石油钻井是一项投资巨大，发展前景广阔，关系国家战略资源开发的工程。石油钻井具有不可控、模糊性、随机性因素多、地下情况未知等特点，钻井安全直接影响着钻井进度和钻井成本。保障钻井安全是实现安全钻井，节省钻井成本的关键，也是钻井提速提效和保障钻井人员和设备安全的重要基础。 本文针对噪声环境下石油钻井数据流趋势异常检测困难、时效约束条件下微小趋势异常检测精度不高、滑动窗参数最优化设置困难三个方面的问题，提出了相应的解决模型和方法，基于塔里木油田石油钻井真实数据验证了本文所研究检测方法的有效性，主要完成的工作如下。 (1)针对噪声环境下石油钻井数据流趋势异常检测准确率低的问题，提出了滑动嵌套窗体数据流异常检测模型(Sliding Nest Window Chart Anomaly Detection based on the Data Stream，SNWCAD-DS)。首先根据噪声与趋势异常在异常检测初始具有相同的变化特征，提出采用短期窗体弱化当前点的影响；其次以长短窗体差值作为趋势异常计算的基础；最后根据出界率作为趋势异常检测的门限值屏蔽噪声干扰。该方法可以有效滤除噪声的影响，提高趋势异常检测的准确率。 (2)针对时效约束和噪声干扰下石油钻井数据流微弱趋势异常无法检测问题，提出了两次计算累积和控制图模型(Double CUSUM based on Data Stream，DCUSUM-DS)。首先采用滑动嵌套窗模型弱化噪声的影响；其次提出把原始值映射到两次累积和空间，提取新的特征量；最后提出新的微小趋势异常检测方法。该方法可以提高噪声环境下，微小趋势异常检测的准确率。 (3)针对滑动窗参数在石油钻井数据流异常检测模型中参数最优化难以设置问题，提出以PSO、NSGAII和SPEA2相结合的集成算法作为滑动窗参数确定方法。首先以1-TPR和FPR为两目标，使用PSO进化算法，找到两目标最小的帕累托前沿。其次在找到的帕累托前沿集合中，再次寻找帕累托前沿，进而确定异常检测模型最佳参数设置。该方法可以提高趋势异常检测算法检测的准确率，且不增加算法计算复杂度。 (4)本文以SNWCAD-DS和DCUSUM-DS趋势异常检测模型为基础，以进化算法确定滑动窗参数为原则，建立一套石油钻井预警系统。通过实际钻井数据的试验结果表明：钻井工程预警系统能够根据趋势异常检测对钻井事故进行准确预报，极大的提高预警系统的时效性。 本文研究成果已经应用于塔里木油田石油钻井数据流的趋势异常检测，据2015到2018年统计数据表明，采用本文设计算法的钻井工程预警系统，参数趋势异常检测准确率为92.65％，系统预警综合决策准确率达到了86.31％。 关键词：石油钻井；趋势异常检测；控制图；趋势异常；优化算法；滑动窗

 Oil drilling is a project with huge investment and broad development prospects, which is related to the national strategic resource development Oil drilling has the characteristics of uncontrollable, ambiguous, random factors and unknown underground conditions. Drilling safety directly affects drilling progress and drilling costs. Studying drilling safety is the key to achieving safe drilling and saving drilling costs. It is also an important basis for improving drilling speed and ensuring the safety of drilling personnel and equipment. In this thesis, we study the problem of abnormality of oil drilling data stream anomaly detection under noise environment, the low precision detection of small trend anomaly under time constraint, and the difficulty of optimal setting of sliding window parameters. The corresponding solution model and method are proposed, based on Tarim The actual data of oilfield oil drilling verified the effectiveness of the detection method studied in this thesis. The main work is as follows: (1)Aiming at the difficulty of numerical trend anomaly detection in noise environment, a sliding nested window chart anomaly detection model SNWCAD-DS (sliding nested window chart anomaly detection based on the data stream) is proposed. Firstly, according to the characteristics that the noise and trend anomaly have the same performance at the beginning, the short-term form is used to weaken the influence of the current point; secondly, the difference between the long and short forms is proposed as the basis of trend anomaly calculation; finally, the processing mechanism of shielding noise interference is proposed based on the number of out-bound points as the threshold value of trend anomaly detection. The method can effectively filter out the influence of noise and improve the accuracy of the trend abnormality detection. (2)Aiming at the problem that the weak trend anomaly of oil drilling data stream can not be detected under time constraints and noise interference, the double CUSUM based on data stream is proposed. Firstly, a sliding nested window model is used to weaken the influence of noise. Secondly, the original value is mapped to two cumulative sum spaces to extract new features. Finally, a new method for detecting small trend anomaly is proposed. This method can improve the accuracy of microtrend abnormality detection under noisy environment. (3)Aiming at the problem that the parameters of sliding window are difficult to set in the anomaly detection model of numerical industrial data stream, an ensemble algorithm combining PSO, NSGAII and SPEA2 is proposed to determine the parameters of sliding window. Firstly, taking 1 -TPR and FPR as two objectives, we use PSO and other evolutionary algorithms to find the Pareto frontier with the smallest two objectives. Secondly, in the Pareto frontier set found, the Pareto frontier is searched again, and then the optimal parameter setting of the anomaly detection model is determined. The method can improve the accuracy of the detection of the trend anomaly detection algorithm without increasing the computational complexity of the algorithm (4)Drilling engineering early warning system has been implemented in this thesis. Based on the trend anomaly detection model of SNWCAD-DS and DCUSUM-DS, an oil drilling early warning system is established on the principle of determining sliding window parameters by evolutionary algorithm The real drilling data experiments show that the drilling engineering early warning system can accurately predict drilling accidents according to trend anomaly detection, greatly improving the timeliness of the early warning system The research results of this thesis have been applied to the trend anomaly detection of oil drilling data stream in Tarim Oilfield. According to the statistical data from 2015 to 2018, the accuracy of the parameter trend anomaly detection using the design algorithm of this paper is 92.65%, and the system early warning comprehensive decision The accuracy rate reached 86.31%. Keywords : Industrial data stream; Trend anomaly detection; Control chart; Evolutionary algorithms; Sliding window