Search Results

The paper presents a solution to the DX'26 Competition LiU-ICE Track. The presented approach merges physics insights with data-driven residuals. The applied Grey-Box approach, involving the custom implementation of a recurrent LSTM network with state memory, is well-suited to environments with high inertia. The system does not directly estimate specific values of individual physical quantities, but focuses on determining their rate and direction of change, eliminating the problem of exploding residuals. Furthermore, this approach offers an advantage over classical LSTMs due to its physical representation of states. Structural analysis and the determination of the MSOs enabled the isolation of relationships between the observed variables, whilst simultaneously decoupling the training data from physical quantities unrelated to the variable under investigation. Data decoupling was leveraged to selectively incorporate faulty data into training. The solution implements four virtual sensors, each equipped with its own pair of neural networks. The proposed fault isolation approach is based on a fault diagnosis matrix, fixed alarm thresholds, and scalar product analysis. Additionally, adaptive covariate shift compensation mechanism was introduced to handle changes in external conditions. The resulting algorithm proved to be highly resilient to measurement noise and capable of accurately analysing the process’s variability, scoring True Detection Rate TDR = 92.26%, False Alarm Rate FAR = 7.86%, and True Isolation Rate TIR = 79.62% on hidden test data.