Bridging Neural Perception And Symbolic Reasoning For Intelligent Industrial Process

Abstract

Automating industrial processes is becoming indispensable in contemporary manufacturing and production. Even though neural perception models utilising deep learning have demonstrated promising results in applications of pattern recognition, anomaly detection, and visual inspection, frequently characterised by a deficiency in interpretability and reasoning ability. In contrast, symbolic AI systems that utilise knowledge representation and formal logic can ensure transparency and logical inference but lack the ability to process real-world sensory information. In this paper, a hybrid neuro-symbolic framework that aims at realising intelligent automation by connecting neural perception and symbolic reasoning is proposed. Framework combines CNNs for feature extraction and perception with a knowledge graph-based symbolic reasoning engine for decision-making and process control. Evaluate the framework on four benchmark industrial data sets, which include MVTec AD, SECOM, Steel Plates Faults, and MIMII, and demonstrate that method outperforms pure deep learning and rule-based methods on all datasets with an average classification accuracy, precision, recall, and F1 score of 95.7%, 94.2%, 93.8%, and 94.0%, respectively. It is also interpretable, which can be ideal in safety-critical industrial environments.