Temperature monitoring is one of the easiest and frequent way to support maintenance procedures in the industry. In this paper, a procedure of long term temperature data processing for anomaly detection is proposed. The data obtained from a commercial SCADA (supervisory control and data acquisition system) covers a couple of months of gearbox operation. During that period, repair action has been performed. The gearbox is part of the belt conveyor drive unit used in a deep underground mine. Due to time-varying load conditions, changing environment (ambient temperature), frequent starting and stopping of the machine simple decision-making rules (as if the temperature is higher than the threshold) are not appropriate. To detect anomaly in the data describing the process, we proposed a procedure based on the statistical testing methodology applied to long term data analysis. From a signal processing perspective, the proposed algorithm leads to data segmentation. As a result, one receives a moment (called the structure break point) where the statistical properties of the signal change. As criteria of segmentation we utilize the classical test statistics of the known statistical tests (parametric Student's t and non-parametric Wilcoxon tests) for equal means of two samples, however, we applied them not the raw signal but to its characteristics (here means and medians calculated for the work shifts). The proposed procedure is very effective for long term temperature data, however, it indicates the shift detected anomaly, not particular samples. Assuming that the degradation process is rather slow and reaction on anomaly takes some time, it is enough in practice to provide information that the machine can work for the next shift. The proposed procedure could be considered as a training process to establish a “threshold” in the statistics domain for complex decision making.

温度监控是支持行业维护程序的最简单且频繁的方法之一。本文提出了一种长期的温度数据异常检测处理程序。从商业SCADA（监控和数据采集系统）获得的数据涵盖了数月的变速箱操作。在此期间，已执行维修措施。变速箱是用于深井地下的皮带输送机驱动单元的一部分。由于时变的负载条件，变化的环境（环境温度），频繁启动和停止机器的简单决策规则（好像温度高于阈值）是不合适的。为了检测描述过程的数据中的异常，我们提出了一种基于统计测试方法的程序，该方法适用于长期数据分析。从信号处理的角度来看，所提出的算法导致了数据分割。结果，人们会收到一个信号的统计特性发生变化的时刻（称为结构断裂点）。作为分割的标准，我们利用已知统计检验（参数Student t和非参数Wilcoxon检验）的经典检验统计量对两个样本进行均等化，但是，我们不是将原始信号应用于特征，而是将其应用于特征（此处的均值和为工作班次计算的中位数）。所提出的过程对于长期温度数据非常有效，但是，它指示了检测到的异常位移，而不是特定的样本。假设降级过程相当缓慢，并且对异常的反应需要花费一些时间，那么在实践中足以提供信息以使机器可以为下一个班次工作。提议的程序可以被视为一种培训过程，可以在统计领域为复杂的决策制定“阈值”。