Monitoring, machine learning, hydraulic pumps, Embedded systems. 

Most processes in the productive sector begin with the generation of kinetic energy performed by electric motors. In operations involving fluids, motors associated with pumps move and vary the pressure of the fluids. Due to the relevance of these equipment, malfunctions can negatively impact the entire production chain, resulting in problems such as decreased efficiency, unexpected downtime, and potential safety hazards to workers. To prevent such losses, strategies for monitoring industrial assets are applied, assessing factors such as the condition and performance of motors and pumps, assisting in decision-making regarding corrective and preventive maintenance. Vibration and current sensors are commonly used for monitoring the condition of electric motors and pumps. However, these sensors, which require contact, have associated installation costs that may make some applications unfeasible. A low-impact alternative in operation is the use of acoustic sensors, which have gained prominence in monitoring systems for their ability to detect anomalies even in early stages. This assessment does not require physical contact with the equipment, allowing for a mobile and portable approach to asset condition monitoring. Regarding the models used for detecting and diagnosing defects in the operation of electric motors and pumps, machine learning techniques are predominantly used, notably neural networks. Therefore, given the relevance of the topic, this article investigates the feasibility of using miniaturized machine learning, known as TinyML, using acoustic sensors for detecting defects in electric pumps. More specifically, for defect detection, an audio classification model based on convolutional neural networks (CNN) was developed, implemented, and embedded in the Arduino Nano platform. To validate the proposal, the "MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection" benchmark was used, containing audio samples of defects from different pump models. To assess the feasibility of the embedded system, the cost-benefit relationship between the complexity of the neural network model and its accuracy in anomaly detection was analyzed. For this purpose, a quantization technique was used to reduce the CNN neural network by approximately 50%. The results obtained indicate an accuracy of over 96% in defect detection, establishing a solution of simple implementation, low cost, and efficiency, potentially applicable in various industrial areas, significantly improving asset monitoring and reducing maintenance-related costs.