This study has developed an effective anomaly detection model that overcomes the limitations of traditional methods and includes various production paths and execution time information, in response to complex production processes within a smart factor...
This study has developed an effective anomaly detection model that overcomes the limitations of traditional methods and includes various production paths and execution time information, in response to complex production processes within a smart factory environment. For this purpose, it proposes a methodology that utilizes dynamic recurrent neural networks to accurately model changing processes and predict potential anomalies.
Particularly, this methodology contributes to process prediction and optimization by learning patterns from real-time data, and it can identify specific issues such as exceeding the average time required for assembly processes. The effectiveness of this methodology has been validated through experiments conducted under various production processes and conditions in smart factories, and it is expected to contribute to enhancing the competitiveness and product quality in the manufacturing industry.
The potential for enhancing process transparency and efficiency has been identified through the application of process mining techniques and process mining project methodology (PMPM). In the process mining analysis phase, the focus was on analyzing event log data to identify bottlenecks and observe exceptional cases.
The process data extracted in the analysis phase were subsequently applied to three major deep learning algorithms adopted in this study: DNN (deep neural network), LSTM (long short-term memory), and Bi-LSTM (bidirectional long short-term memory). Each algorithm's strengths and weaknesses were considered in the analysis, enabling the prediction of the total duration of the manufacturing process. Deep learning learns complex data patterns through multi-layer artificial neural networks, while LSTM and Bi-LSTM effectively handle long-term dependencies in sequence data using gate mechanisms. Particularly, Bi-LSTM is capable of recognizingmore complex patterns through bidirectional information.
The performance of these models was evaluated using various metrics including accuracy, precision, recall, F1 score, and loss function. DNN and LSTM showed excellent performance, while Bi-LSTM demonstrated somewhat lower performance. It was revealed that the differences in performance could be influenced by various factors, including data complexity, noise, and sequence length.
This research represents performance under limited conditions, therefore, further studies in various environments are necessary. Particularly, Bi-LSTM requires improvements in performance through hyperparameter tuning and structural optimization. Moreover, as optimal performance is not guaranteed in all situations, research on the appropriate model selection and parameter tuning is crucial. Future research directions include applying this study's approach to predicting various process behaviors and performance indicators.
This will include research on solving the overfitting problem, improving prediction performance through the integration of unstructured data, enhancing the stability of predictions with cross-validation and ensemble techniques, and studying early detection and response strategies for anomalies.
The development of a real-time data streaming anomaly detection and early warning system using the model developed in this study will be an important direction for future research. The results of this research have high applicability in various industries, including manufacturing, and can be used for anomaly prediction and root cause analysis, necessitating further in-depth research.