An unmanned aerial vehicle form design method and device based on emotional driving and a medium

Abstract

The application belongs to the field of product generative design, and discloses a UAV form design method and device based on emotion driving and a medium. The method comprises the following steps: collecting user comment data, and extracting and screening core emotional words by using a BERTopic model; obtaining visual attention data of users on different UAV form components through an eye movement experiment, and calculating objective importance weight of each form component by combining an entropy weight-TOPSIS method; introducing the objective importance weight as prior knowledge into a deep learning model, constructing a Transformer-BiLSTM model fused with the prior weight, establishing a mapping relationship between user emotion and key UAV form, and predicting and generating an optimal form combination sketch; generating the sketch into a high-fidelity rendering image by using a stable diffusion model, combining physiological signals collected through eye movement and electroencephalogram multi-modal experiments, objectively verifying and optimizing the generated scheme, and determining an optimal design scheme, so that intelligent conversion from abstract emotional semantics to concrete product form is realized, and product design efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of product design, and in particular to an emotion-driven method, device and medium for drone form design. Background Technology

[0002] As artificial intelligence continues to expand into the design field, emotional factors are shifting from subjective aesthetics to calculable design variables. The key to achieving emotion-driven intelligent design lies in how to digitally characterize users' emotions and effectively map them into tangible product forms. However, accurately translating users' emotional needs remains a core challenge for designers.

[0003] Currently, existing emotional design and evaluation methods still have the following shortcomings in practical applications: 1. Limitations of traditional methods for acquiring and transforming emotional needs. Most methods still rely on questionnaires or semantic evaluation, which can only obtain limited explicit preference data and are difficult to capture users' visual attention and emotional fluctuations during the form perception process. This emotional mapping process, lacking multi-dimensional information, makes it impossible for designers to clearly identify which form features trigger specific emotional responses, resulting in untraceable and unexplainable design decisions, exhibiting typical "black box" characteristics. Kansei Engineering (KE), as a mature method, has been widely used in many fields. However, user emotions have significant randomness and ambiguity, and traditional KE methods relying on linear regression often fall short in high-dimensional semantics, long-range dependency analysis, and fine-grained emotional feature capture. 2. Insufficiency of existing deep learning models in form mapping. More and more research is beginning to introduce nonlinear machine learning techniques, such as convolutional neural networks (CNN) and generative adversarial networks (GAN), to improve the efficiency of KE in emotion capture. However, these methods often can only choose between global features or local sequence features, failing to take both into account. Currently, Transformer, with its self-attention mechanism, can efficiently model long-range dependencies in sequences, demonstrating excellent feature representation capabilities. However, Transformer still falls short in characterizing local dependencies and sequential features under limited sample conditions, making it difficult to fully capture fine-grained temporal information in emotional semantics. Relying solely on a single model makes it difficult to simultaneously integrate forward and reverse sequence information, failing to strengthen the representation of contextual information and local dynamics. 3. The selection and evaluation of generative design solutions lack objectivity. Generative AI (GAI) technologies such as the Stable Diffusion Model (SDM) have been widely used in visual creative expression, capable of efficiently and massively generating large-scale design solutions in a short time. However, how to evaluate and select high-quality design solutions remains a crucial issue. Traditional evaluation methods mainly rely on subjective means such as questionnaires or interviews, resulting in a lack of objectivity and real-time feedback in design feedback. Although EEG technology has been widely applied in the design field, it is mostly limited to single-modal applications, lacking deep integration and systematic exploration of multi-source data. While eye-tracking technology can accurately locate a user's visual attention, there is still a lack of mature solutions for applying multimodal fusion of eye tracking and EEG to design evaluation, making it difficult to achieve a two-layer analysis from explicit behavior to implicit cognition.

[0004] In view of the shortcomings of existing technologies in product emotional design, such as strong subjectivity in demand acquisition, "black box" mapping of form features, and lack of objective physiological basis for selecting generation schemes, this invention aims to propose an effective technical solution. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an emotion-driven method, device, and medium for drone form design. This method constructs a closed-loop design architecture of "emotion mining - weight quantification - intelligent mapping - generation and objective evaluation," achieving precise transformation and scientific verification from vague user emotional needs to concrete product appearance, greatly improving the objectivity of design decisions and user satisfaction.

[0006] In a first aspect, the present invention provides an emotion-driven method for designing the form of a drone, comprising the following steps:

[0007] S1: Collect and process user online comment data, and use the BERTopic model to extract and filter core sentiment words;

[0008] S2: Collect eye-tracking data of the user while observing sample images of the drone. Based on the eye-tracking data, use the entropy weight-TOPSIS method to calculate the objective importance weight of each part of the drone.

[0009] S3: Construct and train a Transformer-BiLSTM sentiment mapping model that integrates prior weights, incorporate the objective importance weights as prior knowledge into the Transformer-BiLSTM sentiment mapping model, input the UAV morphological feature data and the core sentiment words, and output the optimal morphological feature combination corresponding to each sentiment word.

[0010] S4: Using a stable diffusion model, the optimal combination of morphological features is transformed into a high-fidelity rendering image; for the high-fidelity rendering image, eye movement signals and electroencephalogram signals are collected simultaneously when the user views it, objective physiological indicators are extracted, and subjective scores are combined to comprehensively evaluate and rank each rendering image to determine the final drone design scheme.

[0011] As an optional implementation of the first aspect of this application, step S1 is specifically implemented as follows: Each user comment data is mapped to a vector E containing semantic information using the Sentence-BERT model; the UMAP algorithm is applied to convert the vector E from a high-dimensional space to a low-dimensional space; the HDBSCAN algorithm is used to perform density-based clustering on the dimensionality-reduced vector to discover potential sentiment themes; the core sentiment words are selected by calculating the class-based TF-IDF score, wherein the formula for calculating the class-based TF-IDF score is: ;in, For words In cluster The frequency in For words Total frequency across all clusters The average number of words per cluster.

[0012] As an optional implementation of the first aspect of this application, step S2 is specifically implemented as follows: acquiring eye-tracking index data on preset morphological components when the user observes a sample image of a drone, and constructing an original evaluation matrix X; normalizing the original evaluation matrix X using a weighting method to obtain a probability matrix P; and calculating the information entropy of the j-th eye-tracking index of each morphological component. and their corresponding weighting coefficients Construct a weighted TOPSIS decision matrix and normalize it into a standardized decision matrix. In order to determine the ideal solution With negative ideal solution Based on the weighting coefficients Calculate the sample of each morphological component and the positive ideal solution. and negative ideal solution Weighted Euclidean distance and ; Calculate the relative closeness of samples of each morphological component. and the relative closeness The objective importance weight of the aforementioned morphological components is output, and the calculation formula is as follows: ; where m represents the total number of drone morphological component samples participating in the evaluation.

[0013] As an optional implementation of the first aspect of this application, the specific method for constructing and training the Transformer-BiLSTM sentiment mapping model that integrates prior weights in step S3 is as follows: the objective importance weight vector is converted into a Hadamard product. With the normalized UAV morphological feature matrix By performing element-wise multiplication, a weighted feature matrix is ​​constructed. The weighted feature matrix The input is fed into a Transformer encoder, where a weighted self-attention mechanism captures the global dependencies between features. This weighted self-attention mechanism introduces the objective importance weight vector during the calculation of the query matrix Q and the key matrix K. Modulation is performed; the global feature sequence output by the Transformer encoder is input into a two-layer BiLSTM network to extract local temporal features, wherein the final feature representation of BiLSTM is... From the second layer of forward hidden state and reverse hidden state It is assembled; the predicted sentiment rating value is output through a fully connected layer, and the optimal combination of morphological features is selected and output based on the prediction results.

[0014] As an optional implementation of the first aspect of this application, in step S4, the specific implementation of converting the optimal morphological feature combination into a high-fidelity rendering image is as follows: the line drawing of the optimal morphological feature combination is used as a constraint and input to the ControlNet-Canny edge detection module; a stable diffusion model is used to generate an image of the content described by the input text prompt under the constraint of the ControlNet-Canny edge detection module, wherein the forward diffusion process of the stable diffusion model converts the original image... Gradually perturb to noisy image Its closed form is: ;in, For noise dispatch coefficient, The noise is standard Gaussian noise; the reverse denoising process of the model estimates the denoising result by fitting a Gaussian distribution and uses the KL divergence between the two Gaussian distributions as the training loss, and finally outputs the high-fidelity rendering image.

[0015] As an optional implementation of the first aspect of this application, in step S4, the method further includes performing time-frequency analysis on the acquired EEG signal, specifically by using a complex Morlet wavelet as the mother wavelet to perform continuous wavelet transform for time-frequency feature extraction. The mathematical expression of the complex Morlet wavelet is: ;in, For time, For the center frequency, The normalization constant is The bandwidth parameter of a wavelet is typically determined by the number of periods. The event-related spectrum perturbation (ERSP) is calculated based on the wavelet transform coefficients to quantify the power change relative to the baseline. The inter-trial coherence (ITC) is calculated based on the phase angle information of the wavelet transform coefficients to measure the consistency of neuronal firing phase across different trials. The results of the ERSP and ITC analysis are output.

[0016] As an optional implementation of the first aspect of this application, the specific method for extracting objective physiological indicators in step S4 is as follows: the scalp electrodes are divided into three regions of interest: the forebrain, midbrain, and hindbrain; the electrode signals in each region of interest are spatially averaged, and the event-related potential (ERP) components are extracted using a superposition averaging technique; the average waveform of the event-related potential (ERP) is... The calculation formula is: Where N is the total number of trials, The time-domain signal of the EEG in the kth valid trial is used; the N200 and P300 component amplitudes are extracted from the average waveform of the event-related potential (ERP), and combined with the total fixation duration, total access duration and fixation count indicators extracted from the eye movement signal to form and output the objective physiological indicators.

[0017] As an optional implementation of the first aspect of this application, the specific method for comprehensively evaluating and ranking the rendering images in step S4 is as follows: constructing a multimodal decision matrix that includes subjective scores and the objective physiological indicators; calculating the objective weights of the subjective scores and each objective physiological indicator using the entropy weight method; calculating the relative closeness between each rendering image and the ideal solution using the TOPSIS method; ranking all high-fidelity rendering images according to the relative closeness, and determining the design scheme corresponding to the highest-ranked rendering image as the final UAV design scheme.

[0018] In a second aspect, embodiments of this application provide an electronic device, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of the method described in the first aspect.

[0019] Thirdly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. This invention proposes an objective quantification method for form weights based on eye-tracking data. Addressing the subjectivity of traditional design methods that rely on designer experience to assign component weights, this invention acquires visual attention data through eye-tracking experiments and combines this with the entropy-weighted TOPSIS method to calculate the objective importance of form components. This method accurately quantifies the actual weights of different components such as the fuselage and arms in the user's perception, providing scientific prior knowledge for subsequent models.

[0022] 2. This invention constructs a Transformer-BiLSTM sentiment mapping model that integrates prior weights. Addressing the challenge of existing deep learning models struggling to balance global and local features with small sample sizes, this invention innovatively embeds morphological weights into the Transformer's attention mechanism and combines it with the temporal extraction capabilities of BiLSTM. This hybrid model significantly reduces the risk of overfitting and achieves accurate mapping of sentiment semantics to product form.

[0023] 3. This invention establishes a multimodal objective evaluation system based on eye-tracking and EEG coordination. It overcomes the "explicit" limitations of traditional questionnaire evaluations by simultaneously collecting visual attention (eye movement) and neurocognitive (ERP) signals, enabling quantitative analysis of users' implicit emotional responses. This system can objectively verify from a physiological perspective whether the design scheme evoked the expected emotional resonance, providing scientific and reproducible physiological evidence for design decisions.

[0024] 4. This invention realizes a highly efficient and automated generation process based on SDM. By utilizing generative AI technology (SDM+ControlNet), the predicted morphological sketches are quickly converted into high-quality rendered images, solving the problem of long rendering times in traditional modeling and significantly improving the efficiency of design iteration. Attached Figure Description

[0025] Figure 1 This is a flowchart of an emotion-driven drone form design method according to an embodiment of the present invention;

[0026] Figure 2 These are images of 102 UAV image samples collected in this embodiment of the invention;

[0027] Figure 3 This is a heatmap of emotion-related keywords and topic similarity in an embodiment of the present invention;

[0028] Figure 4 This is a structural diagram of the UAV morphology in an embodiment of the present invention;

[0029] Figure 5 These are heatmaps and gaze sequences of UAV eye tracking in embodiments of the present invention;

[0030] Figure 6 These are 20 UAV design sample diagrams from embodiments of the present invention;

[0031] Figure 7 These are the spectrum diagram and time-frequency analysis diagram of the EEG channel in this embodiment of the invention;

[0032] Figure 8 These are average ERP amplitude maps of different brain regions in embodiments of the present invention;

[0033] Figure 9 These are renderings of each UAV design scheme generated by Hunyuan 3D in this embodiment of the invention.

[0034] Figure 10 This is a simplified optimal UAV design diagram in an embodiment of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0037] Example 1

[0038] Please see Figure 1 This is a flowchart illustrating an emotion-driven drone morphology design method provided by an embodiment of the present invention. The method may include the following steps:

[0039] S1: Collect and process user online comment data, and use the BERTopic model to extract and filter core sentiment words.

[0040] Specifically, step S1 is implemented as follows:

[0041] Step S11: Morphological Deconstruction and Dataset Construction

[0042] Consumer-grade drone sample images were collected using web crawling technology and subjected to standardized preprocessing (background removal, perspective correction, and normalization). Based on expert focus group discussions, the drone's morphology was deconstructed into five core components (fuselage, arms, propellers, landing gear, and cameras), and a drone morphology decomposition table was constructed.

[0043] Step S12: Sentiment word extraction based on BERTopic

[0044] First, a pre-trained Sentence-BERT model is used to map each user comment to a vector E containing semantic information. Then, the UMAP algorithm is applied to transform the high-dimensional vector E into a low-dimensional space. The HDBSCAN algorithm is used for density-based clustering to discover potential topics. Finally, core sentiment words are selected by calculating class-based TF-IDF (c-TF-IDF) scores. This algorithm measures the vocabulary... In cluster The frequency of words in the c-TF-IDF index and their inverse frequency across all clusters are used to quantify the contribution of words to the sentiment topic: the c-TF-IDF score calculation formula is as follows:

[0045]

[0046] in, For words In cluster The frequency in For words Total frequency across all clusters The average number of words per cluster.

[0047] S2: Collect eye-tracking data of the user while observing sample images of the drone. Based on the eye-tracking data, use the entropy weight-TOPSIS method to calculate the objective importance weight of each part of the drone.

[0048] Specifically, step S2 is implemented as follows:

[0049] Step S21: Data Acquisition and Normalization

[0050] Using visual attention data obtained from eye-tracking experiments, the entropy-weighted TOPSIS method is applied to comprehensively evaluate the importance of different components, obtaining a ranking of the importance of morphological components. These morphological elements are then quantified as prior knowledge for constructing the prediction model. First, an original evaluation matrix is ​​constructed. ,in Indicates the first The sample at the th The original values ​​for each indicator are given, where n represents the total number of samples and m represents the total number of evaluation indicators. To calculate the information entropy, the data is normalized using a weighted average method to obtain a probability matrix. ;

[0051]

[0052] Step S22: Calculate the weights using the entropy weight method.

[0053] Calculate the first Information entropy of each indicator and their corresponding weighting coefficients :

[0054]

[0055]

[0056] Where k represents the adjustment coefficient.

[0057] Step S23: Calculate the importance of morphological components using the TOPSIS method.

[0058] To eliminate the influence of indicator dimensions on distance calculation, a standardized decision matrix is ​​constructed using vector normalization. :

[0059]

[0060] Determine the ideal solution and its approximation.

[0061] First, define the positive ideal solution. (Optimal Vector) and Negative Ideal Solution (Worst-case vector):

[0062]

[0063]

[0064] Based on entropy weight Calculate the weighted Euclidean distance between each sample and the positive and negative ideal solutions. and :

[0065]

[0066]

[0067] Finally, the relative similarity of each sample is calculated. The objective importance weight of the morphological components:

[0068]

[0069] S3: Construct and train a Transformer-BiLSTM sentiment mapping model that integrates prior weights. Integrate the objective importance weights as prior knowledge into the Transformer-BiLSTM sentiment mapping model. Input the UAV morphological feature data and the core sentiment words, and output the optimal morphological combination corresponding to each sentiment word.

[0070] Specifically, step S3 is implemented as follows:

[0071] Step S31: Feature Definition and Normalization

[0072] The original morphological feature matrix was normalized using the Min-Max normalization method. and emotional tags Standardization is performed to map the numerical range to [0, 1] to ensure training stability and convergence speed.

[0073]

[0074] in, and These represent the minimum and maximum values ​​of the feature in the dataset, respectively. and These represent the minimum and maximum values ​​of the sentiment label in the dataset, respectively.

[0075] Step S32, Multi-scale weighted fusion

[0076] To incorporate morphological component importance information into the feature input, the calculated entropy weight-TOPSIS objective importance weight vector is used. Normalized features Weighting is performed. A weighted feature matrix is ​​constructed using the Hadamard Product. This achieves multi-scale fusion of physical scale features and cognitive scale weights:

[0077]

[0078] in This represents the element-wise multiplication method (Hadamard product).

[0079] Step S32: Construction of Weighted Self-Attention Mechanism

[0080] The weighted feature sequence is input into the Transformer encoder to capture global dependencies between features. Self-attention is calculated as follows:

[0081]

[0082] in, , and These represent query, key, and value matrices, respectively. It is the dimension of the key.

[0083] To further enhance the model's focus on key components, this invention improves upon the traditional scaled dot product attention mechanism by proposing a weighted self-attention mechanism. Traditional attention mechanisms rely solely on learning weights from the data, while the improved mechanism... In the process of matrix calculation, explicit introduction Features are modulated to guide attention allocation based on prior knowledge:

[0084]

[0085] Multi-head attention output uses standard splicing mapping:

[0086]

[0087] Step S33: BiLSTM Local Feature Extraction and Prediction

[0088] The global feature sequence output by the Transformer is input into a two-layer BiLSTM network to extract local temporal features. The final feature representation of BiLSTM... From the second layer of forward hidden state State and reverse hidden state The data is pieced together and finally output through a fully connected layer to determine the predicted sentiment rating. Based on this prediction result, the optimal combination of morphological features for the drone is selected.

[0089] Step S331: LSTM cell core calculation.

[0090] For any direction of LSTM, at time The gating update can be written as:

[0091] Forgotten Gate Input gate Output gate Cell state and hidden state renew:

[0092]

[0093]

[0094]

[0095]

[0096]

[0097] Step S332, Double-layer bidirectional structure

[0098] The Transformer-BiLSTM sentiment mapping model in this step uses a two-layer BiLSTM structure:

[0099] The first layer contains 64 hidden units, which process the input sequence through both forward LSTM and forward LSTM networks. (i.e., the output of the previous layer):

[0100]

[0101] The output of this layer is composed of hidden states concatenated in two directions:

[0102]

[0103] in This represents the forward hidden state of the first layer of BiLSTM. This represents the reverse hidden state of the first layer of BiLSTM;

[0104] To perform higher-level feature abstraction, the second layer (128 hidden units) further processes the output of the first layer. The final feature representation is taken from the combination of the forward state of the second layer at the last time step and the backward state at the first time step, to preserve complete sequence information:

[0105]

[0106] S4: Using a stable diffusion model, the optimal morphological combination is transformed into a high-fidelity rendering image; for the high-fidelity rendering image, eye movement signals and electroencephalogram signals are collected simultaneously when the user views it, objective physiological indicators are extracted, and subjective scores are combined to comprehensively evaluate and rank each rendering image to determine the final drone design scheme.

[0107] Specifically, step S4 is implemented as follows:

[0108] Step S41: High-fidelity image generation based on the stable diffusion model (SDM)

[0109] The line drawing with the optimal morphological feature combination output from step S3 is transformed into a realistic rendered image using a stable diffusion model. To ensure that the generated result strictly follows the predicted morphological structure, the ControlNet-Canny edge detection module is introduced as a constraint. The specific process is as follows:

[0110] Step S411, Diffusion Process (Forward Noise Addition)

[0111] During training, noise is continuously added to the images to corrupt them; during inference, the images with added noise are denoised to restore the original images. This process is repeated during the training phase. "In the iteration, at any time..." It can be done and This means that the value at any given time can also be obtained through... and noise get:

[0112]

[0113] in, and Indicates the intensity of noise. and They represent the first The signal preservation coefficient of the second iteration and the first iteration The preset noise variance for the next iteration. The noise scheduling coefficients are used. Forward diffusion will affect the original image. Gradual perturbation Its closed form is:

[0114]

[0115] in, It's the initial, clean image. Standard Gaussian noise. Add noise. Then, an image with added noise was generated. The second formula is derived from the first formula, indicating that by subtracting a certain intensity of noise from the image after adding noise, the image is obtained. .

[0116] Step S412, Reverse Denoising

[0117] When the original sharp image is known, the backdiffusion process can be transformed into a forward process using Bayes' theorem, thus allowing the calculation of the distribution of the backdiffusion process:

[0118]

[0119] in , Let represent the mean of the posterior distribution. This represents the variance of the posterior distribution. Represents the identity matrix; this means that given At this time, the reverse process follows a Gaussian distribution, and the Bayesian formula can be used to transform the reverse process into the forward process:

[0120]

[0121] in, and They can be converted to each other, and through... By subtracting different noise levels, the mean can also be obtained.

[0122] Step S413: When fitting the distribution, the estimation results of the stable diffusion model follow a Gaussian distribution:

[0123]

[0124] Step S414: Compare the distribution of the reverse process with the predicted distribution of the model.

[0125] Since the KL divergence can be used to describe the difference between two distributions, and the KL divergence of the multivariate Gaussian distribution has a closed-form solution, it can be seen that the loss function becomes the calculation of the KL divergence between two Gaussian distributions:

[0126]

[0127] in, The distribution is the distribution of the reverse process. The distribution is the estimated distribution of the model. This represents the Körbek-Leibler divergence, abbreviated as KL divergence. This represents the learnable parameters of the neural network. The training loss is the KL divergence between two Gaussian distributions, i.e., the difference between these two distributions.

[0128] Step S42: Eye-tracking electroencephalography (ET-EEG) multimodal signal acquisition and analysis

[0129] A multimodal evaluation experiment was constructed, and eye movement trajectories and electroencephalogram (EEG) signals of users were collected simultaneously while viewing the generated scheme. Objective physiological indicators were extracted through signal processing.

[0130] Step S421: EEG Time-Frequency Analysis Based on Continuous Wavelet Transform (CWT)

[0131] Given the non-stationary nature of EEG signals, complex Morlet wavelets are used as the mother wavelet for time-frequency feature extraction. The mathematical expression for the complex Morlet wavelet is:

[0132]

[0133] in, For time, For the center frequency, The normalization constant is The bandwidth parameter of the wavelet is determined (usually by the number of cycles). For the complex sine component.

[0134] For any number of trials of EEG signal Its time and frequency wavelet transform coefficients at Obtained by convolving the signal with wavelet functions of different scales:

[0135]

[0136] Convolution results It is a complex number, and it also contains the amplitude and phase information at that moment and frequency.

[0137] Step S422, Event Correlation Spectrum Perturbation (ERSP)

[0138] The ERSP is calculated using wavelet transform coefficients to quantify the power variation relative to the baseline (i.e., the ERS / ERD phenomenon). Assumptions For the total number of trials, the first... The wavelet coefficients of each trial are The average spectral power change (in decibels dB) is calculated as follows:

[0139]

[0140] in, For the first The instantaneous power of each trial This represents the baseline average power before stimulus presentation. Positive values ​​indicate enhanced energy response (ERS), and negative values ​​indicate diminished energy response (ERD).

[0141] Step S423, Inter-trial coherence (ITC)

[0142] ITC (or Phase Lock Value, PLV) is used to measure the consistency of neuronal firing phase across different trials. It utilizes the phase angle information of wavelet coefficients. ITC is defined as:

[0143]

[0144] The value of ITC is between 0 and 1. When The time interval indicates that the EEG activity at that time point has a strict phase-locked relationship across all trials, which usually corresponds to the generation mechanism of evoked potential (ERP) components.

[0145] Step S424: Extraction of Event-Related Potentials (ERPs) and Eye-Tracking Metrics

[0146] To improve the signal-to-noise ratio, this invention divides the scalp electrodes into three regions of interest (ROIs): the forebrain, midbrain, and hindbrain, based on brain anatomy. Spatial averaging of the electrode signals within each region is performed, and a superposition averaging technique is used to extract the ERP components. Assuming... For the first The time-domain signal of EEG from N valid trials, with a total of N trials, then the average waveform of ERP. Defined as:

[0147]

[0148] Step S43: Multimodal Integrated Evaluation and Decision Making

[0149] To eliminate the bias of a single evaluation dimension, a comprehensive evaluation system was constructed, incorporating subjective scoring, visual attention (eye-tracking metrics), and cognitive emotion (EEG metrics). First, a multimodal decision matrix was constructed, and the objective weights of each physiological indicator were calculated using the entropy weight method. Second, the TOPSIS method was employed to calculate the closeness of each generated solution to the ideal solution. Based on the closeness ranking, the UAV design scheme with the best performance in visual attractiveness (high gaze duration) and emotional arousal (high P300 amplitude, low N200 amplitude) was selected, thus completing the final design decision.

[0150] Taking a drone as an example, the design process of the present invention will be further explained in conjunction with the accompanying drawings and product renderings, thereby verifying the effectiveness of the method proposed in this invention.

[0151] Step S10: Deconstruction of UAV morphological features and extraction of core sentiment vocabulary;

[0152] This invention utilizes web crawling technology, using "drone" and "aerial drone" as search keywords, to collect raw image samples from major e-commerce platforms such as JD.com and Taobao, covering multiple perspectives including frontal and 3 / 4 side views. To eliminate interference from irrelevant variables, strict standardized preprocessing was performed on the collected images, including background removal, perspective correction, and resolution normalization to 1280×1280 pixels. Subsequently, an expert focus group consisting of 3 industrial designers with more than three years of experience and 4 design graduate students was formed. Based on morphological clarity and typicality criteria, invalid data with duplication, blurring distortion, and lack of significant UAV features were eliminated. Finally, a dataset containing 102 representative UAV morphological data with uniform perspective and high resolution was selected and constructed. Figure 2 ).

[0153] This study collected 6293 valid user reviews from brands such as DJI and Sima on mainstream e-commerce platforms. In the preprocessing stage, HanLP was used with a LARGE_BERT model for word segmentation and part-of-speech filtering, preserving adjective semantics that reflect user perception of word form. In the modeling stage, Sentence-BERT was used to generate dense word vectors, which were then input into the BERTopic framework for topic clustering. By setting an adaptive number of topics (auto) and calculating c-TF-IDF weights, the study successfully extracted four core sentiment dimensions: "refined, simple, compact, and high-end." Figure 3 Figure (a) verifies the clustering accuracy of the method when processing unstructured emotional corpora. To further verify the effectiveness and independence of the extracted sentiment dimensions, this invention calculates the cosine similarity between the topic embedding vectors and generates a topic similarity heatmap, see [reference needed]. Figure 3(b) This independence verification ensures that the subsequently constructed Transformer-BiLSTM prediction model can receive diverse and mutually exclusive sentiment feature inputs, thereby effectively improving the model's ability to interpret complex user needs.

[0154] Step S20: Calculation of morphological component weights based on eye tracking and entropy weight-TOPSIS method;

[0155] Morphological deconstruction is a key method in KE (Key Opinion Framework) for uncovering the emotional characteristics of products. Following expert focus group discussions, a morphological composition system for drones was established, dividing it into five core categories: fuselage, arms, propellers, landing gear, and cameras. Figure 4 Therefore, through in-depth observation and feature extraction of 102 representative samples, a drone morphology decomposition table was constructed. This experiment utilized eye-tracking technology to capture the user's visual attention focus and cognitive path. Eight participants with drone operation experience were recruited, and their gaze heatmaps and visual scanning sequences during sample observation were collected and analyzed. In the formal experimental phase, the experimental procedure and precautions were explained to the participants (instructions), and gaze calibration and pre-testing were performed after wearing the device to ensure gaze capture accuracy. Subsequently, participants sat in front of a 27-inch monitor, focused their gaze on the center area of ​​the screen, and started the experimental program by clicking the mouse. During the visual stimulus presentation phase, drone sample images appeared every 15 seconds, with a single image display time set to 10 seconds and a 5-second buffer interval between samples. This observation sequence was repeated 8 times before the experiment automatically ended.

[0156] After the experiment, the collected data was processed using Tobii Pro Lab software, and superimposed heatmaps and visual scan sequences of eight drone samples were exported. Heatmap ( Figure 5 The distribution shows that the visual attention of the 8 participants was highly focused on the fuselage and propeller area of ​​the drone, and their observation frequency and scanning path also showed a significant convergence.

[0157] Based on the morphological deconstruction table, the samples were divided into 5 independent regions of interest. Subsequently, 7 key eye-tracking metrics were selected to quantify the user's visual perception characteristics, including: total fixation duration, average fixation duration, number of fixations, first fixation duration, total access duration, average access duration, and number of accesses. Specific statistical results are summarized in Table 1. Furthermore, to eliminate dimensional differences, the above metrics were normalized, and the comprehensive mean of each morphological category was calculated.

[0158] Table 1. Values ​​of various eye-tracking metrics within each AOI (Region of Interest).

[0159]

[0160] Five industrial designers with backgrounds in drone and intelligent aircraft design were invited to rate the impact of five drone form factors on user emotions and a Kansei mean matrix of a seven-level Likert scale was established (Table 2).

[0161] Table 2. Sentiment word rating matrix for drone components

[0162]

[0163] First, the average matrix of expert ratings was normalized to obtain a standardized matrix. Then, the entropy and weight coefficients of the drone sentiment evaluation terms were calculated, yielding the information entropy, information utility, and weight values ​​of each sentiment term. Based on this, the weighted coefficients of the sentiment terms were integrated into the TOPSIS analysis model. This invention uses five drone morphological components as evaluation objects, constructing a decision matrix for each component's four sentiment evaluation terms. Next, the attribute data of each morphological component was vector normalized, calculating the positive and negative ideal solutions for each component, and finally, the distances between each morphological component and the optimal and worst solutions were calculated (Table 3).

[0164] Table 3. Calculation results of the weights of each indicator

[0165]

[0166] By comparing eye-tracking experimental data with the entropy-weighted TOPSIS comprehensive evaluation results, this invention observes a significant consistency between users' visual perception characteristics and expert decision-making opinions regarding the attention given to key components. Specifically, the fuselage, as the core form carrier, dominates both visual dwell time and expert comprehensive scores, while auxiliary components such as landing gear have secondary weight.

[0167] This high degree of coupling between subjective and objective evaluation verifies the reliability of the extracted feature weights. Based on this, the present invention uses this set of objective importance weights, which have been verified through multimodal analysis, as prior knowledge and embeds them into the subsequent Transformer-BiLSTM generative model.

[0168] Step S30: Construction of Transformer-BiLSTM sentiment mapping model with prior weights and optimal morphology prediction;

[0169] To achieve accurate predictions of both the geometric parameters of a drone and the user's emotional perception (refined, simple, lightweight, high-end), this invention constructs a Transformer-BiLSTM hybrid prediction model that integrates prior knowledge. Traditional deep learning models often process all input features equally, making it difficult to automatically focus on key components (such as the fuselage and propellers) that have the greatest impact on user emotions under small sample conditions.

[0170] To address this, the proposed model innovatively embeds the importance weights of morphological components into the network's input layer and attention mechanism. The model's construction process mainly includes the following three key steps:

[0171] 1. Data Normalization and Prior Weight Embedding: First, Min-Max standardization is used to eliminate dimensional differences. Then, the objective morphological weight vector is embedded using the Hadamard Product. Injecting into the feature space. This operation enables multi-scale fusion of physical features and cognitive weights, effectively enhancing the saliency of key morphological regions and suppressing unnecessary noise.

[0172] 2. Weighted Transformer-BiLSTM network construction: A "weighted self-attention mechanism" is proposed, utilizing... Explicit modulation query ( AND key The matrix enhances the model's ability to capture global dependencies on key elements. A two-layer BiLSTM (64 / 128 units) is used in the backend to extract local temporal features.

[0173] 3. Regularized training and loss optimization: To suppress overfitting, the model employs a combination of mean squared error (MSE) and L2 regularization term (…). The composite loss function is composed of 0.1 and 0.2. Combined with the Adam optimizer and dynamic Dropout mechanism (0.1 / 0.2), it significantly improves the model's generalization ability and robustness in predicting complex sentiment dimensions.

[0174] This invention invited 100 participants with a design background to score 102 drone samples on four emotional dimensions: "refined, simple, lightweight, and advanced," using a 7-point Likert scale. The average scores were then used to construct an emotional evaluation matrix (Table 4), which served as the standard emotional labeling data for subsequent model training.

[0175] The constructed user sentiment evaluation matrix was imported into the Transformer-BiLSTM model for training. The dataset was divided into a training set (first 70 samples), a test set (middle 15 samples), and a validation set (last 15 samples) in sequence. The model learned the mapping relationship between morphological features and sentiment semantics through the training set and used the test set to validate the reliability of the predictions. Experimental results show that the model exhibits high goodness of fit across the four core sentiment dimensions, and the test set shows high coefficients of determination for refinement, simplicity, lightweight, and advanced aspects. The results showed that the Transformer-BiLSTM fusion model achieved prediction accuracy of 0.7902, 0.8511, 0.9413, and 0.9672, respectively. Experiments demonstrate that this model possesses extremely high prediction accuracy and can effectively establish a deep mapping relationship between morphological features and user emotions. Based on this, to comprehensively explore the relationship between morphological features and customer emotions, this invention used an exhaustive method to calculate all possible combinations of morphological features, totaling 32,768. Prediction results show that the feature combination (7,3,8,8,7) achieved the highest emotional evaluation value in the "refined" dimension; the combination (1,2,3,4,6) obtained the highest "simple" emotional score; the combination (7,2,8,7,8) performed best in the "lightweight" dimension; and the combination (7,5,8,8,8) was associated with the highest score in the "advanced" emotion dimension.

[0176] Table 4. User Sentiment Evaluation Matrix

[0177]

[0178] Step S40: Image generation and eye-tracking-EEG multimodal assessment based on SDM

[0179] Based on the prediction results output by the Transformer-BiLSTM model, the optimal combination of morphological components for the four core emotional dimensions of "refined, simple, lightweight, and sophisticated" is generated using SDM. To ensure that the generated results not only conform to the predicted morphological features but also possess high-quality visual aesthetics, a "Text-to-Image" generation strategy combined with "ControlNet" is adopted to construct a structured prompt word template, with descriptive text prompts such as "drone, emotional words (e.g., concise)" as input. The ControlNet-Canny edge detection module is used, with the line drawing of the morphological combination as a constraint input, strictly controlling the geometric contour of the generated image to be consistent with the prediction results, with a weight set to 0.8 to balance morphological accuracy and detail richness. Stable-diffusion-v1-5 is used as the base model, Euler a is selected as the sampling method, the number of iterations is set to 30, and the prompt word guidance coefficient is set to 7.5. Based on the above process, multiple design schemes are generated for the four core emotional dimensions. To ensure the diversity and representativeness of the subsequent multimodal evaluation experiments, after initial screening, five distinctive high-quality design schemes were ultimately retained for each emotional dimension, totaling 20 drone design samples. Figure 6 These schemes will serve as visual stimulation materials for subsequent eye-tracking-electroencephalography (EEG) multimodal evaluation experiments, objectively verifying the effectiveness of the prediction model and assisting in the final design decision.

[0180] Furthermore, neurophysiological measurement techniques were introduced to construct a multimodal evaluation experiment using simultaneous eye-tracking and electroencephalography (EEG) acquisition. Compared to traditional subjective questionnaires, multimodal physiological signals can more objectively reveal the user's visual attention allocation and emotional arousal mechanisms during cognitive processing. The experiment recruited 27 university students (aged 18-28) with design backgrounds or drone usage experience. =23.5, =2.1). All participants had normal or corrected visual acuity, no color blindness or color weakness, no history of neuropsychiatric disorders, and had not taken any psychotropic drugs or consumed alcohol or caffeine within 24 hours prior to the experiment. Twenty high-fidelity renderings of drones were selected as the stimulus materials for the experimental group. To eliminate extraneous variables, all images underwent standardized preprocessing: the resolution was uniformly adjusted to 2480×2480 pixels, background noise was removed (uniform white background), and the main viewpoint of the drones (e.g., 45° side view) was kept consistent to ensure that the differences in induced physiological signals stemmed solely from the morphological features themselves.

[0181] This invention employs the Emotiv Flex 2 Gel wireless EEG acquisition system. This device is equipped with 32-channel wet electrodes, with electrode layout strictly adhering to the international 10-20 system standard. The data acquisition software uses the EmotivPRO platform to record raw EEG signals in real time and monitor electrode impedance to ensure data quality. The eye-tracking system uses the Tobii Pro Fusion screen-based eye tracker, with a sampling rate set to 250 Hz. This device has extremely high temporal resolution, enabling precise capture of microsaccades and fixation trajectories. The experimental program was written using E-Prime 3.0 and ran on a 23.8-inch high-resolution display, sending synchronization triggers to the EEG and eye-tracking devices via a parallel port to ensure temporal alignment of multimodal data.

[0182] This experiment employed a single-factor experimental design (drone design group and blank control group). Experimental data consisted of behavioral data (subjects' preference ratings of stimulus samples) and physiological data (electroencephalogram and eye-tracking data). Furthermore, a seventh-order Likert scale was used to collect subjective ratings from subjects for each drone design scheme.

[0183] Experimental Procedure. (1) The screen displays the experimental instructions. After fully understanding the task requirements, the subject presses the space bar to start the formal experimental procedure. (2) Each trial begins with a 1000 ms "+" fixation point to guide the subject's gaze to center and establish a baseline. Subsequently, the system randomly presents a drone design image from the generated scheme set as a visual stimulus, with a presentation duration set to 8000 ms. (3) After the stimulus disappears, the screen enters a 4000 ms blank page to eliminate visual persistence from the previous round and separate motion artifacts. Then, a scoring interface is presented. The subject needs to press a button to score the image preference conformity of the scheme based on a 7-point Likert scale. The system records the subject's behavioral data simultaneously.

[0184] Power spectral density (PSD) analysis was performed on the preprocessed EEG signals, and spectrograms of all channels were plotted in the range of 0–25 Hz. Figure 7As shown in (a) of the figure, the curves show that the EEG signal energy is mainly concentrated in the low-frequency band (Delta / Theta band) and the Alpha band around 10 Hz. Further observation of brain topography at different frequencies revealed that at 10.0 Hz (Alpha band) and 6.0 Hz, the energy is significantly concentrated in the occipital lobe region at the back of the scalp. This indicates that during cognitive activities, the occipital lobe brain region associated with visual processing is highly activated and has a large power distribution. In-depth time-frequency domain analysis was performed on the P3 channel of the key electrode located in the parietal / occipital lobe region, and the results are as follows: Figure 7 As shown in (b) of the figure, the analysis included event-related spectrum perturbation (ERSP), inter-trial coherence (ITC), and event-related potential (ERP) waveforms. Immediately after stimulus presentation (0 ms), the EEG signal exhibited significant energy changes. Particularly within the 200 ms to 400 ms time window, a significant event-related synchronization (ERS) phenomenon was observed in the 20 Hz to 40 Hz (Beta / Gamma band) range (highlighted in red in the figure), indicating that the brain underwent intensive cognitive processing and attentional resource allocation during this period. The ITC heatmap in the middle shows strong phase coherence (phase-locked phenomenon) in a specific frequency band around 300 ms after stimulus presentation. This coincides with the corresponding ERP waveform (black curve) below, where a significant peak shift (corresponding to cognitive potential components) was observed between 300 ms and 400 ms, confirming that the visual stimulus induced a stable and significant cognitive neural response.

[0185] Based on brain functional regions and experimental data characteristics, electrode channels were divided into three categories for comprehensive analysis: the anterior brain regions (Fp1, Fp2, Fz, F3, F4, FC1, FC2), the midbrain regions (C1, C2, C3, C4, Cz, CP1, CP2), and the posterior brain regions (P3, P4, P7, P8, Pz, O1, O2). Data from all subjects were superimposed to plot the average ERP waveform for each of the three brain regions. Figure 8 Significant N200 and P300 components are present in the anterior, middle, and posterior regions of the brain. Significant P200 components are present in the anterior and posterior regions of the brain.

[0186] This paper conducts a transparent and reproducible comprehensive evaluation and ranking of UAV design schemes based on multimodal evidence (questionnaires, eye tracking, and ERP). Considering the differences in dimensions and subjective disputes regarding weights among multiple indicators, this paper adopts the "Entropy Weight-TOPSIS" multi-indicator decision-making method to rank the design schemes. To ensure the simplicity, interpretability, and avoidance of redundant scoring in the fusion ranking, this paper retains only three representative eye tracking indicators: total fixation duration, total visit duration, and number of fixations. First, a decision matrix is ​​constructed. To reduce subjective weighting disputes, the entropy weight method is used to estimate the weights based on the indicator dispersion (Table 5). After obtaining the weights... Subsequently, the ideal and negative ideal solutions of TOPSIS were obtained, and a comprehensive evaluation was performed according to the classic TOPSIS (Table 6). Based on the calculation results, the sample rankings were categorized by sentiment words. Sample 4 was correlated with the highest score for the "simple" sentiment. Sample 11 achieved the highest ranking for "compact". Sample 5 performed best in the "refined" sentiment dimension. Sample 17 ranked highest among the "advanced" sentiment words.

[0187] Sample 4's "simplicity" stems primarily from the refined form and details: the body uses a continuous, integrated curved shell with a clear outline and simplified block hierarchy; surface decorations and dividing lines are minimal, maintaining overall unity and order; the color scheme is predominantly white, with black used for functional differentiation and supplemented by small-scale embellishments, creating a low-color, low-contrast visual expression. Sample 17's body uses a relatively complete integrated rectangular block with rounded corners, resulting in a simple yet stable outline; the saturated metallic gray color scheme combined with cool-toned light strips successfully creates a highly recognizable technological feel. Sample 11 compresses the volume into a small and compact core body, while the four arms extend outward in the form of a slender frame, creating ample white space and giving the overall design a stronger sense of transparency and suspension; Sample 5 has a compact overall shape with uniform rounded corners, appearing visually refined; the black functional window at the front transitions naturally with the body, presenting a neat and complete appearance.

[0188] Based on the 2D concept image generated from the SDM model, Hunyuan 3D technology was introduced to achieve rapid construction of the 3D model. The resulting image is shown below. Figure 9 As shown, this automated path not only significantly shortens the iteration cycle of model verification and adjustment, but also effectively reduces the technical threshold and labor costs of complex surface modeling, empowering designers to participate in design decisions more efficiently. As a typical case, Figure 10 It showcases the optimal design solution and its 3D generation results under the "simplicity" emotional dimension.

[0189] Table 5. Calculation results of entropy weight method

[0190]

[0191] Table 6. TOPSIS Complete Sorting

[0192]

[0193] Optionally, this application embodiment also provides an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the various processes of the above-described embodiment of an emotion-driven drone form design method and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0194] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiment of an emotion-driven drone form design method and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0195] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0196] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0197] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0198] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. An emotion-driven method for designing the form of a drone, characterized in that, Includes the following steps: S1: Collect and process user online comment data, and use the BERTopic model to extract and filter core sentiment words; S2: Collect eye-tracking data of the user while observing sample images of the drone. Based on the eye-tracking data, use the entropy weight-TOPSIS method to calculate the objective importance weight of each part of the drone. Specifically, this includes: acquiring eye-tracking index data on preset morphological components when a user observes a sample image of a drone, and constructing an original evaluation matrix X; normalizing the original evaluation matrix X using a weighted method to obtain a probability matrix P; and calculating the information entropy of the j-th eye-tracking index for each morphological component. and their corresponding weighting coefficients Construct a weighted TOPSIS decision matrix and normalize it into a standardized decision matrix. In order to determine the ideal solution With negative ideal solution Based on the weighting coefficients Calculate the sample of each morphological component and the positive ideal solution. and negative ideal solution Weighted Euclidean distance and ; Calculate the relative closeness of samples of each morphological component. and the relative closeness The objective importance weight of the aforementioned morphological components is output, and the calculation formula is as follows: Where m represents the total number of drone morphological component samples participating in the evaluation; S3: Construct and train a Transformer-BiLSTM sentiment mapping model that integrates prior weights, incorporate the objective importance weights as prior knowledge into the Transformer-BiLSTM sentiment mapping model, input the UAV morphological feature data and the core sentiment words, and output the optimal morphological feature combination corresponding to each sentiment word. S4: Using a stable diffusion model, the optimal combination of morphological features is transformed into a high-fidelity rendering image; for the high-fidelity rendering image, eye movement signals and electroencephalogram signals are collected simultaneously when the user views it, objective physiological indicators are extracted, and subjective scores are combined to comprehensively evaluate and rank each rendering image to determine the final drone design scheme. The specific method for extracting objective physiological indicators is as follows: the scalp electrodes are divided into three regions of interest: the forebrain, midbrain, and hindbrain; the electrode signals in each region of interest are spatially averaged, and the event-related potential (ERP) components are extracted using a superposition averaging technique. The average waveform of the ERP is... The calculation formula is: Where N is the total number of trials, The time-domain signal of the EEG in the kth valid trial is used; the N200 and P300 component amplitudes are extracted from the average waveform of the event-related potential (ERP), and combined with the total fixation duration, total access duration and fixation count indicators extracted from the eye movement signal to form and output the objective physiological indicators. The specific method for comprehensively evaluating and ranking the renderings is as follows: construct a multimodal decision matrix that includes subjective scores and the objective physiological indicators; calculate the objective weights of the subjective scores and each objective physiological indicator using the entropy weight method; calculate the relative closeness between each rendering and the ideal solution using the TOPSIS method; rank all high-fidelity renderings according to the relative closeness, and determine the design scheme corresponding to the highest-ranked rendering as the final UAV design scheme.

2. The method according to claim 1, characterized in that, The specific implementation method of step S1 is as follows: The Sentence-BERT model is used to map each user comment data into a vector E containing semantic information; The UMAP algorithm is applied to transform the vector E from a high-dimensional space to a low-dimensional space; The HDBSCAN algorithm is used to perform density-based clustering on the dimensionality-reduced vectors to discover potential sentiment themes; The core sentiment words are selected by calculating class-based TF-IDF scores. The formula for calculating the class-based TF-IDF score is as follows: ； in, For words In cluster The frequency in For words Total frequency across all clusters The average number of words per cluster.

3. The method according to claim 1, characterized in that, The specific method for constructing and training the Transformer-BiLSTM sentiment mapping model that integrates prior weights in step S3 is as follows: The objective importance weight vector is represented by the Hadamard product. With the normalized UAV morphological feature matrix By performing element-wise multiplication, a weighted feature matrix is ​​constructed. ; The weighted feature matrix The input is fed into a Transformer encoder, where a weighted self-attention mechanism captures the global dependencies between features. This weighted self-attention mechanism introduces the objective importance weight vector during the calculation of the query matrix Q and the key matrix K. Modulation; The global feature sequence output by the Transformer encoder is input into a two-layer BiLSTM network to extract local temporal features, wherein the final feature representation of BiLSTM is... From the second layer of forward hidden state and reverse hidden state It is pieced together; The predicted sentiment rating is output through a fully connected layer, and the optimal combination of morphological features is selected and output based on the prediction results.

4. The method according to claim 1, characterized in that, In step S4, the specific implementation method for converting the optimal morphological feature combination into a high-fidelity rendering image is as follows: The line drawing of the optimal combination of morphological features is used as a constraint and input into the ControlNet-Canny edge detection module; A stable diffusion model is used to generate an image from the input text prompt description under the constraints of the ControlNet-Canny edge detection module. The forward diffusion process of the stable diffusion model converts the original image... Gradually perturb to noisy image Its closed form is: ； in, For noise dispatch coefficient, Standard Gaussian noise; The model's reverse denoising process estimates the denoising result by fitting a Gaussian distribution and uses the KL divergence between the two Gaussian distributions as the training loss, ultimately outputting the high-fidelity rendered image.

5. The method according to claim 1, characterized in that, In step S4, the method further includes performing time-frequency analysis on the acquired electroencephalogram (EEG) signals, specifically as follows: Complex Morlet wavelets are used as the mother wavelet for continuous wavelet transform to extract time-frequency features. The mathematical expression of the complex Morlet wavelet is as follows: ； in, For time, For the center frequency, The normalization constant is The bandwidth parameter of a wavelet is determined by the number of periods. For complex sine components; The event-related spectral perturbation ERSP is calculated based on wavelet transform coefficients to quantify the power variation relative to the baseline; The inter-trial coherence (ITC) is calculated based on the phase angle information of the wavelet transform coefficients to measure the consistency of the neuronal firing phase across different trials. Output the results of Event Correlation Spectrum Perturbation (ERSP) and Inter-trial Coherence (ITC) analysis.

6. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein when the program or instructions are executed by the processor, they implement the steps of an emotion-driven drone form design method as described in any one of claims 1-5.

7. A readable storage medium, characterized in that, The program or instructions are stored on the readable storage medium, and when the program or instructions are executed by the processor, they implement the steps of the emotion-driven drone morphology design method as described in any one of claims 1-5.

Images