STUDY OF DISTRIBUTED/FEDERATED LEARNING APPROACHES WITH RANDOM FOREST FOR SENSOR FAILURE DETECTION IN MANUFACTURING ENVIRONMENT

Sammanfattning: In a manufacturing environment, the continuous operation of production lines is of most importance to ensure optimal productivity and profitability. To maintain efficiency and safety, production lines often rely on systems that utilize data from sensors placed around the manufacturing process. These sensors provide valuable information about various parameters, such as temperature, pressure, and volume, which are critical for monitoring and controlling the equipment. However, if these sensors fail and report incorrect data to the system, it can lead to people being hurt, and cause damage to the machines and the product. It is therefor critical that sensor failures are detected before an accident happens. Machine learning is highly capable of detecting anomalies in sensor data. The problem is that it requires all the data to be centralized to make a prediction. An alternative is to use Distributed Learning (DL) or Federated Learning (FL) approach, which overcomes the limitations of traditional machine learning. DL and FL enables multiple clients to train on data subsets, and send the model parameters to the server for aggregation, minimizing the data needed to be sent. This paper addresses which DL or FL approaches, and which machine learning algorithm can be used together to detect sensor anomalies in data.  The Random Forest machine learning algorithm is selected for its proven usability in detecting anomalies in machine learning. Random Forest is evaluated on an dataset called Modular Ice cream factory Dataset on Anomalies in Sensor (MIDAS) in combination with two distributed approaches: horizontal FL and vertical DL. In the experiments, the horizontal FL server achieves a slightly lower accuracy than the clients do. The vertical DL server performs better than all its clients, but its accuracy is lower than the horizontal FL server.