As the electric vehicle (EV) industry rapidly grows, the reliability of EVs is an ongoing challenge to the automotive industry. Among them, due to the introduction of electric motors, the aging and degradation of power electronics in EVs have a direct influence on the overall system safety and may lead to total failure. Therefore, extending the lifetime of power electronics has been a focus in the last decades, where reducing the temperature swings is the key to achieving the goal. However, temperature optimization usually requires future information on power loads, which is not available in classic approaches. Therefore, in this paper, we propose a baseline framework for lifetime-oriented temperature control with reinforcement learning (RL). Focusing on long-term prediction, the framework integrates various physical modelings of EV modules (sensors, actuators, vehicle dynamics and temperature management) and utilizes real-time route information to train an agent for driving behavior prediction. Based on further interaction with the EV model, future temperature development can be estimated in advance, thus enabling better swing optimization. Experiments demonstrate the effectiveness of our approach and the lifetime of power electronics can be extended by up to 63% on a representative test route. Compared to the classic approach, our predictive temperature control shows impressive energy efficiency by achieving up to 2.8x power loss reduction with better lifetime optimization.

As a well-developed technology, spectral analysis is intensively utilized in enormous application domains. Despite the variety of spectrometry and spectrometers, classic approaches to spectral analysis typically exploit complex physical modeling to solve the tasks. Although such procedures are powerful and have been widely acknowledged in the academic as well as industry world, they still face many limitations. To address these issues, artificial intelligence (AI)-based approaches have been proposed in the last decades and proven to be successful. Previous work in this emerging interdisciplinary topic enables a fast, accurate and efficient spectral analysis compared to the classic approaches. However, in the meantime, new challenges arise with the introduction of AI methods due to its model-free, data-driven and blackbox behavior. Firstly, as spectral analysis is a highly physics-related topic, there are scenarios where the task is by definition unresolvable but the AI models still try to make predictions, which leads to undesired behavior. Besides, since AI approaches lack interpretability and reliability by nature, in real-world (especially commercial and medical) applications, abnormal results are inevitable and will potentially result in huge losses. In addition, data privacy issue also becomes crucial for cloud-based AI solutions. Thus, in this paper, we thoroughly explain these challenges that current state-of-the-art (SOTA) faces and discuss opportunities for possible outlets in future work.

Recently, neural-network-based approaches for spectral analysis have been proposed and are proven to be successful. However, since neural networks require huge computational as well as storage resources, the application of neural networks may face latency, memory footprint, storage size and power consumption issues, especially on low-power edge devices. In this paper, we propose a first hybrid approach in spectral analysis to optimize and accelerate the neural network execution while consistently preserving the final performance. First, feature selection is performed to reduce the data dimension and guarantee an efficient input throughput. Then, after the neural network training, we further prune the network to obtain a more compact network architecture. Finally, the network will be quantized to reach a higher compression ratio with low-cost operations for edge devices. We conduct extensive experiments on various target hardware platforms to demonstrate the effectiveness of our method, and results show that our approach can achieve up to 52x mode size compression and 600x speedup with even better performance in most cases. As a representative example, we successfully deploy a DenseNet with only a 0.1 MB model size and 0.9 ms inference time on a Raspberry Pi, which enables real-time on-site spectral analysis for industrial and commercial applications.

In many academic and industrial domains, the utilization of spectral analysis is a crucial procedure to extract relevant element information qualitatively or quantitatively. To determine the element concentrations of a measured sample, the classic approaches build physical models and utilize an iterative solver to obtain a precise result of the inverse problem. In the meanwhile, deep-neural-network-based approaches have been introduced to address the limitations of the classic methods and have proven to be successful. However, to deploy the models in industry to achieve a commercial level performance on a vast number of devices, both classes of methods need to conduct device calibration at high costs. Therefore, in this paper, we propose a meta-learning-based approach to achieve excellent calibration results at minimal cost by learning to calibrate. First, we formulate the general spectral analysis problem as a multi-device multi-configuration task that consists of various basic tasks. Then, we train a meta network based on large-scale datasets with a basic task-aware design. Finally, the network is calibrated with a few measurements (few-shot) on an unknown device to optimize the device-specific performance. Extensive experiments show the effectiveness and efficiency of our approach over baseline methods by achieving the best pre-and after calibration error rate across different unknown devices. Besides, compared to previous work after calibration, our approach performs on par even without calibration, it thus makes the zero-shot setting feasible, which is practical in the real-world scenario where an unknown device needs to be deployed without reference samples available for calibration. Moreover, the resource analysis shows that our approach requires significantly less expenditure to deploy large-scale devices in industry, which contributes to a huge saving and growth potential.

In the broad domain of spectral analysis, the identification of present elements is a major task for qualitative analysis and a crucial preliminary for the following quantitative evaluation. Classic approaches require manual work with prior knowledge, which is time-consuming. To improve this process, neural-network-based methods have been introduced in the last decades. However, the scope of work is limited since usually only a small number of elements are covered. Compared to previous work, in this paper, we set up a new baseline by proposing a comprehensive framework capable of identifying the most common (up to 28) elements precisely and efficiently. Various neural network architectures are evaluated on large-scale simulation datasets and real measurements from the industry. Besides, to reduce the computational and data storage cost under big data industrial settings, our approach utilizes our previous work to select important features to reduce the data dimension while maintaining the prediction performance and interpretability. Compared to other dimension reduction baseline methods, our approach outperforms by achieving the best prediction accuracy and providing an intuitive data reduction result. Overall, results on real measurement data prove the feasibility of our approach as a general framework for large-scale spectral identification, and the application of the feature selection method can reduce 80% parameters and 96.9% FLOPs of CNN networks with even better test accuracy on real-world data.