Abstract Anomaly detection is a crucial aspect for both safety and efficiency of modern process industries. This paper proposes a two-steps methodology for anomaly detection in industrial processes, adopting machine learning classification algorithms. Starting from a real-time collection of process data, the first step identifies the ongoing process phase, the second step classifies the input data as “Expected”, “Warning”, or “Critical”. The proposed methodology is extremely relevant where machines carry out several operations without the evidence of production phases. In this context, the difficulty of attributing the real-time measurements to a specific production phase affects the success of the condition monitoring. The paper proposes the comparison of the anomaly detection step with and without the process phase identification step, validating its absolute necessity. The methodology applies the decision forests algorithm, as a well-known anomaly detector from industrial data, and decision jungle algorithm, never tested before in industrial applications. A real case study in the pharmaceutical industry validates the proposed anomaly detection methodology, using a 10 months database of 16 process parameters from a granulation process.

中文翻译：

---

用于异常检测和过程阶段分类的机器学习，以改善安全和维护活动

摘要 异常检测是现代过程工业安全和效率的重要方面。本文提出了一种工业过程异常检测的两步方法，采用机器学习分类算法。从过程数据的实时收集开始，第一步确定正在进行的过程阶段，第二步将输入数据分类为“预期”、“警告”或“关键”。当机器在没有生产阶段证据的情况下执行多项操作时，所提出的方法非常相关。在这种情况下，将实时测量归因于特定生产阶段的难度会影响状态监测的成功。论文提出了有和没有过程阶段识别步骤的异常检测步骤的比较，验证其绝对必要性。该方法应用决策森林算法，作为众所周知的工业数据异常检测器，以及决策丛林算法，之前从未在工业应用中进行过测试。制药行业的一个真实案例研究验证了提议的异常检测方法，使用来自制粒过程的 16 个过程参数的 10 个月数据库。