Over the years, numerous analytical and empirical methods have been developed to prevent machine failure, reduce costs, and increase production capacity. In the present age, companies like M2M Data Corporation and Senseye started out by developing physical models of machines based on data collected from their customers’ facilities. The data consists of parameters such as pressure, speed of motors, particle concentration in lubricants, acoustic data, temperature in the friction generating part of the equipment, and machine unique data sheets. For example, each pump has its unique “pump curves” that correlate pump rotation speed with discharge pressure and flow. A deviation of these above parameters is used to flag the plant personnel of possible machine issues, and could also suggest a specific malfunction.
Looking into the future, startups like Otosense, 3DSignals, Predikto, and Mtell are using machine learning algorithms to look for patterns in machine data and associate them with specific machine defects. Though the algorithms themselves may not be based on any physical models of machine operation, they detect anomalies with respect to an accepted baseline among machines and operation methodologies.
The advantage of data-agnostic models like machine learning algorithms is that a single parameter could be used to derive multiple behaviors in a single piece of equipment. This does not mean it is superior, but it speaks to the ability of these type of algorithms to find patterns in data. Astronomers use machine learning techniques to separate light data from different sources in a unit of space to determine objects such as galaxies, quasars, planets, and galaxy clusters. Companies today in predicative maintenance (PdM) are doing something similar on the factory floor. Instead of light data, they use similar machine learning techniques used by astronomers on sound data from acoustic sensors to differentiate between different sources of sound on the factory floor.
Broadly, machine learning algorithms fall into two categories: supervised and unsupervised. Supervised refers to the need for an algorithm to have an man-in-the loop to manually annotate data, which is similar to physics-based models. The limit is dependent on the amount of annotated data that is available for training, and the features (types of data: pressure, power, flow) that are available. The accuracy of the system is highly variable, since the models have to be fine tuned such that they can be accommodative of machine data and features other than that was used for training. Unsupervised methods do not need annotated data and hence are more “blind” compared to supervised algorithms because they are heavily data agnostic, and rely on being able to distinguish patterns in data to generate clusters based on their similarity. This would not explicitly indicate or can understand a machine issue but instead brings attention to an anomaly. For instance, Otosense uses unsupervised learning methods to provide its predictive maintenance (PdM) analytics. It has experts on staff and their readers to identify clustered sound data from unsupervised clustering and trains a rule-based algorithm layer – which would be the sound recognition models. The sound recognition models are similar to supervised learning algorithms. The promise of unsupervised learning is that these algorithms can monitor machine data for minor changes in machine behavior to indicate a causation of failure that was not recognized in the earlier stages of PdM.
In the following insights, we will explore more about machine learning in PdM. We think that this is critical because machine learning is used to bring advanced analytics for insights in multiple industries, and the PdM market won’t be left out, especially considering the birth of “Industry 4.0”.
By: Venkatesh Konanur