Abstract: Multivariate time series data often exhibit multi-scale characteristics and complex interdependencies, posing challenges for anomaly detection. To address this issue, an innovative unsupervised multivariate time series anomaly detection model was developed, namely M3AD (Multi-scale Mamba Multi-Layer Perceptron Anomaly Detection). Firstly, a multi-scale feature extraction method was employed, segmenting the time series across different temporal scales to capture both short-term and long-term patterns. Secondly, a MultiLayer Perceptron (MLP) and convolutional layers were utilized for feature learning, extracting local and high-level feature representations. Subsequently, a selective state-space model Mamba was introduced to efficiently capture key information in long sequences. Finally, sensitive anomaly detection was achieved through a loss function based on KL (Kullback-Leibler) divergence and anomaly score calculation. To validate the model's effectiveness, M3AD was compared with seven models, such as AnomalyTransformer and Memto, across four public datasets. Experimental results showed that M3AD demonstrated significant advantages and leading performance compared to existing methods in key metrics such as precision, recall, and F1 score. The results demonstrated the effectiveness and superiority of M3AD in the task of multivariate time series anomaly detection.

Key words: anomaly detection, multi-scale feature extraction, time series, state-space model, unsupervised learning

摘要： 当前多变量时间序列数据常表现出多尺度特征和复杂的相互依赖性，给异常检测带来了挑战。为了解决这一问题，提出一种创新的无监督多变量时间序列异常检测模型M3AD（Multi-scale Mamba Multi-Layer Perceptron Anomaly Detection）。首先，模型采用多尺度特征提取方法，通过在不同时间尺度上分割时间序列，捕捉短期和长期模式；其次，利用多层感知机（MLP）和卷积层进行特征学习，以提取局部和高级特征表示；再次，引入选择性状态空间模型Mamba，通过它高效的处理能力捕捉长序列中的关键信息；最后，通过基于KL（Kullback-Leibler）散度的损失函数和异常分数计算，实现对异常的敏感检测。为了验证模型的有效性，将M3AD与AnomalyTransformer、Memto等7种模型在4个公共数据集上进行了比较。实验结果表明，M3AD在精确率、召回率和F1分数等关键指标上相较于对比方法展现出显著优势和领先性能，验证了M3AD在多变量时间序列异常检测任务中的有效性和优越性。

关键词: 异常检测, 多尺度特征提取, 时间序列, 状态空间模型, 无监督学习