Existing zero-shot fault diagnosis methods are typically system-specific and numerically sensitive, which lack adaptive deployment capabilities across heterogeneous photovoltaic (PV) system scales and topologies. Multimodal large language models (MLLMs) emerge as a powerful solution in cross-system generalization, but their adoption in PV fault diagnosis has been limited by the lack of PV knowledge integration and challenges in processing diverse operating conditions. To bridge this gap, an MLLMs-empowered framework for zero-shot PV fault diagnosis is proposed for the first time, which jointly integrates data-driven and knowledge-driven schemes. First, a chain-of-thought-based data augmentation pipeline is constructed to achieve data-knowledge alignment and interpretable results. Second, a two-stage adaptation strategy is specifically designed for PV data to overcome system scales, diverse topologies, and numerical differences. It consists of a Kolmogorov–Arnold networks-based condition adaptive layer embedded in vision transformer and a low-rank adaptation-based PV domain fine-tuning. Third, we design a microservices-based architecture for PV-MLLM deployment that enables flexible component decoupling and adaptive inference, significantly reducing hardware requirements and resource consumption. The proposed method achieves 99.66% and 97.25% diagnostic accuracy on simulated and real-world datasets.

In microservice architecture, each industrial application is decomposed into multiple microservices and deployed on cloud servers to provide timely services to users. However, existing methods rarely optimize service deployment strategies from the user's perspective to reduce the leasing costs of cloud servers. Furthermore, accurately estimating the required number of instances and resource utilization for microservices remains challenging, and decision-making in large-scale scenarios also faces significant timeliness constraints. To solve this problem, this article proposes a Column Generation deployment strategy, which decomposes the microservice deployment problem into a master problem for scheme selection and subproblems for scheme generation and proves the gap between its convergent solution and the optimal solution. A Distributed Column Generation strategy is further introduced to enable efficient problem-solving. Experimental results based on real-world server pricing demonstrate that the proposed method exhibits a high degree of consistency with the theoretically optimal solution. Compared to the baseline methods, it reduces the average total cost of ownership (TCO) for users by 11.3%, while the decision-making time is only 21.5% of that of the comparative methods. At the same time we used our approach to make deployment decisions for real industrial microservices and deployed them on real cloud servers. Compared to the baseline approach, it reduces TCO by 4%, but decision-making is 97% faster.

Data-driven fault diagnosis methods for photovoltaic modules often encounter issues of sample scarcity and insufficient generalization, whereas knowledge-driven multimodal large language models (MLLMs) can comprehend human-summarized prior knowledge for logical reasoning, significantly enhancing the model’s general applicability. This paper proposes a high-generalization diagnostic framework based on knowledge-driven approaches, featuring a state correction vision transformer on MLLMs to solve the problem of sample scarcity, thereby enabling fault diagnosis grounded in photovoltaic knowledge. To mitigate the high computational cost of the large model inference, the framework incorporates an edge-based small model using Support Vector Machines to filter faulty samples. Additionally, a carefully designed photovoltaic knowledge datasets, along with an adaptive fine-tuning method, facilitates efficient domain-specific refinement of the pre-trained model. To address the potential issue of frequent model updates, each framework component is encapsulated as a microservice for easy orchestration. The framework has been deployed and tested on real cloud-edge-end cluster, achieving a diagnostic accuracy of 97.25% and reducing the invocation of large models by 90%.

Voltage prediction is critical for ensuring both safety and operational efficiency of vanadium redox flow batteries (VRFBs) in long-duration energy storage. In this paper, we propose a physics-data fusion model framework for long-term voltage prediction of VRFBs. First, a parameterized Nernst equation is introduced to construct a physically meaningful latent space. Second, the temporal dynamics of hidden state variables are modeled based on electrochemical principles to precisely capture their time-dependent behavior. Subsequently, auxiliary variables are constructed using a physics-guided approach based on the polarization. Finally, a deep neural network is employed as the output layer of the model to establish the mapping between multidimensional feature variables and voltage. Experimental results demonstrate that this approach effectively captures the long-term voltage aging trends under diverse operational conditions, improving both accuracy and generalization performance.

Intelligent fault diagnosis of photovoltaic (PV) arrays in small-sample scenarios remains challenging due to poor model accuracy and generalization. Existing methods fail to simultaneously address issues of varied operation conditions and insufficient samples, leading to the limited applicability of models built by few-shot learning. In addition, factors, such as data transmission and computation costs, also need to be considered. Therefore, this article proposes a cloud-edge collaborative self-correcting-guided generalized contrastive learning framework for small-sample PV fault diagnosis. First, an end-to-end self-correcting model is proposed to eliminate the influence of variable environments. Then, a self-correcting scheme is integrated with contrastive learning to achieve model generalization, and a type screening method is designed to improve model accuracy. Furthermore, a fast fault filtering mechanism is proposed to enhance the algorithm efficiency with cloud-edge collaboration. Both simulation and real data are utilized to validate the proposed method.

With the rapid development of smart manufacturing, existing microservice deployment strategies based on the TCP protocol tend to encounter problems such as network latency, excessive header length, and port exhaustion in high concurrency scenarios. MQTT Broker can relay data between microservices, effectively alleviating these issues. However, the communication overhead of the Broker is significantly related to its deployment and the dependency of microservices. An appropriate deployment solution can significantly improve service quality. To address this issue, this paper models the deployment of MQTT Broker and microservices as a graph with cycles. By introducing auxiliary variables, the NP-hard problem is transformed into a mixed-integer quadratic programming problem, improving problem-solving efficiency and optimizing the microservices deployment strategy to reduce communication overhead. Meanwhile, a real-world heterogeneous cluster was constructed to experiment on fault detection applications, elaborating on the superiority of the microservice deployment strategy designed in this paper. Experimental results demonstrate that our approach reduces communication overhead by 20% in comparison to both the greedy strategy and the K8s strategy. This suggests that achieving an optimal balance between the number of Brokers and microservices is vital for ensuring efficient resource utilization and maintaining low communication latency, particularly as the scale of services expands.

The microservice architecture has gained widespread adoption in smart manufacturing, enabling the collaborative completion of production tasks through the integration of multiple microservices. However, migrating microservices in dynamic environments poses challenges for maintaining production quality and service efficiency. First, there are complex dependencies between microservices, such as layered and chain structures, making microservice migration a difficult process. Second, large-scale production scenarios require rapid decision-making based on high-dimensional variables to adapt to the dynamic environment. Third, microservice migration can cause interruptions, so careful selection of microservices is crucial to minimize production stagnation during migration. To tackle these challenges, we develop an efficient and economical migration strategy ($E^{2}MS$). This approach considers the complex dependencies between microservices and optimizes the system cost by selecting appropriate microservices for migration. We formulate an integer non-convex quadratic programming problem and employ techniques such as variable reduction, penalty functions, and successive convex approximation (SCA) to solve it. The proposed strategy enables efficient decision-making for microservice migration in dynamic production environments and exhibits strong scalability. Our experimental results demonstrate the exceptional dynamic performance of the proposed method, significantly reducing the time required to obtain migration strategies and achieving a 90% reduction in microservice interruptions compared to other methods.

Many computation-intensive industrial applications need to be run in the cloud, which relies on a lot of sharply varying data transmitted from the industrial field. To save the communication bandwidth and ensure data with required accuracy obtained by the cloud, we design a data transmission architecture based on dual prediction scheme and cloud-edge collaboration and a dual-mode algorithm based on deep learning. With the proposed architecture, a deep learning model is deployed and synchronized on the edge and cloud sides. Further, the proposed algorithm can help the cloud for computation with locally predicted data or real-time data from the edge, depending on whether the predicted data are adequately accurate. A physical validation platform including a sensor, an edge gateway, and a cloud server is built, and drastically changing real vibration data are collected to validate the proposed scheme. The results show that the proposed scheme can reduce 88.66% of data transmission while guaranteeing deviations less than 0.1.

In photovoltaic (PV) systems, machine learning-based methods have been used for fault detection and diagnosis in the past years, which require large amounts of data. However, fault types in a single PV station are usually insufficient in practice. Due to insufficient and non-identically distributed data, packet loss and privacy concerns, it is difficult to train a model for diagnosing all fault types. To address these issues, in this paper, we propose a decentralized federated learning (FL)-based fault diagnosis method for insulated gate bipolar transistor (IGBT) open-circuits in PV inverters. All PV stations use the convolutional neural network (CNN) to train local diagnosis models. By aggregating neighboring model parameters, each PV station benefits from the fault diagnosis knowledge learned from neighbors and achieves diagnosing all fault types without sharing original data. Extensive experiments are conducted in terms of non-identical data distributions, various transmission channel conditions and whether to use the FL framework. The results are as follows: 1) Using data with non-identical distributions, the collaboratively trained model diagnoses faults accurately and robustly; 2) The continuous transmission and aggregation of model parameters in multiple rounds make it possible to obtain ideal training results even in the presence of packet loss; 3) The proposed method allows each PV station to diagnose all fault types without original data sharing, which protects data privacy.

The inverter is the most vulnerable module of photovoltaic (PV) systems. The insulated gate bipolar transistor (IGBT) is the core part of inverters and the root source of PV inverter failures. How to effectively diagnose the IGBT faults is critical for reliability, high efficiency, and safety of PV systems. Recently, deep learning (DL) methods are widely used for fault detection and diagnosis. Different from traditional diagnosis methods, DL methods use deep neural networks which can automatically extract the useful representative features from raw data. However, DL methods require large amounts of data, which leads to the high cost of communication, storage, and computation. To tackle these issues, a data-driven fault detection and diagnosis method for IGBT open-circuit faults based on compressed sensing (CS) and convolutional neural networks (CNN) is proposed in this paper. CS is adopted to compress raw signals, and the optimal value of compression ratio (CR) is determined by considering the trade-off between classification accuracy and model training time. The overlap sampling method is adopted for data segmentation. Meanwhile, overlap sampling can also increase the number of training samples and improve the sample correlation. The compressed signals are segmented and reconstructed into two-dimensional feature maps for model training. Finally, compared with CNN of the same structure, the developed CS-CNN model can compress 85% of data without accuracy loss. The performance comparison with the state-of-the-art networks demonstrates that the test accuracy is 98.68% and the model training time is much shorter than other methods.