大型衡器是大宗货物称重计量的关键设备，是典型的复杂非线性多传感器系统，因不具备故障在线检测与容错功能，任一传感器故障都将导致称重系统失效，是目前大型称重领域亟待解决的难题；受计算复杂性和样本数量等限制，基于传统神经网络方法的大型衡器故障检测与容错建模将遭遇若干瓶颈问题。本项目提出一种新型调制窗-RBF神经网络及优化算法，在此基础上研究大型衡器故障诊断与容错方法，主要包括：①构造新型调制窗激励函数，构建基于先验知识的约束条件，建立调制窗-RBF神经网络及优化算法，减少运算量、提高泛化能力；②构建一种基于无附加传感器的称重传感器状态在线检测方案，建立大型衡器故障诊断和容错测量模型，实现故障状态下的准确称量。这种故障诊断与容错测量方法可解决大型衡器称重难题，提高称量可靠性；形成的调制窗-RBF神经网络及优化算法，可望突破传统神经网络计算量与泛化能力的局限，为神经网络优化与实用性研究开辟新思路。

A large-scale weighing instrument (LSWI) is the key apparatus of measuring bulk goods, and it is a typical complex and nonlinear multi-sensor system. LSWI is not a fault-tolerant system, and its faults cannot be detected online. Therefore, it does not work if any of its load cells is faulty, which is one of the most urgent problems of the large weighing filed. The model of fault detection and fault tolerance for LSWI based on the conventional neural network(NN) will encounter several bottlenecks because of the NN’s complexity and the number of the samples for training an NN. The aim of this project is to study a new modulation window radial basis neural network (MW-RBFNN) and its optimization algorithms, and an approach of fault diagnosis and fault tolerance for LSWI with this new NN. The major works include: ①constructing a small-calculation and fast-response modulation window function to meet the NN’s activation function, and the constraints for optimizing an NN based on the system’s prior knowledge, and then creating the modulation window radial basis neural network and its optimization algorithm; ②designing a programme of detecting the LSWI’s load cells online without state sensors, and establishing the model of the fault diagnosis (FD) and fault-tolerant measurement (FTM) for LSWI that is applied to accurately weighing loads when the LSWI is faulty. Based on this project, a new theory and method of fault diagnosis and fault-tolerant measurement for LSWI will be obtained, which will solve this proposed problem of LSWI, and improve the LSWI’s reliability. In addition, a new modulation window radial basis neural network (MW-RBFNN) and its optimization algorithms will be established, which will break through the limitations of the calculation and the generalization ability in a convention neural network, and will provide a new ideal for optimizing, simplifying and implementing neural networks.