Machine learning optimises drive system performance through condition monitoring

If the reliable operation of a compact drive system is critical for productivity or the protection of the host machine, ensuring its health is crucial. And if performance of the application risks human safety, this elevates the demand to essential.

To achieve this, maxon's new condition monitoring service, called maxon MIND, uses motor signals to provide diagnostics - in advance - about the condition of the drive system including the mechanics driven by the motor, based on machine learning. With the capability to detect early deviations in operation, this enables taking informed maintenance decisions, long before failures cause operational challenges.

As a result, maxon MIND could be key to ensure safety for machines that have direct contact with people, such as tools and robots used in surgical procedures. Alternatively, for applications where breakdowns must be prevented to avoid the financial cost of downtime and machine damage, including many uses across industrial automation, maxon MIND could be a vital service. The system can also save time and cost by basing drive system replacement on accurate data, rather than an estimation of system health, to prevent premature replacement.

A few lines of code - the only requirement

The basic architecture requirement to integrate maxon MIND includes a motor and position sensor connected to any master controller, plus a cloud connection.

maxon MIND operates with either a brushless DC motor with Hall sensors or an encoder, or a brushed DC motor with an encoder, and the system can analyse data from a single motor or a multi-axis system. From this architecture, slight changes in the data will be recognised to alert to potential problems in advance.

Any master controller capable of a current reading every 1ms-5ms (milliseconds) can be used, ranging from a PLC or other proprietary master controller, or even a Raspberry Pi, with a standard internet network connection as the output, using a protocol such as MQTT.

All the data analysis and model predictions happen in the cloud. To achieve this, the only requirement for the user is to add new code to their control environment from a specification provided by maxon, with no additional hardware or software necessary.

Machine learning specific to each application

When data is transferred to the cloud, the machine learning model analyses the data and compares it with the reference state, and any deviation or anomaly is immediately detected and flagged with a warning. While development of the model is based on the physical relationships of a mechatronic system, for each new project, it also learns from the specific detail, based on the real data from the application.

Crucial to the effectiveness of the system's machine learning capability is the approach to generate training data to teach failure detection. The value of maxon MIND is understanding failure potential in advance, which gives time to take action before challenges emerge for real.

In the real world, drive system maintenance issues typically take many months or several years to develop, and developers don't have time to test all permutations of failure over this extended timespan.

To resolve this challenge, the first stage is defining the precise parameters specific to each application. This requires data analysis in partnership with the customer lasting several weeks. But, once this step is defined, the model can be trained with machine learning in a process taking as little as 10 minutes. This thorough approach develops a failure detection model that verifies the precise control parameters, and crucially, identifies deviations to alert.

Qualifying the value of data analysis

The machine learning approach is fast and simple, and application-specific models can be developed with less than 100 reference cycles, taking less than 10 minutes, while diagnostics reports can be created with less than 20 cycles. As a result, the system can be implemented within days.

However, as data quality is vital to develop a model that can deliver reliable insights even under varying operating conditions, the close analysis of the application to define measurement cycles is key. This takes place within a feasibility study that sets out the potential for condition monitoring before any service agreement has been reached.

The Preliminary Data Analysis is the initial stage where the customer provides a CSV file of application data, such as current readings relative to time and position, plus wider variables if available such as temperature or velocity. Within 4-8 weeks, the maxon MIND system provides a prediction on the health status of the drive, which can be verified against actual values to validate the accuracy of the test.

Adding a greater layer of validation, an optional Proof of Concept develops specific models that the customer can install and run live on their own system. Over 8-12 weeks, this phase provides clarification on the accuracy of the condition monitoring data within a realistic setting.

The final stage is the Lifetime Test that simulates and records the operation of a drive across its lifetime but sped up to take place over just 6-12 months, thanks to machine learning. This allows the user to experience maxon MIND, installed on their actual system and using real data, to deliver measurable results over the long term.

Understanding the data

To use the data, plus the insights and alerts the system provides, maxon MIND empowers the user to manage condition monitoring to fit their own metrics. Via an API (Application Programming Interface) the user can integrate the data into their own eco-system.

As a result, maxon MIND has the power to prevent all drive system failures and drive system-related failures, silently monitoring system health in the background, with no additional time or resources required by the user.