The use case: development and deployment In one refinery plant, a significant incident occurred when a cracked gas compressor tripped due to a flow transmitter malfunction, leading to a catastrophic plant shutdown. The compressor tripped on a ‘Common Excessive Surge Trip Interlock’ from its control system. Investigations revealed that the impulse line of one of the two third-stage flow transmitters had plugged, causing its flow indication to drop to zero while the other transmitter showed a steady flow (see Figure 3 ). Since the anti-surge control system calculates surge flow as an average of these two transmitters, the calculated surge flow erroneously dropped, pushing the compressor's operating point into the surge limit region. Although the anti-surge valve attempted to open, the sluggish and erratic response from the faulty flow transmitter ultimately triggered the trip. Crucially, prior to the development of this solution, there was no existing system capable of alerting stakeholders to sensor malfunctions caused by impulse line choking sufficiently in advance for preventative action. This absence of an early warning mechanism underscored the urgent need for a robust and intelligent solution. This led to the development of this solution. Data used for model training The success of any machine learning model heavily relies on the quality and representativeness of its training data. For
Surge controller
1(FTs)
FT
FT
Intercooler
Intercooler
1st stage
2nd stage
3rd stage
Figure 3 Process block diagram
a metric like Mean Squared Error (MSE) for continuous data, defined as: he loss function quantifies the difference between the original input and ally, this is measured using a metric like Mean Squared Error (MSE) for as: 𝐿𝐿𝐿𝐿 ( 𝑋𝑋𝑋𝑋 , 𝑋� �𝑋𝑋 )= 1 𝑁𝑁𝑁𝑁 ൫𝑋𝑋𝑋𝑋 𝑋 𝑋��𝑋𝑋൯ 2 𝑁𝑁𝑁𝑁 where X represents the original input vector, X ̂ is the reconstructed output vector, and N is the dimensionality of the input. When trained on vast amounts of ‘normal’ flow transmitter data from multiple, redundant sensors, AE learns a compressed, low-dimensional representation of normal behaviour. When presented with anomalous data, such as a reading from a choked impulse line, it struggles to reconstruct it accurately, resulting in a high reconstruction error. By setting a threshold on reconstruction error, any data point where error exceeds this threshold can be confidently flagged as an anomaly, allowing for early detection of even subtle deviations from expected operational behaviour. 𝑖𝑖𝑖𝑖 𝑖𝑖 riginal input vector, 𝑋� �𝑋𝑋 is the reconstructed output vector, and N is the ut. When trained on vast amounts of "normal" flow transmitter data from ensors, Autoencoder learns a compressed, low-dimensional behaviour. When presented with anomalous data like a reading from a ruggles to reconstruct it accurately, resulting in a high reconstruction old on reconstruction error, any data point whose error exceeds this ntly flagged as an anomaly, allowing for early detection of even subtle operational behaviour. ment and deployment significant incident, a cracked gas o flow transmitter malfunction ant shutdown. The compressor xcessive Surge Trip Interlock" from tigations revealed that the impulse Stage Flow Transmitters had indication to drop to zero while wed a steady flow. Since the Anti- culates surge flow as an average , the calculated surge flow erroneously dropped, pushing the point into the surge limit region. Although the Anti-Surge Valve uggish and erratic response from the faulty FT ultimately triggered the elopment of this solution, there was no existing system capable of ensor malfunctions caused by impulse line choking sufficiently in action. This absence of an early warning mechanism underscored the and intelligent solution. This lead to development of this solution. ining ine learning model heavily relies on the quality and representativeness s anomaly detection model, it was crucial to identify a period of tion for the flow transmitters. Historical data analysis revealed that Compressor 3rd Stage Flow Transmitters frequently exhibited arying degrees Recognizing n is minimum d as the 25-09-2022 15:25:02 07-10-2022 03:01:21 Figure 3: Process Block Diagram
Normal: Training data
294.442
19.000
17.000
53.000
18-10-2022 14:33:40
30-10-2022 02:05:53
10-11-2022
13:38:17
Figure 4 Example of sensor data identified as ‘normal’ (training data) after impulse line flushing, showing stable readings
. This period absence of both
Refining India
68
Figure 4:Example of sensor data identified as 'Normal' (training data) after impulse line flushing, showing stable readings.
Powered by FlippingBook