Prédiction du vieillissement des condensateurs dans les alimentations de polymérisation UV

Prédiction du vieillissement des condensateurs dans les alimentations de polymérisation UV

Abstract

A UV curing line stops for sudden faults and for gradual ones that have already damaged product quality. The gradual kind costs more, because a 2% lamp-power drift produces marginal cures for weeks and then scrap. This article sets out the health indicators, from capacitance drift and ESR rise to the partial-discharge trend, then gives the alarm tiers and six design steps that turn telemetry into a maintenance schedule.

1. Introduction

A UV curing line stops for two kinds of failures. The sudden kind includes a blown IGBT, a shorted capacitor and a tripped breaker.

The gradual kind is noticed only after product quality has already degraded. A 2% drift in lamp power can run for weeks, producing marginal cures that appear downstream as adhesion failures, gloss defects or delamination.

By then the whole production run is scrap. Equipment builders have responded by instrumenting the power supply, so voltage, current, temperature and partial discharge feed a controller that recognises failure signatures and estimates remaining life.

This article covers the health indicators, the telemetry arithmetic, four field pain points, and six design steps that make the prediction reliable.

2. Mécanisme de base

2.1 Capacitance drift and ESR rise

Two classical signatures carry most of the diagnostic weight. A healthy Class I ceramic capacitor holds its capacitance within ±2% over life, and a degrading one drifts 5–10% as electrodes corrode or cracks grow.

The second signature comes from the dissipation factor:

ESR = tanδ / (2π × f × C)

where ESR is the equivalent series resistance, tanδ the dissipation factor, f the measurement frequency, and C the capacitance.

On a 4700 pF part at 1 MHz, tanδ of 0.1% gives 34 mΩ and tanδ of 0.3% gives 102 mΩ. A tripled loss figure moves the impedance of the part by a factor that any in-circuit measurement can see.

The change is measurable with an AC impedance bridge, or in circuit from the ripple voltage.

2.2 Ripple-based health monitoring

At a fixed operating point the ripple voltage across the capacitor follows its impedance:

ΔV = I_ripple × |Z(f)|

where ΔV is the ripple voltage across the capacitor, I_ripple the ripple current through it, and |Z(f)| the impedance magnitude at frequency f.

That impedance has three terms:

|Z(f)| = √(ESR² + (1/(2πfC) − 2πf × ESL)²)

where ESL is the equivalent series inductance of the capacitor. At a constant ripple current, the measured ripple voltage tracks the impedance trend directly.

A 20% rise in that trend is a reliable early-warning threshold, and it appears long before the capacitance leaves its ±2% band.

2.3 Partial discharge as the most sensitive indicator

For high-voltage capacitors the partial-discharge monitor reacts earlier than any impedance trend. A part screened below 5 pC can be monitored against its own baseline.

PD activity that climbs from under 5 pC toward 20–50 pC at constant operating voltage signals void growth or surface degradation, months before dielectric breakdown.

The same reasoning sets the trigger-stage thresholds. For a capacitor rated at 100,000 ignition cycles, a 5% drift in pulse amplitude is the warning level and 10% is the action level.

2.4 The three-tier alarm structure

Thresholds come from the datasheet and the failure mode analysis. The tiers are watch, warn and act:

T_watch = 0.5 × T_warn

where T_watch is the watch threshold and T_warn the warning threshold. With a 20% impedance rise as the warning level, the watch tier sits at 10%.

Remaining life comes from the drift slope, not the instantaneous value:

L_remaining = (D_limit − D_now) / slope

where L_remaining is the estimated life in operating hours, D_limit the action threshold, D_now the present drift, and slope the drift rate per 1000 operating hours.

Fig. 1 — A Class I part stays under 3% impedance rise over 20 kh, while the degrading part crosses the 10% watch tier near 11 kh and the 20% warning tier near 18 kh.

3. Points faibles en matière d'ingénierie

Pain Point 1: Closed-loop masking

The regulator compensates for capacitor aging by extending duty cycle or raising bus voltage. The lamp keeps running at nominal power, so nothing looks wrong.

The compensation range is finite. When it runs out, the failure appears as a trip or a dim lamp with no prior alarm, which is the failure the predictive system was meant to prevent.

Pain Point 2: Sensor placement and bandwidth

The ripple voltage across a DC-link capacitor is small, in the hundreds of millivolts, and it sits at hundreds of kilohertz. An ADC sampling at 1 kHz sees the DC value and misses the signature entirely.

The sense path therefore has to be designed into the board from the start, with wideband routing and a calibration step in production.

Pain Point 3: Data without context

A raw stream of temperatures and currents carries no meaning without baselines. A machine with no commissioning baseline drifts its own normal value, and the first real anomaly looks like a sensor offset.

Baselines must be captured at commissioning and re-anchored on a schedule, or the trend line measures the sensor and not the capacitor.

Pain Point 4: Maintenance execution gap

The system identifies a failing part, and the replacement still needs an authorised technician, the right spare and a maintenance window. Typical MTTR in a densely packed high-voltage cabinet runs 4–8 hours.

If the alarm does not reach the scheduling system, the prediction is read and ignored, and the failure arrives on schedule anyway.

4. Étapes de conception

Strategy 1: Instrument the highest-value signatures

Monitor DC-link ripple voltage on a wideband path, capacitor hot-spot temperature, trigger pulse peak amplitude and leakage current. Those four signatures cover about 80% of UV power supply failures.

A 400 V DC link switching at 200 kHz develops 250 mV rms across the capacitor, which a 12-bit converter resolves to about 0.1 mV per count. Keep the sensing chain short and matched, and use a résistance d'échantillonnage of known temperature coefficient in the leak path so drift in the sensor is separable from drift in the capacitor.

Strategy 2: Anchor baselines at commissioning

Record every signature at first power-up and again after 100 operating hours, and store both sets in the controller.

Re-anchor after each major component replacement. Without that step the fleet data stops being comparable, and a new part reads as an anomaly against the baseline of the old one.

Strategy 3: Trend the slope, not the value

Store the moving average and the slope of each signature. Act on the slope, such as capacitance drift per 1000 hours, because the instantaneous value is noisy and load-dependent.

Drift slope gives remaining life from L_remaining = (D_limit − D_now)/slope, which is the number a maintenance planner can schedule against.

Strategy 4: Choose capacitors with a narrow healthy envelope

Use parts whose behaviour over life is characterised. N4700 Class I capacitors with ±30 ppm/°C and tanδ below 0.1% hold a narrow envelope, so a 20% impedance rise is unambiguous.

Class II parts drift under bias and temperature as a normal behaviour, and that normal drift hides the anomaly you are trying to detect. A condensateur à film bank in the same cabinet has its own envelope, and the thresholds have to be set per dielectric.

Strategy 5: Integrate the maintenance workflow

Link the alarm output to the ERP. An action-tier crossing should generate a work order with the spare part number, the suggested maintenance window and the estimated remaining life.

That step converts a prediction into a scheduled job, which is the only form a busy maintenance department can act on.

Strategy 6: Validate the prediction loop annually

Once a year, compare predicted failures against actual ones and tune the thresholds with the field data. Track the false-alarm rate and the missed-detection rate separately.

Both have to stay below their targets. Thresholds tuned only against false alarms eventually miss real failures, and thresholds tuned only against misses produce alarms nobody reads.

Paramètre de composant	Valeur typique	Condition de test
Tension nominale (série HVCT8G)	10 kV DC	haute tension 1.5× pendant 60 s
Capacitance	4700 pF ±20% (472M)	1 MHz, 1 V efficace
ESR à 1 MHz	inférieur à 50 mΩ	à tanδ 0.1 % ou mieux
Inductance de boucle (accord E40)	en dessous de 8 nH	Espacement des bornes de 40 mm
La resistance d'isolement	supérieure à 10 GΩ à 500 V CC	25 °C, 60 % HR
Plage de température	−40 °C à +125 °C	Série IEC 60068-2
Capacité dv/dt d'impulsion	au-dessus de 50 kV/µs	à la tension nominale

5. HVC N4700 capacitors for monitored supplies

Three requirements drive the choice at this node: a healthy envelope narrow enough that a 20% impedance rise means something, aging data that supports a remaining-life model, and PD low enough that a rise above 10 pC is a genuine anomaly. HVC builds Class I parts for that set, from a 6,000 m² owned factory in Dongguan, across 1 kV to 100 kV.

Six paramètres importants ici

• N4700 Class I dielectric with ±30 ppm/°C and tanδ below 0.1% holds capacitance and ESR inside tight limits over life, so the diagnostic thresholds stay unambiguous and false alarms stay rare.
• HVC publishes typical life-versus-temperature and life-versus-voltage data, which gives a remaining-life algorithm the physics-based model it needs.
• PD below 5 pC at 1.2× rated voltage is guaranteed per part, so a PD monitor can treat any rise above 10 pC as a real anomaly.
• A 1–100 kV range with screw-terminal and disc packages allows the same monitoring technique at the 400 V DC link and at the 20 kV trigger stage.
• tanδ below 0.1% keeps the temperature signature meaningful, because a hot-spot reading then tracks the electrical stress of the part.
• HVC's 40 kV / 441 screw-type capacitor replaced a first-tier international brand in field-service programmes. condensateurs poignée de porte et les disques correspondent aux paramètres de Vishay et les gammes haute tension TDK.

Livraison et support

Standard parts ship in 4–8 weeks, and characterised matched sets in 6–10 weeks. HVC applications engineering reviews the sensor set, the baseline schedule and the alarm thresholds against your failure history.

Send the sensor set, the controller architecture and the failure history. HVC returns a diagnostic design review with aging data, telemetry guidance for ripple, temperature and PD, and samples within two weeks.

6. Questions fréquentes

Q1: Which signal catches capacitor aging first?

Partial discharge on high-voltage parts, and the ripple-voltage impedance trend on low-voltage ones. From ΔV = I_ripple × |Z(f)|, a 20% impedance rise is detectable long before the capacitance leaves its ±2% band, and a PD level climbing from under 5 pC toward 20–50 pC appears months before breakdown.

Q2: What sampling rate does ripple monitoring need?

High enough to see the switching term, which sits at hundreds of kilohertz in a UV supply. An ADC sampling at 1 kHz captures the DC value and nothing else. Put a wideband sense path on the DC-link capacitor at design time, because a retrofit rarely achieves the same coverage.

Q3: How do I set the watch, warn and act thresholds?

Start from the failure mode analysis. Derive the warning threshold from the signature, set the watch tier at half of it with T_watch = 0.5 × T_warn, and set the action tier where the part must be replaced at the next planned stop. For a capacitor rated at 100,000 ignition cycles, 5% pulse-amplitude drift warns and 10% acts.

Q4: Why trend the slope instead of the value?

Because the instantaneous value moves with load and temperature, while the slope of the drift reflects the aging mechanism. From L_remaining = (D_limit − D_now)/slope, the slope also yields the remaining life in operating hours, which is what a maintenance planner schedules against.

Q5: How often should baselines be re-anchored?

At commissioning, again at 100 operating hours, and after every major component replacement. Without re-anchoring, the machine drifts its own normal value, and the first real anomaly becomes indistinguishable from a sensor offset.

Q6: What does an uninstrumented supply cost in downtime?

It appears as unscheduled downtime. Typical MTTR in a packed high-voltage cabinet runs 4–8 hours, and a 2% lamp-power drift can run for weeks before anyone sees the adhesion or gloss failures downstream. A line that loses four hours to a fault a trend would have predicted has paid for the diagnostic system several times over.

7. Références

1. E. Kuffel, WS Zaengl, J. Kuffel, Principes fondamentaux de l'ingénierie haute tension, 2e éd., Newnes, 2000.
2. Mohan, N., Undeland, TM et Robbins, WP, Électronique de puissance : convertisseurs, applications et conception, 3e éd. Hoboken, NJ : John Wiley & Sons, 2003.
3. Avis MIL-HDBK-217F 2, Prédiction de la fiabilité des équipements électroniques 1995.
4. CEI 60384-1:2021, Condensateurs fixes pour équipements électroniques - Partie 1 : Spécifications générales, 6e éd. Genève : CEI, 2021.
5. CEI 60270:2025, Techniques de test haute tension - Mesure des décharges partielles basée sur la charge, 4e éd. Genève : CEI, 2025.
6. GB/T 7354-2018, 高电压试验技术局部放电测量. Publié : 中国标准出版社, 2018.
7. H. Nakao, Y. Yonezawa, T. Sugawara, Y. Nakashima et F. Kurokawa, Méthode d'évaluation en ligne de la dégradation des condensateurs électrolytiques pour la prédiction des défaillances des alimentations à découpage à commande numérique, IEEE Transactions on Power Electronics, vol. 33, no. 3, pp. 2552-2558, 2018. https://doi.org/10.1109/tpel.2017.2691048.
8. Z. Li et al., Méthode de surveillance intelligente de l'état de dégradation des condensateurs à film métallisé par le signal ultrasonique de la décharge auto-cicatrisante, dans les actes du Symposium international CPSS & IEEE sur le stockage et la conversion de l'énergie (ISESC) 2024, p. 206-211, 2024. https://doi.org/10.1109/isesc63657.2024.10785409
9. États-Unis 6,760,215, Condensateur à seuil de claquage à haute tension, DF DeVoe (lié au Presidio), accordé en 2004, https://www.freepatentsonline.com/6760215.html .
10. Utilisation des condensateurs céramiques 3 kV (Discussion sur la stabilité C0G/NP0 vs X7R/Y5S), Forum Retrotechnique, https://forum.retrotechnique.org/t/utilisation-des-condensateurs-ceramiques-3kv/115858 .
11. Équipe d'ingénierie des applications HVC, Condensateur céramique haute tension Ressources techniques, https://www.hv-caps.com