Realize an anomaly detection system for data collected from different sensors located inside an industrial machine for the production of pasta, with the aim to prevent malfunctions and down times. In particular, data is coming from sensors placed on electric motors of fans used to dry pasta, that are subject to frequent failures.
Currently maintenance is carried out on the electric motors of fans when a fault is detected (i.e. after a downtime on the production line). In this context, the possibility to identify anomalies in advance and to predict malfunctions becomes extremely relevant.
Interpreting the sensor data isn’t easy because anomalies are detectable only through the interaction of multiple data sources. Moreover, the frequent change in production, related to different pasta shapes, make the data non homogeneous.
The project was divided into two phases:
- the first concerned the preliminary analysis of quality and quantity of available data through detectiv.ai. The aim was to identify anomalies and to understand the relationship between data and covariates mostly involved in the process;
- the second involved the development of two machine learning algorithms used in sequence to classify the data recorded by the sensors.
The first algorithm classifies data as “normal” (i.e. conforming to the standard behavior of the machine) or “potentially abnormal” (i.e. possible evidence of failure). At all times the system generates a vector that defines the “engine status”. If the vector contains at least one anomalous value, it is given as input to the second algorithm that allows to classify the state as “never seen before” or “similar to one already recordered”. In the second case, the system (through SIMILARITY) automatically calculates a similarity score and labels the anomaly with one of the tags previously inserted by the operator. The system identifies the most similar anomalous case already faced, proposing a solution to the problem based on past experiences (indicating for example “who” and “how” the maintenance was performed in the past) reducing considerably the time needed to identify a solution.
The system implemented allowed to predict malfunctions, reducing the amount of downtime due to failures and collateral damage associated with them. It also allowed to speed up the maintenance operations through the proposal of successful solutions already adopted on similar cases, becoming an important tool for business continuity and digitization of the operators’ practical experience.
The availability of the production machines also resulted in increasing overall plant productivity.