Self-Learning of Decision Rules for Process Control

本项目的目标是开发一种广泛适用的自动化方法，从传感器网络中的数据中学习决策规则。当既没有预先指定正常操作也没有预先指定要检测的异常时，所学习的决策规则将检测来自正常操作环境的异常。学习者还将整合分类数据和数值数据，识别决策规则信号发生时的贡献者，并随着传感器系统需求的变化而调整规则。基本上，它将适应正常环境的特征，处理分类和数值数据，自我训练，然后检测异常。将开发可变重要性的测量，以提供诊断检测到的信号的指导。这些措施也将用于自适应地提高可以纳入解决方案的变量的数量。初步实验将扩大到测试与模拟和真实的数据的学习者。学习者将与传统的最佳解决方案进行比较，然后在传统解决方案失败的压力更大的条件下进行评估。该项目的成功将提供廉价且易于安装、维护和管理的网络智能。这些结果可用于最终开发可在传感器网络基础设施中使用的嵌入式模块（智能）。衍生的决策规则的灵活性，可以纳入高度非线性模型。这种方法的灵活性和概念简单性，沿着现在广泛可用的计算资源，可以产生对这种自动化方法的兴趣激增。该方法可用于制造以及生物、民用和运输监测等领域。此外，变量重要性的度量可以应用于询问复杂应用中的其他学习算法。

The objective of this project is to develop a widely-applicable, automated method to learn decision rules from the data in a sensor network. The decision rules that are learned will detect anomalies from the normal operating environment when neither the normal operations nor the anomalies to be detected are pre-specified. The learner will also incorporate categorical as well as numerical data, identify contributors to a decision-rule signal when it occurs, and adapt the rules over time as requirements for the sensor system change. Basically, it will adapt itself to the characteristics of the normal environment, handle categorical and numerical data, self-train, and then detect anomalies. Measures of variable importance will be developed to provide guidance to diagnose a detected signal. Such measures will also be used to adaptively to improve the number of variables that can be incorporated into a solution. Preliminary experiments will be expanded to test the learner with simulated and real data. The learner will be compared to traditional, optimal solutions when they exist and then evaluated under more stressful conditions in which traditional solutions fail.Success of this project will provide network intelligence that is inexpensive and easy to install, maintain, and mange. These results can be used to eventually develop an embedded module (intelligence) that can be made available in sensor-network infrastructures. The flexibility of the derived decision rules can incorporate highly nonlinear models. This flexibility and the conceptual simplicity of the approach, along with the computational resources now widely available, can generate a surge of interest in this automated approach. The methods can be useful to manufacturing and in areas such as biological, civil, and transportation monitoring. Also, the measures for variable importance can be applied to interrogate other learning algorithms in complex applications.