Fault diagnosis plays a critical role in ensuring industrial safety, production efficiency, and product quality. In practical industrial processes, various process variables exhibit distinct sensitivities to different fault types. Traditional methods that directly model and analyze multivariate process data tend to introduce redundant information and ignore the dynamic interdependencies among variables, thereby deviating the fault diagnosis mechanism from actual industrial operating laws. Such issues inevitably impair the model interpretability and degrade the overall fault diagnosis performance. Therefore, to effectively utilize the multivariate information of industrial processes and enhance the model interpretability and fault diagnosis performance, in this paper, a dynamic dependencies of fault-related variables based interpretable fault diagnosis framework is proposed. Specifically, all process variables are screened based on the distribution differences between normal and fault samples for selecting fault-related variables. Subsequently, a new graph structure learning module is constructed to capture the dynamic dependencies among fault-related variables and build a graph, which is incorporated into feature enhancement and fusion process of fault-related variables to improve the representation capability of fault features. After that, fault diagnosis is performed based on the fused features, improving the diagnosis accuracy and interpretability. Finally, three case studies based on the hot rolling process are designed to validate the effectiveness, feasibility and generalization performance. Experimental results show that the proposed method achieves fault diagnosis accuracies of 97.67%, 95.5%, and 98.12% in three cases, respectively.

In the increasingly integrated and complex industrial processes, it is difficult for a single type of data to adequately characterize faults due to the correlation and multi-source heterogeneity of process data. Therefore, to achieve accurate fault diagnosis, multi-source heterogeneous data is necessary to be integrated to obtain comprehensive and multi-scale fault information. However, significant differences in dimensions and structures are exhibited by multi-source heterogeneous data, which may make the high-quality data fusion difficult to achieve. Responding to the above issues, in order to achieve a unified representation and fusion of multi-source heterogeneous data, and enhance the fault diagnosis performance, a unified representation and fusion framework is proposed for multi-source heterogeneous data based fault diagnosis in industrial processes. Specifically, the optimal region containing useful information in the fault image data is first selected as the representative region. Subsequently, tensor fusion is performed on the time series data and the selected region of image data respectively, and the low-rank decomposition is used to obtain the low-dimensional unified vector representation for data fusion. Furthermore, a fault diagnosis module based on the multi-scale lightweight convolutional neural network is constructed in which the multi-scale features are extracted and fused from the fusion vector and then used for fault diagnosis. Finally, to validate the effectiveness and superiority of the proposed framework, comparative experiments based on actual data from the hot rolling process are conducted.

This paper presents a generative data augmentation approach for human activity recognition (HAR) to close the distribution gap between laboratory training and real-world deployment. Despite the recent success of deep learning methods in wearable sensor-based HAR tasks, performance degradation occurs during real-world deployment due to training data scarcity and the vast variability in human activities. In light of this, we aim to enhance the diversity of training datasets by generating new data points within the vicinity of existing samples, as informed by domain expertise. Unlike the commonly utilized methods that augment data by interpolating in data space or feature space, we innovate by applying interpolation in the condition space of a conditional generative model to augment HAR datasets. We use domain-specific knowledge to extract statistical metrics from sensor data, which serve as conditions to direct the generation process. We demonstrate how a conditional generative diffusion model, steered by interpolated conditions, can synthesize realistic new data with various high-level features that benefit the robustness of the downstream HAR models. Our methodology advances the use of interpolation in data augmentation by exploring the capability of a state-of-the-art generative model, offering novel perspectives for bolstering the robustness and generalizability of HAR systems. Experimental results demonstrate that condition space interpolation outperforms the conventional interpolation-based and generative model-based augmentation methods across various datasets and downstream classifier combinations.

Internet of Things (IoT) sensing models often suffer from overfitting due to data distribution shifts between training dataset and real-world scenarios. To address this, data augmentation techniques have been adopted to enhance model robustness by bolstering the diversity of synthetic samples within a defined vicinity of existing samples. This paper introduces a novel paradigm of data augmentation for IoT sensing signals by adding fine-grained control to generative models. We define a metric space with statistical metrics that capture the essential features of the short-time Fourier transformed (STFT) spectrograms of IoT sensing signals. These metrics serve as strong conditions for a generative model, enabling us to tailor the spectrogram characteristics in the time-frequency domain according to specific application needs. Furthermore, we propose a set of data augmentation techniques within this metric space to create new data samples. Our method is evaluated across various generative models, datasets, and downstream IoT sensing models. The results demonstrate that our approach surpasses the conventional transformation-based data augmentation techniques and prior generative data augmentation models.