融合注意力机制的LSTM职工心理压力状态评价方法*

doi:10.16265/j.cnki.issn1003-3033.2026.03.0432

摘要/Abstract

摘要：

为确保职工身心健康,提升心理压力评价方法的准确性与可解释性,以多模态生理时间序列数据为研究对象,提出一种融合注意力机制(AM)的长短期记忆(LSTM)网络方法(LSTMA),实现职工心理压力状态的准确评价。首先,以数据集WESAD中多模态生理时间序列数据(血容量脉搏(BVP)、心电图(ECG)、皮肤电活动(EDA)、肌电图(EMG)、呼吸(RESP)、体温(TEMP)和三轴加速度(ACC))为研究对象,通过分模态LSTM模块的门控记忆机制,精准捕获跨时间步时序依赖特征,有效保留与心理状态强相关的关键生理特征,并过滤短期随机噪声,确保生理特征数据能真实表征职工心理状态的动态演化;然后,在特征融合后引入AM,基于各模态、各时间步生理数据的特征重要性自适应分配注意力权重系数,强化对心理压力状态敏感的关键特征与微小响应特征,同时抑制冗余信息干扰;最后,通过全连接神经网络完成心理压力状态准确评价。结果表明:LSTMA在中性、压力、愉悦、冥想4分类任务中,心理压力状态评价准确率达94.56%;经留一法交叉验证后,准确率提升至98.08%;消融试验验证了分模态LSTM与AM的协同增强效应,模型解释性分析进一步佐证LSTMA方法设计的科学性与合理性。

关键词: 注意力机制(AM), 长短时记忆(LSTM)网络, 心理压力, 状态评价, 生理时间序列数据, 多模态

Abstract:

To safeguard employees' psychological health and improve the accuracy and interpretability of psychological stress evaluation methods, taking multimodal physiological time-series data as the research object, a LSTM with AM(LSTMA) method was proposed to accurately evaluate employees' psychological stress states in the paper. Firstly, using the multimodal physiological time-series data (Blood Volume Pulse (BVP), Electrocardiogram (ECG), Electrodermal Activity (EDA), Electromyogram (EMG), Respiration (RESP), Body Temperature (TEMP), and three-axis Acceleration (ACC)) from the WESAD dataset were adopted as the research carrier, the gating memory mechanism of the modal-specific LSTM modules was utilized to accurately capture cross-time-step temporal dependency features, effectively retain key physiological features strongly associated with psychological states, and filter out short-term random noise, thereby ensuring that the physiological feature data could truly characterize the dynamic evolution of employees' psychological states. Secondly, after feature fusion, the attention mechanism was introduced to adaptively assign attention weight coefficients based on the feature importance of physiological data across different modalities and time steps, enhancing key features and micro-response features sensitive to psychological stress states while suppressing the interference of redundant information. Finally, the accurate evaluation of psychological stress states was accomplished through a fully connected neural network. Experimental results show that the LSTMA method achieves an evaluation accuracy of 94.56% for the four-classification task (neutral, stress, pleasure, and meditation) of psychological stress states. After Leave-One-Out Cross-Validation (LOOCV), the accuracy is improved to 98.08%. Ablation experiments verify the synergistic enhancement effect of the modal-specific LSTM and the attention mechanism, and model interpretability analysis further confirms the scientificity and rationality of LSTMA.

Key words: attention mechanism (AM), long short-term memory (LSTM) network, psychological stress, status evaluation, physiological time-series data, multimodal

中图分类号:

X911安全心理学

曹海青, 姚志英, 吕淑然, 姚翠友. 融合注意力机制的LSTM职工心理压力状态评价方法^*[J]. 中国安全科学学报, 2026, 36(3): 229-237.

CAO Haiqing, YAO Zhiying, LYU Shuran, YAO Cuiyou. Evaluation of employee's psychological stress status using LSTM with attention mechanism[J]. China Safety Science Journal, 2026, 36(3): 229-237.

图/表 7

图1

图2

表1

表2

表3

图3

图4

参考文献 21

[1]	ELZEINY S, QAEAQE M. Stress classification using photoplethysmogram-based spatial and frequency domain images[J]. Sensors, 2020, 20(18):4987-4999. doi: 10.3390/s20174987
[2]	BOUTER Y, BRZOZKA M M, RYGULA R, et al. Chronic psychosocial stress causes increased anxiety-like behavior and alters endocannabinoid levels in the brain of C57BL/6J mice[J]. Cannabis and Cannabinoid Research, 2020, 5(1):51-61. doi: 10.1089/can.2019.0041 pmid: 32322676
[3]	FLOVIO D M, FRANCA D. High-resolution physiological stress prediction models based on ensemble learning and recurrent neural networks[C]. Proceedings of the IEEE International Symposium on Intelligent Systems and Computing, 2020:1-6.
[4]	FITRL I I, SRI W. Stress detection from multimodal wearable sensor data[C]. Proceedings of the 2^nd International Conference on Engineering and Applied Science, 2020:1-7.
[5]	GHANAPRIYA S, ORCHID C P, RAVINDER K. Stress recognition with multi-modal sensing using bootstrapped ensemble deep learning model[J]. Expert System, 2023, 40(6):1-16.
[6]	PHILIP S, ATTILA R, ROBERT D, et al. Introducing WESAD, a Multimodal dataset for wearable stress and affect detection[C]. Proceedings in 2018 International Conference on Multimodal Interaction, 2018:400-408.
[7]	PRERNA G, JAYASANKAR S, ANDREAS D, et al. Stress detection by machine learning and wearable sensors[C]. Proceedings in the 26^th International Conference on Intelligent User Interfaces, 2021:43-45.
[8]	MOHAMED A A A, ROAA M, NANCY M, et al. A deep learning approach using WESAD data for multi-class classification with wearable sensors[C]. Proceedings in the 6^th Novel Intelligent and Leading Emerging Sciences Conference, 2024:194-197.
[9]	SRIRAM K P, PRAVEEN K G, ABDUL A S G, et al. Deep learning-based automated emotion recognition using multimodal physiological signals and time frequency methods[J]. IEEE Transactions on Instrumentation and Measurement, 2024,73: DOI: 10.1109/TIM.2024.3420349.
[10]	CHOI H S. Emotion recognition using a Siamese model and a late fusion-based multimodal method in the WESAD dataset with hardware accelerators[J]. Electronics, 2025, 14(4):1-16. doi: 10.3390/electronics14010001
[11]	ANUBHAV B, BEHNAM B, DIRK R, et al. Attentive cross-modal connections for deep multimodal wearable-based emotion recognition[C]. 9^th International Conference on Affective Computing and Intelligent Interaction Workshop and Demos, 2021:1-5.
[12]	ZHAO Minghui, ZHAO Lulu, GAO Hongxiang, et al. Cross-modal attention fusion of electrocardiogram emotion and voiceprint for depression detection[C]. Proceeding of Computing in Cardiology Conference, 2024:1-4.
[13]	WANG Ruiqi, JO W, ZHAO Dezhong, et al. Husformer: a multi-modal transformer for multi-modal human state recognition[J]. IEEE Transactions on Cognitive and Developmental Systems, 2024, 16(4):1374-1390. doi: 10.1109/TCDS.2024.3357618
[14]	HOCHREITER S, SCHMIDHUBER J. Long short-term memory[J]. Neural Computation, 1997, 9(8):1735-1780. doi: 10.1162/neco.1997.9.8.1735 pmid: 9377276
[15]	陈小乾, 尹亮, 展宗辉, 等. 基于注意力机制和RCN-BiLSTM融合的风电机组故障识别[J]. 中国电力, 2025, 58(8):94-102.
	CHEN Xiaoqian, YIN Liang, ZHAN Zonghui, et al. Fault identification for wind turbine based on attention mechanism and RCN-BiLSTM fusion[J]. Electric Power, 2025, 58(8):94-102.
[16]	ITTI L, KOCH C, NIEBUR E. A model of saliency-based visual attention for rapid scene analysis[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1998, 20(11):1254-1259. doi: 10.1109/34.730558
[17]	吴思, 张旭光, 方银锋. 基于注意力机制的人群计数方法[J]. 中国安全科学学报, 2022, 32(1):127-134. doi: 10.16265/j.cnki.issn1003-3033.2022.01.017
	WU Si, ZHANG Xuguang, FANG Yinfeng. Method of crowd counting based on attention mechanism[J]. China Safety Science Journal, 2022, 32(1):127-134. doi: 10.16265/j.cnki.issn1003-3033.2022.01.017
[18]	曹淑超, 戈伟斌, 里聪慧, 等. 基于注意力Seq2Seq网络的人群安全风险评估[J]. 中国安全科学学报, 2025, 35(12):196-203. doi: 10.16265/j.cnki.issn1003-3033.2025.12.1477
	CAO Shuchao, GE Weibin, LI Conghui, et al. Crowd safety risk assessment based on Seq2Seq-attention network[J]. China Safety Science Journal, 2025, 35(12):196-203. doi: 10.16265/j.cnki.issn1003-3033.2025.12.1477
[19]	HENDRICKS L A, AKATA Z, DARRELL T. Generating visual explanations[C]. 14^th European Conference on Computer Vision (ECCV), 2016:3-19.
[20]	KANG Tianyu, DING Wei, CHEN Ping. CRESPR: modular scarification of DNNs to improve pruning performance and model interpretability[J]. Neural Networks, 2023,172: DOI: 10.1016/j.neunet.2023.12.021.
[21]	MICHAL B, HENRIK S. Generating visual explanations from deep networks using implicit neural representations[C]. 2025 IEEE/CVE Winter Conference on Applications of Computer Vision, 2025, 7:3310-3319.

模型分评价	标签值
模型分评价	中性	压力	愉悦	冥想
中性	96.81	0.42	5.76	4.86
压力	0.23	97.8	0.62	2.76
愉悦	2.02	0.11	92.72	1.47
冥想	0.94	1.67	0.9	90.91

方法	AB^[6]	RF^[7]	DNN^[8]	LSTMA
分类	三分类	三分类	三分类	四分类
准确率/%	80.34	67.56	90.26	94.56
参数量/×10⁶	12.8	18.5	215.3	156.7
FLOPS/×10⁹	0.08	0.12	8.6	5.3
单轮训练时间/min	3.2	4.7	14.5	10.3
加噪后准确率/%	64.74	55.26	81.36	91.36

模型	四分类准确率	冥想状态准确率
LSTMA	94.56	90.91
模型1(无LSTM)	85.28	76.42
模型2(无注意力)	90.32	84.55