Open AccessArticle Multi-Sensor Data Fusion for Early Warning of Corrosion-Prone Conditions in Closed Zones of a Medical Rescue Aircraft 1 Faculty of Mechatronics, Military University of Technology, 2 gen. Sylwestra Kaliskiego Street, 00-908 Warsaw, Poland 2 Air Force Institute of Technology, 6 Księcia Bolesława Street, 01-494 Warsaw, Poland 3 Institute of Condensed Matter Chemistry and Technologies for Energy, National Research Council, Corso Stati Uniti 4, 35127 Padova, Italy 4 Institute of Condensed Matter Chemistry and Technologies for Energy, National Research Council, Via de Marini 6, 16149 Genova, Italy * Author to whom correspondence should be addressed. Appl. Sci. 2026, 16(12), 5807; https://doi.org/10.3390/app16125807 (registering DOI) Submission received: 10 May 2026 / Revised: 28 May 2026 / Accepted: 5 June 2026 / Published: 9 June 2026 Featured Application The proposed monitoring system is intended for hidden and difficult-to-access aircraft zones, where a first-stage early warning of corrosion-prone conditions can support targeted inspection, local drying, cleaning, and maintenance planning. In this study, the in-service validation concerns the environmental, conductance, and electrochemical sensing layer installed on a medical rescue helicopter. PZT and eddy-current sensing are retained as higher-level follow-up diagnostics within the proposed hierarchy: they have been validated separately under laboratory or controlled conditions, but their installation on the flying helicopter has not yet been performed because the system has only recently entered operational use and additional aircraft instrumentation requires an approved maintenance-access window, dedicated downtime, and a justified operational trigger. The concept may inform other closed-zone monitoring applications, including civil engineering, but the numerical thresholds and sensor configuration require case-specific calibration. Abstract Identifying corrosion-prone conditions early is a major maintenance challenge in closed, hard-to-access structural zones. This paper reports an in-service validation of the first monitoring layer of a multi-sensor data fusion approach for early warning of such conditions in selected closed zones of a medical rescue aircraft. The work covers sensor selection, installation in restricted-access compartments, and analysis of data from helicopter operations. Environmental, conductance, and electrochemical channels are combined to identify persistent conditions favorable to long-term corrosion development and to assign warning levels linked to maintenance actions. The thresholds proposed here are empirical screening criteria from the 82-day campaign, not universal damage thresholds or proof of existing corrosion. PZT and eddy-current sensing are planned as follow-up diagnostic layers in the overall architecture. These technologies have been validated separately under laboratory or controlled conditions but were not installed on the flying helicopter during this initial period. Although persistent severe early-warning episodes were detected, they did not coincide with an approved maintenance-access window suitable for additional PZT/EC hardware installation. The present results therefore characterize the corrosion-prone environment and the likelihood of corrosion initiation, not the type, exact location, pit depth, mass loss, or crack initiation of actual damage. Field inspection evidence of corrosion in hidden zones supports the practical relevance of early warning, while full end-to-end validation of localization and damage-growth monitoring remains future work. 1. Introduction Machine-learning methods are increasingly applied in SHM and corrosion monitoring to combine heterogeneous sensor features, classify damage states, predict degradation trends, and support inspection planning [ 12, 13, 16, 18]. Common approaches in aerospace and infrastructure SHM include supervised classifiers, random forests, and convolutional or recurrent neural networks, particularly where direct experimental data are limited [ 12, 13]. In adjacent engineering fields, CFD-generated data combined with optimized convolutional neural networks have been used to build fast surrogate models for complex coupled fields, showing how physics-based simulation and learning algorithms can extend sparse experimental observations [ 19]. Recent machine-learning-guided screening workflows similarly illustrate how data-driven models can explore large, high-dimensional parameter spaces when reliable training data are available [ 20]. These developments point toward virtual corrosion-scenario generation and threshold optimization in the longer term. The present 82-day two-location campaign is not, however, sufficient to train and validate a robust autonomous ML classifier; the warning logic here is therefore deterministic and threshold-and-persistence-based, while data-driven models are left for future work. 2. Materials and Methods 2.1. Closed-Zone Corrosion Risk in Medical Rescue Aircraft The most critical corrosion-prone zone was identified as the underfloor space of the medical cabin. The aircraft corrosion literature consistently identifies hidden, poorly ventilated, and difficult-to-access compartments as vulnerable to moisture accumulation, contaminant retention, and undetected corrosion progression. Underfloor regions are especially problematic because limited inspectability combines with strong potential for long-term electrolyte retention, which favors localized and crevice-type corrosion [ 1, 2]. For a medical rescue helicopter, the underfloor cavity is exposed not only to water ingress and condensation but also to mission-specific liquids that can penetrate through floor discontinuities, including panel joints, riveted connections, assembly gaps, cable and pipe penetrations, and locally degraded seals. In operational practice, these liquids may include blood, saline, infusion fluids, glucose solutions, drug solutions, disinfectants, and residues from cabin cleaning and decontamination. Once trapped beneath the floor, such fluids may persist much longer than on exposed cabin surfaces, creating a local environment with elevated humidity, surface contamination, increased electrolyte conductivity, and prolonged time-of-wetness. These are the primary drivers of corrosion initiation and propagation identified in aircraft corrosion research [ 1, 2]. Direct inspection of this zone requires removing floor panels—a labor-intensive process that typically withdraws the helicopter from service for a period. Continuous, indirect monitoring is therefore well-suited here: it can indicate whether inspection is necessary and whether floor removal can be deferred within a condition-based maintenance strategy [ 1]. 2.2. Sensor Suite Selection Reliable corrosion assessment cannot rely solely on observing already developed damage. It requires combining information on environmental severity, electrolyte presence, surface contamination, and electrochemical activity, with structural consequences detectable by SHM/NDT methods as a complementary layer. Recent reviews confirm that no single sensing category is sufficient to characterize corrosion risk in hidden and complex service environments [ 2]. Based on the sources reviewed here, the most deployment-ready option for onboard corrosion monitoring is an aircraft-qualified integrated sensor suite developed by Luna, comprising an air-temperature and relative-humidity sensor, a surface-temperature sensor, a surface-conductance sensor, a free-corrosion sensor, and a galvanic-corrosion sensor. This conclusion is based on the combination of functional coverage and formal implementation maturity: the manufacturer explicitly states that the system is qualified for flight safety, operates independently of the aircraft’s electrical power system, and continuously records both environmental and corrosion-related parameters over extended periods [ 6, 7]. The aircraft-qualified sensor suite used in the present study was selected for immediate deployability and flight-safety maturity. Nevertheless, several emerging sensing technologies are relevant for future generations of CHM systems in enclosed compartments. Flexible iontronic sensors exploit ionic dielectrics and electrical double layers to obtain high sensitivity, adaptability, durability, and reduced susceptibility to electromagnetic interference [ 24]. Bio-inspired ion-gating and iontronic neuromorphic concepts offer low-power adaptive signal processing at the sensor interface [ 25]. Multimodal artificial-skin and stretchable-sensor studies demonstrate how heterogeneous, conformable sensors can be combined with learning architectures for pattern recognition and perception-oriented data fusion [ 26, 27]. Recent electret and heterointerface-based sensing concepts also show potential for resilient and adaptive perception under changing mechanical and environmental conditions [ 28]. These technologies are not yet direct replacements for aircraft-qualified corrosion-monitoring hardware, but they point toward future conformable, wireless or drift-mitigated, pH/chloride-sensitive, and thin-liquid-film sensing layers that could help distinguish condensation from mission-related fluids in closed helicopter zones. 2.3. Sensor Installation in Underfloor and Closed Zones The installation strategy was based on the assumption that the underfloor space was the primary critical zone. Sensors were therefore prioritized for the lowest fuselage regions and for locations where liquid migration and retention were most likely, including floor-panel joints, riveted interfaces, fastening regions, penetrations, and local structural discontinuities. The objective was not merely geometric coverage but rather representation of the most aggressive local microenvironments within the closed compartment [ 1, 2]. 2.4. Data Acquisition During Operation The monitoring concept relied on in-service data acquisition rather than laboratory-only or post-maintenance observations. Atmospheric corrosion conditions in aircraft depend strongly on actual mission histories, operational states, thermal transitions, and contamination events, and TM0416-2016 explicitly emphasizes the value of continuous electrochemical and conductance records over time for capturing temporal variability that occasional inspections cannot resolve [ 3]. 2.5. Data Preprocessing and Feature Extraction After retrieval, the raw sensor data were preprocessed: completeness checks, time-stamp synchronization, identification of missing or non-physical records, filtering of short-term disturbances, and segmentation into intervals representing distinct operational states. This stage was treated as the basis for maintenance-relevant interpretation, not a routine data-cleaning step [ 1, 3]. Relative humidity alone does not define a corrosion-relevant state. The local microclimate and atmospheric corrosivity were assessed from the combined behavior of humidity, surface conductance, electrochemical activity, and state duration. A high-humidity episode without electrolyte formation or electrochemical response was not treated as equivalent to a sustained state with elevated conductance and measurable corrosion signals. The threshold values are screening criteria for early warning, not direct damage thresholds. Their role is to indicate that the monitored zone has entered conditions that, if sustained, could favor corrosion initiation. Direct confirmation and localization of damage are addressed by the higher-level follow-up sensing layers [ 2, 3]. 2.6. Multi-Sensor Data Fusion and Warning Logic For traceability, a persistent episode was defined algorithmically rather than visually. An episode starts when all conditions assigned to a given alert level are satisfied in the synchronized time series. It ends when the same conditions are no longer satisfied for longer than one nominal sampling interval. Because the median sampling interval was approximately 30 min, events shorter than 1 h were treated as transient and were not used to define persistent corrosion-prone states. A single missing or isolated sub-threshold point not exceeding one nominal sampling interval may be merged with the adjacent episode if the threshold condition resumes immediately afterwards; longer interruptions define separate episodes. This definition was used to distinguish short environmental fluctuations from sustained wetness or electrochemical activity relevant for maintenance decisions. In the third stage, eddy-current sensing is used for local verification and follow-up monitoring of metallic degradation. When alternating current flows through the probe coil, it generates a time-varying magnetic field that induces eddy currents in the conductive substrate. These induced currents generate a secondary magnetic field that modifies the impedance of the sensing coil. Corrosion-induced material loss, pitting, cracking, subsurface discontinuities, or changes in lift-off alter the eddy-current distribution and, consequently, the measured amplitude and phase response. Once a suspect region has been identified, eddy-current sensing is intended to assess whether material loss or corrosion-related discontinuities are present and to monitor their growth over time [ 5]. Within this hierarchy, the environmental and electrochemical channels define the overall corrosion-prone state of the monitored zone, PZT sensing is intended to narrow the assessment to a suspect structural subregion, and eddy-current sensing is intended to provide local confirmation and follow-up observation at a specific conductive detail. Direct metrics such as pit depth, mass loss, crack initiation, or damage-growth rate cannot be derived from the currently installed environmental and electrochemical sensor sets alone. They require either physical inspection or a local NDT/eddy-current layer installed at the relevant detail. This distinction is particularly important for hybrid interfaces, fastening regions, and local joints between composite panels and the aluminum helicopter structure, where corrosion may initiate in a highly localized manner and then evolve under repeated service exposure. 3. Results Results focus on Sensor sets 1 and 2, the two most informative horizontal installations in the campaign. Sensor set 1 was installed beneath the cabin floor in a closed underfloor zone; Sensor set 2 was installed below the main gearbox. Their schematic locations are shown in Figure 1. Sensor set 1 captured prolonged underfloor wetting episodes; Sensor set 2 recorded the clearest condensation-driven activation sequence. Interpreting these results requires supplementary context, including basic flight-operation data (when and how the helicopter was operated, typical flight profiles, mission frequency, and flight altitudes) and meteorological data from nearby weather stations. This information allows reconstruction of the environmental and operational conditions during the campaign and supports attribution of increased moisture signals to ambient humidity, temperature effects, or liquid ingress from operational fluids such as water, blood, or other agents used during missions. 3.1. Sensor Deployment Feasibility Both sensor sets remained operational throughout the campaign (24 January 2026 to 16 April 2026), recording 3935 records (Sensor set 1) and 3938 records (Sensor set 2). The median sampling interval was 30 min for both; maximum observed gaps were 35.0 min and 33.8 min, respectively; and 99.87%/99.90% of intervals fell within 30 ± 5 min for Sensor sets 1 and 2, respectively. Autonomous data logging was therefore feasible for long-duration helicopter operations without significant loss of continuity. The signal ranges were sufficient for structural health monitoring ( Figure 2). Sensor set 1: RH = 21.4–100.0%, air temperature = −10.9 to 24.7 °C, Cond Lo = 0.005–1.0 µS, and Cond Hi = 5.0–241.4 µS. Sensor set 2 covered RH = 6.6–98.7%, air temperature = −11.7 to 47.5 °C, Cond Lo = 0.005–1.0 µS, and Cond Hi = 5.0–1894 µS. Repeated returns to the dry-state baseline (Cond Lo ≈ 0.005 µS) after wet episodes confirm that the sensors did not remain artificially latched in a high-wetness state. The signal-continuity metrics and final corrosion totals for both sensor sets are summarized in Table 2. 3.2. Representative In-Service Signals The two sensor sets captured two distinct classes of in-service response ( Figure 3). Sensor set 2 showed a condensation-driven activation sequence. Between 5 and 14 February 2026, RH rose from dry conditions to sustained levels of 80–90%, while Cond Lo rose from the dry baseline to values approaching 1.0 µS. The dominant galvanic-corrosion episode occurred from 7 February 14:50 to 8 February 08:20, lasted 18 h, and accumulated 0.001066 C, i.e., 50.7% of the final total galvanic charge. The dominant free-corrosion episode followed from 12 February 00:20 to 13 February 12:50, lasted 37 h, and accumulated 0.000798 C, i.e., 57.8% of the final total free-corrosion charge. This pattern reflects the gradual build-up of a persistent electrolyte film on the horizontal surface below the gearbox. Sensor set 1 showed a far more persistent underfloor wetting signature. Over the full campaign, it reached Tot Free Corr = 0.007995 C—well above Sensor set 2 (0.001381 C)—while its galvanic total was comparable (0.002231 C vs. 0.002103 C). The dominant free-corrosion episode (22 February 12:09 to 24 February 10:39, 47 h) contributed 0.002710 C—33.9% of the final free-corrosion charge. The dominant galvanic episode (27 February 14:39 to 28 February 08:39, 18.5 h) contributed 0.000907 C—40.7% of the final galvanic charge. A notable event on 18 March 2026 illustrates the underfloor trapping mechanism. Sensor set 1 remained near RH = 98–100%, and its conductance proxy increased rapidly to Cond Hi = 241.4 µS, whereas the other sensors in the aircraft simultaneously dropped to roughly 19–35% RH by about 09:00. According to the operational notes accompanying the analysis, these abrupt excursions occurred during flight missions and dropped rapidly after the floor access hatch was opened. The signal shape is more consistent with localized liquid ingress or trapping beneath the floor than with ambient humidity alone. Accidental spillage or leakage of transported fluids is a plausible cause, but the present sensor set does not identify the fluid chemistry directly. The dominant corrosion-related charge-accumulation episodes are summarized in Table 3. 3.3. Corrosion-Prone Condition Indicator The RH–conductance space offers a practical way to distinguish merely humid conditions from truly corrosion-prone conditions ( Figure 4). Sensor set 2 is useful for identifying onset behavior because it experienced the full dry-to-wet-to-corrosive transition without the pronounced underfloor liquid trapping seen on Sensor set 1. For persistent episodes of at least 1 h, free corrosion on Sensor set 2 began at RH = 84.3–90.2% and Cond Lo = 0.245–0.830 µS, while galvanic corrosion began at RH = 85.7–98.2% and Cond Lo = 0.328–1.0 µS. By comparison, persistent activity on Sensor set 1 shifted toward higher wetness: free-corrosion episodes began at RH = 85.1–100% and Cond Lo = 0.585–1.0 µS, whereas galvanic episodes began only at RH = 92.5–100% and Cond Lo = 1.0 µS. Conductance, rather than RH alone, is the more direct indicator of an electrochemically active electrolyte film. A conservative corrosion-prone condition indicator (CPCI) was defined as RH ≥ 80% and Cond Lo ≥ 0.05 µS. This mild threshold captured 95.5% of free-corrosion activity and 100% of galvanic activity on Sensor set 2, and 98.8% of free-corrosion activity and 98.9% of galvanic activity on Sensor set 1. These results confirm the indicator’s sensitivity for the present data set, but should not be read as universal statistical limits. The 82-day campaign covered only the initial operational period and did not include full seasonal, mission-profile, or geographic variability. CPCI is a screening indicator—its purpose is to identify conditions that, if sustained, may promote future corrosion initiation—not a quantitative damage metric. For other aircraft, installation positions, climates, or mission profiles, the physical logic applies, but numerical thresholds and persistence windows should be recalibrated using local baseline data and, where available, inspection or NDT feedback. A stronger regime (RH ≥ 90% and Cond Lo ≥ 0.5 µS) corresponded to 133.5 h in Sensor set 2 and 627.0 h in Sensor set 1. This higher threshold captures the most severe and long-lasting wet states, especially in the closed underfloor volume, and provides a practical escalation level for maintenance-oriented warning assessment. Its significance is prospective: it marks persistent exposure severity that may justify preventive intervention and activation of higher-level diagnostic layers, but does not by itself prove that structural corrosion damage is present. 3.4. Warning-Level Assessment for Closed Zones For closed or difficult-to-inspect aircraft zones, the data support a three-level warning logic: Watch = RH–conductance framework: Watch = RH ≥ 80% and Cond Lo ≥ 0.05 µS; Warning = RH ≥ 90% and Cond Lo ≥ 0.5 µS; Severe = RH ≥ 95%, Cond Lo = 1.0 µS, and Cond Hi ≥ 20 µS. The Severe state was defined to capture not only high humidity but also a persistently conductive film with clear evidence of strong wetting or replenishment. Each state is prospective, not confirmatory: it indicates that the monitored zone has experienced persistently unfavorable conditions warranting increased surveillance, preventive action, or higher-level diagnostics—not that corrosion damage is already present. When this logic was applied to the two focus sensor sets ( Figure 4), Sensor set 2 spent 179.0 h in Watch, 82.0 h in Warning, and 51.5 h in Severe. It produced 6 warning episodes and 3 severe episodes longer than 6 h, with the longest warning episode lasting 52.5 h and the longest severe episode lasting 21.5 h. This is consistent with repeated, yet episodic, wetting below the gearbox. The response of Sensor set 1 was substantially more critical. It spent 223.5 h in Watch, 220.5 h in Warning, and 406.5 h in Severe, and generated 20 warning episodes longer than 6 h and 14 severe episodes longer than 6 h. The longest warning episode lasted 96.0 h, while the longest severe episode lasted 70.5 h. These persistence times are too long to be explained by short ambient-humidity excursions alone and instead point to wetness retention within a closed volume, very likely associated with trapped liquid or repeated replenishment. The key observation for maintenance is that persistence matters more than peak amplitude in closed zones. Short excursions were separated from persistent episodes using the duration rules from Section 2.6; warning and severe episodes longer than 6 h were counted separately as maintenance-relevant events. When the severe state persists for several hours, inspection should cover not only corrosion products but also drainage paths, local ventilation, and fluid-ingress routes. In the proposed architecture, such states would justify PZT-based localization and local eddy-current or NDT verification when access permits. The operational note that abrupt Sensor set 1 excursions occurred during flight missions further supports inspecting for underfloor liquid ingress or spillage after severe alarms. Sensor set 2 is best suited for deriving the empirical corrosion-onset envelope; Sensor set 1 is the more relevant sensor for warning-level assessment in closed aircraft zones. Together, the two signals show that CHM sensors can distinguish between broad humidity-driven corrosion risk and the more critical local wetting events that require direct maintenance action. When interpreted through the proposed maintenance-support layer, these results translate into different operational responses for the two focus locations. Sensor set 2 would predominantly support watch- or warning-type actions, including intensified trend review and targeted inspection planning, whereas Sensor set 1 would repeatedly escalate to severe alarms linked to direct maintenance intervention in the underfloor compartment. From the maintenance crew’s perspective, the system’s significance lies not in autonomous AI decision-making, but in presenting a traceable graded alert that shows which thresholds were exceeded, for how long, and which predefined maintenance action is recommended, as summarized in Table 1. 3.5. Proposed PZT and Eddy-Current Follow-Up of Persistent Warning Episodes The campaign focused on environmental, conductance, and electrochemical channels because the system had only recently entered operation. The severe and long-duration events identified for Sensor set 1 define the most relevant time windows and structural zones for future higher-level follow-up diagnostics. No PZT or eddy-current sensors were installed on the flying helicopter during this initial campaign—not as a methodological omission, but because installing additional sensors on a real aircraft requires maintenance access, aircraft downtime, integration with service procedures, and an operational justification. Although severe early-warning episodes were detected, they did not coincide with an approved maintenance-access window or dedicated flight-test opportunity during this initial period. 4. Discussion 4.1. Relevance for SHM and Condition-Informed Maintenance The proposed multi-sensor approach extends SHM beyond structural response to include the environmental and electrochemical conditions that promote corrosion in enclosed aircraft zones. This is especially relevant in underfloor spaces, where direct inspection is limited and corrosion can develop unnoticed between maintenance actions. Combining humidity, temperature, surface conductance, and corrosion-related signals gives a more realistic representation of the monitored zone’s condition than any single-parameter observation [ 2]. The main practical advantage is earlier identification of conditions that may lead to hidden corrosion, allowing action before corrosion progresses to damage requiring repair. Monitoring output informs decisions about whether the compartment should be opened, cleaned, inspected, or checked for drainage issues or fluid ingress—preventive action against corrosion-promoting conditions rather than corrective action after damage has occurred [ 2, 16]. A practical consequence is that closed zones need not be opened solely because a fixed maintenance interval has been reached. Opening and cleaning can be prioritized when the measured local environment indicates persistent wetness or electrochemical activity. This is condition-informed maintenance support, not a complete prognostic CBM solution. The system does not yet estimate remaining useful life or predict corrosion-growth kinetics; it identifies periods and locations where the probability of corrosion initiation is elevated and inspection is better justified. The method is not limited to aerospace structures. The same early-warning, localization, and follow-up logic can be transferred to civil engineering components with hidden or difficult-to-access zones—closed steel details, box sections, joints, anchorages, cavities—where moisture retention and delayed inspection allow degradation to develop unnoticed. The helicopter application is a demanding real-world demonstrator of a general SHM methodology based on persistence analysis and staged diagnostic escalation. 4.2. Traceable Maintenance-Support Interface and Transition Toward CBM A key practical extension is integrating fused sensor interpretation with a traceable maintenance-support interface. The system first classifies warning states using deterministic thresholds and persistence rules. The GPT-type module then prepares a human-readable alert that explains the event, lists the exceeded criteria, and links the warning level to a predefined maintenance recommendation. This separation is important for aviation use: the diagnostic classification must remain reproducible from sensor data and the rule base, while natural-language generation serves only to improve communication with maintenance personnel. This logic supports a gradual transition from fixed-interval inspection toward condition-informed maintenance and, in the longer term, condition-based maintenance. At the present stage, however, the system is not a full CBM or prognostic solution. It does not estimate remaining useful life, model corrosion-growth kinetics, or autonomously optimize intervention timing. Its demonstrated value is narrower but operationally useful: it provides early evidence that a hidden zone has experienced persistent corrosion-promoting conditions and helps maintenance personnel decide whether continued observation, ventilation, drying, cleaning, opening, or targeted NDT is warranted. The decision-support logic proposed in this study is summarized schematically in Figure 5, which illustrates the sequence from multi-sensor acquisition and AI-assisted interpretation to alert generation, maintenance action, and localized follow-up inspection. The scheme integrates sensing, data interpretation, and maintenance decision support into a single operational workflow. In the first stage, sensor inputs from environmental and corrosion-related channels are acquired from the monitored helicopter zone and, after preprocessing and feature extraction, transformed into a fused representation of the local condition. This information is then interpreted by a traceable rule-based layer that recognizes abrupt increases in humidity, persistent surface wetness, conductance excursions, and combined electrochemical responses. The resulting alert level is linked to maintenance actions, which may range from trend review to opening the hatch for ventilation and drying, local wipe-down cleaning, or more direct intervention. In a future operational extension, PZT sensing would support localization of the most suspect structural subregion, whereas eddy-current sensing and targeted NDT would provide local verification and follow-up monitoring of degradation growth in critical metallic details. 4.3. Support for Targeted NDT 4.4. Limitations and Practical Implementation Issues First, the CPCI and the associated Watch/Warning/Severe thresholds are not direct measures of corrosion damage. They are early-warning criteria designed to flag conditions that, if sustained, may promote corrosion initiation and growth. Derived from an 82-day initial campaign and two sensor sets, they should be treated as empirical screening values rather than universal statistical limits. Assessing threshold drift and recalibrating the warning logic for other aircraft types, structural details, climates, and service environments will require additional seasonal data, different mission profiles, and additional installation sites. The same applies to potential civil engineering use: the methodology transfers conceptually, but threshold calibration must be performed for the specific material system, exposure scenario, and inspection strategy of the target structure. Second, the current sensor layout should not be interpreted as proof of complete spatial coverage of all high-risk closed zones. The two focus sensor sets were installed in priority locations selected from engineering judgment and expected liquid-retention paths, but corrosion risk in enclosed aircraft compartments is spatially heterogeneous. Unmonitored blind spots may remain, especially near drains, penetrations, lap joints, fasteners, seal discontinuities, and local debris or contaminant traps. A mature fleet implementation should therefore combine the present warning logic with a risk-based sensor-placement map, additional measurement points where justified, and periodic feedback from inspection findings. Fourth, the present in-service validation covers the early-warning layer based on environmental, conductance, and electrochemical sensing. PZT and eddy-current sensing are defined here as staged follow-up diagnostics: PZT sensing localizes probable corrosion-affected subregions, and eddy-current sensing confirms and monitors local damage progression. Both layers have been validated separately in laboratory or controlled studies but were not installed on the flying helicopter during this initial period. Installing additional sensors on a real aircraft depends on aircraft availability, access to the monitored compartment, maintenance planning, approval of the installation procedure, and a justified operational trigger. Persistent severe early-warning episodes were observed during the campaign, but none coincided with an approved maintenance or flight-test window suitable for PZT and EC installation. Their full integration into the operational workflow remains a planned next step. In operational aircraft, complete end-to-end validation of corrosion-warning logic is difficult to achieve on a case-by-case basis. Hidden zones cannot be opened repeatedly without cost, downtime, and maintenance burden, and long-term corrosion evolution cannot be intentionally reproduced under operational conditions for validation purposes alone. The present work should therefore be treated as a first in-service validation stage: the early-warning thresholds are supported by in-service measurements and by field evidence showing that prolonged corrosion-prone conditions in hidden zones can eventually lead to costly corrosion damage [ 1, 2]. Correlating warning episodes with pit depth, mass loss, crack initiation, or local damage-growth rate will require targeted inspection and installation of the EC/NDT follow-up layer. Further validation should progressively link warning episodes with direct inspections, PZT-based localization, eddy-current follow-up, and maintenance findings. The same limitation applies to civil engineering assets with hidden zones, where repeated intrusive opening for one-to-one validation may be technically possible but economically unjustified. Fifth, the economic benefit of the proposed system was not quantified in this study. The expected operational value lies in reducing unnecessary floor-panel removal, shortening troubleshooting time after wetting or contamination events, and preventing late discovery of extensive hidden corrosion. A rigorous cost–benefit assessment requires longer fleet use, records of actual maintenance decisions, labor hours, aircraft downtime, installation costs, sensor replacement costs, and corrosion-repair costs—data not yet available because the system had only recently entered service. Future work should compare the warning-supported workflow with conventional scheduled access and with simpler single-sensor strategies, using indicators such as avoided disassembly hours, false-alarm rate, confirmed findings, and repair-cost reduction. 4.5. Field Inspection Evidence Supporting the Maintenance Relevance of Persistent Corrosion-Prone Conditions Field inspection evidence supports the maintenance relevance of the proposed warning logic. Figure 6 shows examples of corrosion observed in a hidden aircraft zone after prolonged exposure to moisture-retaining and corrosion-prone conditions, identified in riveted and lap-joint-type details representative of the closed structural locations considered in this monitoring concept. These findings support the practical interpretation of the adopted warning thresholds, but should not be read as a one-to-one temporal and spatial validation of specific Sensor set 1 or Sensor set 2 severe events. The photographs provide qualitative maintenance context: persistent corrosive conditions, even when detected only through environmental, conductance, and electrochemical indicators, can evolve into actual corrosion damage requiring costly corrective maintenance [ 1, 2]. Corrosion removal and repair in the observed case were labor-intensive—requiring access to a closed zone, removal of corrosion products, cleaning, restoration of protective layers, and local repair. This justifies the early-warning approach and the need for staged follow-up diagnostics; direct localization and damage quantification require PZT, eddy-current sensing, targeted NDT, or physical inspection in future campaigns. 4.6. Outlook: Advanced Sensors and Computational Data Fusion The next step is combining longer operational data collection with direct inspection feedback and higher-level diagnostic layers. In the near term, this means installing PZT and eddy-current sensors at the next approved maintenance window or dedicated aircraft-access opportunity. In parallel, the alert thresholds should be recalibrated using seasonal data, different mission types, and additional aircraft locations—allowing sensitivity, false-alarm behavior, and threshold drift to be quantified rather than inferred from a single 82-day campaign. 5. Conclusions This paper reports an in-service validation of the first monitoring layer of a multi-sensor data fusion approach for early warning of corrosion-prone conditions in closed aircraft zones. Environmental, conductance, and electrochemical sensing identified persistent unfavorable conditions in hidden compartments during aircraft operations and distinguished general humidity exposure from more critical long-duration wetting events. Within the proposed architecture, these first-stage signals indicate increased long-term corrosion risk; they do not constitute direct proof of existing structural damage. From a maintenance perspective, the approach supports condition-informed decision-making in areas where routine inspection is difficult. The installed sensors characterize the general environmental and electrochemical conditions of the monitored zone and indicate whether corrosion-favorable conditions have persisted. They do not determine corrosion type, exact location, pit depth, mass loss, or crack initiation. Their outputs should inform decisions about whether a closed zone should be monitored further, aired, dried, opened, cleaned, or inspected—not serve as substitutes for NDT or engineering assessment. Complete case-by-case validation in operational aircraft is inherently limited by access, cost, availability, and airworthiness constraints. The present study is the first operational validation stage toward a fully integrated corrosion-warning and follow-up monitoring framework. Further maintenance opportunities are required before the PZT and eddy-current follow-up layers can be installed and validated on the flying helicopter. The system is not intended to replace operational NDT or engineering judgment. Its purpose is to support and better target inspection activities by identifying locations and periods of increased corrosion risk and by linking warning levels to maintenance actions. PZT and eddy-current sensing remain important follow-up layers, validated separately in laboratory or controlled conditions and ready for aircraft implementation when a persistent warning state or scheduled access opportunity justifies installation. The same logic may be adapted to civil engineering structures and infrastructures with hidden or difficult-to-inspect zones, but sensor selection, thresholds, persistence criteria, and maintenance actions must be recalibrated for the relevant material system and exposure environment. Author Contributions Conceptualization, P.C.; methodology, P.C., L.V.-G., A.K. and M.D.; investigation, P.C., L.V.-G., L.M., A.B. and A.N.; validation, P.C., L.V.-G., A.K., M.D., L.M., A.B. and A.N.; formal analysis, P.C., M.D. and A.B.; data curation, P.C., L.V.-G. and A.N.; writing—original draft preparation, P.C.; writing—review and editing, L.V.-G., A.K., M.D., L.M., A.B., A.N. and A.L.; visualization, P.C. and A.B.; supervision, M.D., A.K. and A.L. All authors have read and agreed to the published version of the manuscript. Funding This research was funded by the statutory work no. 31-1188-1 of the Air Force Institute of Technology (ITWL). The research was also partially supported by Polish National Center for Research and Development within the scope of the project LIDER13/0132/2022. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results. Institutional Review Board Statement Not applicable. Informed Consent Statement Not applicable. Data Availability Statement The data presented in this study are available from the corresponding author upon reasonable request. The authors agree to provide the research data for scientific and editorial verification. Acknowledgments During the preparation of this manuscript, the authors used a GPT-5.1-based generative AI tool for translation support and for adapting the text between the working manuscript and the Applied Sciences journal template. The authors reviewed and edited the output and take full responsibility for the content of this publication. M.D. would like to acknowledge support from the Polish National Center for Research and Development within the scope of the project LIDER13/0132/2022. Conflicts of Interest The authors declare no conflicts of interest. References Figure 1. Schematic location of Sensor set 1 and Sensor set 2 on the medical helicopter. Sensor set 1 was installed beneath the cabin floor in a closed under-floor zone, whereas Sensor set 2 was installed below the main gearbox in a horizontal configuration. Figure 1. Schematic location of Sensor set 1 and Sensor set 2 on the medical helicopter. Sensor set 1 was installed beneath the cabin floor in a closed under-floor zone, whereas Sensor set 2 was installed below the main gearbox in a horizontal configuration. Figure 2. Full-campaign signals for Sensor set 1 and Sensor set 2: ( a) air temperature, ( b) relative humidity, and ( c) low-frequency surface conductance. Continuous operation for ~82 days, with repeated returns to the dry-state baseline, demonstrates that autonomous deployment was feasible under in-service helicopter conditions. Figure 2. Full-campaign signals for Sensor set 1 and Sensor set 2: ( a) air temperature, ( b) relative humidity, and ( c) low-frequency surface conductance. Continuous operation for ~82 days, with repeated returns to the dry-state baseline, demonstrates that autonomous deployment was feasible under in-service helicopter conditions. Figure 3. Representative in-service signals. ( a, c) Sensor set 2 during the February condensation-driven activation sequence, showing RH and Cond Lo build-up followed by galvanic and free-corrosion accumulation. ( b, d) A localized wetting event occurred near Sensor set 1 on 18 March 2026, during which RH remained near saturation, and conductance rose sharply, while the other aircraft sensors simultaneously dried out. Figure 3. Representative in-service signals. ( a, c) Sensor set 2 during the February condensation-driven activation sequence, showing RH and Cond Lo build-up followed by galvanic and free-corrosion accumulation. ( b, d) A localized wetting event occurred near Sensor set 1 on 18 March 2026, during which RH remained near saturation, and conductance rose sharply, while the other aircraft sensors simultaneously dried out. Figure 4. Empirical activation maps in RH–conductance space. Gray points indicate no increment in cumulative corrosion charge; red points indicate free-corrosion increments; purple points indicate galvanic-corrosion increments. Dashed lines show RH = 80% and 90%; dotted lines show Cond Lo = 0.05 and 0.5 µS. Sensor set 2 occupies a broader activation corridor, whereas Sensor set 1 corrosion increments are concentrated at near-saturated RH and high conductance. Figure 4. Empirical activation maps in RH–conductance space. Gray points indicate no increment in cumulative corrosion charge; red points indicate free-corrosion increments; purple points indicate galvanic-corrosion increments. Dashed lines show RH = 80% and 90%; dotted lines show Cond Lo = 0.05 and 0.5 µS. Sensor set 2 occupies a broader activation corridor, whereas Sensor set 1 corrosion increments are concentrated at near-saturated RH and high conductance. Figure 5. Block diagram of the proposed traceable multi-sensor corrosion-warning and maintenance-support architecture for closed helicopter zones. Sensor inputs are first preprocessed and then interpreted using threshold- and persistence-based data fusion. The GPT-type module is limited to natural-language reporting of the predefined alert class and maintenance recommendation. PZT and eddy-current sensors are shown as planned follow-up diagnostic layers rather than as field-validated outputs of the present 82-day campaign. Figure 5. Block diagram of the proposed traceable multi-sensor corrosion-warning and maintenance-support architecture for closed helicopter zones. Sensor inputs are first preprocessed and then interpreted using threshold- and persistence-based data fusion. The GPT-type module is limited to natural-language reporting of the predefined alert class and maintenance recommendation. PZT and eddy-current sensors are shown as planned follow-up diagnostic layers rather than as field-validated outputs of the present 82-day campaign. Figure 6. Field inspection evidence of corrosion in a hidden aircraft zone after prolonged persistence of corrosion-prone conditions. ( a) Close-up view of corrosion damage developing along a riveted and lap-joint-type detail. ( b) Wider view of corrosion products and degradation in a hidden structural zone. These photographs provide qualitative maintenance context for the proposed early-warning logic. They do not establish a direct one-to-one temporal or spatial correlation with the specific severe events recorded by Sensor sets 1 and 2 during the present campaign; rather, they show why persistent corrosion-prone conditions in closed zones should be treated as operationally important and followed, when justified, by localization and direct NDT/EC verification. Figure 6. Field inspection evidence of corrosion in a hidden aircraft zone after prolonged persistence of corrosion-prone conditions. ( a) Close-up view of corrosion damage developing along a riveted and lap-joint-type detail. ( b) Wider view of corrosion products and degradation in a hidden structural zone. These photographs provide qualitative maintenance context for the proposed early-warning logic. They do not establish a direct one-to-one temporal or spatial correlation with the specific severe events recorded by Sensor sets 1 and 2 during the present campaign; rather, they show why persistent corrosion-prone conditions in closed zones should be treated as operationally important and followed, when justified, by localization and direct NDT/EC verification. Table 1. Alert level vs. maintenance action in the proposed maintenance-support logic. Table 1. Alert level vs. maintenance action in the proposed maintenance-support logic. Alert Level Indicative Condition Maintenance Action Typical Crew/System Instruction Watch Short or mild increase in humidity/wetness; no persistent structural confirmation. Continue observation and review trends; no immediate opening required. Keep the zone under surveillance; optionally open the hatch for airing at a convenient service opportunity. Warning Persistent rise in RH/conductance and/or intermittent corrosion activity. Plan targeted inspection and preventive drying/cleaning at the next suitable maintenance opportunity. Open the hatch, air out the compartment, check the drains and seals, and dry any visible moisture if present. Severe Near-saturated humidity, high conductance, persistent electrochemical activity, or repeated long-duration events. Immediate intervention with local follow-up diagnostics. Open the hatch promptly, remove trapped liquid, dry or wipe the area with a clean cloth if necessary, then trigger PZT localization and EC/NDT verification. Table 2. Summary of signal continuity and final corrosion totals for the two focus sensor sets. Table 2. Summary of signal continuity and final corrosion totals for the two focus sensor sets. Sensor Records Median Interval Max Gap Final Totals Sensor set 1 3935 30 min 35.0 min Tot Free 0.007995 C; Tot Galv 0.002231 C Sensor set 2 3938 30 min 33.8 min Tot Free 0.001381 C; Tot Galv 0.002103 C Table 3. Dominant in-service corrosion episodes captured by Sensor set 1 and Sensor set 2. Table 3. Dominant in-service corrosion episodes captured by Sensor set 1 and Sensor set 2. Sensor Dominant Mechanism Time Window Duration Charge Share of Final Total Sensor set 2 Galvanic 7 February 14:50–8 February 08:20 18 h 0.001066 C 50.7% Sensor set 2 Free 12 February 00:20–13 February 12:50 37 h 0.000798 C 57.8% Sensor set 1 Free 22 February 12:09–24 February 10:39 47 h 0.002710 C 33.9% Sensor set 1 Galvanic 27 February 14:39–28 February 08:39 18.5 h 0.000907 C 40.7% Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. © 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license. Share and Cite MDPI and ACS Style Ciężak, P.; Dziendzikowski, M.; Kurnyta, A.; Vázquez-Gómez, L.; Mattarozzi, L.; Benedetti, A.; Nidzgorska, A.; Leski, A. Multi-Sensor Data Fusion for Early Warning of Corrosion-Prone Conditions in Closed Zones
