工地施工过程中需要判断浇筑区域内部的混凝土与泥浆的分类, 目前大部分使用人工检测法, 为保证工程质量, 浇筑的混凝土桩基高度往往远高于设计值, 从而造成混凝土的极大浪费。 为此, 提出了基于多光谱的泥浆分界自动检测技术。 首先, 利用光谱仪分析了11种混凝土、 泥浆不同比例混合情况下的可见光反射光谱, 结果表明混合比例系数 K与光谱反射率之间存在函数关系, 并利用反射率的计算公式确定用反射光强度代替反射率。 为此, 基于AS7341光谱芯片设计了由8个可见光光谱波段和STM32单片机以及JDY-31蓝牙通信模块组成的监测系统, 并设计了透明外壳。 对上述11种不同比例泥浆混凝土混合溶液进行了多光谱反射光强度数据采集, 采集不同比例黄土和黑土与混凝土混合的反射光强度值组成数据集, 发现反射光强度值随着时间进行波动, 因此提出了基于多光谱的注意力机制融合长短记忆卷积神经网络(CNN-LSTM-Attention)的泥浆混凝土分界预测算法, 利用CNN网络对输入的8个通道反射光强度的关键参数进行特征提取, 获取更多的局部特征, 提高后续预测精度。 再通过LSTM层对CNN层增加接口和反向门来实现反向传播, 避免数据梯度消失和梯度爆炸的问题。 最后使用Attention函数, 在众多波长反射值输入中关注更重要的光谱信息, 减少关注或者忽视其他光谱信息, 从而解决信息过量问题并提高效率。 仿真结果表明该算法的精确率为0.952, 召回率为0.944, F1分数为0.938。 相较于其他算法的仿真结果具有更高的精度和稳定性, 能够达到工地施工的国家要求标准并且该方法不需要人工测量, 直接上位机远程实时控制光谱数据采集过程, 得到混凝土和泥浆的分界检测结果。

During construction, it is necessary to determine the classification of concrete and mud within the pouring area. Currently, most projects use manual detection methods. To ensure engineering quality, the height of poured concrete piles often far exceeds the design value, leading to significant waste of concrete. To address this issue, an automatic mud boundary detection technology based on multispectral imaging is proposed.First, a spectrometer was used to analyze the visible light reflection spectra of 11 mixtures of concrete and mud at different ratios. The results show a functional relationship between the mixing ratio coefficient K and spectral reflectance. The calculation formula for reflectance was used to determine that reflected light intensity can be used as a substitute forreflectance. Based on this, a monitoring system was designed using the AS7341 spectral chip, which consists of 8 visible light spectral bands, an STM32 microcontroller, and a JDY-31 Bluetooth communication module, all enclosed in a transparent housing. Multispectral reflected light intensity data were collected for the 11 concrete-mud mixtures at different ratios, forming a dataset of reflected light intensity values for mixtures of loess and black soil with concrete at varying proportions. It was found that the reflected light intensity values fluctuated over time. Therefore, a mud-concrete boundary prediction algorithm based on a Convolutional Long Short-Term Memory Network with Attention Mechanism (CNN-LSTM-Attention) was proposed. The CNN network extracts key features from the input reflected light intensity data of 8 channels to capture more local features and improve subsequent prediction accuracy. The LSTM layer adds interfaces and reverse gates to the CNN layer to enable backpropagation, avoiding gradient disappearance and explosion. Finally, the Attention function focuses on more critical spectral information among the input reflectance values at multiple wavelengths, thereby reducing or ignoring other spectral information to address the problem of information overload and improve efficiency.Simulation results show that the algorithm achieves a precision of 0.952, a recall of 0.944, and an F1 score of 0.938. Compared with other algorithms, it demonstrates higher accuracy and stability, meeting national standards for construction sites. This method eliminates the need for manual measurement. It enables real-time remote control of spectral data collection via a host computer, providing direct detection results for the boundary between concrete and mud.