LNG tank cryogenic pumps are installed vertically inside a storage tank, with the suction (foot) valve positioned at the bottom. This design eliminates the risk of major tank leaks caused by issues with pipes or connections.
Both the pump and motor units are submerged, with the column serving as a guide during installation and also functioning as the discharge pipe that channels the pumped liquid to the top of the tank. Each pump is equipped with an inducer, an axial flow impeller positioned at the lowest possible level in the tank to enhance the Net Positive Suction Head Required (NPSHR). A suction valve is used to isolate the tank contents from the pump column; it is flanged to the lower end of the pump column and is closed by coil springs as well as the hydrostatic pressure of the liquid inside the tank.
Plant operators rely on information from various plant information systems to monitor the pump’s operating conditions and identify any potential issues. However, despite the wealth of information provided, it can be challenging for operators to make clear judgments about possible pump malfunctions due to the numerous operating variables, such as varying tank levels, flow rates, pressure, LNG density, and pump operating hours. Additionally, given the pump's functional setup, condition monitoring is restricted to operational variables, meaning data such as vibration levels, bearing temperatures, and motor temperatures are not available.
In these conditions, an Artificial Neural Network (ANN) can offer valuable support by predicting the motor's energy consumption over time based on any operating condition, and comparing it to actual energy usage to detect potential malfunctions.
The ANN dynamic simulation model considers seven inputs: flow, pressure, inlet temperature, output temperature, tank level, LNG density, and operating time since the last overhaul. The model has a single output, which is the energy consumption per hour. To achieve the desired accuracy in predictions, the hidden layer was set to contain a minimum of 20 neurons.
14,988 data points (simulation steps) were used for analysis.
-
10,082 data points were collected during the pump’s operation leading up to its overhaul. The first data point corresponds to the pump running for 2,828 hours, and the last before the overhaul was logged at 16,126 operating hours.
-
4,857 data points were recorded after the overhaul, starting from zero operating hours (i.e., as-new condition) until a sudden and catastrophic failure occurred at 5,171 hours.
The unexpected nature of the failure, with no prior warning signs, prompted strong interest from plant managers in applying machine learning techniques to explore the potential for predictive analytics.
For training the Artificial Neural Network (ANN), at least 70% of the pre-overhaul data was selected. During this period, the pump was operating under a wide range of conditions, yet performed well, with only normal wear detected during the overhaul—no significant damage was found.
Once the training phase achieved a satisfactory level of accuracy, the ANN model was applied to test pump performance both before and after the overhaul.
The outcomes revealed several noteworthy insights. During the testing phase prior to the overhaul, the accumulated error showed only a slight upward trend. It's important to note that during this period, the pump had already accumulated a significant number of operating hours—much higher than those seen during the training period. This discrepancy could account for the mild increase in the slope of the error curve.
In contrast, after the overhaul, there is a noticeable and consistent rise in the slope of the error curve. This steep increase persists over time, even though the pump's operating hours are comparable to those within the training dataset. Such behavior strongly indicates an abnormal condition or malfunction occurring immediately after the overhaul.
To evaluate the robustness of the ANN algorithm, its results were benchmarked against those produced by other established machine learning techniques used in predictive analytics. This comparative analysis helps support informed decision-making for asset management (Olivencia et al., 2015). For this purpose, two additional algorithms were implemented and tested on the same dataset: Support Vector Machines (SVM) and Random Forest.
The findings revealed that all three models—ANN, SVM (non-linear), and Random Forest—demonstrated comparable predictive accuracy, with Random Forest achieving the best performance in terms of error reduction. Specifically, the Root Mean Square Error (RMSE) values for the testing phase before the overhaul were:
-
ANN: 0.00043
-
SVM (non-linear): 0.00045
-
Random Forest: 0.00038
For the testing phase after the overhaul, the RMSE values were:
-
ANN: 0.0022
-
SVM (non-linear): 0.0023
-
Random Forest: 0.0020
These results confirm that while all models are effective, the Random Forest algorithm slightly outperforms the others in both scenarios.