KDD时间序列异常检测，使用matrix profile准确率只有

方法参考自https://blog.csdn.net/BulletTech2021/article/details/119080716 @BulletTech2021

使用与文中相同的 matrixprofile lib 计算异常数据位置

使用Fourier Transform计算周期，作为matrix profile计算过程中的window_size

准确率仅有50%， 达不到文中88%的准确率

计算周期函数如下：

def getWindowSize(timeseries, n=5, fmin=0.02):
    yf = abs(fft(timeseries))  # 取绝对值
    yfnormlize = yf / len(timeseries)  # 归一化处理
    yfhalf = yfnormlize[range(int(len(timeseries) / 2))]  # 由于对称性，只取一半区间
    yfhalf = yfhalf * 2   # y 归一化

    xf = np.arange(len(timeseries))  # 频率
    xhalf = xf[range(int(len(timeseries) / 2))]  # 取一半区间

    fwbest = yfhalf[signal.argrelextrema(yfhalf, np.greater)]
    xwbest = signal.argrelextrema(yfhalf, np.greater)

    xorder = np.argsort(-fwbest)
    xworder = list()
    xworder.append(xwbest[x] for x in xorder)
    fworder = list()
    fworder.append(fwbest[x] for x in xorder) 

    if len(fwbest) <= n:
        fwbest = fwbest[fwbest >= fmin].copy()
        x = len(timeseries)/xwbest[0][:len(fwbest)]
        y = fwbest
    else:
        fwbest = fwbest[fwbest >= fmin].copy()
        x = len(timeseries)/xwbest[0][:len(fwbest)]
        y = fwbest
    res_x = -1
    max_y = -1
    for i in range(len(x)):
        if y[i] > max_y:
            max_y = y[i]
            res_x = x[i]
    return int(res_x)