Fault diagnosis and prognosis (FDP) tries to recognize and locate the faults from the captured sensory data, and also predict their failures in advance, which can greatly help to take appropriate actions for maintenance and avoid serious consequences in industrial systems. In recent years, deep learning methods are being widely introduced into FDP due to the powerful feature representation ability, and its rapid development is bringing new opportunities to the promotion of FDP. In order to facilitate the related research, we give a summary of recent advances in deep learning techniques for industrial FDP in this paper. Related concepts and formulations of FDP are firstly given. Seven commonly used deep learning architectures, especially the emerging generative adversarial network, transformer, and graph neural network, are reviewed. Finally, we give insights into the challenges in current applications of deep learning-based methods from four different aspects of imbalanced data, compound fault types, multimodal data fusion, and edge device implementation, and provide possible solutions, respectively. This paper tries to give a comprehensive guideline for further research into the problem of intelligent industrial FDP for the community.

The existing data-driven methods for the state of health (SOH) prediction of lithium-ion battery are limited by the data quantity, resulting in insufficient generalization performance, prediction accuracy and prediction stability. To tackle this challenge, this paper proposes a novel model, namely the CNN-MMHA, for SOH prediction. The CNN-MMHA incorporates mask mechanisms and multi-head attention mechanisms (MHA) to effectively capture the interdependencies within time series data (i.e. MMHA). Simultaneously, the utilization of the data conversion module and convolutional neural network (CNN) enables the capture of temporal sequences and local characteristics of battery states, accordingly enhancing the predictive ability of MMHA. In this study, This paper execute a sequence of simulation experiments on NASA and CALCE data sets, where this paper compare the performance of a variety of models: Multilayer Perceptron (MLP), Long Short-Term Memory Network (LSTM), CNN, CNN-LSTM, MMHA, and CNN-MMHA respectively. The results demonstrate that the proposed model yields optimal convergence and generalization performance. Furthermore, it has the utmost prediction accuracy and stability. Ultimately, the exceptional performance and real-world applicability of the proposed model are corroborated through its experiments on the real battery.

When a fault occurs, driven by the reconfiguration strategy, the topology of ship power system (SPS) will be changed to isolate the fault area and restore the lost load. However, traditional optimization methods have limitations, such as getting stuck in suboptimal solutions or not offering real-time solutions. This paper proposes a novel deep reinforcement learning (DRL) method to address the dynamic fault reconfiguration problem in real-time for SPS. While considering the load priorities, the fault reconfiguration model is formulated with the goal of the maximum weighted load power and the minimum switching action. Then, a deep Q network (DQN) combined with action mask mechanism, what is called DQN-Mask algorithm, is applied to optimize the switch action. The proposed method enables end-to-end control from the fault node data to the switch sequences, and has migration capability in different scenarios. Two case studies are analyzed based on historical fault datasets of medium voltage DC (MVDC) SPS. The numerical results verify the effectiveness, real-time performance, migration and scalability of the proposed DQN-Mask algorithm.

Ship optimization scheduling, using deep reinforcement learning (DRL), has been extensively researched and implemented. Notably, the deep Q-learning algorithm (DQN) has achieved successful deployment within the optimization scheduling domain. However, there is currently almost no research on compressing and accelerating DQN-based All-Electric Ships (AES) energy scheduling models. This paper proposes a DQN knowledge distillation (DQN-KD) compression algorithm that incorporates the teacher replay memory pool (T-rpm) learning mechanism for compression problem of the DQN-based optimization scheduling model of AES. The DQN-KD algorithm can effectively transfer the knowledge of teacher agent to student agent, and further improve the training efficiency and performance of student agent using the T-rpm learning mechanism. The experimental results conduct on the AES system demonstrate that our proposed compression method is highly effective. Comparing with the teacher model, the Parameters, FLOP and Memory of the student model are significantly reduced by 87.7%, 92.61% and 88.3% respectively. Interestingly, despite these significant reductions, the student agent only experiences a marginal increase of 0.33% in economic consumption compared to the teacher agent. Furthermore, when the Parameters of the student agent are further reduced by 47.5%, and FLOPs by 50.4%, along with a 47.3% reduction in Memory, the resulting increase in economic consumption is only 0.59% compared to the teacher agent. Importantly, even with these notable reductions, the compressed agent maintained strong generalization performance.