Blind failure prediction
Blind failure prediction is a machine learning method that can predict equipment failure without having to rely on historical failure data. It is based on the creation of a nominal operating space for the device when implementing the solution.
The software uses the equipment’s healthy data to establish a normality space, with any result outside this space being considered as an anomaly and reported to the maintenance teams: this is what we call anomaly detection.
DiagFit software, powered by advanced algorithms derived from years of scientific research, can anticipate failures with high accuracy and low error rates.
Reducing machine downtime, optimizing performance and equipment lifespan, as well as maximizing profitability through real time inventory management are all benefits that encourage manufacturers to invest in failure and wear prediction to stand out from their competitors.
The emergence of “product as a service” allows manufacturers to offer guaranteed availability to their customers. This change in business model is possible thanks to on-board sensors coupled with a failure prediction software, allowing the manufacturer to guarantee a constant quality of service throughout the life of the equipment.
Even if preventive maintenance based on equipment condition and manufacturer’s guidelines has proven itself, it is undeniable that predictive maintenance is attracting more and more industrial companies. This enthusiasm is mainly due to cost optimization regarding spare parts, maintenance interventions and machine downtime. These gains are made possible by anticipating failures and monitoring the actual use of the equipment.
Thanks to the early detection of failure signals, the uptime of the equipment is significantly improved. Moreover, an anticipated failure is one that does not occur in an untimely manner. For many sectors, especially in transport and energy, where human lives are at stake, this parameter is crucial for user safety.
From preventive to predictive maintenance
Even if preventive maintenance based on equipment condition and manufacturer’s guidelines has proven itself, it is undeniable that predictive maintenance is attracting more and more industrial companies. This enthusiasm is mainly due to cost optimization regarding spare parts, maintenance interventions and machine downtime. These gains are made possible by anticipating failures and monitoring the actual use of the equipment.
Improve equipment uptime and safety
Thanks to the early detection of failure signals, the uptime of the equipment is significantly improved. Moreover, an anticipated failure is one that does not occur in an untimely manner. For many sectors, especially in transport and energy, where human lives are at stake, this parameter is crucial for user safety.
Stand out from the competition
Reducing machine downtime, optimizing performance and equipment lifespan, as well as maximizing profitability through real time inventory management are all benefits that encourage manufacturers to invest in failure and wear prediction to stand out from their competitors.
Switching to a service approach
The emergence of “product as a service” allows manufacturers to offer guaranteed availability to their customers. This change in business model is possible thanks to on-board sensors coupled with a failure prediction software, allowing the manufacturer to guarantee a constant quality of service throughout the life of the equipment.
Blind mode, an asset for failure prediction in Industry 4.0
At Amiral Technologies, blind mode has several meanings when it comes to our DiagFit failure prediction software
DiagFit does not need to know the type of equipment or sensor it is monitoring. All this information can be anonymized.
DiagFit does not need to be enriched by the equipment failure history. Learning is based on the nominal operation of the equipment (unsupervised approach).
DiagFit does not need the imported training data to be labeled beforehand.
“ These advantages allow our customers to quickly deploy a failure prediction solution for their equipment. ”
Chalk & talk
Antoine Vignon, Data Scientist at Amiral Technologies
Blind failure prediction
“Creating a blind failure prediction model comes down to characterizing the set of operating points [of the equipment] as best as possible.”
Technological and scientific expertise
DiagFit is based on unique know-how and scientific intellectual property from the French National Centre for Scientific Research (CNRS).
This is why Amiral Technologies’ DNA is made up of a clever mix of innovation, fundamental research on Incremental Machine Learning and operational development expertise.
FAQ
Frequently Asked Questions
An observation is made by the entire industry:
- there is a wide variety of industrial equipment and sensors
- the data generated by the sensors measuring the physical quantities of the equipment are mainly industrial time series, by nature complex to analyze
- and finally there are few or no occurrences of failures in the historical data, thus making the predictive models long to deploy
The only approach to meet these three constraints is the one practiced by Amiral Technologies with its DiagFit software. It consists of implementing a failure prediction software solution based on the creation of a space of normality specific to the equipment, created from healthy data from its sensors. It is a process commonly called “anomaly detection” in machine learning which is based on so-called unsupervised artificial intelligence algorithms. Amiral Technologies implements this approach with its own scientific methods. This leads to models being produced quickly, failure detection very early, and the detection of weak signals never seen before by manufacturers.
It is essential to be able to detect anomalies already known as well as those never observed before. This is the case on test benches, on new equipment put into operation, and on aging equipment.
Blind mode means that Amiral Technologies’ DiagFit software does not need to know the type of equipment it is monitoring, and does not need historical failure data to produce predictive models.
