How IoT technologies and hardware experts help implement predictive analytics: Factory5 experience
What is on the market and what do we offer
Consider the market proposals for the same three groups of barriers: data, business processes and expertise.
Data
As a rule, typical solutions work with existing data from industrial control systems, historical data storage systems (historian) or files, such as Excel spreadsheets. It happens that development companies say: “Additional sensors are not needed, we work with existing ones.” On the one hand, this is convenient, on the other hand, it is sometimes inefficient, since the quality of the data may not be suitable for predictive analysis.
Our solution, F5 PMM, an intelligent system for monitoring and predicting the technical condition of equipment, uses the IoT subsystem. It provides connection of IoT sensors, gateways, their configuration and management. We also work with partners who help create a complete solution with sensors. If you produce sensors, gateways or network infrastructure for industry, please write to e-mail info@factory5.ai
Business processes
Most often, vendors offer workflow managers as a basic process management tool that allow you to manage emerging events: assign a priority, category, change status, etc. Integration with EAM systems is poorly described, although it is claimed. This is partly due to the fact that such solutions are usually part of the Asset Performance Management (APM) systems package, and vendors are interested in selling as many of their own products as possible. In addition, AWP systems are expensive and difficult to implement.
F5 PMM is different in that we have both native integration with our own MRO automation system (F5 EAM) and an open integration service for any ERP / EAM systems, and these two products form a comprehensive asset management solution. Dashboard and report designers also allow the user to create visualization panels that are convenient for him, consisting of maps, diagrams, graphs and tables.
Expertise
Often vendors have their own closed and universal algorithms to cover the maximum possible range of equipment. The system for each sensor compares the real value with the calculated one at the current time, then the diagnostic rules analyze the value of the total deviation and the contribution of the deviation of each sensor, and an alarm is triggered when the threshold is exceeded. Forecasting is usually limited, this is also due to the specifics of the models – for example, the value of an individual sensor can be predicted for a short period of time.
The concept of F5 PMM is that for simple analysis, the product has an anomaly detector and an expert rules service, and for more advanced analysis, it has customizable predictive models. The detector, using universal algorithms, will help to detect “suspicious” areas in the data, which may indicate pre-failure states. Predictive models are external plug-ins for the system; they can be developed individually for the customer, including by external experts, which provides additional flexibility.
How the F5 PMM equipment condition monitoring and forecasting system works
Our flagship product is the F5 PMM system. It pulls data from different systems: both from SCADA / APCS archives via SQL, and via low-level industrial protocols (for example, OPC). This data is pre-processed, stored and linked to the hardware registry to make it convenient to work with. The system analyzes them both with the help of predictive models and with the help of expert subsystems and rules. The resulting deviations pass through the defect classifier, and the system provides information to the user in the form of reports and dashboards on the state of the equipment, and also transfers the identified incidents to the EAM and ERP systems for assigning bypasses or repair work.
F5 PMM. Equipment condition monitoring and forecasting system
How Factory5 overcomes barriers to predictive analytics adoption
Data
Data is the most important fuel for the development of predictive diagnostic algorithms. To obtain quality data, it is important to ensure that it meets certain criteria. In our opinion, the most important are:
discreteness – an indicator of the amount of data recorded per unit of time: per second or minute. Data about processes that change infrequently, such as the mode of operation of a machine, can be obtained once a second. But when there is a need to track, for example, a change in motor current, there is a need to receive data a thousand times a second, and sometimes more often.
Number of signals (tags) – metric of data completeness and sufficiency. It happens that the equipment simply does not have enough sensors. In this case (we have had such experience), you can add a virtual sensor-sensor to calculate the value of the parameter. Sometimes this is not possible and it is necessary to equip the unit or equipment with additional data acquisition systems.
Volume shows time periods of telemetry, availability of maintenance and repair data, failures. The more often we get data, the more accurately we can develop a diagnostic algorithm.
It is important to determine whether there is an intersection of telemetry time intervals with data on equipment failures, with MRO data and other external information? The more often repair and failure data intersect with telemetry, the more accurate diagnostic algorithms can be developed.
Data sources for predictive analytics and their limitations
How to get quality data?
1. Sensors. They collect information about processes, describe diagnostics and external conditions. Each source has features that sometimes prevent the implementation of predictive analytics in full.
Process data may not contain enough tags, may have delays, small resolution and rounding. For diagnostic systems, portable complexes are used, the data from which are contained in insufficient quantities and are difficult to interpret. Data from external systems is important in answering the question: “Why do the same systems behave differently in different places?”. Influencing factors can be weather and climate, and with any changes in them, it is important to record deviations in the operation of the equipment.
2. Events. Failure and MRO data are often stored in the format of logs and tables, so the human factor plays a negative role here: data is easy to lose, distort or not be recorded at all.
All industries are different: somewhere there is a lot of accumulated data, but somewhere they have to be additionally mined. Each case requires its own approach.
Let’s analyze two illustrative cases: in the first, the already existing data turned out to be enough to build a prognostic model, and in the second, it was necessary to equip the equipment with IoT sensors.
Case 1: Diagnostics of screw and centrifugal compressors based on big data
Given: Air compressors in an assembly plant that need early detection of pre-failure condition of the assembly – so that service work can be scheduled in time and production downtime can be prevented. The data were of good quality: a continuous archive of telemetry from 43 sensors was accumulated from each compressor for 1.5 years with a resolution of 1 minute. Also, the customer provided maintenance and repair logs in excel tables, and in 95% of cases, information about failures intersected with data from sensors.
Solution: It was possible to build a predictive model based on historical data, which used 4 key tags, and predicted pre-failure states for several days on test data with an accuracy of 85%.
Case 2: Traction motor of a locomotive, the failure of which causes great losses
Given: The electric motor that drives the locomotive is a critical node. Its failures entail a number of costly consequences: transportation costs, disruption of production plans, refurbishment. Despite scheduled preventive repairs, engines regularly broke down due to harsh operating conditions – the engine resource was exhausted until the locomotive entered the depot for repairs. Monitoring and forecasting of the state were complicated by the fact that the data from the current sensors were rounded up to 10 A (passport parameters of the motor 800 V / 900 A), the update frequency was insufficient – from 1 to 10 Hz, and the historical archive was not continuous. Some locomotives had no sensors at all. However, the data from the logs on failures and maintenance work performed were collected by the customer over a long period.
Solution: Our technique, based on motor current spectral analysis, required high frequency data (25 kHz), for which the motors were equipped with new current sensors and an IoT gateway to transmit data to the F5 PMM system. Analysis of the incoming data helped to identify failures at an early stage, which made it possible to plan preventive work and avoid costly remedial repairs. Since preventive repairs for this type of equipment are 3 times cheaper than restorative ones, the introduction had a specific economic effect. We talk more about the methodology here.
Data is the fuel for predictive analytics. To obtain the maximum effect, it is desirable to have not only telemetry data, but also logs of failures and maintenance activities carried out, as well as data from external systems (for example, climate). As our practice shows, in some cases the existing data is sufficient to build a predictive model, in others it is not. We can evaluate the suitability of your data for predictive analysis, and if necessary, partners will help with the installation of IoT sensors.
Expertise
Successful implementation of predictive equipment maintenance requires not only data, but also predictive models. “Naked math” in such cases does not work well, to develop a really working model, expertise in equipment is required when:
selection of tags (sensors) – when selecting features that affect the state of the equipment;
setting threshold values;
interpretation of the results of the model.
We can say that a predictive model is a digitized experience of experts.
There are no built-in predictive models in F5 PMM, they are all external – loadable, open and changeable. We see that mature companies have their own models, but they need automation to make these models work. This approach allows you to attract external resources for the development of models: industry companies, research institutes, experts. In world practice, such a trend is already developing – marketplaces of predictive models, where they can be selected and bought according to their needs.
Case: Predictive model for an oil pump with the help of external experts
Given: The oil pump for which you want to develop a predictive model. Fault and production data are available, as well as telemetry for 1+ year.
Solution: Together with industry experts from the St. Petersburg State University of Railway Transport (PSTU), we came up with a description of the engineering model of the process. Hence the concept of the MX-model (Math & eXperience) was born, which combines physical and mathematical models.
When modeling, we took into account the oil pressure in the gearbox, its temperature, the temperature of the oil in the can, the pressure drop across the oil filter, the position of the fuel rails, traction power, and the pressure drop across the assembly. The anomaly was calculated as the difference between the actual pressure and the calculated one. The model revealed anomalies, the frequency of which was used to determine the health index (index of technical condition) and a repair plan was drawn up. The model development process is described in more detail. in another article.
If data is the “fuel” of predictive analytics, then the model is its “motor”. The role of expertise of industry experts is so significant that without it, the first prototypes of models would not give the desired result. Together with the customer’s experts and the university’s specialists, we actually obtained a digital twin that correctly describes the operation of the oil pump.
In any incomprehensible situation, contact Factory5 experts – we will assess the potential of your data, your company’s readiness to implement predictive analytics, and work out the TOR. We have a strong expertise in creating integrated solutions for EAM and MRO automation. The Data Science F5 team will be happy to help you methodologically build your model.
