Condition monitoring and maintenance of hydraulic systems are essential to ensure their proper working in various applications. This paper uses multi-task transfer learning, which employs shared information between the datasets and creates a machine learning model that is accurate and generalized for more than one dataset. This paper uses multi-layer perceptron (MLP) as the main model architecture for multi-task transfer learning because MLP can find complex features and non-linear interactions in the dataset. After finding satisfactory correlations between the internal pump leakage and valve condition datasets, the multi-task transfer model was trained and tested for the predictive maintenance of pump leakage and valve condition. An accuracy of 94.33% for internal pump leakage and 93.65% for the switching behaviour of hydraulic valves was found. This paper demonstrates that knowledge transfer between two related datasets can play a vital role in significantly improving data-driven models’ performance.