LST-MADNet: Learnable scattering transform network for multi-modal time-series anomaly detection in industrial manufacturing systems

Robust anomaly detection in heterogeneous, multi-channel time-series data is essential in modern manufacturing systems. Traditional rule-based methods and many existing AI-driven approaches often fall short in capturing the complex temporal, spectral, and cross-modal patterns inherent in such data, frequently struggling with varying sampling frequencies and the inability to identify anomalies arising from inter-sensor correlations. To address this, we propose LST-MADNet, a novel neural network architecture designed for multi-modal time-series anomaly detection. At its core is LSTL-1D, a neural layer that replaces the fixed wavelet filters of conventional scattering transform networks (STNs) with parameterized wavelets. This enables the extraction of invariant time–frequency representations suitable for industrial anomaly detection. LSTL-1D serves as the foundational stage of LSTN, a single-channel anomaly detection model. We extend LSTN to LST-MADNet using various fusion schemes to integrate data from multiple sensors, treating each distinct sampling frequency as a separate modality and capturing critical cross-channel correlations that may indicate anomalies. We validate the proposed framework on a real-world industrial dataset collecting time-series data from three modalities in a robotic screwing process. Our results demonstrate superior performance compared to state-of-the-art deep learning baselines. It is observed that the network is capable of effectively adapting its frequency focus depending on the spectral content of the time-series data, extracting representations that are well-suited for anomaly detection. The introduction of a learnable scattering-based representation and a modular, fusion-oriented architecture represents a significant advancement toward scalable monitoring systems in industrial manufacturing.