Ruichu Cai

Foundation models enable the reuse of pretrained knowledge across time series classification (TSC) tasks, but transfer to a new task still requires target information to establish the task specific decision rule. Existing surveys mainly organize time series foundation models by architecture, pretraining objective, training data, or source modality, while the target classification procedure receives less attention. We instead organize foundation models for TSC into three transfer mechanisms: representation transfer, fine-tuning and adaptation, and in-context learning. Alongside the survey, we develop and release the TSC-FM Benchmark, a unified evaluation infrastructure covering 198 TSC datasets and multiple target settings. We use the benchmark to study how downstream classifiers, input settings, and task characteristics affect the assessment of pretrained models. The results show that the same pretrained representation can receive substantially different assessments under different downstream classifiers, that the best reported low shot model differs between univariate and multivariate settings, and that method groups lose accuracy at different rates as the number of classes increases. These findings suggest that transfer capability depends on the match between pretrained knowledge, target task structure, and the target procedure, rather than on the pretrained backbone alone. Future research should both expand the effective range of transfer and reliably identify when a target task falls outside that range.

LEARNING POINTS The structural causal model (SCM) is a next‐generation artificial intelligence method. It can transparently reveal the causal graph of a system, make predictions based on causal mechanisms, enable individualised inference and facilitate virtual intervention experiments. Each SCM method has its own advantages and limitations that should be considered in practice. As the field develops rapidly, further breakthroughs in both SCM methods and precision medicine are anticipated. Currently, there are only a few applications of SCM in psychiatry research. We encourage the use of it as a tool for advancing precision medicine in psychiatry.

In recent years, Source-Free Domain Adaptation (SFDA) has gained significant attention in the field of cross-domain fault diagnosis, as it effectively addresses domain discrepancies while preserving the privacy of sensitive industrial data. However, existing SFDA methods typically rely on offline adaptation using unlabeled target domain mechanical data and struggle to fully leverage domain-invariant information in time-series data, posing challenges for real-time cross-domain fault diagnosis. To address these issues, we propose a Time-Frequency Consistency-guided Test-time Domain Adaptation (TFCTA) method. Specifically, we propose a novel Time-Frequency Consistency (TFC) strategy that leverages the frequency domain to extract cross-domain invariant features, thereby more effectively guide the adaptation process in the temporal domain. For test-time adaptation, we design a set of Instance-aware Batch Renormalization (IBR) layers to replace conventional batch normalization layers. First, we employ Batch Renormalization (BR) to normalize features using cross-batch statistics from the target domain, effectively mitigating data distribution shifts during testing. Second, we propose a Distance-based Instance Weighting (DIW) strategy to emphasize representative instances through by assigning higher weights, thus improving the reliability of the statistical computation process. Extensive experiments on multiple public datasets demonstrate that TFCTA significantly improves online cross-domain diagnostic accuracy and outperforms existing test-time adaptation methods, highlighting its potential for practical fault diagnosis applications. Note to Practitioners—In real-world industrial processes, data distribution shifts caused by varying operating conditions (academically referred to as domain shifts or discrepancies) can degrade the performance of fault diagnostic models. Furthermore, the need to protect sensitive industrial data and provide real-time diagnostic results poses significant challenges for cross-domain fault diagnosis. This paper proposes a Time-Frequency Consistency-guided Test-time Domain Adaptation (TFCTA) method to address these issues. The proposed TFC strategy in TFCTA fully leverages domain-invariant information from the frequency domain to better guide the adaptation process in the temporal domain, allowing TFCTA to maintain high robustness even under varying operating conditions. TFCTA only utilizes source data during the source model training phase, and in the adaptation phase, it only requires the pre-trained source model, which helps protect sensitive industrial data. Through the proposed IBR layers, TFCTA can fine-tune the model in real-time based on the input data, providing accurate and reliable fault diagnosis results. As a result, TFCTA is well-suited for cross-domain fault diagnosis in advanced manufacturing, where both data privacy and online diagnostic capabilities are critical.