多模态感知驱动下高堆石坝施工仿真参数集成深度学习模型

Multimodal perception-driven high rockfill dam construction simulation input modeling using an ensemble deep learning model

现有施工仿真参数建模方法主要依靠单一模态数据,且现有多模态数据采集过程存在一定的滞后性,导致仿真的实时性和准确性不足。针对上述问题,本文提出了多模态感知驱动下高堆石坝施工仿真参数集成深度学习模型。首先,在SpringBoot框架下开发了基于移动智能手机传感器的运动学和声学数据实时采集云平台,并采用低通滤波器和梅尔频谱等方法实现堆石坝施工机械多模态数据的实时采集与预处理;其次,提出了用于自动提取多模态数据特征的堆石坝施工机械精细活动状态识别深度学习模型。该模型集成了改进深度卷积长短期记忆循环神经网络(Improved DeepConvLSTM,IDeepConvLSTM)与深度卷积神经网络的优势,前者可精确感知施工机械运动方向,后者可从声音模态中感知施工机械振动状态。其中,IDeepConvLSTM在卷积层中间加入批量归一化层以提高收敛速度,并采用梯度缩放和剪裁以避免梯度爆炸的问题;进一步地,在云平台中采用大窗口移动过滤器在线处理机械活动识别结果,实现堆石坝施工仿真参数的实时建模。工程案例表明,相比于单一的运动学或声学模态的机械活动识别方法,本研究所提方法的识别精度分别提高了9.22%和23.62%。研究成果为提高堆石坝施工仿真的准确性和实时性提供了新的思路和技术手段,具有一定的应用和推广价值。

The real-time and accurate modeling of construction simulation parameters is the key to ensure the accuracy of construction simulation.The existing construction simulation parameter update methods mainly rely on single modal data while the existing multi-modal data acquisition process has a problem of lagging, resulting in insufficient real-time and accuracy of simulation.To solve those problems, this paper proposes an ensemble deep learning model for high rockfill dam construction simulation input modeling driven by multimodal perception.First, a cloud platform for real-time acquisition of kinematics and acoustic data based on mobile smartphone sensors was developed under the SpringBoot framework, relying on this, the multi-modal data of rockfill dam construction machinery is collected in real-time and pre-processed using low-pass filters and mel spectrograms.Secondly, an ensemble deep learning model for rockfill dam construction machinery activity recognition is proposed to automatically extract multimodal data features and accurately identify mechanical fine activity.The model integrates an improved deep convolution long short-term memory neural network(Improved DeepConvLSTM,IDeepConvLSTM) accurately perceives the motion direction of construction machinery and the advantages of deep convolutional neural network to perceive the vibration state of construction machinery from sound modes.Among them, IDeepConvLSTM adds a batch normalization layer in the middle of the convolution layer to improve convergence speed, and gradient scaling and clipping are used to avoid the problem of gradient explosion;further, a large window moving filter is used to process the recognition results of mechanical activity online, and real-time modeling of rockfill dam construction simulation parameters is realized.The engineering case shows that, compared with the single kinematic or acoustic mode mechanical activity recognition method, the recognition accuracy of the proposed method is improved by 9.22% and 23.62%,respectively.The research results provide new ideas and technical means for improving the accuracy and real-time performance of rockfill dam construction simulation, which have significant application and popularization value.