Knowledge Center

Industrial PSU EMI Troubleshooting: A New Filter Peak

  • 16 Nov 2026
  • Powernexu Team

An added EMI filter can reduce switching-frequency noise while introducing a new peak elsewhere. That peak is not automatic proof that the switching power supply has become unstable. It may reflect a resonant gain peak in the filter, or it may accompany an interaction between the filter’s output impedance and the converter’s input behavior. The two problems require different evidence even though damping can be relevant to both.

For EMI troubleshooting in industrial switching power supplies, this distinction matters when a modification improves one part of the spectrum and worsens another. The analysis below concerns input-filter resonance and interaction with a regulated converter. A DC-input buck-converter example helps explain the mechanism; it does not supply component values or an impedance model for an unidentified industrial AC-DC mains input.

One filter has an attenuation response and an output impedance

A filter’s attenuation response describes how a disturbance is transferred between specified ports under defined source and load conditions. An LC low-pass network can have a gain peak near resonance before its attenuation increases at higher frequencies. If the original disturbance contains energy in that region, the filtered observation can be larger there even while the switching fundamental is reduced.

The filter’s output impedance answers another question: how strongly does voltage at its output change when the connected converter changes the current it draws? A pronounced impedance peak can make the source presented to the converter quite different from the stiff supply used during the converter’s original evaluation. This concerns the interaction between two connected networks, rather than just noise transmission through a passive network.

Analog Devices separates these questions in its EMI filter design analysis. Its LTpowerCAD buck-converter example shows a modeled filter resonance peak at 245 kHz, below the example’s switching frequency. The article then separately compares filter output impedance with supply input impedance. The modeled spectrum and the impedance comparison are related design views, but neither substitutes for the other.

A troubleshooting report should preserve that separation. A lower noise marker at the original switching frequency demonstrates improvement at that observation point and condition. It does not describe the new resonant region or establish stable regulation. Conversely, a new narrow peak does not establish a sustained autonomous oscillation. The report needs to say which behavior was actually observed.

A resonant response can amplify disturbance without proving self-oscillation

Consider a hypothetical converter whose original emissions have a switching-related peak and lower-level content at nearby frequencies. An added filter suppresses the first feature but has a resonant gain peak at a lower frequency. The resulting spectrum can contain a new prominent feature where the filter amplifies existing excitation. Nothing in that description alone requires an unstable control system.

A lightly damped network can also ring after a disturbance. A startup event or load transition can excite its natural response, producing several cycles that decay. A spectrum captured during those events may show energy near the resonant frequency. The decay envelope and the conditions that triggered it matter; calling every ringing waveform oscillation conceals the difference between a transient response and a sustained instability.

Now suppose a periodic disturbance repeatedly excites the network before the preceding response has disappeared. The trace can look persistently busy while the excitation remains present. That observation still does not independently establish that the connected converter-filter system sustains an oscillation without that drive. Engineers need the time record and operating context, not only a frequency marker.

These distinctions explain why a frequency below the switching fundamental should prompt investigation rather than an immediate control-loop redesign. Resonance, modulation, operating-mode changes, and an unstable interaction can occupy overlapping spectral territory. Frequency is a useful clue about a possible energy-storage network; it is not a unique label for the cause.

Regulated input behavior makes the source impedance relevant

Over an operating interval in which a converter holds approximately constant input power, its current rises when its input voltage falls. That produces negative incremental input resistance: the small change in current has the opposite sign from the small change in voltage. Incremental describes behavior around an operating point. It does not mean that the equipment is generating net power back toward its source.

For an idealized constant-power input, current equals input power divided by input voltage. Differentiating that relationship gives an incremental resistance of minus input voltage squared divided by input power. The expression helps explain why lower input voltage or higher power can make the magnitude smaller. It is a simplified operating-point model, not a complete frequency-dependent impedance measurement.

Imagine a hypothetical DC-input stage taking 100 W at 24 V. The idealized resistance magnitude would be 24 squared divided by 100, or 5.76 ohms. At an assumed 20 V and the same power it becomes 4 ohms. These are invented conditions for illustrating the relationship, not specifications or stability limits for a commercial industrial supply. They show why a filter evaluated at one operating point may face a different connected load at another.

Texas Instruments’ SSZT993 discussion of input-filter oscillations emphasizes that converter input behavior and filter output impedance must be considered together. Damping can reduce a filter’s impedance peak. However, comparing one calculated resistance with one component’s nominal reactance cannot establish the stability of the installed system. Control dynamics, the actual filter network, and operating conditions remain part of the problem.

An industrial AC-DC input needs its own model

The DC-input explanation should not be carried unchanged onto the mains terminals of an AC-DC unit. A rectifier, bulk storage, power-factor-correction stage, internal filtering, and control loops can all change the input behavior seen at different frequencies. The relevant network is the one actually presented at the location where the additional filter connects.

TI specifically distinguishes the line-frequency behavior of an offline PFC stage from its behavior beyond the current-control-loop crossover. The controller makes the input behave like a positive resistance at line frequencies so current follows the mains waveform, while different incremental behavior can occur at higher frequencies. Treating the entire AC input as one frequency-independent negative resistor would discard that distinction.

Consequently, a DC-bus instability calculation does not produce a mains filter design by changing the voltage number. The product engineer needs the applicable manufacturer guidance or an appropriate measured or validated model at that interface. A purchased AC-DC supply may already contain an input network whose details are not public. An external filter then interacts with an assembly, rather than with the simplified buck stage shown in a teaching diagram.

The practical scope is also different. A low-voltage converter development board can expose design nodes intended for engineering evaluation. A closed industrial mains supply has safety, compliance, and service constraints. This explanation does not recommend bypassing an installed filter, shorting inductors, opening a supply, or moving protective-earth connections as a diagnostic experiment.

Use operating changes to test a hypothesis, not declare a verdict

A new peak that follows a filter modification makes the filter a plausible participant. To strengthen that inference, the development team can compare its predicted resonant region with the actual observation under a documented configuration. Agreement is useful, but the prediction must use effective component behavior and the relevant connected impedances. A calculation with unrelated nominal values can match by accident.

Load dependence adds another piece of evidence. The converter’s operating point, control mode, and effective input behavior can change as demand changes. If the suspicious feature grows only in a particular load range, that narrows the conditions of interest. It does not independently distinguish an impedance interaction from a mode change or another load-dependent disturbance.

The time-domain behavior provides a different view. A decaying response after an identifiable event supports a different description from a persistent periodic variation in otherwise steady operation. Correlating input-node behavior with output regulation can reveal whether the phenomenon is confined mainly to the filter observation or accompanies a larger operating disturbance. Such measurements require an appropriately equipped development laboratory and a safe, documented method.

Differential voltage probe beside a closed server power module

A useful result remains conditional: for example, the modified assembly develops a repeatable feature near the modeled resonance under a defined load, with a recorded time response. That statement is narrower and more actionable than declaring the whole PSU unstable from a single plot. It tells the designer what the model must explain.

Effective component values can move the suspected resonance

A nominal LC value is only an initial description of the network. Capacitance can change with applied bias and temperature, and inductance can change with current. Component resistance and parasitic elements affect damping and high-frequency behavior. These variations can move a resonant feature or alter its amplitude without any change in the part number printed on the bill of materials.

This is particularly relevant when the fault appears after a component substitution advertised as electrically equivalent. A replacement capacitor may have the same nominal capacitance but different effective capacitance or loss in operation. A replacement inductor may have a different current-dependent behavior. The assembly can therefore retain its nominal filter label while its attenuation and impedance curves change.

Analog Devices discusses DC-bias reduction of ceramic capacitance and current-dependent inductance in the context of its filter model. The useful engineering inference is to use the component behavior at the actual operating point. It is not a reason to assume every ceramic capacitor has the same derating or that every ferrite produces the same resonance shift.

Damping also creates trade-offs. A network that reduces a resonant peak may add loss, change transient behavior, or affect attenuation elsewhere. The designer has to retain the required noise suppression while improving the relevant dynamic behavior. Selecting the largest capacitor or adding arbitrary series resistance does not establish that both objectives have been met.

The repair must answer both questions the filter introduced

A modification can solve a resonant emissions peak while leaving the converter-filter interaction unresolved, or improve operating stability while sacrificing attenuation in a required band. Those are different outcomes. The final engineering comparison should therefore retain both the noise observation and evidence about the connected system’s operating behavior.

If the issue was primarily resonant transmission of an existing disturbance, a credible explanation connects the excitation, transfer response, and observed peak. If a filter-converter instability was identified, the explanation also needs the relevant input and output impedances and their operating dependence. One explanation may involve both mechanisms, but neither should be inferred solely from the other.

The industrial integration boundary remains important after the circuit change. A corrected development-board model does not independently establish emissions conformity of the finished cabinet, and an acceptable emissions trace does not certify regulation or startup under every load. Preserve the demonstrated result at the configuration and operating point where it was obtained.

Industrial control cabinet beside a machine enclosure

That is the useful stopping point for this troubleshooting problem: identify whether the new filter peak is a transmission response, part of an unstable interaction, or still unresolved. The next component choice can then address the supported mechanism instead of repeating an ineffective cycle of adding more filtering whenever the spectrum becomes worse.

Share:

Leave a Reply

Your email address will not be published. Required fields are marked *