PHM in manufacturing—prognostics and health management—uses equipment data to assess current condition, identify developing faults and, where possible, estimate how equipment may degrade. For industrial automation teams, its value is practical: earlier warning can give maintenance and operations teams more time to investigate problems and plan interventions before a fault causes an unplanned stop.
Examples from semiconductor and automotive manufacturing illustrate different PHM challenges. Semiconductor applications can focus on monitoring equipment such as vacuum pumps, while automotive applications can examine degradation in power electronics. In both cases, PHM depends on connecting relevant measurements with fault knowledge and a process for turning analysis into maintenance decisions.
Why PHM matters in high-investment manufacturing
Manufacturing facilities depend on capital-intensive equipment and coordinated production processes. An interruption can therefore affect more than the asset that fails. The presentation described two power-outage examples from Korea in 2018: a 30-minute outage associated with a loss of 60,000 wafers and an estimated loss of about US$46 million, and a two-hour outage associated with a production loss of 120 vehicles. These examples show the potential scale of disruption; they are specific historical cases, not a universal estimate of downtime costs.
PHM is one element in a wider manufacturing-control environment. In semiconductor production, it can complement practices such as fault detection and classification, advanced process and equipment control, chamber-to-chamber matching, statistical process control and integrated metrology. The aim is not to replace these methods, but to provide information about equipment health that can support operational decisions.
PHM in manufacturing shifts monitoring closer to the equipment
Traditional equipment monitoring often concentrated on mission-critical assets, used high-performance, high-cost sensors and relied on wired connections and manual inspection. The direction described in the presentation broadened monitoring to less mission-critical equipment, with satisfactory-performance, lower-cost sensors, wireless connections and automatic data collection. That shift can make it more feasible to collect condition data across a wider range of assets.
Automated collection alone does not create a useful PHM program. Teams still need to choose suitable measurements, interpret them in the context of the machine and define how alerts or diagnoses will be reviewed. The objective is a practical chain from sensor data to health monitoring and fault diagnosis—not simply more data in a historian or data lake.
Semiconductor equipment monitoring: vacuum pump condition
Vacuum pumps are one example of equipment that can be monitored using vibration data. A typical workflow places an accelerometer on or near the pump, sends measurements through a gateway for storage and analysis, and presents the results through visualization tools. This creates a basis for checking equipment condition over time and investigating changes.
Use time- and frequency-domain features
Vibration analysis can examine signals in both the time and frequency domains. Time-domain measures such as root mean square (RMS) vibration severity summarize signal magnitude, while kurtosis can help indicate changes in the distribution of vibration peaks. These values need to be interpreted against the machine and its operating context; no single feature identifies every fault.
Frequency-domain analysis looks for patterns associated with particular mechanical conditions. Examples in the presentation include a dominant component at running speed (1X) that may indicate unbalance, harmonic components that can be associated with looseness, and bearing-related frequencies and sidebands that can help reveal bearing defects. Misalignment is another condition considered in the testbed and field-data work. These patterns support diagnosis, but should not be treated as interchangeable or conclusive without suitable validation.
The work described used vibration signals from a small testbed and from a vacuum pump in the field. Field monitoring compared vibration severity and kurtosis over time, illustrating how trends can be reviewed rather than relying only on a single reading. The presentation also reported a strong positive correlation between the two measures in that field dataset; that finding applies to the reported data and does not establish a general relationship for every pump.
Early gas monitoring was another cited PHM application. Real-time monitoring was presented as a way to help prevent gas leaks or explosions. The presentation noted that losses in the event of an accident could reach US$200 million to US$500 million, emphasizing the potential safety and operational stakes without specifying that monitoring alone eliminates the risk.
Automotive PHM: tracking inverter and IGBT degradation
In automotive manufacturing, the presentation focused on inverter power electronics. An inverter converts direct current from the DC link into alternating current for the motor; the system includes components such as diodes and insulated-gate bipolar transistors (IGBTs). The PHM objective described was to develop degradation models for electrostatic-discharge (ESD) damage to IGBTs, supporting prognostics rather than waiting only for a component to fail.
A general remaining useful life (RUL) workflow starts with accelerated testing and sensing. Engineers identify a failure precursor in the electrical signals, combine it with domain knowledge, and develop an empirical or physics-based degradation model. A prediction algorithm is trained using one dataset, then tested on separate data. Its predictions must be checked for accuracy before they can be relied on to inform maintenance planning.
This process makes validation essential. A model that fits test data is not, by that fact alone, proven to predict degradation reliably in operating equipment. PHM teams need to distinguish measurements used for training from testing data and assess how well predictions match observed behavior.
Scaling diagnosis from testbeds to complex assets
Equipment diagnosis also faces a data and generalization challenge. A testbed may provide labeled examples of known conditions, while a real plant can have more complex behavior and large volumes of unlabeled data—including conditions not represented in the original training set. The presentation explored vibration images and deep-learning methods as a way to learn features and support classification or clustering across testbed and power-plant data.
In the described approach, labeled testbed and plant examples supported recognition of known conditions, while unlabeled plant data could be grouped to identify patterns that did not neatly match those categories. That distinction matters operationally: an unfamiliar cluster is not automatically a diagnosed fault. It is a reason for further investigation, especially when the model encounters behavior outside its known classes.
Practical implications for automation teams
These cases point to several considerations when planning a PHM initiative:
- Start with an operational objective. Define whether the priority is fault detection, downtime reduction, maintenance planning, safety monitoring or degradation forecasting.
- Select measurements that match the failure mode. Vibration can support rotating-equipment diagnosis; electrical signals can be relevant to power-electronic degradation.
- Build a validation path. Separate training and test data, and check model performance against observed equipment behavior.
- Plan for unknown conditions. Monitoring systems should make unfamiliar patterns visible rather than forcing every observation into a known fault category.
- Connect analysis to action. A useful result needs to reach the people responsible for assessing equipment and deciding what to do next.
PHM can help manufacturing organizations move from periodic or manual inspection toward data-informed equipment health monitoring. The semiconductor and automotive examples show that approaches differ by asset and failure mechanism, but share a need for appropriate sensing, meaningful analysis and validation. For automation and maintenance leaders, the practical goal is not prediction for its own sake: it is credible equipment-health information that supports timely decisions and helps manage the consequences of unexpected failure.