A SENTIMENT ANALYSIS MODEL FOR ELECTROENCEPHALOGRAM SIGNALS OF STUDENTS IN UNIVERSITIES USING A CONVOLUTIONAL NEURAL NETWORK AND SUPPORT VECTOR MACHINE MODELS
Abstract
Sentiment analysis in teaching evaluation has significant implications. By analyzing students’ sentiments toward instructors, educational institutions can gain valuable insights into teaching effectiveness. These data can guide curriculum development, instructional improvements, and faculty training initiatives. Positive sentiment indicates effective teaching methods, engagement, and student satisfaction; negative sentiment flags areas that need attention. Sentiment analysis can help identify patterns, trends, and outliers, aiding in targeted interventions and personalized support. It also enables comparisons across different courses, instructors, and departments. However, it is crucial to ensure the accuracy and fairness of sentiment analysis algorithms, considering potential biases and the contextual nature of the feedback. This study proposes a sentiment classification model CNN–SVM that combines a convolutional neural network (CNN) and a support vector machine (SVM). Taking students majoring in art in comprehensive colleges and universities as the research object, by collecting the electroencephalogram (EEG) signals of students during teaching evaluation. CNN–SVM is used as the emotional analysis model to obtain the emotional analysis of teaching evaluation results. EEG is a typical physiological signal, and data based on this signal can more truly reflect student emotions. The adaptive CNN feature extraction function and the super generalization classification performance of SVM can reduce the individual differences and data noise between data, thereby improving sentiment classification performance. The experimental results demonstrate that using technology to analyze sentiment can assist educational institutions in more properly comprehending the feedback and opinions of students on instruction. With regard to sentiment analysis, the CNN–SVM method that is derived to produce the fusion algorithm has solid performance.
1. Introduction
Teaching evaluation is a crucial activity in education, playing a significant role in improving teaching quality, meeting student needs, and fostering professional development for teachers.1,2 Teaching evaluation helps teachers understand their instructional effectiveness, identify their strengths and weaknesses, and make instructional improvements based on this feedback. Simultaneously, teaching evaluation serves as an important avenue for student engagement in the educational process, as their feedback helps teachers better understand their learning experiences and needs, thereby enabling adjustments in teaching strategies and content. Sentiment analysis is a method that utilizes natural language processing and machine learning techniques to identify and understand emotions and sentiments expressed in text.3,4 By analyzing factors such as emotional vocabulary, tone, and context, sentiment analysis determines the emotional states, such as positive, negative, or neutral, conveyed in the text. In the context of teaching evaluation, sentiment analysis can be combined with evaluation results to extract additional information about students’ emotional experiences and states. There exists a close relationship between teaching evaluation and sentiment analysis. Sentiment analysis provides a more comprehensive and in-depth understanding of teaching evaluation. By analyzing the emotions expressed in student evaluations, teachers can better comprehend students’ attitudes and emotional responses toward instructional activities, gaining further insight into their levels of satisfaction, optimism, doubts, or concerns about the instructional content. This emotional information can assist teachers in adjusting their teaching methods and strategies to better meet students’ needs and enhance teaching effectiveness. In summary, teaching evaluation and sentiment analysis are interconnected and mutually beneficial in the education field. Sentiment analysis offers additional dimensions and depth to teaching evaluation, helping teachers gain a better understanding of students’ emotional experiences and states. This, in turn, facilitates instructional strategy improvements, meets student needs, and enhances overall teaching quality.
The data used in sentiment analysis include text data, audio data, and image data. Text data are a common type of data for sentiment analysis, as text can be analyzed for vocabulary, tone, and context to infer emotional states. Audio data can also be analyzed for features such as pitch, tone, and speech rate to infer emotional states. Image data utilize computer vision techniques to analyze facial expressions and body language, among other features, to infer emotional states. In the context of sentiment analysis, electroencephalogram (EEG) holds an important position as a type of biological data.5,6,7 EEG measures electrical activity on the scalp and reflects changes in neural activity. Compared to other types of data, EEG offers several advantages in sentiment analysis. First, EEG directly measures brain activity, providing direct observation of emotional processing rather than relying on external observations or subjective reports. Second, EEG has a high temporal resolution, capturing millisecond-level changes in brain electrical activity. This enables EEG to provide rapid and dynamic information about emotional responses, helping to study temporal characteristics and emotion sequences. Third, EEG can provide multiple EEG features, such as EEG spectra and event-related potentials (ERPs), which can be correlated with emotional states. For example, enhancements or suppressions in specific frequency bands can be associated with certain emotional states. Fourth, compared to other biological data (such as facial expressions), EEG exhibits relatively low interindividual variability. This means that EEG can provide more stable and consistent results in emotion recognition without being heavily influenced by individual differences. In conclusion, EEG possesses unique advantages in sentiment analysis, providing direct brain activity information and high temporal resolution, which aids in studying the dynamic processes and neural mechanisms of emotions. It offers a reliable and objective biological data source for sentiment analysis, facilitating a deeper understanding of emotional states and emotional processing.8,9
Emotion analysis based on EEG is a technique that utilizes brainwave signals to identify and understand individual emotional states. By analyzing the spectral and temporal features of EEG signals, it is possible to infer the brain activity patterns associated with specific emotional states, thus achieving the goal of emotion analysis. First, emotion detection based on EEG is an important application area. By monitoring an individual’s brainwave activity, it is possible to identify and classify different emotional states, such as happiness, sadness, and anger.10,11 This is crucial for diagnosing psychological disorders, emotion regulation, and mental health management. Second, EEG-based emotion analysis can be applied to evaluate users’ emotional responses to products, services, or virtual experiences. By monitoring a user’s brainwave activity while using products or engaging in virtual experiences, it is possible to understand their emotional experiences and cognitive load, thus aiding in the design of products and experiences that better meet user needs.12,13 Third, EEG-based emotion analysis can be used to assess the effectiveness and impact of advertisements. By monitoring brainwave activity while watching advertisements, it is possible to understand the audience’s emotional responses and cognitive processes. This helps advertisers optimize ad creativity, increasing its appeal and influence.14,15 Fourth, EEG-based emotion analysis can be applied in the field of education and training, helping educators understand students’ emotional states and learning experiences. By monitoring students’ brainwave activity during the learning process, it is possible to identify their interests, attention, and cognitive load, thus providing personalized learning support and feedback.16,17 Fifth, EEG-based emotion analysis can also be used in disease diagnosis and treatment. In conclusion, EEG-based emotion analysis has wide-ranging applications in emotion detection, user experience evaluation, advertisement effectiveness assessment, education and training, and disease diagnosis and treatment.18,19 By analyzing patterns and features of individual brainwave activity, valuable information about emotional states, cognitive processes, and mental health can be obtained, providing decision-makers and professionals with deeper insights and support. This work, inspired by the research above, uses sentiment analysis to improve teaching assessment results. This will assist educational institutions in comprehending student input and improving teaching quality. This paper proposes CNN–SVM sentiment analysis by examining comprehensive university art majors and this field’s teaching assessments. The trials show that the model used in this research can categorize attitudes more accurately, which has real-world relevance.
2. Related Information
2.1. Sentiment analysis method
Currently, the most common types of techniques for analyzing sentiment are those that are based on sentiment lexicons20 and machine learning.21 The first step is to segment the text and extract keywords, the second step is to use the sentiment dictionary to calculate the sentiment value of the text, and the third step is to use the overall sentiment value as the foundation for determining the sentiment tendency of the text. Among these methods, the first is the most common. Figure 1 depicts the basic idea behind this.

Fig. 1. Principle of sentiment analysis based on a dictionary.
The accuracy of sentiment analysis based on a lexicon significantly depends on the accuracy of the sentiment lexicon. However, the quality of a sentiment lexicon is significantly affected by background knowledge, and it is typically difficult to develop a high-quality sentiment lexicon from scratch. In addition, this approach treats the text as if it were only a collection of words; as a result, it disregards the information regarding the order, grammar, and syntax of the words. However, it is unable to convey the semantic information contained within the text in its entirety. Because machine learning-based sentiment analysis does not rely on dictionaries, its accuracy is far higher than that of traditional approaches. Figure 2 illustrates the fundamental idea of sentiment analysis, which is based on machine learning.

Fig. 2. Principles of sentiment analysis based on machine learning.
Although this strategy is straightforward and useful, it places excessive reliance on feature engineering. The words used in the remarks are not merely a superposition of other words, and the sentiments conveyed by various grammatical patterns are entirely different from one another. Sentiment classification in complicated sentences is not ideal, and the robustness of the classification model is inadequate. Word-level features are not accurate enough to describe the context of the text. Moreover, the categorization of sentiments in simple phrases is not perfect.
2.2. CNN
The only thing that the convolutional neural network (CNN) input is related to is the output of the neurons in the layer below it. Currently, CNNs are utilized for the processing of network structure data, particularly in image recognition.22 Figure 3 provides an illustration of the construction of CNN. Its sparse interaction and the fact that it shares parameters make its feature extraction function its most significant achievement. This function is defined by the nature of the system. The concept of sparse interaction proposes that the image’s local features should first be learned using the convolution kernel operation, and then the image’s local features should be integrated using the fully connected layer to produce a global feature. This is the basic theory behind sparse interaction. The convolution operation has a property known as parameter sharing, and the parameters of the convolution kernel have a trait known as translation equivariance. Both of these properties are maintained throughout the movement of the convolution operation.

Fig. 3. CNN structure diagram.
Feature extraction requires the convolution layer. The convolution kernel convolved with the input picture from the layer below generates feature maps. The convolution process involves a weighted summation of pixel information and continuous convolution operations. The convolution kernel’s matching weight value indicates the domain points’ contribution to the output. Convolution can highlight the original signal’s qualities, reducing noise interference.
Downsampling, such as secondary extraction of feature maps and data dimensionality reduction, is important to pooling. Pooling is feature invariant and prioritizes image quality over location. Most pooling operations do not create new parameters, which reduces spatial dimensionality, reduces computing, and prevents overfitting. Downsampling the local receptive field of the layer before it makes the network more resistant to input distortion. Regional features drive pooling. Despite the output feature map being much smaller than the original, the total number of feature maps will not change. Most pooling processes are average, maximum, or random. Maximum pooling obtains the maximum feature value from the pooled rectangular frame. This reduces the average value divergence caused by network parameter inaccuracy and helps extract texture information.
The fully connected layer transforms all hidden layer characteristics into a one-dimensional feature vector for its input. The fully linked layer also connects all nodes from the previous layer to the next. The completely linked layer has end-to-end outputs for input and output. The fully connected structure illustrates how output categories and convolutional layer features correspond one-to-one. The softmax function activates the fully linked layer. This function requires that the output layer nodes match the total classifications. The softmax function translates the anticipated value of the fully linked layer into a probability between 0 and 1.
2.3. SVM
Support vector machine (SVM) is one of the most classic algorithms among intelligent algorithms. It is a type of supervised learning algorithm that separates support vectors using planes. The plane used for separation is referred to as the optimal hyperplane. As shown in Fig. 4, the support vectors comprise several sample points, the condition of which is that they are located in close proximity to the separation plane. Finding the hyperplane that produces the greatest distance between the SVM and the hyperplane is the primary goal of the SVM algorithm. This is accomplished by locating the optimal hyperplane. This distance is referred to as the margin in SVM, and the boundary that offers the greatest possible distance for the support vector is referred to as the ideal hyperplane. The ideal hyperplane guarantees that the model will fit the data and be robust.

Fig. 4. SVM schematic diagram.
3. Sentiment Analysis Model Based on Teaching Evaluation
3.1. Teaching evaluation sentiment analysis process
Figure 5 depicts the procedure for conducting sentiment analysis using the CNN–SVM model as the basis. The graphic clearly demonstrates that it can be broken down into three distinct sections at a glance. The first step is the acquisition of data and the subsequent preparation of that data. Second, the data that have been preprocessed are fed into the CNN so that the sentiment data features can be extracted. After the features have been retrieved, they are fed into an SVM classifier, and the final result of sentiment classification is then output.

Fig. 5. Process for sentiment analysis of teaching evaluation.
3.2. CNN–SVM model
In this study, CNN is utilized to extract the sentimental data, and SVM is chosen to serve as the classifier for the CNN–SVM model that is ultimately derived. Both the CNN adaptive feature extraction function and the SVM super generalization classification performance have the ability to reduce the individual differences and data noise between data while simultaneously increasing the differences between the various categories. Figure 6 provides an illustration of the CNN–SVM model structure. The characteristics of the fully connected layer are extracted after the samples have been processed through the six-layer CNN structure, and then they are delivered to the SVM classifier.

Fig. 6. CNN–SVM model structure.
The CNN model can be used as a classifier on its own, while the softmax method is employed in the output layer of the network to make decisions on classification. In this study, a CNN is the only method utilized for feature extraction, and the following explanation will justify why an SVM was chosen for classification: Any sample can have an effect on the classification method used by the CNN’s softmax classifier, which requires all of the output data to participate in decision-making. The hyperplane of an SVM is the only thing that depends on the support vector, and the number of samples is kept to a minimum. Both of these factors contribute to the model’s ability to classify data quickly and accurately. It should be noted that the objective of SVM optimization is not to reduce empirical risk but rather to decrease structural risk. It is possible to improve the model’s capability for generalization by basing it on design aspects that pose the least risk to the structure. As the objective function of the SVM is a convex function, solving the optimization problem posed by the SVM is a convex optimization problem. Because of this, it is possible to find a global optimal solution rather than a local optimal solution.
When performing sentiment analysis, the gathered sentiment dataset is first split into two parts in proportion: the training set and the test set. These two parts are named. To fit the model, the general training set is utilized. The validation set can be used to modify the hyperparameters of the model, while the test set is used to assess the model’s capacity for generalization. In most cases, the CNN model is utilized as a feature extractor to automatically extract features, while the SVM model is utilized in the role of a feature classifier. The SVM classifier receives the features extracted by the CNN to begin the training process. Following the completion of the SVM model classification, the classification result is then used as the final classification result. Figure 7 provides an illustration of the training procedure for the CNN–SVM model.

Fig. 7. CNN–SVM model training process.
4. Experimental Design and Results Analysis
4.1. Dataset introduction
In this paper, two experiments are designed to analyze the performance of the algorithm used. First, we compared the classification performance of BP,23 SVM,24 PSO–SVM,25 CNN,26 and CNN–SVM on the public DEAP dataset. Public dataset experiments can more objectively compare the performance of the models used in this paper. The DEAP dataset includes data from 32 participants (22 males and 10 females) aged between 19 and 37 years. The dataset was collected using a variety of physiological sensors, including EEG, galvanic skin response (GSR), peripheral skin temperature, blood volume pulse (BVP), and respiration. EEG signals were recorded from 32 electrodes placed on the participants’ scalps. Second, the subject of this paper is sentiment analysis of teaching evaluation. By collecting homemade teaching evaluation datasets, each model is used for sentiment analysis. The evaluation index used in the experiment is the classification accuracy, and the parameters of each model are the parameters set in the references. The parameter settings of CNN–SVM are shown in Table 1. The activation function is selected as ReLU.
Convolutional layers | Number of cores | Kernel size | Step size |
---|---|---|---|
Convolution_1 | 32 | (6,6) | (1,1) |
Max_pooling_1 | 32 | (2,2) | (2,2) |
Convolution_2 | 32 | (6,6) | (1,1) |
Max_pooling_2 | 32 | (2,2) | (2,2) |
Convolution_3 | 64 | (3,3) | (1,1) |
Max_pooling_3 | 64 | (2,2) | (2,2) |
Convolution_4 | 64 | (3,3) | (1,1) |
Max_pooling_4 | 64 | (2,2) | (2,2) |
Convolution_5 | 128 | (3,3) | (1,1) |
Max_pooling_5 | 128 | (2,2) | (2,2) |
Convolution_6 | 128 | (5,5) | (1,1) |
Max_pooling_6 | 128 | (2,2) | (2,2) |
4.2. Model performance verification experiment
To verify the classification performance of the algorithm used in this paper, this paper conducts experiments on the DEAP public dataset mentioned earlier. The number of categories is 4, which are valence, arousal, dominance, and liking. The experimental results are the average of 10 experiments. The specific experimental results are shown in Table 2.
Model classification criteria | Potency | Arousal | Dominance | Like degree | |
---|---|---|---|---|---|
BP | Mean | 0.8243 | 0.8381 | 0.8586 | 0.8449 |
Std | 0.0452 | 0.0485 | 0.0503 | 0.0587 | |
SVM | Mean | 0.8320 | 0.8581 | 0.8676 | 0.8605 |
Std | 0.0459 | 0.0486 | 0.0428 | 0.0490 | |
PSO–SVM | Mean | 0.8426 | 0.8643 | 0.8779 | 0.8834 |
Std | 0.0378 | 0.0401 | 0.0375 | 0.0326 | |
CNN | Mean | 0.8975 | 0.9007 | 0.9054 | 0.8999 |
Std | 0.0279 | 0.0310 | 0.0278 | 0.0287 | |
CNN–SVM | Mean | 0.9242 | 0.9288 | 0.9305 | 0.9328 |
Std | 0.0265 | 0.0247 | 0.0297 | 0.0256 |
According to the experimental results, the following three conclusions can be drawn:
(1) | Compared with various classification models, the sentiment classification results obtained by the CNN–SVM model used in this paper are the best, and the classification accuracy obtained in the four classification standards of valence, arousal, dominance, and likability all exceeds 0.9. This fully validates the high performance of the used model. In addition, the standard deviations obtained based on this model are all within 0.03, which is lower than those based on the other models. This shows that the model used in this paper is relatively more stable. | ||||
(2) | BP, SVM, and OSP–SVM are all machine learning algorithms. The results obtained by these three models are generally lower than those of the latter two models, which shows that in the classification task of the DEAP dataset, the performance of machine learning algorithms is lower than that of deep learning algorithms. Machine learning algorithms tend to underperform compared to deep learning algorithms in EEG classification due to several factors. EEG signals are complex and nonlinear, requiring the ability to capture intricate patterns and relationships. Deep learning models excel in this regard by automatically learning hierarchical representations and extracting meaningful features from raw EEG data. Traditional machine learning algorithms often rely on manual feature engineering, which is challenging given the high dimensionality and complexity of EEG signals. Deep learning models also offer scalability and flexibility, handling large volumes of data efficiently through parallel computing and distributed training. They excel in capturing temporal and spatial dependencies inherent in EEG signals, utilizing specialized architectures such as recurrent and CNNs. Additionally, deep learning enables end-to-end learning, optimizing all layers simultaneously, while traditional machine learning algorithms require separate preprocessing and feature extraction steps. Last, deep learning benefits from large-scale training on publicly available EEG datasets, allowing them to learn diverse patterns and generalize better. Overall, the inherent capabilities of deep learning models make them more suitable for accurate EEG classification compared to traditional machine learning algorithms. | ||||
(3) | In the deep learning model, the classification accuracy obtained by CNN is slightly lower than that of CNN–SVM, which shows that this paper uses CNN for feature extraction and SVM as a classifier can improve the classification performance of the model. CNNs are well suited for analyzing spatial dependencies in data, and EEG signals exhibit spatial characteristics due to multiple electrode channels. By leveraging convolutional layers, CNNs can effectively capture local patterns and spatial information present in EEG signals. The hierarchical nature of CNN architectures allows them to learn increasingly complex features, enabling the extraction of discriminative representations from the raw EEG data. These learned features serve as meaningful representations that enhance the discriminative power of the SVM classifier. SVMs, known for their ability to find optimal decision boundaries, can effectively leverage the informative features extracted by the CNN to classify EEG data accurately. The combination of the CNN’s feature extraction capabilities and the SVM’s classification process enables the model to capture the underlying patterns and variations in the EEG signals, leading to improved classification results compared to using either method independently. |
4.3. Teaching evaluation sentiment analysis experiment
To analyze the sentiment recognition effect of the model used on teaching evaluation, we first collected 32,000 pieces of evaluation data from art majors in a certain comprehensive university on teachers. Teachers are given an overall quantitative score on a 10-point scale. After data cleaning, data preprocessing is performed on the original comment data. In the preprocessing stage, each piece of data needs to be sentimentally labeled, and those with a score of less than 4 are marked as negative, those with a score greater than or equal to 7 are marked as positive, and the others are marked as neutral. The preprocessed dataset is shown in Table 3.
Sentimental label | Teaching attitude | Teaching content | Teaching method | Teaching effect |
---|---|---|---|---|
Negative | 2,574 | 2,210 | 2,038 | 2,187 |
Neutral | 2,632 | 2,587 | 2,384 | 2,732 |
Positive | 2,893 | 2,867 | 3,026 | 2,868 |
Total | 8,099 | 7,664 | 7,448 | 7,787 |
To verify the effectiveness of the model used in the sentiment analysis of student teaching evaluation, approximately 15% of the comments about the same teacher in the dataset are used as the verification set, and the rest of the data are used as the training set for experiments. The experimental results are the average of 10 experiments. The classification accuracy of each model on the dataset is shown in Table 4.
Model classification criteria | Teaching attitude | Teaching content | Teaching method | Teaching effect | Total | |
---|---|---|---|---|---|---|
BP | Mean | 0.7432 | 0.7628 | 0.7356 | 0.7643 | 0.7515 |
Std | 0.0345 | 0.0302 | 0.0365 | 0.0312 | 0.0331 | |
SVM | Mean | 0.7375 | 0.7586 | 0.7389 | 0.7587 | 0.7484 |
Std | 0.0273 | 0.0300 | 0.0263 | 0.0278 | 0.0279 | |
PSO–SVM | Mean | 0.7561 | 0.7508 | 0.7434 | 0.7786 | 0.7572 |
Std | 0.0301 | 0.0287 | 0.0265 | 0.0258 | 0.0278 | |
CNN | Mean | 0.7878 | 0.7776 | 0.7679 | 0.7865 | 0.7800 |
Std | 0.0283 | 0.0252 | 0.0263 | 0.0238 | 0.0259 | |
CNN-SVM | Mean | 0.8006 | 0.7987 | 0.7889 | 0.7905 | 0.7947 |
Std | 0.0233 | 0.0253 | 0.0262 | 0.0256 | 0.0251 |
According to the experimental results, the following two conclusions can be drawn:
(1) | The results obtained by each model in the teaching evaluation data are basically consistent with the public DEAP dataset sentiment analysis results. Overall, the sentiment classification results based on machine learning algorithms are worse than those based on deep learning algorithms. The primary reason for the inferior performance of machine learning algorithms compared to deep learning algorithms in classification tasks can be attributed to the ability of deep learning models to automatically extract complex and abstract features from raw data. Machine learning algorithms typically rely on manually engineered features, which can be limited in capturing intricate patterns and relationships in the data. In contrast, deep learning models utilize multiple layers of interconnected neurons, enabling them to learn hierarchical data representations. This hierarchical representation allows deep learning models to effectively capture both low-level and high-level features, leading to improved classification accuracy and generalization capabilities. | ||||
(2) | The CNN–SVM model used in this paper has the highest sentiment classification accuracy. The reasons are as follows. One advantage of using SVM as a classifier in classification tasks is its ability to handle high-dimensional feature spaces effectively. While CNNs excel at feature extraction, SVMs can efficiently process these extracted features and make accurate predictions. SVMs utilize a margin-based approach, seeking to find an optimal decision boundary that maximizes the separation between different classes. This results in robust classification performance, even with limited training data. Moreover, SVMs have a solid theoretical foundation and can handle both linear and nonlinear classification problems through the use of kernel functions, making them versatile and reliable for a wide range of classification tasks. |
5. Conclusions
Sentiment analysis plays a crucial role in teaching evaluation by analyzing students’ opinions and sentiments toward their instructors and courses. By automatically classifying text as positive, negative, or neutral, sentiment analysis can provide valuable insights into the overall satisfaction and perception of teaching quality. It helps identify areas of improvement, gauges the effectiveness of instructional methods, and detects potential issues that impact student engagement and learning outcomes. By leveraging sentiment analysis, educational institutions can make data-driven decisions to enhance teaching strategies, address student concerns, and ultimately improve the overall teaching and learning experience. Based on this background, this paper proposes to use the CNN–SVM model to classify the sentiment tendency of teaching evaluation data. The CNN in this model has the function of adaptive feature extraction, and the fusion of SVM with strong generalization performance can reduce the individual differences and data noise between data, thereby improving sentiment classification performance. The experimental data in this paper adopt EEG signals, which are the most typical representatives of physiological data. Experiments show that the model used has certain classification advantages. However, the model used in this paper has certain limitations as well. First, the combination increases the overall model complexity, making it more challenging to interpret and understand the learned features. Additionally, the training time of such a combined model can be significantly longer due to the training requirements of both CNN and SVM components. Moreover, hyperparameter tuning becomes more intricate as it involves optimizing parameters for both the CNN and SVM, which adds to the complexity of the model selection process. Balancing these factors and achieving optimal performance requires careful consideration and experimentation. This is also the content that this research will continue to study in the future.
Acknowledgment
This work was supported by the Central University Basic Research Business Fee Special Fund Project (NJ2023043) and the Research Fund for Education and Teaching Reform of Nanjing University of Aeronautics and Astronautics (2017JG1126Y).
ORCID
XUEZHI FAN https://orcid.org/0009-0000-7965-7942
JIE ZHANG https://orcid.org/0000-0002-6608-8207
MENGTING YANG https://orcid.org/0009-0000-4459-1656