Fault diagnosis method for metering devices based on improved deep reinforcement learning

Abstract A fault diagnosis method for electric energy metering devices based on improved deep reinforcement learning is proposed to address the problem that traditional fault diagnosis methods are not good at handling complex time-series data and weak fault features, to improve the accuracy and generalization ability of diagnosis. First, through an analysis of the structure and operational parameters of electric energy metering devices, five fault types are identified, laying the foundation for constructing a diagnostic task with clear fault semantics. Then, a classified Markov decision process (CMDP) is defined to model the diagnostic procedure, establishing a fault diagnosis simulation environment. This effectively characterizes the uncertainty in state transitions and the discriminative relationships among multiple fault types, providing an interpretable decision-making framework for the algorithm. Meanwhile, an agent architecture integrating a joint attention mechanism and a 1DCNN-BiCGLSTM network is constructed. The 1DCNN is responsible for extracting local temporal features, while the BiCGLSTM captures long-term bidirectional dependencies, enabling deep representation of multi-scale temporal patterns in fault signals. Furthermore, the incorporation of a joint attention mechanism adaptively allocates feature weights, significantly enhancing the focus on critical fault segments and sensitive indicators, thereby improving the model’s ability to identify subtle anomalies and complex fault patterns. Finally, through interactions between the agent and the diagnostic simulation environment, the optimal diagnostic strategy is learned autonomously via reinforcement learning, resulting in highly accurate fault diagnosis for metering devices. This approach eliminates the reliance on extensive labeled data and manual feature engineering typical of traditional methods. Based on the PyTorch deep learning framework, experimental analysis is conducted on the proposed method, and the results shows that its fault diagnosis accuracy and time are 95.83% and 12.69 s, respectively, which are superior to other comparative methods.