融合多源因素的编码器-解码器沉降长时预测模型  

Encoder-decoder settling long-time prediction model incorporating multi-source factors

作　　者：徐浩钧 顾敏明[1,2,3] 程洪福 李晨露 胡伏原 XU Haojun;GU Minming;CHENG Hongfu;LI Chenlu;HU Fuyuan(School of Electronic and Information Engineering,Suzhou University of Science and Technology,Suzhou 215009,China;Jiangsu Industrial Intelligent and Low-carbon Technology Research and Engineering Center,Suzhou 215009,China;Suzhou Key Laboratory of Intelligent Low-carbon Technology Application,Suzhou 215009,China;Suzhou Humble Administrator's Garden Management Office(Suzhou Garden Museum),Suzhou 215009,China)

机构地区：[1]苏州科技大学电子与信息工程学院,江苏苏州215009 [2]江苏省工业智能低碳技术工研究程中心,江苏苏州215009 [3]苏州市智能低碳技术应用重点实验室,江苏苏州215009 [4]苏州市拙政园管理处(苏州市园林博物馆),江苏苏州215009

出　　处：《微电子学与计算机》2025年第2期39-49,共11页Microelectronics & Computer

基　　金：江苏省研究生科研与实践创新计划(SJCX22_1588);苏州市科技发展计划(SS202133)。

摘　　要：石质文物假山长期曝露于室外,受多源因素影响易形成不均匀沉降,因此假山沉降传感器监测与长时精准预测对石质文物保护十分必要。现有沉降长时预测方法难以有效解决噪声和瞬时波动造成的精度降低与应用可靠性问题。为此,本文提出一种融合多源因素的编码器-解码器沉降长时预测模型。在多源因素编码器中设计动态多源因素融合模块将深度特征进行融合并实时计算沉降、温度、振动、裂缝等多源因素与目标数据的动态相关性;在时域增强解码器中构建多头自适应平滑模块,通过多头注意力的方法自适应学习各时间步的平滑指数,保留时间序列长期趋势,减少传感器带来的噪声和瞬时波动。本模型以环秀山庄沉降监测系统的实测数据集进行验证,结果表明该模型相较于基线方法在评价指标均方根误差(Root Mean Squared Error,RMSE)指标、平均绝对误差(Mean Absolute Error,MAE)指标以及连续排序概率评分(Continuous Ranked Probability Score,CRPS)最高分别提升了19.1%、19%以及16.3%,且符合实际应用需求。Stone cultural relics rockery exposed to outdoor for a long time,and it is easy to form uneven settlement under the influence of multiple source factors,so rockery settlement sensor monitoring and long-term accurate prediction of stone cultural relics protection is very necessary.It is difficult to solve the problems of accuracy reduction and application reliability caused by noise and instantaneous fluctuation.To solve the above problems,a long-term prediction model of encoder-decoder settlement based on multi-source factors is proposed in this paper.The dynamic multi-source factor fusion module is designed in the multi-source factor encoder to fuse the depth features and calculate the dynamic correlation between the multi-source factors such as settlement,temperature,vibration and crack and the target data in real time.The multi-head adaptive smoothing module is constructed in the time domain enhanced decoder,and the smoothing index of each time step is learned adaptively by the method of multi-head attention,the long-term trend of time series is preserved,and the noise and instantaneous fluctuation caused by the sensor are reduced.The measured data set of the settlement monitoring system of Huan-xiu Mountain Village was verified,and the results showed that compared with the baseline method,the evaluation indexes of Root Mean Squared Error(RMSE),Mean Absolute Error(MAE)and Continuous Ranked Probability Score(CRPS)increased by 19.1%,19%and 16.3%respectively,which met the practical application requirements.

关 键 词：沉降长时预测 多源因素 编码器-解码器 注意力机制 多头自适应平滑 石质文物保护 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]