近年来，应用人工神经网络（ANN）模型提升工业传感精度已成为热门研究方向。然而，神经网络模型是纯数据驱动的多元“黑箱”模型，其隐藏层提取的特征缺乏实际物理意义，导致基于ANN的传感模型性能不稳定，难以在流程工业现场实际应用。针对上述挑战，本文提出一种名为偏最小二乘辅助优化网络（PLSaoNET）的通用ANN模型。PLSaoNET利用PLS模型辅助确定网络的初始化权重和隐藏层神经元数量，后续训练过程作为PLS回归结果引导的再优化过程，使网络具备统计约束特性，从而降低对数据的依赖。此外，针对工业现场样本标签分布不均的问题，本文设计了一种用于网络再训练的分层采样方法。通过两个工业传感应用验证了所提方法的有效性和优越性：基于激光诱导击穿光谱（LIBS）数据的铁精矿浆铁品位监测，以及基于近红外（NIR）光谱数据的柴油品质评估。与PLS回归模型和基于Xavier初始化的反向传播神经网络（BPNN）模型相比，PLSaoNET展现出最优的建模精度和泛化性能。本文设计了完整的理论框架来指导超参数确定并明确网络的求解路径，满足了工业过程对精度、鲁棒性和易用性的三重要求。所提模型在提升生产过程工业传感的精度和可靠性方面具有巨大潜力。

The application of artificial neural network (ANN) models to achieve higher accuracy in industrial sensing has become a popular research topic in recent years. However, neural network models are purely data-driven multivariate “black-box” models, and the features extracted from the hidden layer have no actual physical meaning, making the performance of ANN-based sensing models unstable and difficult to practically apply at process industry sites. To address these challenges, this paper proposes a generalized ANN model called the partial least squares (PLS)-assisted optimization network (PLSaoNET). PLSaoNET employs the PLS model to assist in determining the initialization weights of the network and the number of hidden-layer neurons. The subsequent training serves as a reoptimization process guided by the PLS regression result, enabling the network to incorporate statistical constraints and thereby reducing its reliance on data. In addition, to address the problem of uneven distributions of sample labels at industrial sites, this paper designs a stratified sampling method for network retraining. The efficiency and superiority of the proposed method are verified via two industrial sensing applications: the monitoring of iron grade in iron ore concentrate slurry samples based on laser-induced breakdown spectroscopy (LIBS) data, and the assessment of the quality of diesel fuels based on near-infrared (NIR) spectroscopy data. In comparison with a PLS regression model and a Xavier initialization-based backpropagation neural network (BPNN) model, PLSaoNET exhibits the best modeling accuracy and generalization performance. This work designs a complete theoretical framework to guide the determination of hyperparameters and specify the solution paths of the network, thereby satisfying the triple requirements of accuracy, robustness, and ease of use in industrial processes. The proposed model holds great potential for improving the accuracy and reliability of industrial sensing in production processes.