Experimental Technology and Management

[Objective] In semiconductor manufacturing, the remaining useful life(RUL) of equipment must be accurately predicted to ensure production efficiency and minimize economic losses. However, this task is fraught with substantial challenges. The heterogeneity of multisource sensor data, which encompass various signal types and measurement scales, poses a complex data integration problem. Meanwhile, the scarcity of key failure samples makes it arduous to train reliable prediction models. Traditional prediction methods, which are based on single-dimensional modeling, struggle to capture the intricate physical coupling relationships among different components of the equipment. Moreover, they cannot adequately reproduce the evolution laws of cross-temporal and spatial states during equipment degradation. Hence, these methods have limited prediction accuracy, lack interpretability, and cannot meet the high demands of modern semiconductor manufacturing processes. This study addresses these issues by developing a spatiotemporal joint modeling method that integrates a temporal convolutional network(TCN) with a graph convolutional network(GCN). The joint modeling aims to achieve an in-depth multiscale analysis of the dynamic degradation laws of equipment. [Methods] First, a learnable GCN is constructed. Based on the physical topology of sensors installed on the equipment, the GCN is designed to model the spatial relationships among different sensor nodes. Through a multi-order neighborhood information aggregation mechanism, the GCN effectively extracts the hierarchical spatial correlation features of the equipment. This process allows the model to understand the interactions among different components and their influences on each other in the spatial domain. Next, the TCN with a hierarchical dilated convolution architecture plays a vital role in handling time-series data. The dilated convolution layers capture the long-term trend features of equipment degradation without sacrificing the ability to detect local fluctuation patterns. By employing stacks of multiple layers with different dilation rates, the TCN analyzes the time-series data at various time scales to reveal hidden patterns and trends in equipment degradation. Finally, a spatiotemporal interaction module deeply fuses the spatial and temporal feature data extracted by the GCN and TCN. This integration enables the model to comprehensively examine the spatial and temporal aspects of the state of the equipment, facilitating accurate regression prediction of the RUL. [Results] The performance of the proposed model was evaluated via rigorous experiments on the dataset of an industrial ion etching system. The results demonstrated that the proposed model performed remarkably better than traditional models. Specifically, in terms of the root mean square error(RMSE) prediction index, the proposed model achieved a reduction of 20.6% compared to the long short-term memory(LSTM) model. The LSTM model, which is widely used in time-series prediction, often struggles in capturing complex spatial relationships. In contrast, the proposed model integrating the GCN effectively overcomes this limitation. Compared to that of a convolutional neural network(CNN), the RMSE of the proposed model was lower by 23.3%. CNNs are mainly designed for extracting spatial features in images, and their application in equipment RUL prediction without considering the unique spatiotemporal characteristics of equipment data leads to suboptimal performance. The specialized spatiotemporal architecture of the proposed model provides a more suitable solution. Compared to that of the multilayer perceptron(MLP), the RMSE of the proposed model was lower by 29.9%. The simple fully connected structure of the MLP is incapable of effectively modeling spatial and temporal dependencies, highlighting the advantage of the architecture of the proposed model. Additionally, when compared with those of the classical TCN and the Transformer network, the RMSE of the proposed model was lower by 11.0% and 13.7%, respectively, further validating the effectiveness of the joint spatiotemporal modeling approach. [Conclusions] This study not only provides an innovative and effective method for the predictive maintenance of industrial equipment but also has far-reaching implications for the semiconductor manufacturing industry. Enabling highly accurate RUL predictions, it helps enterprises optimize their maintenance strategies. Such optimization can substantially reduce unplanned downtime losses during wafer manufacturing, improving production efficiency and reducing costs. Furthermore, it paves the way for shifting the paradigm of intelligent manufacturing from a reactive post-failure response approach to a proactive pre-failure intervention strategy. This transition can enhance the overall reliability and competitiveness of the semiconductor manufacturing industry, driving it toward a more intelligent and sustainable future.