随着智能技术与教育的紧密结合，智慧学习得到了前所未有的发展。在学生脱离实体教室与教师直接监督的情况下，评估其学习沉浸体验成为了重要的研究内容。对学习沉浸体验的评估有助于教师把握学生学习状态和反思教学手段，有助于学生探索“互联网+教育”模式下的学习方法，有助于从沉浸式学习的角度完善新型学习空间的建设。传统的评估方法（量表、访谈等）存在受主观偏差影响大、实时性差、人力统计效率低等局限，本文基于人体多元生理特征提出了一系列深度学习网络模型，围绕学习沉浸体验的客观化、智能化评估展开深入的研究。本文的主要研究成果和创新点如下：

（1）设计了包含生理信号采集方案的学习沉浸体验实验范式。本文基于可穿戴智能传感设备构建了人体EEG和PPG生理信号采集系统，并设计了基于VR学习场域的高等级沉浸体验、基于互动学习场域的中等级沉浸体验和基于传统学习场域的低等级沉浸体验实验范式，以此采集到了本文所需的生理特征数据集，共得到14553组样本，包含189189个特征。

（2）提出了基于深度学习的学习沉浸体验评估模型。考虑到传统机器学习模型难以处理大体量数据以及在高维度输入数据上的预测性能表现较弱等的问题，本文针对一维时序生理特征数据集以1DCNN-LSTM为主要结构构建了多元特征深度网络模型，实现了对学习沉浸体验的三分类预测评估，取得了81.00%的准确率，相比于问卷、量表等传统评估方法，极大地缩短了数据分析的时长，印证了客观生理特征评估学习沉浸体验的可行性。

（3）实现了模型在输入特征层面、网络结构层面及模型训练层面的优化。在输入特征层面，针对1DCNN-LSTM模型在各标签之间混淆度较大的问题，本文基于软注意力机制构建了1DCNN-Attention-LSTM模型，将注意力聚焦到对沉浸体验分类有更多贡献的生理特征上，使得准确率提高至84.65%，标签混淆度降低11.33%。针对模型在网络结构与模型训练层面上的超参数优化问题，本文基于粒子群优化算法设计了PSO-1DCNN-Attention-LSTM模型，搜寻重要超参数的融合最优解，使得模型的准确率提高至87.98%，标签间的混淆度再次降低了9.89%。

本文提出了多元生理特征驱动的深度学习模型，实现了对学习沉浸体验客观、智能的评估，在实验范式设计、深度模型构建及超参数融合优化等方面展现出了一定的实用价值。本文希望能为生理信号与智能传感的研究提供新思路，为智慧教育的发展提供新动力，为教师对学生的多层面考察提供新视角，为沉浸环境的搭建提供新评估手段。

With the close integration of smart technologies and education, smart learning has grown like never before. Evaluating students’ learning immersion experiences while they are removed from the physical classroom with direct teacher supervision has become an important research component. The evaluation of the learning immersion experience not only helps teachers to grasp students’ learning status and reflect on teaching methods, but also helps students to explore learning methods under the “Internet plus Education” environment, and helps to improve the construction of new learning spaces from the perspective of immersive learning. Traditional evaluation methods (scales, interviews, etc.) are limited by subjective bias, poor real-time performance, and low statistical efficiency. This paper proposes a series of deep learning models based on multiple physiological features of the human body and conducts an in-depth study on the objective and intelligent assessment of the learning immersion experience. The major research contents and innovation points of this paper are as follows:

(1) Design an experimental paradigm of learning immersion experience with a physiological signal acquisition scheme. In this paper, we build a human EEG and PPG physiological signal acquisition system based on wearable intelligent sensing devices and design an experimental paradigm of high-level immersion experience based on VR learning field, medium-level immersion experience based on interactive learning field, and low-level immersion experience based on traditional learning field, to collect the physiological features dataset required, and obtain a total of 14553 groups of samples, containing 189189 features.

(2) Propose a deep learning-based learning immersion experience evaluation model. Considering the problems that traditional machine learning models have difficulty in handling large volumes of data and weak prediction performance on high-dimensional input characters, this paper constructs a multi-featured deep network model with 1DCNN-LSTM as the main structure for the one-dimensional time-series physiological feature dataset, and achieves a three-classification prediction evaluation of learning immersion experience with 81.00% accuracy, and greatly reduces the time-consuming of data analysis compared with traditional methods such as questionnaires and scales, which confirms the feasibility of objective physiological features to evaluate learning immersion experience.

(3) Achieve optimizations of the model at the input feature stage, network structure stage, and model training stage. At the input feature stage, to address the problem that the 1DCNN-LSTM model has large confusion among labels, this paper constructs a 1DCNN-Attention-LSTM model based on the soft attention mechanism, focusing attention on the physiological features that contribute more to the classification of the learning immersion experience, resulting in an accuracy rate of 84.65% and an 11.33% reduction in label confusion. To solve the hyperparameter optimization problem of the model at the network structure and model training stages, this paper designs a PSO-1DCNN-Attention-LSTM model based on the particle swarm optimization algorithm to search for the fused optimal solution of important hyperparameters, which improves the accuracy of the model to 87.98% and reduces the inter-label confusion by 9.89%.

This paper proposes deep learning models driven by multiple physiological features to achieve an objective and intelligent assessment of the learning immersion experience, and shows   practical value in experimental paradigm design, deep learning model construction, and hyperparameter fusion optimization. This paper hopes to provide new ideas for research on physiological signals and intelligent sensing, new impetus for the development of smart education, new perspectives for teachers to examine students on multiple levels, and new assessment tools for the construction of immersion environments.