A HYBRID KALMAN–DEEP LEARNING FRAMEWORK FOR ANOMALY DETECTION IN INDUSTRIAL SENSOR DATA

Reliable anomaly detection in industrial sensor data is challenging due to stochastic noise, nonlinear system dynamics, and complex temporal dependencies. This study proposes a hybrid multi-stage monitoring framework based on a three-channel anomaly detection pipeline that integrates statistical signal preprocessing, deep learning forecasting, reconstruction analysis, and neural classification. Initially, digital signal preprocessing and Kalman filtering are applied to obtain noise-reduced state estimates. The filtered signals are then processed through two parallel branches: an LSTM prediction model for detecting dynamic deviations and an LSTM autoencoder for identifying reconstruction-based anomalies. The resulting prediction and reconstruction errors, together with filtered signal states, are fused into a multi-channel feature vector and classified using a CNN–LSTM neural network. The proposed framework effectively combines temporal forecasting accuracy, reconstruction-based anomaly sensitivity, and local pattern detection capability, thereby improving robustness against noise and enhancing anomaly detection reliability in complex industrial environments. The novelty of this work lies in the unified three-channel anomaly detection architecture that simultaneously integrates statistical filtering, parallel deep learning analysis, and multi-objective optimization within a single monitoring framework. The proposed approach provides an effective solution for real-time industrial monitoring, early fault detection, and intelligent process control.针对工业传感器数据开展可靠异常检测极具挑战，这是由于场景中存在随机噪声、非线性系统动力学特性以及复杂时序依赖关系。本研究提出一种基于三通道异常检测流水线的混合多阶段监测框架，该框架整合了统计信号预处理、深度学习预测、重构分析与神经分类技术。首先，通过数字信号预处理与卡尔曼滤波（Kalman Filtering）获取降噪后的状态估计值。随后，滤波后的信号通过两条并行分支进行处理：一条为长短期记忆网络（Long Short-Term Memory, LSTM）预测模型，用于检测动态偏差；另一条为长短期记忆自编码器（LSTM Autoencoder），用于识别基于重构的异常。将得到的预测误差与重构误差，结合滤波后的信号状态，融合为多通道特征向量，并通过卷积长短期记忆混合神经网络（Convolutional Long Short-Term Memory, CNN-LSTM）完成分类。所提框架有效融合了时序预测精度、基于重构的异常检测灵敏度与局部模式检测能力，从而提升了对噪声的鲁棒性，并增强了复杂工业环境下异常检测的可靠性。本研究的创新之处在于构建了统一的三通道异常检测架构，可在单一监测框架内同时整合统计滤波、并行深度学习分析与多目标优化技术。所提方法为实时工业监测、早期故障检测与智能过程控制提供了高效解决方案。