Compressors play an essential role in refrigeration. Recent industrial compressors are more efficient, less energy-demanding, and provide higher modularity than ever before. On the other hand, such innovation led to a higher degree of complexity of the equipment, which could undermine the reliability.Every compressor unit has sensors, programmable logic controller (PLC), and human-machine interaction (HMI) modules. The PLC can stop the equipment in case of danger, and the HMI displays the reason why the compressor stopped (e.g. high oil temperature).However, data from the sensors installed on the machinery can be used to obtain more insight to the status of the compressor. Less severe faults can be detected and diagnosed earlier, interrupting in advance the chain of events that leads to more expensive failures.GEA is one of the largest players in industrial refrigeration and produces different models of compressors, such as screw and piston compressors. The company performs maintenance services to their equipment located in different locations worldwide and they made available for this thesis multiple datasets of compressors experiencing different faulty behaviors. The aim of this project is to define methods of fault detection and diagnosis of faulty non-return valves, liquid refrigerant carryover in the compressor crankcase, and an investigation on faulty bearings.Different solutions about fault detection & diagnosis (FDD) in reciprocating compressors have been found in the literature. Among the proposed solutions, two rule-based methods have been developed for the leaking non-return valve. The same fault has then been detected by classifying the data with three supervised machine learning (ML) techniques: decision tree, random forest, and XGBoost. In the end, all the different models have been compared and their respective strengths and weaknesses analyzed. All the ML models showed an advantage in detecting the leaking in non-return valve compared to the rule-based models because the trend of such fault is not always predictable with a series of if-then-else rules.A similar approach has been used for the detection of liquid carryover in the compressor's crankcase. In this case, a rule-based model detected the fault accurately. This was due to the lower grade of complexity in detecting the symptoms of such fault. The ML models required a data augmentation step for the training dataset since the ratio between faulty and non-faulty data was too big. Two data augmentation methods have been used: random oversampling and synthetic minority over-sampling (SMOTE) technique.For the last study case, an investigation and detection of symptoms to detect faulty bearings has been performed. Multiple rule-based algorithms have been defined to detect the symptoms of such fault, while a model-based approach has been proposed to compare the measured power demand with the predicted power demand of GEA's proprietary compressor model.In conclusion, a hybrid approach, rule-based + ML, is the proposed method for the development of GEA's FDD program, since there is no outstanding method that fits all the possible faults. Based on the results of the models developed in this thesis, GEA can choose which algorithm is worth implementing in the FDD program embedded in their compressors.