There are different types of maintenance including Machine Learning in predictive maintenance.

Corrective Maintenance

The machine is in operation until a need for maintenance arises (breakdown, breakage…). The machine is repaired following the breakdown.

Preventive Maintenance

The machine is in operation for a defined period of time (in time, number of cycles). This duration is planned so that the machine does not have time to break down. At the end of this time, the machine is repaired, regardless of its state of wear or its operating cycle.

Predictive Maintenance

The machine is monitored during its operation, and maintenance is planned based on the data collected. Today, this last type of maintenance is the most promising for Factory 4.0. Indeed, it allows to have the lowest possible maintenance costs, since the machines are used at their maximum endurance, but the unexpected and untimely stops are reduced to the strict minimum.

The Prédictive maintenance appeared long before there were any real IT solutions adapted to this need. Today, many industries still use the manual method of sending a technician to go around the machines to collect KPI to detect anomalies.

The appearance of machines and algorithms allowing to automate these tasks prompts the following question: what technique should be used to prédire the state of a industrial machine ?

Artificial intelligence is the discipline of replicating human thinking using an algorithm or other artificial decision making device. The Machine Learning, consists of programmed AI : we will give data to a Machine Learning algorithm, as well as the expected result for these data, and the algorithm will deduce the algorithmic rules: we call this learning.

Within the framework of a predictive maintenance project, several artificial intelligence solutions are available to answer the problem:

The approach Deep Learning, and algorithms of Machine Learning The philosophy behind this is to collect raw, unprocessed data and provide it to an algorithm that is much more complex and powerful than those used in Machine Learning. This algorithm provides models able to analyze raw data and to detect patterns, repeating forms, in order to assemble little by little bits of signal until being able to make a decision.

Each approach has its advantages and its drawbacks. In fact, Deep Learning tends to perform better but in return suffers from a computation time much more important than its competitor (it has to discover the patterns alone, where the Machine Learning model is built based on the patterns).

In Machine Learning, a nice aspect is that the model is not totally a black box, as it has been designed to solve a very well characterized problem, unlike a Deep Learning model, where it is much more complicated to understand the path followed by the algorithm.

The main limitation is the embedded implementation. Indeed, it is much easier to combine embedded, low consumption, with Machine Learning than with Deep Learning, mainly because of the high computation time required by the latter.

Their applications are also different, indeed, Deep Learning is very effective in areas where experts are lacking to build models, these are often very current and complex problems, such as object recognition on images, medical imaging and military, the intelligent word processing, voice recognition… On the other hand́, the Machine Learning approach is confined to well-known domains but in demand of automation, such as predictive maintenance, IoT, fraud detection, etc…

The Tiny ML is a Machine Learning technique implemented in low power systems, such as sensors and others IoT objects. These devices integrate reduced and optimized machine learning applications.  One of the main advantages is that they do not need to be connected to the Internet, all the work is done locally. When the system detects an anomaly, it can then send a simple alert message to the user, we can use a technology radio radio LPWAN low speed to send this message
(LoRa, Sigfox, NB-IoT, LTE-M, 2G…).

ATIM is working on this subject of TinyML and proposes evaluation kits, contact us if you want more information on AI ATIM solutions.