In most industrial facilities, a failure prediction begins with a threshold value and a curve. A temperature sensor reads 68 degrees; the threshold is 70, and the curve has been rising steadily for eleven days. Only when these simple methods reach their limits is it worth the effort to use a machine learning model.
Providers of machine learning platforms advertise failure probabilities to two decimal places. For a fleet of twenty machines with three years of failure history, this is rarely the right place to start. The question to ask before any investment is whether threshold values and trend analyses can actually solve the problem.
Threshold values and trend analyses are sufficient when the cause of the failure is known, a single measurement reliably indicates the symptom, and the progression becomes apparent over the course of days or weeks. They reach their limits when multiple parameters interact simultaneously or a failure occurs without any discernible warning signs. For most mechanical components, such as bearings, pumps, and drives, this simple method covers the vast majority of cases. A machine learning model is only worthwhile if there are enough documented failure cases.
What Distinguishes a Threshold Value from a Trend Analysis
A threshold is a fixed limit. If the storage temperature exceeds 70 degrees, the system triggers an alarm. A trend analysis also considers the rate of change. 58 degrees is not cause for concern if this value has remained stable for months. The same value is a warning sign if it has risen from 42 to 58 degrees within three weeks.
The international standard DIN ISO 17359 describes this exact procedure, from the selection of the measured variable to the definition of alarm criteria (DIN ISO 17359, 2018 Edition). It does not require a learning model. Condition monitoring using thresholds and trends has been standard practice for decades; it is not a preliminary step toward a model.
When a Failure Prediction Without Machine Learning Is Sufficient
Four conditions support the simple approach. First, the cause of the failure is known or can be deduced from the experience of the organization’s own maintenance team, for example, bearing damage due to a lack of lubrication. Second, a single measurement reliably indicates the symptom. Third, the damage develops over days or weeks, not seconds. Fourth, the number of historical failures is insufficient for training purposes. If these conditions are met, a well-calibrated threshold value provides much of the benefit promised by a more complex model. This article describes what predictive maintenance can fundamentally achieve Predictive Maintenance: What It Can and Cannot Do.
The Three Maturity Levels of Condition Monitoring
In practice, condition monitoring develops in stages. The first stage involves simple threshold monitoring, with an alarm triggered when a limit is exceeded. The second stage adds trend analysis, which evaluates the rate and direction of a change. Pattern recognition, which considers multiple parameters simultaneously and detects atypical combinations, does not occur until the third stage.
Most companies achieve measurable improvements in their maintenance before they even reach the third stage. Those who skip this sequence and order a model right away end up paying for computing time and data science without gaining any advantage over Stage Two. This position contradicts the sales pitch of many AI providers, but aligns with experience from plant monitoring projects, in which a model is usually only retrofitted once the data situation allows for it.
When Machine Learning Provides Added Value
A VDMA survey conducted in early 2026 shows how many mechanical engineering companies in Germany are already using machine learning today: 43 percent are using it in production, and another portion are in the planning stages (produktion.de, February 2026). According to the provider NaveSight, combined approaches using rules-based systems and pattern recognition achieve a hit rate of 75 to 85 percent in predicting critical events at companies without their own data science department after approximately one year of documented maintenance history (navesight.com, 2026). This order of magnitude demonstrates the impact of the foundational work on the database and rules rather than serving as a standalone argument in favor of more complex models.
A failure prediction using machine learning provides real added value for components whose wear is influenced by multiple parameters simultaneously, such as gearboxes with interrelated temperature, vibration, and load behavior. This approach also makes sense for systems where failures entail high costs, where data collection and model maintenance quickly pay for themselves. For individual pumps or standard motors with clear failure patterns, however, the effort involved is usually disproportionate to the additional benefit.
So many mechanical engineering companies in Germany are already using AI and machine learning solutions in their operations, including for predictive maintenance, and another portion plans to start using them.
VDMA survey, cited from produktion.de, February 2026
Plan the Implementation: Systems, Data, Responsible Parties
It is rare to be able to implement this across the entire machine fleet all at once. It makes sense to prioritize based on failure frequency and failure costs, usually the ten to twenty most critical pieces of equipment at a site. For each of these pieces of equipment, a designated person is needed to set threshold values and check them regularly. A threshold that made sense at the time of commissioning may be too tight or too loose after three years of operation because bearing clearance and lubrication conditions have changed. Without this calibration, either the number of false alarms increases or damage goes undetected despite monitoring. The article „Condition-Based Maintenance: From Calendar to Condition“ discusses how to transition maintenance from a calendar-based approach to a condition-based one.
| Criterion | Threshold and Trend | Machine Learning |
|---|---|---|
| Required History | None; one threshold value is enough to get started | Incidents documented over a period of usually twelve months |
| Cause of failure | Must be known or presumed | Can we learn from patterns? |
| Lead time | Days to weeks for individual metrics | Even with combined, subtle signals |
| Maintenance requirements | Regular Calibration of Threshold Values | Model Maintenance and Follow-Up Training |
| Typical Introduction | A few weeks per facility | Several months, including data collection |
Threshold values fail when dealing with fault patterns that can only be identified by combining multiple measurement parameters, for example, when temperature and vibration remain unremarkable individually but together form a pattern. For new types of equipment without historical data, there is no reference value against which a meaningful threshold can be established. It is nearly impossible to predict how reliable a trained model will be in one’s own system prior to implementation, as even in research, there is a lack of broadly valid long-term data.
Prioritize assets before deciding on models
Those who first understand which machines are the most critical and what data is available can choose between threshold values and machine learning based on knowledge rather than marketing claims.
Frequently asked questions
When does machine learning become worthwhile for maintenance?
Failure prediction using machine learning is worthwhile as soon as there is a sufficient number of documented failures, typically 12 to 24 months of historical data, and threshold values alone result in too many false alarms or too little lead time. For systems with multiple interacting wear mechanisms, such as gearboxes with coupled temperature and vibration behavior, a trained model detects patterns that a single threshold value cannot capture. For clearly defined failure patterns, the simple method usually remains the more cost-effective choice.
What sensors are needed for basic condition monitoring?
For most mechanical components, one or two measurement parameters are sufficient, such as temperature and vibration for bearings and motors, supplemented by operating hours and cycle count. Many machines already provide these values via their existing control systems, so additional sensors are often unnecessary. More important than the number of measurement points is the question of which value actually signals impending damage. This determination is usually made by the maintenance department in collaboration with the machine manufacturer.
How often do thresholds need to be adjusted?
A threshold value that is appropriate at the time of commissioning becomes outdated over time because bearing clearance, lubrication conditions, and operating conditions change, for example, due to a lubricant change or increased system load. Facilities with a functioning condition monitoring system check their threshold values at least once a year against actual damage incidents and after every unplanned shutdown. If this calibration is not performed, either the number of false alarms will increase, or damage will occur without an alarm being triggered beforehand.
torck designs its equipment monitoring system so that the initial failure prediction does not depend on the availability of a machine learning team. Jouvoli collects machine data and identifies thresholds and trends for maintenance, providing the data foundation on which a machine learning model can later be built. For companies with a diverse fleet of machines and limited in-house data science capacity, this phased approach is the more cost-effective option than immediately implementing AI. The page provides more information on getting started Plant Monitoring.