Gjurmët

Peak Angle: Drift Online Activation Code Offline

ERROR_GETTING_IMAGES-1

Peak Angle: Drift Online Activation Code Offline

to address this challenge, this paper proposes a dual-net-based training and fault-compensation scheme that is capable of updating the ann parameters by leveraging the online data. the proposed scheme is comprised of two ann models, an offline and an online model, which are activated for mpc and/or anomaly detection, respectively. the offline model learns the normal system dynamics from the historical data to compensate for slow-paced degradation/divergence in the plant dynamics. this model is updated using a meta-optimization approach based on the genetic algorithm, which is developed to simultaneously find the optimal ann model hyperparameters and training algorithm. the online model is developed to capture system dynamics in which the ann model has not been updated for the current control cycle. the online ann model is deployed as a reference signal when the plant produces the offline ann model predictions, and the difference between the two models is used to decide which one will be used for mpc. the switch, which dictates which ann model will be used for mpc, is designed to be robust to the detection of anomalies. for this purpose, we employ the trade-off between the model updating accuracy and the number of samples required for a fault detection. if the number of samples required for detection is high, the model update may be incomplete, and the compensation ability for the anomaly will be low. conversely, if the detection is performed with low number of samples, then the updating process may be overly conservative and may require additional samples to complete. the proposed scheme is implemented in a dual-thread decision maker that enables the online ann model to be updated with minimal waiting time and delay.

the online ann training and updating steps are illustrated in fig. 2 b. the input data is collected from the plant over the prediction time window, and the data of each prediction step is collected at every time step. the data sets are then stored in a local database. next, the collected data are fed into the online anns, and the ann models will then be updated. to determine the performance of the anns, the online and offline ann models will be compared to determine whether the updated model outperforms the offline model. the process of the mpc reconfiguration is illustrated in fig. 3 a. the prediction time window is divided into three phases: (i) the offline ann model is updated to predict the output response at the current time step of the plant, (ii) the updated online ann model is compared to the offline model to determine whether the update is successful or not, (iii) the updated online ann model replaces the offline model. it should be noted that the update process is a continuous iteration between the updated online ann model and the offline model. the previous offline model is not completely replaced by the updated online model, since the updated model continues to be compared to the offline model to determine whether the update is successful or not. a historical data set is fed into the two ann models, and the predicted output response is stored in the local database. the prediction for the next prediction step is then performed. the entire process will continue until all predictions are completed. the benefit of this continuous iteration between the offline and online ann models is that when the anomaly is detected, the accuracy of the updated online model can be evaluated and it will be possible to detect the anomaly even before the anomaly is fully developed, i.e., before the prediction error reaches the predefined threshold. moreover, the effectiveness of the updated online model can be determined, and the system can be reconfigured before the anomaly is fully developed. therefore, the updated online model can reduce the probability of system failure or deterioration of system performance as compared to a fixed offline ann model. the process is illustrated in fig. 3 b. 5ec8ef588b

Peak Angle: Drift Online Activation Code Offline

Peak Angle: Drift Online Activation Code Offline

LATEST ARTICLES:

SPONSORED BY: