异常放电可分为癫痫样放电和良性癫痫样变体，是脑电图中轮廓尖锐的放电模式，预示着不稳定的大脑状态，常出现于癫痫或某些神经系统疾病中。异常放电检测在认知神经科学研究、癫痫检测分型与病灶定位、飞行员筛选等方面都具有重要价值。目前识别异常放电以人工标注为主，常需要从数小时脑电图中寻找时长不足0.1秒的异常放电现象，将耗费大量专业医师的时间与精力。为此，存在大量研究从时域、频域等角度提取特征，构建系统检测异常放电。然而受限于有限的公开数据集、模糊的异常放电判断准则以及数据噪声等因素，算法表现常不够理想，且无法提取专家间存在争议的“模糊而微弱”的异常放电。同时，针对性的异常放电检测算法集中于癫痫样放电领域，而良性癫痫样变体的检测算法仍然相当缺乏。
针对异常放电的检测需求，本研究以癫痫样放电和良性癫痫样变体中都大量存在的棘波(spike)为切入点，开展了：异常放电长程检测，自适应模糊推理系统初始化方法改进设计，以及基于自适应模糊推理系统的异常放电检测共三项工作。研究目的是寻找合适的特征、建立稳定且能适应模糊判断准则的分类系统，为将来更微弱、意义更不明确的异常放电检测任务提供稳定的、可解释的、便于参考的分类准则。 
首先，针对原始脑电图文件时程长、异常放电分布稀疏的特点，本研究尝试通过设定概率密度相关的特征阈值直接检测异常放电。根据异常放电与脑电背景在包络中的概率密度差异，研究确定了相关的阈值参数；同时通过均值滤波、样条插值等辅助措施，共同搭建了一个能快速检测并标记异常放电的长程离线检测系统。
其次，减法聚类常用于初始化自适应模糊推理系统，但只能作用于部分目标参数。为初始化更多的系统参数，本研究提出一种改进的减法聚类算法。在鸢尾花数据集的结果表明，基于改进减法聚类算法能得到总体更稳定、规则更稳定、分类准确率高的自适应模糊推理系统。
最终，基于得到的概率密度特征及改进算法，自适应模糊推理系统在异常放电检测中同样得到了稳定的分类效果。为提升分类准确率，研究将非线性能量算子和小波特征纳入特征集，最终实现了准确率95.13%、精确率93.60%、召回率89.17%。该结果优于SVM、MLP、决策树，并与评估器数量大于等于20的随机森林相当。相比之下自适应模糊推理系统参数较少、具有更高可解释性、且具备识别“模糊而微弱”异常放电现象的潜力。
综上，针对异常放电检测需求，本研究发现了有效的概率密度特征，设计了能提升自适应模糊推理系统稳定性的改进减法聚类算法，通过该算法实现了能有效分类且稳定的、可解释的自适应模糊推理系统，为异常放电检测任务提供了可靠的分类准则。

Abnormal discharge, which can be divided into epileptiform discharges and benign epileptiform variants, is a sharply contoured discharge pattern in EEG. It is often seen  in epilepsy or some neurological disorders, indicating an unstable brain state. Abnormal discharge detection is of great value in cognitive neuroscience research, epilepsy detection typing and focal localization, and pilot screening.

Currently, abnormal discharge detection is still based on manual labeling, which often requires searching for abnormal discharges of less than 0.1 second in duration from hours of  EEG. It will consume a lot of time of experts. To this end, A lot of researches extract features from time and frequency domains to build systems for abnormal discharge detection. However, restricted by the limited public data sets, fuzzy abnormal discharge judgment criteria, and data noise, the algorithm performance is often not ideal. It cannot extract the "fuzzy and weak" abnormal discharges which are controversial among experts as well. Meanwhile, abnormal discharge detection algorithms have focused on epileptiform discharges, but the detection algorithms of benign epileptiform variants are still lacking.

For the needs of abnormal discharge detection, we detect spikes which are abundant in epileptiform discharges and benign epileptiform variants. This study has completed three tasks: abnormal discharge long-range detection, improved design of adaptive neuro fuzzy inference system (ANFIS) initialization method, and ANFIS based abnormal discharge detection. This study is to find appropriate features and establish a stable classification system that can adapt to fuzzy judgment criteria, so as to provide stable, interpretable and easy reference classification criteria for "fuzzy and weak" abnormal discharge detection tasks in the future.

Firstly, considering the characteristics of long time EEG and sparse distribution of abnormal discharges, this study attempts to detect abnormal discharges directly by setting probability density-related feature thresholds. Based on the probability density difference between abnormal discharges and EEG background in the envelope, this study defined the relevant threshold parameters. Mean filtering, spline interpolation were also used to build a long-range offline detection system which can detect and label abnormal discharges quickly.

Secondly, subtractive clustering is widely used in ANFIS initialization, but it can only initialize a part of target parameters. To initialize more parameters, an improved subtractive clustering algorithm is proposed in this study. The results on the Iris dataset show that the improved subtractive clustering algorithm can obtain more overall stable, more rules stable and high classification accuracy ANFIS .

Finally, based on the probability density features and the improved subtractive clustering algorithm algorithm, ANFIS also obtained stable classification results in abnormal discharge detection. To improve the classification accuracy, this study recruits nonlinear energy operators and wavelet features into the feature set, achieving 95.13% accuracy, 93.60% precision, and 89.17% recall. The results outperform SVM, MLP, decision tree, and are comparable to random forest with the number of evaluators greater than 20. In contrast, ANFIS has fewer parameters, higher interpretability, and the potential to identify "fuzzy and weak" abnormal discharge.

In summary, for abnormal discharge detection needs, this study finds effective probability density features and designs an improved subtractive clustering algorithm that can enhance the stability of ANFIS. By those, an effective, stable and interpretable ANFIS is constructed, which provides a reliable classification criterion for abnormal discharge.