samples as input to calculate the loss between the model output and the target manipulation quality score; then, backpropagation is performed to obtain the original gradients of each parameter. Then based on the freezing probability Calculate the gradient gating coefficients to obtain the masking gradient. The unfrozen parameters are updated according to equation (11). If the preset iteration termination condition has not been met after the parameter update is completed, the next batch of support set samples is read and the forward propagation step is restarted until the termination condition is met.

[0056] Finally, the encoded features are concatenated with the flight mission performance level and input into the classification module. The predicted probability distribution of the handling quality level is obtained by using the Softmax activation function.

[0057] Step S4: Perform cross-pilot generalization processing on the transformer-based prediction model to obtain the handling quality prediction model. A few-sample fine-tuning method is used to achieve cross-pilot group adaptation, maintain high prediction performance under a small amount of new data, improve the generalization ability and practicality of the method, and thus make the evaluation method applicable to pilot groups with different experience levels and backgrounds.

[0058] Preferably, the cross-pilot generalization process in step S4 employs a few-shot fine-tuning method, specifically as follows: By freezing most of the parameters of the transformer-based prediction model and updating only some weights, the model can be quickly transferred using a small amount of new pilot data, thereby significantly improving the generalization ability and practicality of the method while ensuring prediction accuracy.

[0059] Step S5: Aggregate the predicted handling quality ratings output by the handling quality prediction model. Given an aircraft configuration and mission subject, input the feature sequences of all participating pilots into the handling quality prediction model to obtain the predicted score for each pilot in each test flight. Statistically calculate the arithmetic mean of all predicted scores for all participating pilots under the same mission subject, using this as the mean handling quality perception score for that aircraft under that mission subject, and use it as the final evaluation output.

[0060] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution is provided through a specific embodiment. The specific implementation steps are as follows: The aircraft handling quality assessment method based on the fusion of flight performance and physiological signals described in this invention comprises four parts: mission design and signal acquisition, feature extraction and screening, predictive model construction, and cross-pilot generalization. Figure 1 As shown.

[0061] (1) Task design and signal acquisition; To verify the aircraft handling quality assessment method of the present invention, a typical mission subject primitive (MTE) was designed in the pilot-in-the-loop experimental environment, and experiments were carried out in conjunction with a multimodal signal acquisition system.

[0062] At the mission level, this invention takes eVTOL as the research object and constructs a basic MTE mission covering typical mission subjects such as hovering, heading correction, transition flight, and vertical landing. Figure 2 A schematic design for a "transition-landing" mission is presented: the aircraft enters the deceleration phase from an altitude of 500 feet and an indicated airspeed of 70 knots, gradually transitions to a hovering mode, and achieves a vertical descent to the target point within specified boundaries. Regarding mission evaluation, a grading standard is defined based on the aircraft's flight mission performance. According to the mission objective basic unit (MTE) performance limits, flight performance is divided into four categories: satisfactory, adequate, controllable, and uncontrollable. Among these, SAT and ADQ correspond to stricter performance tolerances, while CON and UNC reflect situations where the aircraft deviates from the mission objective or becomes uncontrollable. The judgment criteria are shown in Table 1.

[0063] Table 1 Examples of Transition-Landing Flight Performance Standards

[0064] At the experimental implementation level, pilots performed tasks within the experimental platform. To ensure experimental safety, the platform was equipped with an emergency stop mechanism, allowing both participants and experimenters to terminate the experiment in real time. Before the experiment, all participants underwent approximately 30 minutes of simulator training to familiarize themselves with the flight environment and mission requirements. During the experiment, pilots wore signal acquisition systems and followed the experimental procedures. After completing each MTE task, pilots completed the Cooper-Harper Scale to provide subjective ratings. Three types of information were recorded simultaneously during the experiment, as shown in Table 2.

[0065] Table 2 Experimental Record Information

[0066] (2) Feature extraction and screening; After acquiring the raw physiological signals, this invention first preprocesses the data to ensure signal stability and usability. Key preprocessing steps are shown in Table 3. Subsequently, the preprocessed physiological signals are used for physiological feature extraction, extracting time-domain, frequency-domain, and nonlinear features from the multimodal signals. These physiological features include, but are not limited to: time-domain statistics (such as mean and standard deviation), frequency-domain energy distribution (such as low-frequency / high-frequency power), and nonlinear indices (such as approximate entropy) of electrocardiogram (ECG) signals; root mean square value and power spectrum characteristics of electromyography (EMG) signals; derivative and skin conductance frequency of electrodermal (EDS) signals; and fixation duration and saccade amplitude of eye movement (EMG) signals. Furthermore, flight performance based on task-specific primitives is quantified into rating indicators, including four categories: ideal, adequate, controllable, and uncontrollable, represented by 1, 2, 3, and 4 respectively, to reflect the quality of flight mission completion.

[0067] Table 3 Physiological signal preprocessing workflow

[0068] To ensure the effectiveness and stability of the extracted features, the importance of candidate features is evaluated at multiple levels. The feature selection process is as follows: Figure 4 As shown.

[0069] (3) Construction of quality prediction model; After obtaining key physiological characteristics and mission performance levels, this invention constructs a transformer-based prediction model for automated evaluation of aircraft handling qualities. This model fully leverages the complementary advantages of convolutional units and multi-head self-attention mechanisms, enabling simultaneous modeling of local signal features and global cross-modal dependencies.

[0070] The model first uses one-dimensional convolution to extract local correlations of the input physiological features, such as... Figure 5 The information fusion module is shown in the diagram. The key physiological feature sequence obtained from step S2 is denoted as Z(z1, z2, …, z). s ), where zs Let represent the feature value of the key physiological feature sequence Z in the s-th feature dimension, where s is the feature dimension, and in this embodiment, s is 20. A one-dimensional convolution operation is applied to the feature dimension, with a kernel size of 3 and a number of kernels (filters) of 64, and a ReLU nonlinear activation function is used. Zero padding is used at the boundaries. The kernel size of 3 is used to extract the local correlation between adjacent features, and the 64 filters are used to capture multiple local patterns in different subspaces. This operation yields the convolutionally encoded feature map E=(e1, e2,…,e…). n Its size is n. 64, where n represents the sequence length; e n This represents the feature vector corresponding to the nth sequence position in the feature map sequence E, used to characterize the local feature information of the input sample at the nth sequence position.

[0071] Subsequently, the transformer mechanism was introduced, such as Figure 5 The transformer module is shown in the diagram. In the overall architecture of this invention, the model first performs local convolutional encoding on the input feature map E to generate a positional encoding sequence P=(p1, p2,…,p) of the same length and dimension. n ), p n Let represent the position encoding vector corresponding to the nth sequence position in the position encoding sequence P, used to characterize the position information of the nth sequence position. Then, element-wise summation is performed to obtain the sequence embedding z containing position information for the nth sequence position. n =e n +p n Let e ​​represent the feature vector e at the nth sequence position. n With position encoding vector p n The position-enhanced embedding vector is obtained by element-wise addition. This yields the embedding sequence input to the transformer unit. Then embed the sequence A series of stacked transformer encoding units (two-layer transformer units in this embodiment) are sequentially fed into the array. Each transformer unit includes a multi-head self-attention sublayer, a feedforward network sublayer, and two residual connections and layer normalization steps. The multi-head self-attention sublayer takes the embedded sequence Z' as input, calculates multi-head self-attention, obtains the global dependencies between time steps / modalities, and outputs the sequence. The first residual connection and layer normalization sums the input and the output of the attention sublayer and performs layer normalization to obtain a stable intermediate representation. Subsequently, the feedforward sublayer independently applies two fully connected and non-linearly activated feedforward layers to the feature vector at each time step, enhancing non-linear modeling capabilities and outputting the sequence. Finally, residual connections and layer normalization are performed again to obtain the final output of the layer. .in, l =1, 2, are the transformer layer numbers; the output of the second layer. The final output of the transformer module is a feature sequence containing global temporal and cross-modal dependency information.

[0072] Subsequently, the second layer output The classification module then proceeds. It first performs batch normalization on both the batch and feature dimensions, standardizing each channel to zero mean and unit variance, and introducing learnable scaling and translation parameters to eliminate statistical bias between different samples, resulting in a normalized feature sequence. Then, perform average-max pooling on the time dimension to normalize the feature sequence. The data is compressed into a single global feature vector u. This global feature vector u is then concatenated with the single feature vector representing flight performance level to obtain a new feature vector. Then, through a fully connected layer, the new feature vector is... The predicted vectors are obtained by mapping them to the output space. Then, a Softmax activation function is applied to the output layer to obtain the predicted probability distribution of each manipulation quality level.

[0073] This design enables the model to integrate both flight performance and physiological signals, effectively capturing complex temporal and cross-modal interactions, thereby improving the accuracy and robustness of control quality prediction. The model structure is illustrated below. Figure 5 As shown.

[0074] (4) Cross-pilot generalization; To improve the model's adaptability across different pilot groups, this invention proposes a few-sample transfer learning method based on dynamic weight freezing, the process of which is as follows: Figure 6 This method first pre-trains on an existing pilot dataset to extract universal cross-subject features; then, a dynamic weight freezing mechanism is introduced during the transfer process to preserve global feature representations through parameter importance calculations and fine-tuning on a small number of samples from new pilots, thereby achieving rapid adaptation and performance maintenance under cross-group conditions.

[0075] This yields the predicted handling quality results (probability distribution of each level) for a single test flight sample. In step S5, the predicted results for all participating pilots under the same mission are statistically analyzed to obtain the mean handling quality perception score, which serves as the final evaluation indicator.

[0076] Among them, the small sample of new pilots includes at least 1–5 labeled samples containing a level of handling quality.

[0077] This invention, through the aforementioned mechanism and a few-sample update strategy, enables the handling quality assessment model to rapidly adapt with a very small amount of new pilot data, while maintaining high classification accuracy and consistency on the query set. Thus, this invention reduces experimental and training costs while achieving broad applicability of the handling quality assessment method under different experience levels and individual differences.

[0078] Through the above steps, this invention achieves the fusion modeling of flight performance and physiological signals, enabling objective and automated evaluation of aircraft handling qualities in standardized mission scenarios, reducing reliance on subjective human scoring, and improving the scientific rigor and consistency of evaluation results.

[0079] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the handling qualities of an aircraft based on a transformer model, characterized in that, Includes the following steps: Design flight missions and collect multimodal physiological signals and flight mission performance of pilots when performing the missions; The performance of the flight missions was graded. Feature extraction and screening of the multimodal physiological signals are performed to obtain key physiological features, which are then combined with the results of flight mission performance grading to form key quantitative indicators of handling quality. Using key indicators of manipulating quality as input, local and global features are jointly modeled by convolutional units and self-attention mechanisms. The global dependencies of features are modeled by positional encoding and multi-head self-attention mechanisms to construct a transformer-based prediction model. The transformer-based prediction model is trained by using few-shot fine-tuning and dynamic weight freezing. The trained transformer-based prediction model is adapted and generalized across pilots to obtain a control quality prediction model. The control quality prediction model is used to predict data from multiple pilots to obtain prediction results for the data from multiple pilots. The final aircraft control quality assessment result is obtained based on the prediction results of the data from multiple pilots.

2. The aircraft handling quality evaluation method according to claim 1, characterized in that, The designed flight missions include pilots constructing Mission Subject Elementary Models (MTEs) in a ring-shaped experimental platform to simulate control mission scenarios; by setting the flight mission completion time, control accuracy, and trajectory deviation, the pilot's flight mission performance when completing the MTEs is converted into multiple levels.

3. The aircraft handling quality evaluation method according to claim 1, characterized in that, The steps for feature extraction and screening of multimodal physiological signals include: Temporal, frequency, and nonlinear features are extracted from multimodal physiological signals to construct a candidate feature set. At least two different feature importance assessment methods are used to evaluate the candidate features, and the features that rank higher in the at least two assessment methods are retained as the key physiological features.

4. The aircraft handling quality evaluation method according to claim 3, characterized in that, The feature importance assessment methods include statistical analysis methods based on Copula entropy, machine learning methods based on Shapley values, and dynamic analysis methods based on channel attention weights.

5. The method for evaluating aircraft handling qualities according to claim 1, characterized in that, The execution steps of a transformer-based prediction model are as follows: Local feature encoding of key physiological features is performed through one-dimensional convolution operation; the encoded features are added with the positional encoding to form a sequence embedding; the sequence embedding is input into at least one transformer encoding layer, which uses a multi-head self-attention mechanism to calculate the global correlation between features; finally, the output of the transformer encoding layer is concatenated with the result of flight mission performance grading, and the predicted probability distribution of the handling quality level is output through a classification layer.

6. The method for evaluating aircraft handling qualities according to claim 1, characterized in that, The specific steps of the few-sample fine-tuning are as follows: Based on the pre-trained transformer-based prediction model, the contribution of model parameters to the output is calculated to determine the importance of the parameters; gradient gating coefficients are generated based on the parameter importance to attenuate or mask the gradients of highly important parameters during backpropagation. Then, using the new pilot data, only parameters that were not frozen or had low gradient decay were updated.

7. The method for evaluating aircraft handling qualities according to claim 6, characterized in that, The contribution of the model parameters is calculated using the integral gradient method, and the importance of the parameters is converted into freezing probabilities based on the contribution, thereby controlling the gradient gating coefficients.

8. The method for evaluating aircraft handling qualities according to claim 1, characterized in that, The multimodal physiological signals include at least three of the following: electrocardiogram, skin conductance, respiration, electromyography, and eye movement signals.

9. The method for evaluating aircraft handling qualities according to claim 6, characterized in that, The new pilot data consists of multimodal physiological signal samples collected from the new pilot during the flight mission and the corresponding flight mission performance ratings.

10. The method for evaluating aircraft handling qualities according to claim 1, characterized in that, The final aircraft handling quality assessment result is the average handling quality perception score obtained by averaging the predicted scores of all participating pilots under the same flight mission.

Images