One-Class Classification Constraint in Reconstruction Networks for Multivariate Time Series Anomaly Detection

Detecting the anomalies in multivariate time series (MTS) data is crucial for maintaining the stability of industrial manufacturing processes and biochemical operations. However, current methods often focus on capturing the normal patterns of training data while overlooking the potential of latent representations. This research introduces the hypersphere constraint network (HSC), an innovative self-supervised model for anomaly detection in MTS. This approach uniquely integrates a one-class classification framework to regulate latent distribution. First, the HSC employs a temporal convolutional network (TCN) and a multilayer perceptron (MLP) to extract latent representations of input data, imposing constraints on the latent distribution to achieve a one-class loss. Second, a self-attention mechanism is applied to reconstruct the input data and calculate the reconstruction loss. Anomalies are identified by combining the one-class loss with the reconstruction loss. By integrating one-class classification with a reconstruction-based method, the HSC significantly increases sensitivity to anomalous data, enhancing the distinction between normal and abnormal data. Evaluations on three real-world datasets and a simulated dataset demonstrate that the HSC model outperforms existing state-of-the-art methods in anomaly detection.