Multi-Label Classification of a Hydraulic System Using Machine Learning Methods¶

C. Molinero Ranera — M.Sc. thesis, Aalto University, July 2022

Industrial hydraulic systems degrade component by component: a cooler loses efficiency, a valve responds more slowly, an accumulator loses pressure. Each failure has a different cost and a different remedy. Condition monitoring asks whether the health of each component can be inferred from the sensor readings of a single operating cycle, so maintenance is scheduled from evidence rather than a calendar. This thesis treats that as a multi-label learning problem.

The learning problem¶

Data points. One data point is a single operating cycle of the hydraulic rig, from a public UCI Machine Learning Repository dataset. There are 2,205 cycles, covering a working circuit and a cooling-and-filtration circuit.

Features. Readings from the test rig's sensor array across the cycle: pressure sensors, a motor power sensor, a vibration sensor, temperature sensors and efficiency sensors. The sensors sample at different rates (roughly 60, 600 and 6,000 values per cycle), so the faster channels are averaged down to a common per-cycle length. Two feature representations are built from the same readings and compared: a flat two-dimensional array with all values per cycle concatenated, and a three-dimensional array keeping each sensor's time series intact. The comparison is deliberate: the flat version suits the classical models, the structured one suits the recurrent network.

A correlation study motivates the preprocessing: four temperature sensors correlate at 1.0, so three of them carry no additional information, and cooler efficiency correlates with cooling power at 0.97.

Labels. Each cycle carries five labels simultaneously, one per monitored element, each taking one of several health-state categories. Four correspond to physical components; the fifth records whether the cycle was stable. The thesis tackles this two ways, as separate multi-class problems, one per element, and as a single multi-label problem. The comparison between those two framings is one of its research questions.

Hypothesis space. Multinomial logistic regression, a decision tree, a random forest, and a long short-term memory (LSTM) recurrent network. The last was chosen to exploit the three-dimensional representation, which the other models cannot use.

Loss function. Two, one per framing. The multi-class networks minimise sparse categorical cross-entropy, the sparse variant because the class labels are integers rather than one-hot vectors. The multi-label networks minimise binary cross-entropy, since there the outputs are binarised and each label is its own yes/no decision. Optimisers differ too: Adam for multi-class, RMSprop for multi-label, both over 100 epochs.

What it found¶

The thesis reports which models reach high accuracy on each element, then asks a question beyond the ranking: whether the high-accuracy models share theoretical structure, or merely happen to fit this dataset. It also answers its own framing question, whether a preliminary per-element multi-class pass is worth doing before the multi-label analysis, and whether maintaining two feature representations improves accuracy enough to justify it.

The correlation finding has the most direct consequence for deployment: if four temperature sensors correlate at 1.0, a future installation can be instrumented with one.

Watch the defence¶

Full text¶

The thesis record is at Aaltodoc. The full text is not open access.